Hey guys, here’s this week’s edition of the Spatial Edge, a place where every CRS is always declared. Anyway, the aim as usual is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Mobile Internet: Cell towers raise household wealth in the Philippines.

• HLS Check: NASA’s harmonised Landsat-Sentinel product closely matches MODIS.

• Crop Calendars: Sentinel-1 pins sowing dates to within eight days.

• Travel Friction: A global 30 m walking and driving surface.

• Water Transfers: Mapping 435 major water transfer projects worldwide.

Subscribe now

Research you should know about

1. How mobile internet raised household wealth in the Philippines

Alright, I’m gonna own this one. This first paper is one I co-authored with Zhiwu Wei and Neil Lee (riding on their coattails of course). It’s one that we’ve been working on for about 2.5 years, so we’re pretty thrilled it’s finally out. Mobile internet is now basic infrastructure across the developing world, but it’s hard to measure its causal effect on living standards because better-off places tend to get connected first. Our paper in the Journal of Development Economics to attempts to examine this for the Philippines.

We built a granular measure of local connectivity from about 270,000 geocoded cell towers, then used distance to submarine cable landing points as an instrument for that connectivity. The idea is that being closer to where undersea internet cables come ashore makes it cheaper to roll out towers, but shouldn’t affect wealth through any other channel. We linked this to household wealth from the 2017 and 2022 Demographic and Health Surveys, and checked the instrument with a placebo test on 2003 survey data, collected before towers were rolled out.

Mobile internet connectivity increases household wealth, with effects across education levels and most pronounced in urban areas. We find that doubling local tower density raises the household wealth index by about 0.04 standard deviations, roughly ten times the naive estimate, and the effect in towns and mid-sized cities is around 3.6 times that in rural areas. Better connectivity also boosts activity in key economic sectors, improves educational outcomes and raises female labour force participation, mostly in occasional work. We conducted an event study, which suggests tower rollout may also make it easier for people to migrate abroad for work and send remittances home.

2. How well NASA’s harmonised Landsat-Sentinel product lines up with MODIS

NASA’s Harmonized Landsat and Sentinel-2 (HLS) product gives 30 m surface reflectance every 2–3 days on average, but it’s important to know how consistent it is with the long-established products people have been using for years. A new study in Remote Sensing of Environment from folks at NASA and others compares HLS Version 2.0 with MODIS’s BRDF-adjusted reflectance (NBAR).

They compared the two products globally across the six spectral bands they share, looking at correlations, mean absolute differences and where the biggest gaps show up. Because MODIS pixels are much coarser, this also shows how resolution affects what each product records over mixed surfaces.

HLS and MODIS NBAR agree closely, with R² values of 0.939 to 0.986 across the six bands. Mean absolute differences are generally under 0.03 for visible bands and under 0.05 for near- and shortwave-infrared, with the largest discrepancies in the blue and SWIR1 bands. The biggest spatial differences are in polar regions, where patches of snow, ice and water get mixed into coarse MODIS pixels but are filtered out of HLS. The authors suggest this could explain some past disagreements between greening trends measured at different resolutions.

3. Pinpointing sowing and harvest dates with Sentinel-1

Knowing when crops are sown and harvested matters for yield estimation and crop modelling, and Sentinel-1 radar can see through clouds, but SAR-based methods tend to work well for some crops and badly for others. A new study in Remote Sensing of Environment from the Hong Kong Polytechnic University works out why, and builds a method that adapts to local conditions.

They found that the cross-polarisation ratio works best for rain-fed crops on rough soil, while VH backscatter works better for irrigated or flooded fields, and built a framework (ASIS) that picks the right indicator automatically. ASIS first clusters Sentinel-1 pixels into groups representing different crops and surface conditions, then uses a new index combining VH and the cross-polarisation ratio to delineate growing seasons. Finally, it uses VH outside the growing season to identify surface conditions and choose the best indicator for detecting sowing and harvest.

ASIS estimated sowing dates with a mean absolute error of 7.62 days and harvest dates with 8.48 days. That cuts sowing-date errors by 32.4% compared with using VH alone and by 52.7% compared with the cross-polarisation ratio alone (7.5% and 9.5% for harvest). It works without prior crop labels or field boundaries, which makes it useful in places where that information doesn’t exist.

4. One model for mapping buildings from any optical imagery

Most building extraction models are trained for one dataset or sensor and struggle when you switch to imagery with a different resolution. A new study from the Technical University of Munich builds a single model that works across optical imagery from 5 cm aerial imagery to 10 m Sentinel-2.

UniBuild trains one model on 12 RGB datasets spanning 0.05 to 10 m resolution, using a DINOv3 backbone with a new high-resolution decoder. The decoder feeds fine shallow features back in to sharpen boundaries and recover small buildings. Two extra training losses keep building edges consistently oriented and stop neighbouring buildings merging in low-resolution imagery, and the model can output georeferenced vector footprints directly.

Across the 12 datasets, UniBuild reached an average IoU of 76.27, against 74.03 for a DINOv3 baseline, with a bigger gain on building boundaries (65.60 vs 61.06). On five completely unseen datasets it averaged an IoU of 63.71, ranging from 89.28 in Waterloo to 32.51 in Pforzheim. The code and model checkpoint are on GitHub.

Geospatial Datasets

1. A global 30 m map of how hard it is to get around

GRUFS, from the World Bank and the University of Twente, is a pair of global 30 m friction surfaces giving travel cost in minutes per metre: one for walking (on roads, paths, ferries and off-road) and one for driving on roads with walking off-road. It’s built from OpenStreetMap, GLAD land cover and FABDEM elevation, with lookup tables and production code, and is designed for accessibility and catchment analysis with planned annual updates. You can find the paper here.

2. Every major water transfer project on Earth

This global inventory of major water transfer projects documents 435 projects with their engineering attributes, plus spatial layers for intake points, transfer routes and the regions each project influences. You can access the data here.

3. Forty years of storm systems tracked from space

This mesoscale convective systems database tracks large organised storm systems between 60°S and 60°N from 1985 to 2024, using satellite infrared brightness temperatures. You can access the data here.

Other useful bits

• ESA has released Earth Virtual Expert (EVE), an open AI assistant for Earth observation built by ESA’s Φ-lab with Pi School, Mistral and Imperative Space. It uses retrieval-augmented generation over ESA, NASA, Copernicus and peer-reviewed sources, and the model, code, training data and benchmarks are public on Hugging Face and GitHub.

• GeoPandas 1.2.0 is out, with rebuilt static plotting (basemaps via a tiles= keyword, colour bars for classified data and a “greedy” scheme so neighbouring polygons never share a colour). It also adds a make_grid function for square or hexagonal grids, reads and writes the upcoming GeoParquet 2.0 spec, and read_parquet can now read directly over HTTP.

• The Copernicus Data Space Ecosystem has launched an openEO Agent that turns a plain-language question into a validated openEO workflow and runs it for you, so you don’t have to pick collections, bands and time periods by hand. It’s a limited first version, and you can request access through VITO.

• Sentinel-3 land surface temperature is now in Copernicus Browser and the Sentinel Hub API, at 1 km with two observations a day (around 10:00 and 22:00 local time) plus uncertainty, cloud probability and NDVI layers. One catch: the Browser shows it without a cloud mask, so cloudy pixels show cloud-top temperature.

• NASA Earth Observatory has mapped severe flooding in Thailand with before-and-after Landsat images from October 2025 and October 2026. Flooding since mid-September has hit 50 provinces, affected more than 4 million people and damaged about 270,000 hectares of farmland.

Jobs

• ESA is looking for an Internal Research Fellow in AI on board for Synthetic Aperture Radar Missions based in Frascati, Italy.

• ICEYE is looking for a Senior Geospatial Software Engineer based in Espoo, Finland (hybrid).

• Oak Ridge National Laboratory is looking for a Postdoctoral Research Associate in Coastal Landscape Geospatial Modeling based in Oak Ridge, United States.

• Isometric is looking for a Geospatial Data Analyst (contractor) based in London, United Kingdom.

Just for Fun

You can’t see a complete auroral oval from the ground because it’s too big a ring around the magnetic pole, but ESA and CAS SMILE spacecraft caught one in ultraviolet light, during daytime. The time-lapse covers about an hour in late July and shows how turbulent auroras really are. SMILE’s job is to study how the solar wind interacts with Earth’s magnetosphere, which helps protect satellites and power grids from solar storms.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge, a place where all your layers line up first time. Anyway, the aim as usual is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Mining Threats: A third of mining land overlaps conservation priorities.

• Food Security: El Niño amplifies the impact of food price spikes.

• Fire Weather: Hourly fire danger reveals sharp local risk gradients.

• Flood Susceptibility: A global 30 m flood susceptibility map.

• Methane Hunting: Planet launches hyperspectral Tanager-2 on Transporter-18.

Subscribe now

Research you should know about

1. A third of the world’s mines sit in conservation hotspots

The UN has committed to halting biodiversity loss by 2030, but mining remains a major driver of decline, and its overlap with conservation priorities has, weirdly, never been properly measured. A new study in Nature Communications led by the Naturalis Biodiversity Center and Leiden University maps that overlap globally, both for today’s mines and for land that could be mined in future.

The team overlaid 145,000 km² of current mining land with conservation priority zones and species ranges, then repeated the exercise for land potentially exploitable for metals, non-metals and coal. This shows which species, biomes and ecosystems are exposed now, and where future mining expansion would collide with conservation goals.

About 34% of current mining land (49,000 km²) overlaps conservation priority zones, intersecting the ranges of more than 2,300 species, many of them threatened. The threats span 14 biomes, with disproportionately high relative impacts in fragile ecosystems like mangroves and the Mediterranean. Looking ahead, 43% of land potentially exploitable for metals (12 million km²), 29% for non-metals (2 million km²) and 37% for coal (4 million km²) also overlaps these zones. The authors pitch the maps as a way to align mining with the Kunming-Montreal biodiversity targets.

2. How El Niño changes the impact of food price spikes

Food price spikes push people into hunger, but the damage isn’t the same everywhere, and early-warning systems rarely account for slow-moving climate patterns like El Niño. A new study from the Universitat de València, Wageningen University and the EU’s Joint Research Centre asks whether El Niño’s influence on vegetation changes how hard price spikes hit food security in sub-Saharan Africa.

They grouped 222 regions in 11 countries by how El Niño affects their vegetation, then estimated the causal effect of price spikes on food insecurity in each group using double machine learning. The groups come from a Granger-causal PCA on 20 years of MODIS NDVI and the Niño 3.4 index, splitting regions into those where El Niño suppresses vegetation, boosts it, or has little effect. The treatment is WFP’s price-spike indicator, the outcome is next month’s share of people in crisis-level coping, and the models adjust for climate, conflict, displacement and economic factors from September 2020 to February 2024.

In regions where El Niño suppresses vegetation, a price spike raised crisis-level food insecurity by about 5.4 percentage points, against roughly 2 points (not statistically significant) elsewhere. For a typical region of about 1.3 million people, that’s around 70,600 more people in crisis the following month. These regions also start from the highest baseline vulnerability. The practical idea is to use the El Niño state to flag vulnerable regions months ahead, then have responses ready for when prices spike. The code is on GitHub.

3. Why fire danger needs an hourly view

Fire danger ratings usually rely on daily values from coarse weather models, which miss the local, fast-changing conditions that drive how wildfires behave. A new study in Scientific Reports from the University of Évora tests what you gain by computing the Fire Weather Index hourly from high-resolution simulations.

They ran the Meso-NH atmospheric model at 1.5 km and 0.5 km over three periods with active fires in Portugal and Spain, then used the output to calculate an hourly Fire Weather Index. This shows how fire danger shifts within a day, including the Fine Fuel Moisture Code, which tracks how dry surface fuels are.

Hourly fire danger follows a clear daily cycle, and the high-resolution runs capture sea breezes and mountain-driven winds that create sharp local gradients invisible in daily, coarse-resolution products. The authors argue this could be built into Portugal’s operational fire danger system, since its AROME weather model shares the same physics as Meso-NH.

4. An AI agent that balances buildings and terrain in 3D maps

Digital surface models built from satellite images matter for urban planning and digital twins, but there’s a trade-off between getting buildings right and getting the whole scene right. A new study from the University of Waterloo, the University of Calgary and Sun Yat-sen University adds an AI agent to manage that trade-off.

They combined Gaussian splatting for satellite imagery with building masks from the Segment Anything Model, giving building pixels more weight during reconstruction. An agentic controller checks mask quality and validation results, picks between standard and building-focused schedules for each scene, and keeps an audit trail of its decisions. They tested it on four areas from the DFC2019 benchmark.

On the Jacksonville JAX_004 scene, building-focused weighting cut building height error from 0.844 m to 0.806 m (about 4.5%). A staged schedule also nudged full-scene error down from 1.362 m to 1.349 m without losing the building gains. Alongside the surfaces, the system outputs building inventories and decision records, so each reconstruction choice can be checked later.

5. Using Geospatial AI to choose where EV chargers go

Charging stations are often sited mainly on cost or travel distance, ignoring the land use, population, roads and traffic around them. A new study treats site selection as a geospatial machine learning problem instead.

The authors used a variational autoencoder to compress a large set of EV-related geospatial attributes, then a graph convolutional network to learn how locations relate to one another. Working in Bryan–College Station, Texas, with data on EV usage, land use, population and traffic, the model looks for sites that resemble successful existing stations.

The model separated existing station locations from other sites with an F1-score of 0.87 and identified 27 candidate locations. Two policy scenarios produced different recommendations, which shows the framework can support different planning goals.

Geospatial Datasets

1. A global 30 m flood susceptibility map

The Global Flood Susceptibility Map (GFSM v1) maps flood susceptibility worldwide at 30 m, using a gradient-boosted tree model trained on about 30 million samples and nine factors like elevation, height above nearest drainage, distance to water and roads, and rainfall frequency. You can access the data here and the code here.

2. Field photos paired with satellite time series across Europe

AgroFusion links 211,483 LUCAS field observations from 2018 and 2022 with field photos, monthly Sentinel-2, Sentinel-1 and Landsat-8 time series, ERA5-Land weather, terrain, soil properties and AlphaEarth embeddings. You can access the data here.

3. How intensively US cropland is farmed

NASA’s LP DAAC and USGS have released the Landsat-derived Global Cropland Intensity Product for the contiguous US, classifying 30 m pixels as single-, double- or triple-cropped, continuous or fallow cropland for a nominal 2020 (2019–2021). You can access the data here.

4. 43 years of hourly European renewables and demand

This European renewables and demand dataset uses neural networks to turn 43 years of weather reanalysis into hourly national time series of solar, onshore wind and offshore wind generation and electricity consumption, trained on quality-controlled ENTSO-E data. You can find the paper here.

Other useful bits

• Planet has launched 20 satellites on Transporter-18, including Tanager-2, a hyperspectral satellite built with Carbon Mapper that has 426 bands at about 5 nm spectral and 30 m spatial resolution for spotting methane. The launch also carried 18 SuperDoves and Google’s Project Suncatcher prototype, bringing Planet’s total to 718 satellites delivered to orbit.

• Landsat 9 has completed its five-year design life, reaching the milestone on 27 September, 54 years after Landsat 1 launched. It routinely collects more than 750 images a day, and with Landsat 8 produces about 1,500 scenes daily.

• NASA Earth Observatory has mapped record flooding on the Gandak River in Nepal and India, comparing MODIS images from 20 and 29 September. The Devghat gauge in Nepal hit 11.94 m on 27 September, the highest since records began in the 1960s.

• NOAA has awarded $67 million over two years to buy commercial radio occultation and ionospheric data for weather and space weather forecasting. Spire receives $33.16 million, PlanetiQ $28.41 million and Ethereal Space $5.48 million, covering 10,000 radio occultation profiles a day.

• ESA has awarded parallel architecture studies to ICEYE and Leonardo for Europe’s Earth Observation Governmental Service, which is due to start operating in 2028.

  Image: ESA.

Jobs

• Saint Louis University is looking for a Geospatial Research Scientist based in St. Louis, United States.

• VITO is looking for an R&D Professional, AI for Earth Observation based in Mol, Belgium (hybrid).

• The Alan Turing Institute is looking for a Senior / Research Associate, AI for Environmental Resilience based in London, United Kingdom.

• Institute for the Study of War (ISW) is looking for an ArcGIS Geospatial Researcher based in Washington, DC.

Just for Fun

Max Lenormand, Geospatial Youtube legend, is going all in on making videos on remote sensing data. Make sure you check out his stuff - we’ve gotta get behind all folks trying their heart out to demystify this space!

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge, a place where every raster gets the resolution it deserves. Anyway, the aim as usual is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Poverty Targeting: ML maps find 83% of Sri Lanka’s poorest areas.

• Location Embeddings: MIND lets you adjust spatial scale after training.

• Warming Acceleration: Regional warming sped up from the 1980s onwards.

• Old-growth Forests: Spatial validation shrinks embeddings’ edge over Sentinel.

• German Farmland: Annual crop maps from 1990 to 2023.

Subscribe now

Research you should know about

1. Testing whether poverty maps find the poorest places

Machine learning poverty maps are usually judged on R², but a new study from LIRNEasia and Carnegie Mellon University tests whether they actually find the poorest places. Governments using these maps to target aid work with a fixed budget, so what counts is whether the areas they pick really are the most deprived. A model can score well on R² and still get that ranking wrong.

They predicted an asset-based poverty index for all 13,985 of Sri Lanka’s smallest administrative units using mobile phone records and satellite data, then scored the models on how many of the poorest units they recovered. Inputs included 11 call-detail-record features (call volumes, mobility, handset type) and 21 remote sensing features like night-time lights, vegetation, accessibility and Landsat embeddings, with the 2012 census as ground truth. The key metric is Recall. This is basically asking the question ‘of the truly poorest areas, what share does the model pick when you can only target the top k%’.

At a 25% targeting budget, random forest and XGBoost models found 83% of the poorest areas, against 45% for a night-time lights baseline. That said, random train-test splits inflated performance (0.870 vs 0.829 with spatially grouped holdouts), and leave-one-province-out recall ranged from 0.944 in Western Province to 0.461 in Northern Province. The census asset index itself only recovered 68% of the consumption-poorest areas, so matching it doesn’t guarantee you reach consumption-poor households. The authors argue for validating against current consumption data, plus community consultation and routes for appeal, before these maps are used to allocate money.

2. A location embedding you can zoom in and out of

Location embeddings turn any latitude and longitude into a feature vector, but they tend to fall apart far from where they were trained. Geographic measurements are sparse, and standard random splits hide this: in the benchmark used here, 93% of test points sat within 25 km of a training point. A new study led by Taylor Geospatial, with UBC, CU Boulder, UCL, TUM, Arizona State and others, tackles these geographic gaps.

They built MIND, a coordinate model that distils four existing location encoders into one embedding whose spatial detail you can dial up or down after training. A SIREN network maps coordinates to a 3,072-dimensional embedding trained to reconstruct AlphaEarth, Climplicit, GeoCLIP and SINR. It’s supervised in nested chunks (64, 128 and so on up to 3,072 dimensions), so early dimensions capture broad patterns (a characteristic distance of about 790 km) and later ones add finer detail (about 200 km). A “chunked penalty” then down-weights the fine-grained dimensions when you need to extrapolate, with no retraining. They also released CoordBench, 52 datasets with 78 climate, land cover and socioeconomic targets.

MIND does best when predicting far from training data, with an R² of 0.356 on 10° regional holdouts against 0.303 for GeoCLIP. On random folds, simple inverse-distance weighting actually did slightly better (0.675 vs 0.643), which shows how forgiving random splits are. At the hardest 40° holdouts R² drops to 0.184, still ahead of GeoCLIP’s 0.151. The checkpoints and CoordBench are on Hugging Face, with a global 1 km embedding grid on Source Cooperative.

3. Where global warming has sped up since 1970

After record-breaking temperatures in 2023 and 2024, there’s an active debate about whether global warming is speeding up, but a global average can hide what’s happening region by region. A new study in Nature Communications from UC Santa Cruz, Lancaster University and NOAA’s Geophysical Fluid Dynamics Laboratory looks at where and when warming has accelerated.

They fitted piecewise linear changepoint models to annual surface temperatures for every grid cell in five global datasets from 1970 to 2024. This picks up when, and by how much, the local rate of warming changes. They regridded the finer datasets (Berkeley Earth and NASA) to 5° to compare like with like, and focused on patterns that showed up consistently across datasets.

Between 4% and 10% of grid cells show a detectable change in their warming rate, and these are mostly accelerations clustered in the Northern Hemisphere mid-latitudes. The timing varies a lot, from the late 1980s in the eastern Mediterranean (in line with falling aerosol pollution there) to around 2010 in the Gulf of Mexico, southeast Mexico and southeast China, which has the largest acceleration of all. Globally, an acceleration still isn’t detectable in surface temperatures, so the authors argue the debate needs a regional lens. The data and code are on Zenodo.

4. How local climate changes what triggers childhood asthma

Childhood asthma hospitalisations spike at certain times of year, but it’s unclear whether the environmental triggers are the same in different climates. A new study in Scientific Reports from Queensland University of Technology compares three high-risk areas of Australia with very different local climates.

The researchers linked 11,177 child and adolescent asthma hospitalisations from 2003 to 2019 with weather, air pollution and vegetation greenness. They used distributed lag non-linear models to estimate region- and season-specific associations, and random forests to rank which environmental factors mattered most.

Greener vegetation was linked to more hospitalisations everywhere, most strongly in the tropical monsoon and rainforest region (incidence rate ratio 1.65). Warmer temperatures went with a 5–8% lower risk per 1 °C, while air pollution effects depended on region and season: SO₂ had the largest effect, in the tropical savannah wet season (IRR 2.11). Greenness and temperature came out as the leading predictors in every climate, which points towards region- and season-specific asthma surveillance.

5. Embeddings lose their edge for old-growth forest under proper validation

Finding Europe’s remaining old-growth forests matters for conservation, and geospatial foundation model embeddings look like an easy win, but a new study from the University of Cambridge shows part of their advantage comes from how models are validated. The team worked in Romania’s Făgăraş Mountains, building 4,825 high-confidence labelled forest parcels (1,016 old-growth and 3,837 confirmed non-old-growth) with Fundaţia Conservation Carpathia.

They compared AlphaEarth and TESSERA embeddings with standard Sentinel-1/2 features, using spatially blocked cross-validation with and without a 10 km buffer between training and test parcels. The buffer distance came from the spatial autocorrelation in the labels (a semivariogram range of about 6.6 km). Each feature set was added to a baseline of terrain and access variables like slope and distance to roads, and they tested both pixel-based XGBoost and CNNs.

Without a buffer, TESSERA led clearly (PR-AUC 0.84 vs 0.76 for both Sentinel features and AlphaEarth), but with a 10 km buffer the gap shrank to 0.03–0.04 and was no longer significant. Satellite features of any kind still added 0.21–0.25 PR-AUC over terrain and access alone, and CNNs didn’t beat pixel-based models at 10 m. The final map flags 55,025 ha (36% of the forested area) as old-growth. The code is on GitHub.

Geospatial Datasets

1. Annual crop maps for Germany, 1990 to 2023

These annual agricultural land-use maps of Germany cover 13 crop classes plus grassland at 30 m for every year from 1990 to 2023, built from Landsat and Sentinel-2 with a deep learning model trained on farmers’ declarations. You can access the data here.

2. The shape of the daily temperature cycle, everywhere on land

This diurnal temperature cycle dataset gives monthly parameters describing how temperature rises and falls through the day (harmonic amplitudes and phases, daily range, and the hours of minimum and maximum) over land at 0.1° for 1990–2024, fitted to ERA5-Land. You can access the data here and the code here.

3. Where China’s grazing livestock are, season by season

CLRD-GLPS maps cattle, sheep and goats across China’s grazing systems at 1 km for 2000–2021, with separate warm-season and cold-season distributions and uncertainty layers. You can access the data here.

4. A benchmark for county-level crop yield prediction

AgroBench is a US crop yield benchmark with over 13 million pixel observations for corn, soybeans, wheat, cotton and sorghum across 5,107 county-years from 2017 to 2024, combining Sentinel-2, Sentinel-1, MODIS land surface temperature, CHIRPS rainfall and SRTM elevation. A random forest baseline reaches R² 0.58 under leave-one-year-out validation. You can access the code here, with the processed data due on Zenodo.

5. US forest attributes linked to field plots

TreeMap 2023-1 from the US Forest Service maps 22 forest attributes (forest type, canopy cover, live and dead biomass, carbon, stand height and more) at 30 m across the contiguous US for 2016, 2020, 2022 and 2023. Each pixel is matched to the most similar Forest Inventory and Analysis plot, so you can pull hundreds of extra variables using the plot ID. You can access the data here.

Other useful bits

• Earth Fire Alliance and Muon Space have released the first images from the FireSat constellation, whose three satellites reached orbit on 7 July and can detect fires as small as 5 × 5 m. First light includes the 290,000-acre Big Grass Fire on the Oregon–Idaho border. Early adopters get twice-daily data by the end of 2026, with hourly global revisit planned by 2029.

• The Copernicus Data Space Ecosystem’s openEO service now serves 109 Copernicus Land Monitoring Service collections, plus nine new Sentinel-5P Level-2 products covering CO, NO₂, methane, SO₂, formaldehyde, ozone, aerosol layer height, aerosol index and cloud. You can load, filter, process and export them in the same openEO workflows as the other CDSE collections.

• USGS has reprocessed about 44,000 Landsat MSS products from Landsat 2 and 3 after finding wrong gain and offset values in Bands 4 and 5 that made scenes look too bright or too dark. Around 1,700 products also had a solar elevation bug that wrongly flagged roughly 27,000 extra pixels as cloud, so if you work with 1970s Landsat, check the processing date in your product IDs.

• NASA has mapped 1,943 valleys hidden beneath Greenland’s ice, about a third of them newly identified, by reading subtle bumps and dips in satellite data of the ice surface. About half of the previously known valleys turn out to extend much further inland, and the work improves the BedMachine Greenland dataset.

  Image: NASA Earth Observatory.

• Hydrosat has unveiled its Osiris thermal constellation, with the first two satellites launching in 2027. Each carries a 25 m thermal sensor (three longwave infrared bands and a 160–340 km swath) plus a 9-band visible and near-infrared sensor at 10 m.

Jobs

• Yale University is looking for a Lead, Geospatial Artificial Intelligence and Machine Learning based in New Haven, United States.

• Planet is looking for a Data Scientist based in Haarlem, Netherlands (hybrid).

• Moody’s is looking for an Analytics & Modeling Analyst, Coastal Flood based in London, United Kingdom.

• Forschungszentrum Jülich is looking for a Research Software Engineer, AI and Earth Observation based in Jülich, Germany.

Just for Fun

This composite of the Harvest Moon rising was taken near Nicosia in Sicily, with ash and gas from Mount Etna drifting in the background. The lower clouds reflect the colours of anti-twilight, and the pink band low in the sky is the Belt of Venus. At full moon the Moon and Sun sit in opposition, so the Moon rises just as the Sun sets.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge, a place where no polygon is ever left unclosed. Anyway, the aim (as usual) is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Displacement Settings: Satellites estimate living standards around refugee camps.

• Forest Carbon: Central Africa’s forests lose 58 Tg carbon yearly.

• Neolithic China: Terrain shaped where early settlements clustered.

• Methane Plumes: Google maps methane plumes from EMIT imagery.

• Overture Maps: Places data grows to 81.5 million features.

Subscribe now

Research you should know about

1. Estimating living standards in displacement settings from space

If you’re in the field of economic geography, then you’ve probably come across heaps of satellite-based poverty and wealth maps. This one (from researchers at UNHCR, Chalmers University and the World Bank) is a bit different though, since it’s focussed specifically on forced-displacement settings. For these groups, welfare can change pretty quickly between surveys, and surveys can obviously be quite costly and infrequent. Camps and settlements also look pretty different from the towns and villages these models learned from.

They took a vision transformer pretrained on 1.2 million DHS households across 36 African countries and adapted it to UNHCR survey data from South Sudan, Cameroon and Zambia. The inputs were Landsat imagery, VIIRS night-time lights and Open Buildings footprints on 6.72 km grid cells. They froze the early layers and fine-tuned the rest on the Forced Displacement Survey (2023–2025) and the Results Monitoring Survey (2024). A sliding window boosted the number of training cells in dense camps from 608 to 1,123, and testing was spatially blocked so test regions were never seen in training.

Out of the box the model failed (R² near zero), but after fine-tuning it reached an R² of 0.40 overall and 0.66 in grid cells overlapping camps. Performance varied a lot by country: 0.62 in South Sudan, 0.44 in Zambia and 0.28 in Cameroon. Projected forward, the model suggests South Sudan’s mean socioeconomic index fell 27% between 2023 and 2025. Anyway, the authors pitch it as a way to flag areas that are getting worse for field checks between survey rounds, rather than a replacement for surveys. This is ultimately the approach I think most people take with deriving remote-sensing based insights on socio-economic issues.,,

2. Synchronised global heat is at a 1,000-year high

So heat extremes that hit several regions at once can do a lot more damage than isolated ones. Interestingly, most prior research has focused on northern summers. So a new study in Nature Communications led by Beijing Normal University looks at globally synchronous warm seasons across all four seasons, going back to 850 CE.

The team built an Extremity Index that combines how intense local land temperature anomalies are with how much land crosses a warm threshold at the same time. They calculated it from instrumental records, paleo-reanalysis products and climate model simulations, then ran detection and attribution analyses to separate natural, greenhouse gas and aerosol influences. The index catches things plain averages miss: in the 2009/10 northern winter, the global land temperature anomaly ranked only 22nd since 1901, but the Extremity Index ranked it 3rd.

These synchronous warm events have intensified sharply since the 1970s in every season, reaching levels not found earlier in records going back to 850 CE. Before 1970, only 18% of December–February seasons and 15% of June–August seasons made the index’s top 5%; since 1970, it’s 89% and 86%. Greenhouse gas forcing is the main driver, and the tropics contribute a lot because their naturally low variability makes warm anomalies stand out. The data and code are on Zenodo.

3. Central Africa’s forests are losing carbon

Central Africa’s dense humid forests hold vast carbon stocks, but they’re increasingly hit by small-scale disturbances that current satellite observations struggle to pick up. That makes it hard to say how much carbon the region is actually gaining or losing, and where. A new study in Nature Communications led by LSCE in France tackles this with much finer maps.

The team used deep learning to map canopy height at 10 m for every year from 2019 to 2022, then turned those maps into 30 m estimates of biomass change. The model combines spaceborne lidar with Sentinel-1 radar and Sentinel-2 optical imagery. They then built carbon budgets for each country and compared them with national inventories and bookkeeping model estimates.

The region’s forests lost a net 58 ± 14 Tg of carbon a year, and small patches did a lot of the damage. 87% of disturbance patches are smaller than 1 hectare, yet they account for 37% of carbon losses from deforestation and 48% of carbon gains in regrowing forests. Gross losses of 126 Tg C a year were only partly offset by 68 Tg C of gains. The Democratic Republic of the Congo is a net source of 46 Tg C a year (79% of the regional loss), mainly through degradation, even though it also has the largest gains in young secondary forests. The estimates broadly line up with country-level inventories, which is a good sign for transparent, map-based carbon monitoring.

4. How terrain shaped settlement in Neolithic China

Cities grew out of settlement networks that existed long before formal roads, states or urban planning, but we know surprisingly little about how those early networks were organised. A new study in Nature Cities from the London School of Economics and colleagues in China looks at how terrain shaped settlement across Neolithic China between 5000 and 2000 BCE.

They combined a newly compiled set of 3,587 georeferenced archaeological sites with a terrain-accessibility index derived from topography. The index captures how easy it is to move overland given the shape of the landscape. They counted sites in 0.25° grid cells, regressed settlement intensity on accessibility across four Neolithic phases, and tracked clustering with Moran’s I and nearest-neighbour distances. The results held after controlling for distance to major rivers.

Settlements clustered more and more along accessible corridors until about 2500 BCE, and then that link weakened sharply. In the core agricultural region, the accessibility coefficient rose from 0.413 in the Early Neolithic to 0.470, then dropped to 0.358 in the final phase (2500–2000 BCE). Across all of China it roughly halved, from 0.163 to 0.083. Clustering kept getting stronger (Moran’s I rose from 0.24 to 0.33), but the main concentration of sites moved into intermediate-accessibility areas. So terrain structured early settlement, but that relationship shifted as societies grew more complex. The data and code are on Zenodo.

Geospatial Datasets

1. A global map of mobile coverage from 1999 to 2030

This mobile network coverage raster maps 2G, 3G and 4G coverage at 1 km across 214 countries and territories from 1999 to 2030, with modelled estimates to 2020, predictions for 2021–2024 and extrapolations to 2030. It combines three models tuned on 2,409 operator-reported coverage maps, and the machine-learning model reaches an AUC of 0.89–0.92 on held-out countries. You can access the data here.

2. US evapotranspiration back to 1985

OpenET has added monthly evapotranspiration estimates for 1985–1999 across the contiguous US at 30 m, using Landsat 5 and an ensemble of six models. You can access the data here, with each individual model also available in the Earth Engine catalog.

3. Methane plumes detected from space

Google Research’s MAPL-EMIT plume dataset uses deep learning to detect methane plumes in hyperspectral imagery from NASA’s EMIT instrument on the International Space Station, at 60 m globally from August 2022 to June 2026. You can access the data here.

4. Where cocoa, coffee, palm and rubber are grown

Google and the Forest Data Partnership have released a 2026 version of their commodity probability models for cocoa, coffee, palm and rubber. The cocoa layer gives the probability that each 10 m pixel contains cocoa trees across the tropics (24°S to 24°N) for 2018–2024, and is designed to help track commodity-driven deforestation. You can access the data here and the documentation on GitHub.

5. City-scale 4G and 5G measurements for Vienna

The Vienna 4G/5G Drive-Test Dataset combines georeferenced LTE and 5G measurements from wideband scanners and phones across Vienna, with estimated base station locations, sector azimuths and antenna heights, plus building and terrain models for ray-tracing work. You can access the data here.

Other useful bits

• Planet has launched Amazon.ia, an initiative combining satellite imagery, ground data and AI to monitor biodiversity across the Amazon, backed by more than $14 million from the Bezos Earth Fund. Interestingly for me though, the website doesn’t actually work. So not sure what’s up with that… Anyway, more than two dozen partners, including Microsoft’s AI for Good Lab, MIT, Cornell and WWF, will produce change alerts, ecosystem condition assessments and evidence of protected-area effectiveness.

• Overture Maps’ September release is out, and it comes with schema v2.0.0 and a breaking change: places now use taxonomy (or basic_category), and the old categories property is gone. Places grew 10.6% to 81.5 million features as BrightQuery coverage expanded into Germany, Italy and Denmark, and buildings now total 2.53 billion features.

• Europe has launched Copernicus Sentinel-3C, which lifted off on a Vega-C from Kourou on 14 September to join Sentinel-3A and 3B. It measures sea temperature, sea level, ocean colour and land temperature, which feeds into monitoring of heatwaves, wildfires and marine pollution.

  Image: EU Agency for the Space Programme.

• Planet has opened a satellite factory in Berlin, a 5,700 m² site that will build its Pelican satellites and add 70 jobs. The first Berlin-built satellites are due to launch on an Isar Aerospace rocket from Norway in early 2027.

  Image: Planet, via Via Satellite.

• Orbify’s warped 3D map demo bends the scene so you can see nearby and distant features at the same time, which could be handy in navigation apps. The Oslo-based company’s warping technology is patent-pending, and the demo runs in the browser with WASD controls.

Jobs

• ABPmer is looking for a Principal Geospatial Consultant based in Southampton, United Kingdom (hybrid).

• Utrecht University is looking for a Postdoc in AI-driven Urban (Re)Design for Health based in Utrecht, Netherlands.

• Utrecht University is looking for a PhD candidate in Spatial Optimization of Land Use in Brazil based in Utrecht, Netherlands.

• National Biodiversity Data Centre is looking for a GIS Officer based in Waterford, Ireland.

Just for Fun

Omega Centauri packs about 10 million stars, all much older than the Sun, into a ball roughly 150 light-years across. At 15,000 light-years away, it’s the largest and brightest of the 200 or so globular clusters in the Milky Way’s halo. Unusually, its stars span a range of ages and chemical make-ups, so it may be the leftover core of a small galaxy that merged with ours.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge, a place where every pixel gets its fair share of ground truth… Anyway, the aim as usual is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Forecast Gaps: Poorer countries get far less accurate weather forecasts.

• Roman Roads: Network science measures the Empire’s 299,000 km network.

• Model Calibration: Geospatial foundation models turn overconfident under clouds.

• WeatherNext 3: Google’s AI weather model now updates hourly.

• Power Outages: Ten years of US outages matched to warnings.

Subscribe now

Research you should know about

1. Why weather forecasts are worse in poorer countries

A new study in Nature Communications finds that weather forecasts are much less accurate in low-income countries, which is kinda a big deal given how much we lean on forecasts to adapt to climate change. Farmers use em to decide when to plant, governments use em to issue heat warnings and energy firms use em to plan supply. But a forecast is only as good as the observations feeding it, and those observations aren’t spread evenly around the world. Anyway, researchers from Columbia University set out to measure how big the gap actually is.

They compared global temperature forecasts against what was actually observed, country by country, going back to 1985. The main forecasts come from ECMWF, verified against weather station records for lead times of one to seven days. They grouped countries by tge World Bank’s income level and checked the pattern held using NOAA’s GFS forecasts and model analysis instead of stations. They also mapped the observing infrastructure itself, which included stuff like land-based weather stations, balloons, etc.

They found that a seven-day forecast in a high-income country is roughly as accurate as a one-day forecast in a low-income country. Accuracy has improved quite a bit since 1985, but the gap between richer and poorer countries hasn’t closed. Part of this is down to some regions simply being harder to forecast. The rest is infrastructure: lower-income countries have fewer stations and radiosondes, the ones they have report less often, and they appear to have less institutional capacity to issue official local forecasts. The upshot is that investing in basic monitoring could narrow the gap.

2. Measuring the Roman Empire’s road network

The Roman road network is the classic example of ancient infrastructure shaping where we live today, but its structure had never been measured across the whole Empire. At its height in 117 CE, the Empire covered almost 5.5 million km² across Europe, the Middle East and North Africa. Most of what we “know” about its roads (that they were straight, that all roads led to Rome and that modern roads follow them) comes from qualitative or regional studies.

A team led by Aarhus University used a detailed digital model of 299,171 km of Roman roads to test these assumptions with terrain analysis and network science. They split the roads into 1 km segments and calculated:

• slope, topographic position and sinuosity (my favourite word)

• road density, and how it correlates with modern roads, ancient sites and today’s population

• travel-time-weighted betweenness centrality using Tobler’s hiking function, (which basically captures how many of the quickest overland routes pass through each segment).

Roman roads are straight where the terrain allows, but still less straight than modern roads, and Rome itself wasn’t the best-connected hub despite pretty much everything you’ve heard about all roads leading there. About 91% of roads stay below a 16% gradient (the upper end of estimates for wheeled traffic) and the median slope is 3.88%. Modern major roads have a median sinuosity of 1.0026, against 1.0035 for Roman main roads and 1.0051 for secondary ones. Roman road density correlates with modern major roads, and 24 of 45 provincial capitals sit in the top two deciles for betweenness. For overland travel, Antioch, Ancyra, Byzantium and Corinth were all better placed than Rome, whose central position came from sea and river routes.

3. Geospatial foundation models get overconfident when imagery is messy

Geospatial foundation models are usually judged on accuracy alone, but a new study from TUM, DLR, MIT and Taylor Geospatial argues we also need to know whether their confidence can be trusted. This is called calibration: when a model says it’s 90% sure, it should be right about 90% of the time. Real imagery is messy (clouds, sensor noise, blur) and a model that’s confidently wrong is a real problem in operational settings.

They tested 16 frozen encoders on 4 classification and 5 segmentation datasets while deliberately degrading the imagery. The line-up included EO-pretrained models like Clay, Prithvi, DOFA, TerraMind and OLMo-Earth alongside general-purpose ones like ResNet, Swin and DINOv3. They then (1) added clouds and shadows, sensor noise and motion blur at five severity levels, (2) cut the training data from 75% down to 1%, and (3) tried three fixes: temperature scaling, deep ensembles and Gaussian-process probes.

On clean data the two groups are indistinguishable, but under distribution shift the EO-pretrained models drift further into overconfidence. At the heaviest cloud level, expected calibration error ranged from 0.32 to 0.82 across encoders. Among high-drift predictions, EO models were confidently wrong about 2.4 times as often (0.14 vs 0.06). Rankings reshuffled too: the rank correlation between clean and cloudy calibration was just τ = 0.07. Temperature scaling didn’t help, and Gaussian-process probes roughly halved the error under heavy cloud but tripled it on clean data. The practical takeaway is to test models across several conditions and metrics before trusting a leaderboard.

4. Using satellite embeddings to map neighbourhood liveability

A new study from the University of Illinois Urbana-Champaign tests whether satellite foundation model embeddings can improve maps of how liveable neighbourhoods are. Socioeconomic indicators are hard to measure in places with little survey data, while satellite imagery is available almost everywhere. Embeddings like Google’s AlphaEarth compress that imagery into ready-made features, but it’s not obvious how much they add on top of the usual inputs.

The team added AlphaEarth, AnySat and TerraMind embeddings to a multimodal transformer that predicts the Netherlands’ Leefbaarometer liveability scores on a 100 m grid. The baseline model already used 2 m imagery, a surface model, night-time lights and points of interest (POIs). It predicts overall liveability plus five dimensions: physical environment, housing, amenities, social cohesion, and nuisance and insecurity. They trained on nine Dutch cities and tested on four others: Eindhoven, Hengelo, Dordrecht and the rural municipality of Beesel.

AlphaEarth gave the most consistent gains, cutting overall liveability error (RMSE) from 0.106 to 0.098, with amenities improving the most. The embeddings helped most where data was patchy. When the aerial imagery was removed at test time, the model with all three embeddings had an error of 0.108 against 0.156 for the baseline. There’s a catch though: in rural Beesel a single embedding made things slightly worse (0.082 vs 0.078), and only combining all three recovered performance (0.073). The code is on GitHub.

5. Shrinking a cropland model 158 times for edge devices

Big vision models are great at mapping cropland from satellite imagery, but they’re far too heavy to run on satellites or devices in the field. Distillation (training a small “student” model to mimic a big “teacher”) is the usual fix. However, most methods assume the two models have similar architectures and stop aligning their features once task training begins.

Researchers at UC Riverside built JEDI, a two-stage method that distils a 639M-parameter I-JEPA vision transformer into a compact SegFormer. First, they project and spatially align the student’s final features to the teacher’s token space, even though the architectures differ. Then they train the student on segmentation while it continues to (1) match the teacher’s softened predictions and (2) keep its features aligned with the teacher’s. They tested it on CalCROP21, a Sentinel-2 cropland dataset from California’s Central Valley.

The smallest student, with just 4.04M parameters, reaches 68.0 mIoU, only 2 points behind the teacher while being 158 times smaller. That’s a 16-point jump over the same student trained without distillation (52.0). It also runs in 7.1 ms per image against 83.5 ms for the teacher. The code is on GitHub.

Geospatial Datasets

1. Ten years of US power outages matched to weather warnings

This harmonised outage and warnings dataset links county-level customer power outage records from the US Department of Energy’s EAGLE-I platform with National Weather Service watches, warnings and advisories for the contiguous United States from 2015 to 2024, indexed hourly. It also includes a consolidated warning archive and daily summaries of missing outage records. You can access the data here.

2. Daily travel between Chinese cities, 2020 to 2025

DIMNet-CPC estimates directed daily trip volumes between 366 Chinese administrative units from 2020 to 2025 by calibrating the Baidu Migration Index against Ministry of Transport totals. Validation errors (WAPE) range from 4.74% at the national level to 26.52% for individual origin–destination pairs. You can find the paper here.

3. Disaster impacts in the Global South from Red Cross reports

ROUGE is a database of non-monetary disaster impacts extracted with large language models from IFRC operational reports, with detail down to subnational level. The EGU version covered 11,370 impacts from 717 reports spanning 2016–2025 across 20 impact types. You can access the code here.

4. Hourly energy demand for 1.69 million Belgian buildings

This urban energy atlas estimates 8,760-hour heat and electricity profiles for about 1.69 million buildings in Wallonia, Belgium, aggregated to 262 municipalities, 20 arrondissements and 5 provinces, and served through an interactive Dash web app. You can find the paper here.

5. Weather grids paired with forecasters’ reasoning

Grid2Text pairs ERA5 meteorological grid data with expert-verified forecast text, including reasoning chains for temperature trends, wind shifts, humidity ranges and precipitation types. It was built by Fudan University and the Shanghai Central Meteorological Observatory for training and benchmarking language models that write forecast discussions. You can find the paper here.

Other useful bits

• Google DeepMind has launched WeatherNext 3, an AI weather model that learns directly from live geostationary satellite data and produces a new forecast every hour. Surface variables like temperature come at 5 km resolution (down from 25 km in WeatherNext 2), and Google reports medium-range precipitation gains of up to 60% when checked against NASA’s IMERG. It’s available in Earth Engine, BigQuery and the Maps Platform Weather API.

  Image: Google.

• Google’s Places Insights historical data is now generally available in BigQuery, with monthly snapshots of place counts going back to January 2024 across 480+ place types. A new PLACES_COUNT_CHANGE function compares two months in a single query, which is handy for tracking openings, closures and turnover.

  Image: Google Maps Platform.

• IBM and NASA have released an open-source lunar foundation model on Hugging Face, trained on over 30 aligned data layers from nine instruments, including NASA’s Lunar Reconnaissance Orbiter and GRAIL missions plus Japan’s Kaguya. It beats an ImageNet-trained baseline by up to 22% at finding likely ice deposits and 19% at crater detection.

  Image: IBM.

• Satellogic has made SynMax the exclusive maritime intelligence channel for its upcoming Merlin constellation, which will remap the planet daily at 1 m-class resolution and co-collect AIS ship-tracking data. The first launch is planned for October 2026, and SynMax’s Theia platform will use the data to track dark vessels and illegal fishing.

• Google Research has released TimesFM-3, a 330M-parameter time-series foundation model pre-trained on over a trillion time points that now handles multivariate forecasting and covariates. The weights are on GitHub and Hugging Face.

Jobs

• EUMETSAT is looking for a Machine Learning Application Developer based in Darmstadt, Germany.

• UNITAR is looking for a Geospatial Analyst – Satellite Imagery and GIS based remotely.

• Laterite is looking for a Developer based in Amsterdam, Netherlands.

• Environmental Investigation Agency is looking for a Climate Data Analyst based in Washington, DC, United States.

Just for Fun

This is the Andromeda Galaxy before any clean-up: a stack of 223 five-minute exposures taken from a garden observatory in Portugal in 2019. The streaks and specks are airplane trails, satellite trails, cosmic rays and bad pixels. Anyone who’s cloud-masked a Sentinel-2 scene will relate, since the polished versions come from reducing these Earthly artefacts with software like DeepSkyStacker and PixInsight.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge, a place where every join is a spatial join. Anyway, the aim as usual is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Map Making: A field guide to turning ML models into maps.

• Slum Mapping: Radar texture separates informal settlements from formal neighbourhoods.

• Foundation Models: ESA pits TerraMind against THOR on ten tasks.

• Generalist VLMs: A general model beats specialist designs with 80,000 samples.

• SAR Reasoning: FUSAR-R1 reasons step by step over radar imagery.

Research you should know about

1. How to turn a machine learning model into a map

Image: Sialelli et al., arXiv

A new paper led by ETH Zurich, with co-authors from Cambridge, WSL, Asterisk Labs and a dozen other groups, sets out best practices for getting from a trained model to a large-scale Earth observation map, and it’s the guide a lot of us wished existed a few years ago. The barrier to producing a wall-to-wall map has dropped a long way, but there’s still no agreed playbook, and decisions made early in the pipeline quietly propagate errors into the final product. It was presented at the ECCV 2026 TerraBytes workshop.

The paper walks through six coupled stages: the data infrastructure landscape, data selection and preprocessing, dataset construction and training, uncertainty quantification, map production and distribution, and validation. The details are where it earns its keep. Google Earth Engine corrects the Sentinel-1 GRD radiometric offset while ESA redistributes the product as is, so switching providers shifts your input distribution. Sentinel-2’s January 2022 processing baseline change, incomplete orbits and Landsat 7’s scan-line gaps all show up as seams if you don’t handle them. Whatever preprocessing you use in training has to be reproduced over the whole mapped domain at inference, so every extra step compounds deployment cost. Spatial autocorrelation inflates apparent accuracy unless you split with blocked cross-validation and buffers wider than the autocorrelation range.

On uncertainty, the authors are blunt that no method is simultaneously cheap, scalable and well calibrated, and that deep ensembles trained on the same spatially autocorrelated data tend to agree where they are collectively wrong. Convolutional encoder-decoders remain the workhorse for wall-to-wall inference because of throughput, even as foundation models gain ground. And a held-out test set only tells you about generalisation within your sampling distribution, so a real map needs a probability-based validation sample where every location has a known, non-zero chance of being checked, with accuracy reported from both the user’s and the producer’s side. There’s an expanded online guide that goes with the paper.

2. Mapping informal settlements with radar texture

Image: Chepkilot, Memar & Gamba, arXiv (Fig. 1)

A new study from Sapienza University of Rome and the University of Pavia tackles a long-running problem in mapping Sub-Saharan African cities: from optical satellites, informal settlements look almost identical to formal, compact low-rise neighbourhoods. Both are roofed in the same oxidised corrugated metal, so reflectance-based classifiers confuse the two. The WUDAPT Local Climate Zone baseline mixes up the two classes 36.4% of the time.

The fix is to bring in Sentinel-1 radar in three tiers: calibrated VV and VH backscatter, grey-level co-occurrence textures over 3 by 3 pixel windows, and a physics-guided index that normalises VH entropy by mean backscatter power to pick out the high-disorder, low-return signature of lightweight low-rise housing. These sit alongside 18 Sentinel-2 indices in a 28-feature random forest, trained on 793 reference polygons in Nairobi and 492 in Eldoret (426,650 training pixels), with separate dry and wet season composites, and then tested on Kigali.

Adding the texture tier lifts overall accuracy from 70.4% for the WUDAPT baseline to 81.6% in the dry season, and the informal settlement F1 score more than triples, from 0.200 to 0.667. It also makes the map far less sensitive to season: the dry-wet F1 gap narrows from 0.084 with optical only to 0.004 with texture, and class confusion drops to 7.3%. The transfer test is the sobering part. Applied zero-shot to Kigali, overall accuracy falls to 0.363, and only with local labels added does the informal settlement F1 reach 0.720, so the features travel but the model still needs some labels from the target city.

3. ESA puts two Earth observation foundation models head to head

Image: Schindlegger et al., ESA Φ-lab, arXiv (Fig. 7)

A new study from ESA’s Φ-lab, IBM Research Zurich and the Norwegian Computing Center systematically compares TerraMind and THOR, the two geospatial foundation models ESA funded through separate consortia, across ten downstream tasks. TerraMind is a multimodal generative encoder-decoder pretrained on 500 billion tokens across nine modalities with a fixed patch size of 16. THOR is a vision transformer that accepts variable patch sizes from 4 to 32 at inference and unifies Sentinel-1, 2 and 3 at native resolutions.

The benchmarks cover flood and burn scar segmentation, artisanal mining, methane leaks, soil property regression, iceberg detection, sea ice, snow and wetland mapping, with ablations over patch size, decoder type and frozen versus fine-tuned encoders, all run on a single RTX A6000. The headline pattern is that THOR leads on SAR-heavy and spatially compact targets while TerraMind leads on optical and multimodal ones, though the fine-grained results are more interesting than the leaderboard.

At matched compute TerraMind wins Sen1Floods11 with 0.910 mIoU against THOR’s 0.894, but THOR at patch size 4 hits an F1 of 0.86 on flood zone detection where TerraMind manages 0.33, and 0.873 mIoU on sea ice against 0.755, at 16 to 44 times the encoder cost. Patch size and decoder choice together explain more performance variance than which model you picked. Sentinel-2 alone beat Sentinel-1 plus 2 on Sen1Floods11 (0.910 versus 0.896), suggesting shallow fusion isn’t enough when optical signal is abundant, and THOR at 5% of the training data matched a UNet trained on 50%. The code is on GitHub.

4. A general-purpose VLM beats the remote sensing specialists

Image: Ailuro et al., INSAIT, arXiv

A new study from INSAIT at Sofia University, with Luc Van Gool among the authors, asks whether remote sensing vision-language models actually need custom encoders, alignment modules and task-specific fusion, or whether a capable general model with a good data recipe does the job. Most remote sensing VLMs lean on specialised architecture, which makes them hard to maintain and slow to benefit from progress in the general-purpose models.

The recipe is deliberately plain: take InternVL3.5 at 8B and 2B parameters, give it SAM3 as a segmentation tool, curate 80,000 training samples from a 2.3 million sample pool balanced across VQA, captioning, detection, segmentation, optical and SAR, multi-temporal, multi-view and very high resolution inputs, then fine-tune with LoRA for 5,000 steps using a group relative tool optimisation objective with task-dependent rewards. No remote-sensing-specific architecture anywhere.

On ten out-of-distribution benchmarks the model ranks first on seven and second on three, with detection precision at 0.5 IoU on GEOBench-VLM rising from 0.24 to 0.56, segmentation IoU on GeoSeg-Bench2 from 0.41 to 0.70, and a 7.6 times relative gain on XLRS-Bench detection. In-distribution it’s competitive or state of the art on DisasterM3, DynamicVL, SARLANG-1M and LaSeRS. The finding worth remembering is that performance tracked the diversity of data sources rather than sample volume: domains with three or more sources kept improving, while temporal VQA with two sources peaked early and then declined.

5. Teaching a model to reason step by step over radar imagery

Image: Yang et al., Fudan University, arXiv

A new study from Fudan University introduces FUSAR-R1, a vision-language model for synthetic aperture radar that shows its working rather than jumping straight to an answer. Existing SAR VLMs output a label with no intermediate reasoning, no logical check and no way to self-correct, which matters when the physics of radar scattering is what separates a ship from a bright rock.

The team rebuilt the FUSAR-GEOVL-1M dataset so that each single-step description becomes a multi-step reasoning chain that folds in geographic context, scale priors, scattering feature analysis and spatial verification, with aircraft detection broken into seven steps and ship detection and land cover classification into six each. Training is a cold start of supervised fine-tuning on those chains followed by group relative policy optimisation with rewards for format, IoU-based detection, counting, classification and land cover recognition.

Against InternVL3.5-4B, FUSAR-R1 lifts object counting accuracy from 45.45% to 67.33%, main land cover category recognition from 44.23% to 85.24% and regional land cover from 6.55% to 64.75%, while cutting the land cover proportion error from 39.58 to 7.67. The ablation is the useful bit for anyone training these things: cold start plus reinforcement learning reaches a detection F1 of 0.676, cold start alone 0.600, and reinforcement learning on its own collapses to 0.197.

Geospatial Datasets

1. Every tree in New York City, by genus

Image: NYC Tree Genus Map (Miller et al., Cornell University)

A Cornell team has classified 1.8 million individual tree crowns across New York City into 18 common genera, which together make up about 78% of the city’s basal area, using an XGBoost classifier trained on time series of 3 m PlanetScope imagery and airborne lidar structure, with labels extrapolated from the city’s street tree inventory. Overall accuracy is 82.0% (kappa 0.794), with Platanus at 95% producer’s and 93% user’s accuracy and Quercus at 88% and 80%, and around 1.4 million of the trees had never been surveyed before because they sit on private land or in natural areas. You can read the paper here, explore the interactive map here, access the data here and the code here.

2. Global evapotranspiration and GPP from 1982 to near present

Image: Earth System Science Data (Fig. 1)

PML-V2.2 is a coupled evapotranspiration and gross primary production record in three flavours: 500 m 8-day from MODIS and VIIRS for 2000 to 2025, 0.1° half-monthly from AVHRR for 1982 to 2020, and a consolidated 0.1° series running from 1982 to 2025. The Penman-Monteith-Leuning model was calibrated at 208 flux sites (about 1,400 site-years) and separates irrigated from rainfed cropland, scoring NSE of 0.72 for ET and 0.76 for GPP at sites and 0.89 to 0.91 against basin water balances. You can read the paper here, access the data here or in Earth Engine, and the code here.

3. Daily temperatures from 17,000 stations on a common UTC day

Image: Earth System Science Data (Fig. 1)

GLBD-FED provides daily maximum, mean and minimum temperature for roughly 17,000 weather stations worldwide from 1981 to 2024, all aligned to a 00:00 to 24:00 UTC window rather than each country’s local observing day. A reconstruction algorithm breaks the sub-daily Integrated Surface Database records into finer intervals and re-aggregates them, which recovers 64% more Tmax and 45% more Tmin values than the conventional approach, and daily coverage grows from about 3,000 stations in the 1980s to around 10,000 today. You can read the paper here and access the data here.

4. Corrected lightning events from China’s FY-4A satellite

Image: Earth System Science Data (Fig. 1)

A geolocation-corrected event dataset from the Lightning Mapping Imager on FengYun-4A, covering the March to September warm seasons of 2019 to 2023 over China, Mongolia, Kazakhstan, Southeast Asia and surrounding seas at 7.8 km nadir resolution. Using the World Wide Lightning Location Network as reference, a spline correction applied across 400 subregions at 10-minute intervals raises the share of events within 15 km of their true position from 36.0% to 51.7%, and within 20 km from 58.2% to 74.8%. You can read the paper here and access the data here.

5. Seventeen years of solar-induced fluorescence from GOME-2

Image: Earth System Science Data (Fig. 1)

A combined solar-induced chlorophyll fluorescence record from the GOME-2A (2007 to 2017) and GOME-2B (2013 to 2023) instruments, gridded at 0.5° between 60°S and 60°N from native 40 km pixels, using the SIFTER v3 algorithm with time, wavelength and scan-angle dependent degradation corrections so the two sensors line up. You can read the paper here and access the data for GOME-2A here and GOME-2B here.

Other useful bits

• NASA has issued the request for proposals for the Landsat 10 spacecraft, with launch no later than 2031. The observatory will carry a single integrated sensor suite called LandIS, fly at 653 km with an 18-day global revisit, and catalogue scenes under a new WRS-3 reference system. Proposals closed on 13 August, the award is expected on 30 December 2026, and there’s a USD 1 million incentive for early delivery.

• Rocket Lab has launched the ninth QPS-SAR satellite for Japan’s iQPS into a 575 km orbit from New Zealand, on Electron’s 93rd flight. iQPS is building out a synthetic aperture radar constellation aimed at 24 satellites by May 2028, and still has nine more dedicated Electron launches booked to fill it out by 2030.

• A Kyoto University researcher has used public Starlink orbital data to map the upper atmosphere, treating roughly 1,200 satellites at 482 km as drag sensors and applying tomography to infer thermospheric density, then checking the result against ESA’s Swarm mission. The paper in Earth, Planets and Space calls it a preliminary, low-resolution first map from a single week of data, but it’s a neat repurposing of a constellation nobody launched for science.

• NASA’s Lunar Reconnaissance Orbiter has imaged the crater left by a Falcon 9 upper stage that hit the Moon on 5 August. Before-and-after shots from 11 and 12 August show a crater about 18 metres wide and under 3 metres deep, with dark rays of space-weathered surface material and brighter ejecta dug up from deeper down. Impact prediction tools, cross-checked with Korea’s Pathfinder Lunar Orbiter, got the location right to within about a kilometre.

Jobs

• UNITAR is looking for an Individual Contractor – Geospatial Analysts – Satellite Imagery and GIS based remotely.

• Esri is looking for an Imagery Product Engineer II for imagery and remote sensing based in Redlands, California.

• Esri is looking for a GIS Product Engineer II for the ArcGIS Maps SDKs for Native Apps based in Cardiff, UK.

• Mapbox is looking for a Machine Learning Engineer III working on routing cost based in Germany.

• Overstory is looking for an Engineering Manager for Machine Learning, remote across the US, Canada and much of Europe.

Just for Fun

Image: Eddie Sgarbossa, via NASA APOD

The Elephant’s Trunk Nebula winds through the emission region and young star cluster complex IC 1396 in the constellation Cepheus, about 3,000 light-years away. The trunk itself, catalogued as vdB 142, is more than 20 light-years long, and its bright ridges and dark, tendril-shaped clouds of cool dust and gas are the raw material for new stars, with protostars already hiding inside. The whole IC 1396 complex spans more than 5 degrees of sky; this view covers close to 1 degree, a little under the width of two full Moons.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Hey guys, here’s this week’s edition of the Spatial Edge, a place where nobody ever loses their projection… Anyway, the aim as usual is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Income Mapping: Google Maps pins predict São Paulo neighbourhood incomes.

• Urban Cooling: Canals cool nearby streets by nearly a degree.

• Wildfire Risk: AlphaEarth embeddings transfer between regions without retraining.

• Crop Forecasts: Sequence models predict wheat canopy a month ahead.

• Vector Mapping: One click draws a building polygon in QGIS.

Research you should know about

1. Estimating neighbourhood income from Google Maps pins

A new study from Bowdoin College and PUC-Campinas asks whether the businesses listed on Google Maps can stand in for a census when it comes to mapping income across São Paulo. Brazil’s census runs every ten years, so for most of the decade policymakers are working from stale numbers at the neighbourhood level. Points of interest, on the other hand, update constantly and cost almost nothing to pull.

The researchers queried Google Places for 47 theoretically chosen anchor categories, which surfaced 386 distinct subtypes across São Paulo’s 26,625 census sectors, with a median of 57 POIs per sector. They reduced the counts with PCA and non-negative matrix factorisation, then tested 77 configurations across seven regression models, from ridge regression to XGBoost, against household income from the 2022 IBGE census.

The best setup, NMF with 50 components feeding gradient boosting, reached a held-out R² of 0.646, so POIs alone explain close to two thirds of the variation in sector income. Churches and bars cluster in lower-income sectors, while doctors, dentists and parking lots track affluence. The model over-predicts the poorest areas and under-predicts the richest, and the residuals are spatially clustered (Moran’s I of 0.334), so there’s still room for spatial modelling. The code is on GitHub.

2. Canals are quietly cooling Britain’s cities

A new study in Nature Communications from the University of Manchester models the thermal effect of every urban canal in Great Britain and Ireland, and finds they cool the air around them by about 0.8°C on a typical day. Urban greenspace gets a lot of attention as a heat fix, with trees and parks shown to cut daytime temperatures by roughly 0.94°C. Bluespace, meaning rivers, lakes and canals, has had far less scrutiny, partly because micrometeorological simulations only ever cover a handful of sites for a day or two, while satellite land surface temperature products are too coarse to see what a 15 metre wide canal does to the street next to it.

The team built an energy balance model that compares the radiative, evaporative and convective fluxes of canal water against an equivalent surface of asphalt, and ran it at 15-minute intervals for the whole of 2022 across 2,356 canal sections. Inputs are canal geometries and depths from the Ordnance Survey and the Canal and River Trust (OpenStreetMap for Ireland), building heights for shading, and hourly OpenWeather data. Modelled water temperatures were checked against Canal and River Trust sensors at six sites, with residuals mostly within 1 to 2°C.

The median annual daytime cooling is 0.77°C within roughly 20 to 50 metres of the banks, rising to 1.2 to 1.4°C in spring and early summer and 1.5 to 2.3°C during the three 2022 heatwaves, including the one that produced the UK’s 40.3°C record. The flip side is a small nighttime warming of about 0.16°C, because water holds its heat. Regionally the daytime effect ranges from 0.32°C to 1.37°C depending on cloud and wind, and the area that benefits is around 12.4 km² in Birmingham, 8.1 km² in London, 4 km² each in Manchester and Liverpool, and 2 km² each in Glasgow and Dublin.

3. Mapping wildfire risk with AlphaEarth embeddings

A new study from UNSW Sydney and the University of Wisconsin-Madison tests whether Google’s AlphaEarth Foundations embeddings can replace the usual stack of hand-built predictors in wildfire susceptibility mapping. The conventional approach means harmonising dozens of layers, from slope and land cover to fire weather indices and soil moisture, and the resulting models tend to fall apart when you move them to a new region.

The team used MODIS active fire detections for Victoria, Australia from 2017 to 2025, giving 18,140 fire cells on a 1 km grid (about 8% of the state) matched 1:1 with pseudo-absences. They trained random forests, XGBoost, LightGBM, an MLP, TabPFN and CNNs on 9, 17 and 25 km patches of embeddings, and compared them against the same models fed 22 physical predictors screened from an initial pool of 35.

The embedding models scored ROC-AUC above 0.92 in Victoria, and when transferred without retraining the AUC rose about 4% in Canberra and fell only about 2% in the Western Sydney and Blue Mountains region, whereas the physical-variable models lost around 25%. High and very high susceptibility concentrated in eastern Victoria along the Great Dividing Range, the Victorian Alps and Gippsland, plus parts of southwestern and central Victoria. The authors pitch it at government agencies and insurers who need risk maps that travel.

4. Forecasting wheat growth a month ahead from Sentinel-2

A new study from Agroscope and the Swiss Data Science Center shows that satellite time series plus weather can forecast how a wheat canopy will develop up to 32 days ahead, at 10 metre resolution, across an entire country. Most crop monitoring is backward looking: you see what the leaf area index did, not what it’s about to do. Clouds make it worse, leaving the satellite record full of gaps that models tend to fill with implausible wiggles.

The team assembled 20.6 million pixel-level Sentinel-2 LAI time series for winter wheat across Switzerland over five seasons (2021 to 2025), paired with daily temperature, sunshine and precipitation. They trained a bidirectional GRU sequence-to-sequence model and a transformer encoder-decoder against MLP and LightGBM baselines, and added a lightweight unimodal shape regulariser that penalises valleys in the predicted trajectory so it keeps the rise-peak-decline shape a wheat season should have.

In leave-one-year-out tests the GRU reached a mean R² of 0.823 and RMSE of 0.731 LAI units, with the transformer close behind at 0.803, and both comfortably beat the non-sequential baselines. The regulariser cost almost nothing in accuracy while making the trajectories biologically sensible. The setup is crop and region agnostic, so the same recipe should port to other crops where LAI is available.

5. One click to draw a building polygon in QGIS

A new paper from LuxCarta, accepted at IGARSS 2026, introduces Click2Poly, a vision-language model that lets a mapper click once on a satellite image and get a clean building polygon or wall line back. Automated building extraction has got good, but production mapping still ends with a human fixing the outputs by hand, and that correction step is where most of the time goes.

Click2Poly extends Microsoft’s Florence-2 with three tools: draw a building, split a connected block into separate buildings, and trace a wall or fence. It was trained on 377,207 building polygons across 202 areas (about 1,344 km²) and 197,234 wall linestrings across 117 areas, all on 30 cm Pléiades Neo imagery cut into 768 pixel patches, with user clicks simulated by sampling random points inside each building during training. It ships as a QGIS plugin talking to a server that caches image encodings.

In a timing study with four operators across four test areas, Click2Poly cut editing time from 303.5 to 199 minutes, a 1.53x speed-up overall, reaching 2.30x in the Philippines and 1.78x in Angola, though it was marginally slower (0.93x) in a dense São Paulo scene. Agreement between operators stayed at 90.7 to 96.5% IoU, and inference runs at about 1.3 seconds to encode an image and 0.8 seconds per polygon on an RTX 3060.

Geospatial Datasets

1. Four decades of global forest fire patches from Landsat

GlobMap FFP is a global inventory of 11.97 million individual forest fire patches mapped at 30 m from Landsat between 1984 and 2022, covering all forest with more than 25% tree cover. You can read the paper here and access the data here.

2. Global forest disturbance regimes inferred from satellite biomass

A Max Planck Institute for Biogeochemistry dataset of four disturbance regime parameters (disturbance rate, gap-size distribution, severity and background mortality) for the world’s forests, derived by inverting the spatial pattern of ESA GlobBiomass 2010 with random forests trained on more than 8 million synthetic forest simulations. It comes as 25 km tiles in CSV and a 0.25° global grid in NetCDF, with reliable predictions for roughly 90% of forest regions. You can read the paper here, access the data here and the code here.

3. Snowmelt runoff onset from Sentinel-1, 2015 to 2024

A University of Washington dataset of the date snowmelt runoff begins each year at 80 m resolution for almost all seasonal snow on Earth (60°S to 81.1°N), for water years 2015 to 2024, plus ten-year median and variability composites. You can read the paper here, access the data here and the code here.

4. Thirty years of paddy rice maps for South and Southeast Asia

Maps of paddy rice extent and cropping intensity (single, double or triple season) at 30 m for India, Pakistan, Bangladesh, Myanmar, Thailand, Laos, Cambodia and Vietnam, for nominal years 1995, 2005, 2015 and 2024, each built from multi-year Landsat and Sentinel-2 composites. You can read the paper here, access the data here and the Earth Engine code here.

5. Using a community wireless network as a rain gauge in New York

OpenMesh is eight months (29 October 2023 to 1 July 2024) of one-minute received signal level measurements from 103 wireless sublinks across 75 hops of NYC Mesh, a community-run network in lower Manhattan and Brooklyn covering about 10 km², in the 5 to 6 GHz, 24 GHz and 58 to 70 GHz bands. You can read the paper here, access the data here and the code here.

Other useful bits

• Europe’s next weather satellite, MTG-I2, launches on 27 August on an Ariane 62 from French Guiana, the first Ariane 6 flight to geostationary transfer orbit. Built by a Thales Alenia Space, OHB and Leonardo consortium for ESA and EUMETSAT, it carries the Flexible Combined Imager and a Lightning Imager, scans the full disc every 10 minutes and Europe every 2.5 minutes, and will pair with Meteosat-12 to produce at least 50 times more data than the Meteosat Second Generation satellites.

• The Copernicus Land Monitoring Service has released 2024 croplands, grasslands and tree cover and forests layers from its High Resolution Vegetated Land Cover Characteristics suite, covering the EEA-38 countries and the UK. The full 2017 to 2023 archive has moved to the Copernicus Data Space Ecosystem at the same time, so the whole time series is now in one place and browsable in Copernicus Browser.

• Schneider Electric is buying AiDASH for USD 350 million in enterprise value. The Palo Alto company uses satellite imagery and AI to flag vegetation, storm and wildfire risk to power lines for utilities, and will slot into Schneider’s One Digital Grid platform. Schneider’s venture arm had put USD 10 million into the company back in 2022.

• Verisk has acquired McKenzie Intelligence Services, a UK geospatial intelligence firm that does real-time post-event damage assessment for catastrophes and conflict events. It joins Verisk’s Catastrophe and Risk Solutions division, where its imagery-derived damage data will feed the existing risk models. The price wasn’t disclosed.

• Intermap has agreed to acquire the rest of PCI Geomatics, the Canadian company behind the CATALYST imagery processing software, for USD 11 million in cash. PCI’s 750-odd algorithms handle imagery from more than 500 satellites across roughly 30,000 licences, and the plan is to pair them with Intermap’s 3D elevation models.

• NASA’s Earth Observatory has a nice piece on reading Adélie penguin diets from the colour of their guano in Landsat imagery. Krill-heavy diets leave pink stains and fish-heavy ones don’t, and across more than 100 Antarctic colonies the colour tracks sea ice: warmer, less icy regions like the Antarctic Peninsula eat krill, while East Antarctic colonies with more ice eat more fish, and fish-fed chicks grow and survive better.

Jobs

• Muon Space is looking for a Senior Remote Sensing Data Scientist for calibration and validation, fully remote.

• Overstory is looking for a Staff Data Scientist working on wildfire, remote in the US or Canada.

• Mapbox is looking for a Senior or Lead Software Data Engineer on its Roads team based in Helsinki, Finland.

• Mapbox is also looking for a Senior Software Engineer for its Rust routing engine based in the United States.

• IUCN is looking for a Director of its Global Information Systems Group based in Gland, Switzerland.

• European Space Agency (ESA) is looking for a Copernicus Sentinel-2 NG Lead Instrument End-to-End Data Chain Engineer based in Noordwijk, Netherlands.

Just for Fun

This all-sky composite stacks the footage from four meteor-monitoring cameras at an observatory in Czechia over the night of 12 to 13 August, when the Perseid shower was at its peak. The cameras logged 1,706 meteors that night, and nearly all of the trails trace back to a single radiant in Perseus at the upper right. Look closer and you can pick out the fainter showers hiding underneath, including the Kappa Cygnids radiating from Cygnus and the antihelion complex near Aquarius, the weak source that sits directly opposite the Sun.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge, a place where the errors are always spatially autocorrelated. Anyway, the aim as usual is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Cyclone Forecasts: Google’s AI buys forecasters an extra day.

• Mobile Networks: Climate hazards threaten millions of base stations.

• Foundation Models: Nearly a third ship code with no weights.

• Image Matching: One model aligns imagery across six continents.

• Riverfronts: A fifth of the world’s riverfronts are modified.

Subscribe now

Research you should know about

1. Google’s AI can now forecast cyclones a day earlier

A new study published in Nature introduces WeatherNext Cyclones, an AI model that produces ensemble forecasts for the track, intensity and size of tropical cyclones anywhere in the world. Cyclones are among the most dangerous and costly weather events there are, and forecasting them well has stayed stubbornly hard. The conventional route is regional physics models run at very high resolution, which is expensive and still leaves intensity forecasts shakier than track forecasts.

The model was trained on a combination of global analysis data and a global database of historical cyclones, and it generates large ensembles of possible global weather and cyclone scenarios out to 15 days. Two things stand out about how it’s built. It runs on inputs that are orders of magnitude coarser than regional models, which suggests high resolution isn’t a strict prerequisite for state of the art intensity forecasting, and it scales to 1,000-member ensembles rather than the conventional 50, which captures rare events much better.

Evaluated on cyclones from 2023 to 2025, the track, intensity and wind radii predictions gave an average of a day or more of extra lead time over leading operational models, which the authors compare to a decade of conventional operational progress. Adding it to a weighted-average consensus ensemble improves that ensemble’s skill too. Google DeepMind says the US National Hurricane Center used the model during the 2025 season on Hurricane Melissa’s rapid intensification and landfall in Jamaica, and the code and weights are now on GitHub under Apache 2.0.

2. What climate change will do to mobile phone networks

A new study in Nature Communications puts numbers on how exposed the world’s mobile networks are to floods and cyclones. Climate risk assessments usually cover roads, power grids and buildings. Telecommunications gets left out, which is odd given that mobile networks are the thing everyone depends on during a disaster.

The researchers built a global, event-based geospatial assessment of roughly 17 million mobile base stations, combining crowdsourced open asset data with global climate hazard models. They ran it under historical conditions and two future scenarios, RCP4.5 and RCP8.5, and estimated both exposure and direct damage across a range of extreme-event return periods.

Under RCP8.5 in 2050, a tropical cyclone event with a 0.1% annual probability affects an estimated 3.7 million base stations and causes USD 4.22 billion in direct damage. Under RCP8.5 in 2080, the equivalent coastal flooding event affects 268,000 stations and causes USD 6.44 billion in damage. Riverine flooding touches more assets globally, around 2.8 million, though damages there stay broadly stable at USD 46 to 47 billion. The practical takeaway is that telecoms assets belong in climate risk assessments as standard, and there’s now open crowdsourced data good enough to do it.

3. Most geospatial foundation models are hard to actually use

A new study from the University of Cambridge asks a question the geospatial foundation model literature mostly skips, which is whether anyone outside the lab can actually run these things. Most evaluations are model-centric and stop at architecture and benchmark accuracy. That tells you nothing about whether an ecologist can pick a model up and get a result out of it.

The team ran a pilot survey with ecology and conservation scientists, turned what they heard into a seven-dimension usability rubric grounded in HCI theory, then had two raters score 89 GeoFMs against it. The dimensions are access and deployment, interaction and customisation, trust and transparency, community and support, scientific permanence, multilingual support, and offline usability.

Nearly 29% of the 89 models scored Level 0 on access, meaning no public weights at all, and another 21% ship weights with minimal guidance, so over half (50.6%) require you to either pre-train from scratch or build a complex environment on your own. Only six models (6.7%) offer a hosted API or a graphical interface: MOSAIKS, Google’s AlphaEarth, Prithvi, AIEarth, SatLas and Tessera. Trust is the bigger gap. Of the accessible models, 96.8% provide replicable benchmarks, exactly one has built-in explainability, and none document uncertainty quantification, which is the feature the surveyed experts ranked highest. It is getting better though: 54% of 2025 releases reached Level 3 or above, against 27% in 2024.

4. Matching satellite images from anywhere on Earth

A new study from Beihang University takes on dense image matching at global scale, where two images of the same place can differ in acquisition time, season, viewpoint, resolution and land cover. Dense matching establishes pixel-wise correspondences and underpins a lot of photogrammetry and change work. The trouble is that large geometric offsets, partial overlap and regions that genuinely can’t be matched make direct dense prediction both unreliable and slow.

The fix is to reformulate matching as localisation-and-registration: first find the matchable overlap and its affine geometry, then refine dense residuals inside the aligned frame. The resulting model is LoRetta. They also released LEVIR-GM, a global multi-temporal optical matching benchmark with matchability labels built in: 103,000 aligned pairs, 827,000 after augmentation, spanning six continents, five years and resolutions from 0.5 m to 1,024 m.

LoRetta scores 83.3% AUC on LEVIR-GM, 1.6 points above the RoMa v2 baseline, with percentage of correct keypoints up 6.5 points at 1 pixel and 8.2 points at 2 pixels, and it runs 47.8% faster. They also demo it on cross-domain geolocalisation, matching astronaut photography and UAV imagery back to satellite scenes. The code is on GitHub and the weights are on Hugging Face.

5. A fifth of the world’s riverfronts have been reshaped by people

A new study in Nature Communications maps 7.52 million kilometres of riverfront, the strip where land meets water along a river, and finds nearly 20% of it has been modified by people. Most river fragmentation research is about dams, which cut rivers lengthwise. The lateral edge, where farming and building press right up against the channel, has been much harder to see at global scale.

The researchers used satellite imagery and deep learning to build a high-resolution global map of riverfront land cover. Agriculture accounts for 13.43% of global riverfront and built-up areas for 6.29%. The rest breaks down as wooded (43.27%), semi-barren (16.03%), grassy or tundra (14.95%) and barren (6.03%).

A distinct modification belt running across parts of Africa and Eurasia accounts for around 60% of all global alterations, with imprint densities six times higher than everywhere else. The authors link this lateral fragmentation to deteriorating water quality and biodiversity threats, and argue it’s a fundamentally different pressure from the longitudinal fragmentation dams cause. Environmental limits and region-specific socio-economic dependencies explain a lot of where the imprints land.

Geospatial Datasets

1. Forty years of gridded hydrology for 9,067 US basins

BASINGRID is a spatially distributed hydrology dataset covering 9,067 basins across the contiguous United States from 1985 to 2024, harmonised to a roughly 1 km Daymet reference grid. You can read the paper here and access the data here.

2. Bias-corrected CMIP6 projections for India

This is a daily, non-stationary, multivariate bias-corrected dataset of precipitation and maximum, minimum and mean temperature for India at 0.25° resolution, derived from 13 CMIP6 global climate models. The paper is here and you can access the data here.

3. A global inventory of landslides studied with DInSAR

A geo-referenced inventory of landslides that have been investigated using satellite differential SAR interferometry, compiled from a systematic review of the literature from 1995 to 2024. The team screened 2,739 contributions and geo-tagged the resulting 1,480 point identification numbers, distinguishing site-specific from area-wide analyses. China, Italy and the US dominate the case studies, and the positions were validated against the global landslide susceptibility map and independent inventories. The paper is here and you can access the data here.

4. A century of Chinese hydropower stations

A century-scale inventory of 2,918 georeferenced hydropower stations in mainland China covering 1909 to 2025, compiled from international databases, national statistics and archival sources. The paper is here and you can access the data here.

Other useful bits

• Google Maps Platform has opened a private preview of Custom Satellite Embeddings, powered by DeepMind's AlphaEarth Foundations. Instead of the annual global embedding layer, you can request 10 m, 64-band embeddings for your own region and time window at intervals as short as five days. Academic researchers can apply for free sample datasets through Earth Engine until 1 September 2026.

• A new open-source Python package called DefoEye wraps GMTSAR into a single time-series InSAR workflow for Sentinel-1, with parallel job execution, interferogram network pruning and a graphical interface, so you don't need to hand-hold every step. Tested across Bologna, Gotland and Houston it matched GNSS stations with RMSE of 4.3 to 11.9 mm and correlations of 0.63 to 0.95, and the code is on GitHub.

• Microsoft has released Aurora 1.5, an update to its open Earth system foundation model that goes from 4 to 26 predicted variables, adds hourly resolution and supports probabilistic ensembles. It beats ECMWF ensemble forecasts on 88.9% of the evaluated variable-and-lead-time targets, and cuts tropical cyclone track error by roughly a third against the original Aurora at day 5. Checkpoints are on Hugging Face and the code is open source.

• PAHO and Esri have wrapped up an advanced GIS training series for public health emergency management across the Americas, covering spatial statistics, risk modelling and geospatial data management. It ran through 2025 and 2026 and finished with a virtual session on 17 June, with 100 participants from 21 countries completing the training.

Jobs

• Planet is looking for a Senior Engineering Manager for its AI Geospatial Assistant team based in San Francisco.

• Muon Space is looking for a Senior Applied Scientist, Geospatial, fully remote.

• Overstory is looking for a Staff Machine Learning Engineer working on wildfire, remote in the US or Canada.

• Esri is looking for a Product Engineer II for 3D Reality Mapping based in Stuttgart, Germany.

• Zipline is looking for a Senior Software Engineer on Maps Routing based in South San Francisco.

Just for Fun

The corona of a total solar eclipse normally looks white, but observers in Spain on 12 August saw a distinctly golden one. Two things did it. Totality happened with the Sun near the horizon, so the light travelled through a lot of atmosphere and lost most of its blue, and smoke from nearby forest fires acted as a second filter that deepened the gold further. The one thing that stayed off-palette was a hydrogen-glowing prominence on the Sun's left edge, which kept its bright pink. This HDR-processed, multiple-exposure image was captured from Benavente, Spain.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge. If you’ve ever considered calling your next kid, ‘WGS84’, we have one thing to say to you: welcome home… In any case, the aim is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Self-Play AI: Self-training model masters geospatial reasoning without labels.

• Crop Mapping: AlphaEarth embeddings replace manual feature engineering.

• Street Generation: Diffusion models create controllable urban street designs.

• Robustness Benchmark: EarthShift tests real-world satellite model generalisation.

• Earth Dataset: WorldTensor unifies global environmental and socioeconomic data.

• Climate Data: Bias-corrected global climate projections through 2050.

Subscribe now

Research you should know about

1. Self-playing AI masters geospatial reasoning without human labels

Training vision-language models to understand satellite and aerial imagery usually requires massive, human-curated datasets. But complex geospatial reasoning (such as analysing object adjacency, counting instances, or comparing areas) creates a combinatorial explosion of possible questions that manual annotation simply can’t cover. To bypass this problem, a new paper introduces GeoX, an autonomous self-play framework that masters spatial reasoning without using any human-annotated question-answer pairs. Instead of relying on static labels, GeoX alternates a single model between two roles: a proposer that programmatically constructs geospatial problems over unlabeled remote sensing images, and a solver that attempts to answer them.

The framework structures these problems as executable Python programs using an underlying library of geometric, topological, and aggregation primitives, alongside an open-vocabulary segmentation tool. To build robust structural comprehension rather than simple pattern recognition, the system forces the model to interrogate each proposal under three distinct reasoning modes: abduction (inferring the input arguments from observed outputs), deduction (predicting outputs from given parameters), and induction (synthesising the underlying program code from input-output examples). A programmatic verifier then executes the code to supply an objective reinforcement learning reward, which incentivises the proposer to create learnable, frontier-pushing challenges and the solver to derive correct answers.

When evaluated zero-shot across multiple remote sensing visual question answering and object counting benchmarks (including RSVQA-HR, EarthVQA, and GEOBench-VLM), GeoX consistently outperformed its base foundation models by an average of up to 5.5 points. Interestingly, despite using zero curated training pairs, the self-play model matched or exceeded domain-specific vision-language models fine-tuned on millions of human-annotated examples, achieving particularly strong gains in complex spatial relation classification and object counting.

2. Ready-made embeddings for field-scale crop mapping

The old school way of doing crop mapping using machine learning involved a bunch of time spent on feature engineering. You’d have to manually clean optical satellite data, mask clouds, and calculate stuff like NDVI over a growing season. To save the pain of all this, a new study explores whether we can skip the manual preprocessing entirely by using Google DeepMind’s AlphaEarth geospatial foundation model embeddings. Instead of raw spectral bands, AlphaEarth provides analysis-ready, annual 64-band “embeddings” (I did a deep dive explainer on this here).

To test this approach, researchers focused on mapping processing tomatoes across California. They extracted AlphaEarth embedding chips for 9,484 agricultural fields (split evenly between tomato and diverse non-tomato crops based on LandIQ 2018 records) and trained a U-Net deep learning segmentation model. On a spatially independent test set of 1,424 fields, the model achieved an exceptional 99.19% pixel accuracy and a 98.11% intersection over union (IoU), cleanly tracing tomato field boundaries with minimal confusion against visually similar annual crops.

The team also used Monte Carlo dropout during inference to map predictive uncertainty. By running each field chip through the model 100 times, they found that uncertainty was consistently quite low within the interiors of well-defined fields and spiked around borders, irregular edges, and mixed margins. This provides a practical safeguard, flagging ambiguous boundary pixels where human review might be needed. While the study relies on reference data from a single year (2018), it shows how foundation model embeddings can be used as an operational alternative to manual feature engineering for field-scale crop mapping.

3. Generating alternative urban streetscapes with diffusion models

While street-view imagery is pretty commonly used by urban planners to measure and evaluate existing built environments, such as calculating greenery, sky, or building view indices, these observational tools can’t generate alternative or non-existing urban scenarios. To address this, a new paper introduces a controllable generative AI framework that combines realistic, alternative streetscapes conditioned on targeted visual metrics and text descriptions. Using a curated multimodal dataset of 32,000 images from Chicago and Orlando, the researchers aligned street-view imagery with text descriptions, object counts, and semantic segmentation maps (specifically road masks) to allow for fine-grained manipulation of urban scenes. By fine-tuning Stable Diffusion and ControlNet models, the framework directly translates abstract planning intentions, such as increasing tree canopy coverage while preserving existing road geometry, into concrete visual street environments.

The results demonstrate that incorporating imagery controls via ControlNet significantly improves both visual realism and semantic consistency compared to text-only diffusion baselines. Using explicit road segmentation masks reduced perceptual dissimilarity (LPIPS) by approximately 6% and boosted overall semantic alignment (mIoU) by 23.7% in Orlando and 46.4% in high-density Chicago. The framework also supports complex, multi-element trade-off scenarios; when language prompts instructed the model to increase tree canopy coverage while reducing sky or building proportions, the generated greenery naturally expanded along sidewalks and foregrounds without encroaching on roadways or distorting building geometry.

Interestingly, when text prompts and visual road masks were intentionally set to conflict, the imagery controls exerted a dominant influence over the generated streetscape. Even when language prompts requested a six-fold variation in road proportions (from 5.71% to 35.71%), the generated streetscapes strictly adhered to the geometry and layout of the input road mask. This reveals a clear control hierarchy where spatial masks anchor the geometric backbone of a scene, while text prompts provide high-level compositional guidance when masks aren’t available.

4. Benchmarking distributional shift in Earth observation

Current satellite machine learning models can be brittle when deployed in out-of-distribution contexts (e.g. new geographies, time periods, or sensors, etc.). However, most benchmarks only test generalisation to new samples drawn from the same in-distribution training set. This leaves a bit of a blind spot for deploying geospatial models in real-world scenarios. To address this, researchers introduce EarthShift, a public testbed for benchmarking distributional robustness across five common natural remote sensing shifts: spatial scale, time, geographic location, sensor modality, and data source. The benchmark pairs established datasets across 11 tasks (e.g. shifting from Sentinel-2 to Sentinel-1 for flood mapping, or testing across countries with vastly different field structures).

To avoid conflating model capability with robustness, EarthShift uses an “effective robustness” metric. Rather than just ranking models by out of distribution performance, effective robustness measures an architecture’s out of distribution performance relative to a baseline predicted by its ID capability. The team evaluated 13 models, including 8 specialized geospatial foundation models (GFMs), generic vision foundation models (VFMs), and standard fully-supervised architectures; across 10,000 experiments. The results actually show that across all architectures and training setups, models suffer an average 15-20% drop in performance when encountering real-world distribution shifts. While models were surprisingly resilient to temporal shifts (such as season-to-season agricultural changes), they degraded severely under sensor and geographic shifts.

Most surprisingly, domain-specific pre-training on Earth observation archives provided no distributional robustness advantage over generic vision models pretrained on ImageNet, or even over randomly initialised fully-supervised models. In fact, generic ImageNet-pretrained ResNet-50 architectures frequently proved to be the most distributionally robust models tested. Full fine-tuning of model encoders often failed to improve, and sometimes even worsened, effective robustness compared to simply freezing the pre-trained backbone…

Geospatial Datasets

1. Earth system foundation model dataset

The WorldTensor dataset harmonises hundreds of global environmental and socioeconomic variables onto a standardised 0.25-degree spatial grid and an annual temporal framework spanning 1900 to 2025. Access the data here and the code here.

2. Bias-corrected high-resolution climate dataset

The BC-HiRMIP dataset provides globally bias-adjusted climate projections derived from four HighResMIP models at a 0.5° spatial and daily temporal resolution, spanning from 1979 to 2050. Access the data here and the code here.

3. Upper reach transboundary river hydrological dataset

The Mainland Southeast Asia International River Upper Reach Hydrological Dataset (MSAIRHD) compiles historical daily streamflow, sediment concentration, and water level records across 119 monitoring stations in the upper reaches of the Irrawaddy, Salween, Mekong, and Red River basins from 1958 to 1987. You can access the data here and the code here.

4. Remote sensing traffic object segmentation dataset

The NWPU-Traffic dataset provides an optical remote sensing collection containing 1,479 high-resolution images and 31,628 instance-level annotated targets across four major transportation categories: cars, airplanes, ships, and trains. Access the data and code here.

Other useful bits

• NASA’s Jet Propulsion Laboratory has completed the first in-orbit test of an untuned vision-language model, running Google’s Gemma 3 on a satellite. By analysing imagery onboard to generate text descriptions, the system bypasses bandwidth limits by downlinking brief summaries rather than massive raw image files.

• A machine learning study used Google Maps Platform’s Places Insights dataset in BigQuery to account for nearly 86% of home value variation across census tracts in Austin, Texas. The analysis revealed that the quality of nearby commercial amenities (gyms, spas, cafes, etc) served as a primary positive predictor of residential property values, whereas industrial sites like storage facilities and auto-repair corridors significantly depressed local valuations.

• The US Federal Communications Commission has introduced new regulations requiring satellite operators to share real-time location data to mitigate collision risks in increasingly congested orbits. The move has reignited a jurisdictional dispute with US lawmakers who argue that authority over space safety and traffic management should reside with the Department of Commerce.

Jobs

• UFZ is looking for an Environmental Computer Scientist in Leipzig, Germany

• The Allen Institute is looking for a Senior Software Engineer (AI For Planet) in Seattle, US.

• The World Bank is looking for an Extended Term Consultant (Associate Geospatial Platform Engineer) in Washington DC, US.

Just for Fun

ESA’s Euclid space telescope has discovered 31 of the most ancient quasars ever found, including two dating back to when the cosmos was just 670 million years old, setting a new cosmic record and more than doubling the known population of these ultra-early supermassive black holes.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge — we put the ‘special’ in geospecial data science… As usual, the aim is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Urban Food Security: Graph AI optimises hydroponic nutrient dosing.

• Crop Yield Prediction: Interpretable deep learning improves harvest forecasts.

• European Mobility: Seven mobility types reshape regional policy.

• Forest Height Mapping: Radar and optical data improve accuracy.

• Macroeconomic Data: Global statistics spanning nearly a millennium.

Subscribe now

Research you should know about

1. A graph-based approach to urban food security

As anyone who’s read the news knows, traditional soil-based agriculture is increasingly vulnerable to various issues like extreme weather, soil degradation, and water shortages. Taken together, this leads to a rise in global food insecurity. In response, ‘soilless’ (or is it ‘soil-free’?) methods like hydroponics and aeroponics are gaining popularity for urban food production. However, these controlled-environment systems rely on precise, real-time nutrient dosing. Current systems often use rigid/static dosing rules that don’t really account for the complex, non-linear interactions between dynamic environmental conditions (e.g. temperature or humidity) and a plant’s physiological response. This inflexibility can essentially lead to nutrient waste and increased crop stress.

A new study from Scientific Reports addresses this by introducing the Optimised Multi-Scale Adaptive Graph Neural Network (OMSA-GNN). The framework uses a Raspberry Pi-based network of sensors to monitor conditions like pH, dissolved oxygen, electrical conductivity, and visual plant health. Rather than treating these as isolated metrics, the system treats them as localised geospatial data points. The OMSA-GNN model, fine-tuned by an Improved Sparrow Search Algorithm (ISSA), builds adaptive graphs to map the relationships between these variables, allowing it to predict plant stress and adjust nutrient dosing vectors dynamically.

When evaluated on a dataset of hydroponic lettuce growth, the OMSA-GNN framework outperformed the older, more traditional static models, as well as several other neural network architectures, in both forecasting accuracy and nutrient efficiency. By treating individual sensor nodes as micro-spatial intelligence units, the model was able to detect early signs of physiological stress and automatically adjust the nutrient mix, reducing waste and stabilising yields. This approach offers a highly adaptable, data-driven solution for securing urban food systems against the unpredictability of climate-stressed environments.

You can access the dataset here.

2. An interpretable deep learning model for crop yield prediction

Accurate crop yield prediction is pretty important for global food security and smart agriculture, but it’s a notoriously difficult problem. Agricultural yields are determined by a complex mix of weather, soil, and management practices that all interact over time. Deep learning models have shown promise in this area, but they do struggle with “gradient instability” in the face of early-season weather anomalies (like a spring drought). I.e. these events lose their impact in the model’s memory by the time it predicts the final harvest.

To tackle this, a new paper from Scientific Reports introduces the MHCNN-LSTM-MHA model, a hybrid architecture that tackles agricultural data in a highly structured way. Rather than mixing weather and soil data together at the start, the model uses parallel Multihead Convolutional Neural Networks (MHCNN) to process them separately, preserving their unique characteristics. This data is then fed into a Long Short-Term Memory (LSTM) network to track temporal changes across the growing season. Finally, a Multi-Head Attention (MHA) mechanism acts as the model’s focus, dynamically prioritising specific time steps to ensure critical events aren’t lost.

When tested on a comprehensive 39-year dataset of US soybean yields, the proposed model achieved an impressive 9.86% improvement over the previous best benchmark. They conducted a SHAP analysis and confirmed that the model’s logic aligned with established agronomic principles, heavily weighting factors like solar radiation, maximum temperature, and precipitation. The conclusion, according to the authors, then is this might be more reliable or ‘trustworthy’ than other approaches.

3. Rethinking spatial mobility across Europe

It’s August in Europe (and for the rest of the world for that matter), which means that everyone is on holidays and travelling around - usually in a caravan or a train of some sort (no gross over-generalisations here…). Anyway, a new study from Nature Communications argues that we need to stop treating all movement as the same. The researchers combined data from the Labour Force Survey, Erasmus+ student exchanges, and geotagged X posts to map out seven distinct types of spatial mobility. This included everything from short-term student exchanges and seasonal agricultural work to cross-border commuting and multilocal living, where people actively split their time between multiple residences.

By running these different mobility flows against a huge variety of regional characteristics, the team found that the factors attracting one type of person often repel another. For instance, an increase in a region’s GDP per capita generally leads to an influx of long-term students and permanent migrants, but it actually causes a drop in seasonal workers and long-distance commuters. Digital infrastructure also plays a fascinating role. Faster mobile internet speeds at an origin location tend to boost cross-border commuting and multilocal living, likely because good connectivity makes flexible work arrangements easier. Conversely, at the destination, faster internet is associated with a decrease in seasonal manual labour.

The findings highlight that a one-size-fits-all approach to regional development just doesn’t work. Rather than simply trying to stop outmigration or blindly attract permanent residents, policymakers might be better off leaning into their region’s specific strengths to attract seasonal workers or digital nomads.

You can access the data here.

4. Combining radar and optical data for better forest mapping

Estimating forest height accurately across large areas is a massive challenge in remote sensing, but it can be pretty important for understanding biomass and carbon storage. While SAR techniques like Polarimetric SAR Interferometry (Pol-InSAR) are widely used for this, their performance can suffer when the underlying physics models struggle with complex terrain slopes and varied canopy structures. A new study from IEEE tackles this by expanding on a recently developed hybrid model that combines machine learning with a physical radar model. The researchers realised that while radar data from satellites like TanDEM-X is great for capturing physical structures, adding multispectral optical data from Landsat could provide the extra missing clues about vegetation types and density needed to resolve tricky structural ambiguities.

To put this to the test, the team took an earlier hybrid model that relied solely on TanDEM-X radar features (Model C) and upgraded it by adding four Landsat optical bands: Red, Near-Infrared (NIR), and two Short-Wave Infrared bands (SWIR1 and SWIR2). They trained and tested this new version (Model D) over the diverse, hilly terrain of the Lopé National Park in Gabon, comparing the results against high-precision airborne LiDAR measurements from the AfriSAR campaign. The optical boost made a remarkable difference. Model D reduced the Root Mean Square Error (RMSE) by 13.5% and the Mean Absolute Error (MAE) by 16.6% compared to the radar-only version.

What’s particularly impressive is how well the new model handled difficult topography. The radar-only approach consistently underestimated forest heights on slopes facing away from the sensor, but the inclusion of Landsat spectral data allowed the machine learning component to correct this bias almost entirely on moderate slopes.

Geospatial Datasets

1. Global macroeconomic statistics dataset

The Global Macro Database compiles 46 macroeconomic variables across 239 countries, harmonising 160 historical and contemporary sources into continuous country-level series. The data runs from 1086 to 2025 with forecasts extending to 2030. Access the data here and code here.

2. Abyssal ocean temperature dataset

The Stratus Ocean Reference Station dataset provides a 13-year continuous time series (2012 to 2025) of deep ocean temperature observations at abyssal depths between 4,191 and 4,526 m in the Southeast Pacific. Access the data here and the code here.

3. Global snow-free leaf area index dataset

The Global Snow-free Leaf Area Index dataset provides monthly global leaf area index (LAI) and stem area index (SAI) fields at a 500-m spatial resolution spanning from 1985 to 2020. Access the code here. The links to the datasets for each year (1985-2020) can be found under Data Records of the article.

4. Hybrid global river discharge dataset

The GLORIF1 dataset provides a monthly global river discharge inventory spanning 1979 to 2019 at a 30-arcminute spatial resolution (approximately 50 km at the equator). Access the data here and the code here.

Other useful bits

• Google’s Isochrones API generates reachable area polygons from an origin point based on actual road network travel times rather than simple straight-line distances. This allows developers to measure true accessibility for practical applications like delivery logistics, real estate commute planning, and urban service mapping.

• Sentinel-2 Paint is an open-source web app that reconstructs user-uploaded images as mosaics using real, geolocated satellite imagery patches from the Sentinel-2 archive. It indexes nearly 93,000 unique surface colours extracted from global imagery.

• NASA has mandated that its 11 billion dollar Habitable Worlds Observatory will be engineered for robotic in-space servicing, assembly, and maintenance. Situated at the Sun-Earth Lagrange point 2, the flagship telescope will feature modular components and standardised interfaces so robotic spacecraft can upgrade and repair its systems in orbit.

• A non-profit is investing in open-source geospatial AI tools to help humanitarian organisations like the WFP predict supply chain disruptions and reroute food aid more efficiently. By identifying structural bottlenecks and road closures in vulnerable regions like Afghanistan before they occur, the initiative aims to mitigate hunger and ensure timely delivery amidst growing global food insecurity.

Jobs

• Mercy Corps are looking for a remote GIS Intern

• The Nature Conservatory is looking for a Natural Climate Solutions Data Scientist in the US

• The FAO is looking for a bunch of Geospatial Data Science Consultants

Just for Fun

A daytime telescopic image captured from British Columbia, Canada, highlights the moment Venus re-emerged from behind the bright limb of a crescent Moon following a one-hour lunar occultation on 17 June.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge. We’re like the geospatial newsletter equivalent of Gemini - we used to be cool last year, but we’re slightly less cool with each passing month... In any case, the aim is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Population Mapping: Foundation models improve population estimates in data-poor regions.

• Annotation-Free AI: Self-training model eliminates manual geospatial labels.

• Mobility Modelling: Hierarchical approach accelerates large-scale movement simulations.

• Foundation Models: Audit exposes major benchmarking and reproducibility issues.

• Settlement Data: Global 10 metre settlement dynamics dataset released.

Subscribe now

Research you should know about

1. Rethinking population mapping with foundation models

When planning infrastructure or responding to disasters, having accurate, up-to-date population estimates is obviously pretty important. However, as anyone who’s used survey data before, it can be outdated or spatially imprecise, especially in rapidly developing countries. Traditionally, demographers have filled these gaps by estimating populations using a bunch of hand-picked geospatial covariates like satellite imagery of nighttime lights, road networks, and building density. While effective, assembling these datasets is a pain. A new paper tests whether we can skip the manual assembly and instead use the Population Dynamics Foundation Model (PDFM). This model generates embeddings by compressing a bunch of different data, such as search trends, maps, and local activity levels, into a single, reusable mathematical representation of a place (btw, I did a deep dive on embeddings here).

To benchmark this, the researchers compared the predictive power of PDFM embeddings against 23 standard geospatial covariates for estimating subnational populations in Brazil, Nigeria, and the United States. They found that PDFM successfully improved predictive fit, reducing unexplained variance by a median of 20.1% across various machine learning models. However, the performance was pretty uneven. PDFM provided the biggest boost in large, less-developed regions where traditional markers like building density and road networks were weak or absent. In highly developed areas where traditional covariates already captured settlement patterns perfectly, PDFM added very little value.

The study also uncovered a major limitation regarding scale. While PDFM embeddings were highly accurate at the specific spatial resolution they were trained for (e.g., zip codes), their performance broke down predictably when the researchers attempted to aggregate them to larger areas like counties. In contrast, standard geospatial covariates proved much more flexible across different spatial scales. The researchers conclude that while embeddings are useful tools for data-poor environments, they shouldn’t outright replace traditional methods. Instead, the future of population modelling likely lies in hybrid approaches that combine the nuanced, human-activity signals of embeddings with physical geospatial covariates.

2. Eliminating manual annotations for geospatial reasoning

Geospatial reasoning (or I guess any reasoning for that matter) requires models to understand fuzzy, abstract user queries and map them to precise physical locations, like finding the best area to set up emergency shelters after an earthquake. Recent advances in Multimodal Large Language Models (MLLMs) have made it possible for AI to autonomously generate the logical steps to answer these queries. However, the big issue is that teaching these models to actually output the correct geographic coordinates still requires massive amounts of human-annotated bounding boxes. This reliance on labels prevents these models from freely learning and self-evolving using the petabytes of unlabelled satellite imagery currently available.

To address this, a new study introduces RemoteZero, a training framework that requires absolutely zero human bounding-box annotations. The researchers noticed that MLLMs are much better at verifying if a given cropped image matches a text description (the “Eye”) than they are at generating raw coordinates from scratch (the “Hand”). So, RemoteZero uses a “Generate-Crop-Verify” loop. The model basicaly guesses a location, crops that area from the image, and then uses a verifier to check if the crop semantically matches the query. This semantic yes/no score is then fed back into the model as a reward using Group Relative Policy Optimization (GRPO), teaching it to find the right locations without ever seeing a human label.

Interestingly, RemoteZero enables the model to teach itself. Instead of relying on a larger, external model to act as the verifier, RemoteZero can use its own frozen policy from a previous training round to verify its current guesses, creating a continuous loop of self-evolution. When tested against fully supervised baselines that rely on thousands of manual annotations, the self-evolving RemoteZero model actually outperformed them, scoring 71.29% accuracy on the EarthReason benchmark. This proves that intrinsic semantic verification is a viable path forward, potentially unlocking the ability to train powerful geospatial AI using entirely unlabelled satellite data.

3. A smarter way to model mobility

Simulating how users move around in large-scale networks, like fog computing environments, can be pretty difficult without good real-world data. When researchers try to build mobility models from scratch, they often run into massive computational bottlenecks because tracking every single access point in a large area is overwhelming. A new study tackles this issue by proposing a hierarchical method that groups Wi-Fi access points based on their geospatial features. Rather than treating every access point as an isolated spot, they’ve organised them into nested levels, such as moving from an entire building down to a specific wing, then a floor, and finally the individual access point. This top-down approach means it’s a lot easier to track user trajectories and simulate realistic movement without getting bogged down by the sheer volume of network data.

The team analysed wireless access logs from the University of the Balearic Islands campus, using tracking data from the Aruba Location Engine. By applying clustering algorithms to profile user behaviour at different granularities, they generated transition matrices that accurately mapped out where people were likely to go next. The results were pretty impressive. Compared to flat, non-hierarchical models, they layered approach massively reduced the computational complexity. It actually slashed the execution time needed for clustering from several hours down to just a few minutes, while drastically cutting the optimal number of clusters required.

4. The messy state of geospatial foundation models

I know, I know. Another Spatial Edge newsletter, another foundation model paper… But as long time readers may know, I’m still a bit skeptical about how useful they are (more info on this here). But a massive audit of 152 papers reveals a pretty big problem, which is that nobody actually knows what the state of the art is. The researchers found that the current literature is a bit of a mess when it comes to standardising evaluations and testing protocols. E.g. they documented 46 cases where different papers reported wildly different scores (sometimes off by tens of points) for the exact same model and benchmark. Without shared controls or a common evaluation harness, comparing these models is nearly impossible.

The issues go deeper than just mixed-up scorecards though. The audit discovered that 39% of papers do not even release their model weights 🤯. On top of that, researchers are frequently tweaking both the model architecture and the pretraining data at the same time without isolating the variables. This means when a new model claims a massive leap in accuracy, you can’t really tell if the improvement came from a new design or just a massive new dataset. It creates a confusing environment where weak models can hide behind massive datasets, and actual architectural progress gets lost in the noise.

Geospatial Datasets

1. Global settlement exposure dataset

The World Settlement Footprint dataset provides global 10-m resolution settlement dynamics published as a cloud-native GeoZarr layer. Access the data here and the platform here.

2. Urban rooftop reflectivity dataset

The Heat Resilience dataset provides high-resolution, building-level rooftop reflectivity data across more than 50 global cities to support urban heat mitigation. Produced by fusing 10-metre Sentinel-2 satellite observations with 30-centimetre commercial satellite imagery, the model reconstructs granular albedo profiles for individual structures. Access the data and documentation here.

3. Small-area global election returns dataset

The Small-Area Global Elections (SAGE) dataset compiles lower-house parliamentary and presidential election returns for 110 countries at the smallest unit published by local electoral authorities. The repository contains roughly ten billion votes across more than eight million geocoded units, with candidate and political party identifiers harmonised across multiple languages. Access the data here and the code here.

Other useful bits

• I really loved this article from Bruno Sanchez. He argues that Europe consistently loses its home-grown space innovations to foreign buyers because public funding focuses on de-risking supply rather than de-risking market demand. To bridge the downstream investment gap, ESA must act as a committed anchor customer through fixed-price contracts that de-risk future commercial demand and keep scaling companies rooted in Europe.

• SkyFi has launched SkyFi for ArcGIS, bringing on-demand satellite imagery directly into Esri workflows via a native Experience Builder widget and live WMTS streaming. The integration allows users to search archives, task satellites, and stream map layers across ArcGIS platforms without switching applications or manually downloading files.

• The 2026–2027 World Bank Country Income Classifications reveal that six countries advanced to higher income categories through different economic mechanisms, including export-led expansion, broad-based domestic growth, and recovery from financial crisis. Non-growth factors such as statistical revisions to national accounts and downward population census adjustments also played decisive roles in pushing certain economies past the per-capita income thresholds.

• UNESCO has launched an open-learning course on Flood Hazard and Exposure Mapping to train participants in flood modeling, risk assessment, and exposure analysis. Featuring hands-on exercises and real-world case studies from Mozambique, the program equips learners to evaluate urban, coastal, and riverine flood impacts for informed decision-making.

Jobs

• Arsenal FC is looking for a Research Engineer under their Data Analytics team.

• The World Bank is looking for a Data Scientist (Consultant) based in Washington.

• WSC is looking for a remote Data Management Specialist based in Canada.

• UNDP is looking for a Data Assistant based in Rio de Janeiro.

• Esri Netherlands is looking for an Analytics Consultant based in Rotterdam.

Just for Fun

A new Smithsonian exhibition marks the US 250th anniversary by mapping the country through its natural contours, such as watersheds, biomes, and fossil beds.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge — a newsletter that’s as anticipated as The Odyssey. The aim, as usual, is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Fine-Grained Damage Assessment: Drone AI improves post-disaster mapping.

• Remote Sensing Agents: RS-Claw streamlines AI tool selection.

• Cross-Border Generalisation: Foundation models struggle across countries.

• Night-Time Dataset: Sentinel-2 releases high-res nighttime imagery.

Subscribe now

Research you should know about

1. Fine-grained damage assessment from drone imagery

Rapid and accurate damage assessment is pretty important after natural disasters, but manual field surveys are often too slow or can be too dangerous to conduct. So automated semantic segmentation models offer a faster alternative, yet they frequently struggle to provide the fine-grained detail required for effective emergency response. When researchers feed high-resolution drone imagery into these models, they typically downsample the images to fit into memory, which actually has the effect of destroying subtle textures that distinguish minor roof damage from major structural collapse. Furthermore, the datasets are heavily imbalanced, with vast amounts of background scenery and very few examples of critically damaged buildings, which often causes standard models to ignore these rare but important hazards.

DA-SegFormer, an adaptation of the SegFormer architecture, is specifically tuned to extract disaster data from high-resolution drone imagery. Instead of aggressive downsampling, the authors use a resolution-preserving protocol where the model processes images in sliding windows, keeping native details intact so it can accurately identify fine-grained damage. To combat the severe class imbalance, they introduced a sampling strategy that forces the model to focus on rare damage cases. They also combined Online Hard Example Mining with Dice Loss, which effectively forces the network to concentrate on the visually subtle decision boundaries between different levels of damage rather than just the easy-to-identify background.

The model shows real promise for emergency responders, as it outperformed the baseline on the RescueNet benchmark with a 74.61% mean intersection over union. Interestingly, it delivered double-digit gains for critical damage classes, identifying minor and major damage far more accurately than standard models. These granular improvements allow for much more reliable automated assessments that help responders allocate help where it is needed most.

2. OpenClaw for Remote Sensing

If you go on X these days, you’re going to be drowning in all the chatter about AI agents. And it’s no surprise that the remote sensing community is also looking at agents. One of the ongoing problems though, is how these agents select their tools. Most current setups just dump the entire list of available tools into the system prompt. Since the remote sensing tool ecosystem is massive, this leads to context bloat and can oftentimes lead to the agent choosing the wrong tool. It creates a major trade-off where agents either struggle to fit everything in their memory or end up omitting critical tools entirely, as they are essentially passive recipients of a flat, static list.

To solve this, researchers introduce RS-Claw. Rather than being passive recipients of a static tool list, RS-Claw agents navigate a hierarchical skill tree. The model essentially discovers and loads tools on demand, moving from a high-level skill summary down to detailed execution documents only when necessary. This progressive disclosure approach keeps the reasoning space clean and free of semantic noise. When tested on the Earth-Bench benchmark, RS-Claw outperformed existing passive baselines across various reasoning tasks, successfully filtering out irrelevant information while slashing input token requirements by up to 86%.

The results show that treating tool acquisition as a dynamic process rather than a static setup significantly improves agent reliability in complex, long-horizon tasks. This is a pretty big shift for geospatial AI, as it allows for much larger, more diverse tool libraries to be used effectively without overwhelming the model.

3. Do foundation models really generalise across borders?

Researchers often hope that foundation models trained on massive satellite archives will provide universal visual features that work anywhere in the world. A new study puts this to the test by using a Leave-One-Country-Out evaluation protocol to see if these models can accurately predict smallholder maize yields across different countries in sub-Saharan Africa. By training models in four countries and testing them on a fifth, the author aimed to find out if geospatial foundation models, like IBM and NASA’s Prithvi-EO, could overcome the usual struggles of cross-country yield transfer.

The results were pretty interesting. While models performed reasonably well when tested on data from the same country, they completely collapsed when asked to predict yields in a new, unseen country, often performing worse than a naive baseline that simply predicts the training set average. Surprisingly, the frozen embeddings from the Prithvi foundation model didn’t provide any meaningful advantage over a standard set of 10-band spectral medians. The author argues that this isn’t necessarily a failure of the foundation models themselves, but rather a reflection of fundamental data distribution shifts in yield across different national contexts that satellite imagery alone cannot easily bridge.

This study serves as a necessary reality check for the field, demonstrating that geospatial pre-training alone can’t close the generalisation gap when yield statistics vary wildly between regions. It establishes a clear, reproducible negative benchmark that should push future research toward more robust domain adaptation and meta-learning techniques.

Geospatial Datasets

1. Sentinel-2 night-time campaign dataset

The Sentinel-2 night-time campaign dataset provides a collection of 56 night-time satellite acquisitions captured during a dedicated observation campaign in December 2025. You can access the data here.

2. Southeast Asian high-resolution climate dataset

The MSEA-Daily dataset provides a high-resolution, gridded daily climate inventory at a 4-km 1/24° spatial scale for mainland Southeast Asia, covering Myanmar, Thailand, Laos, Cambodia, and Vietnam from 1981 to 2014. You can access the data here and the code here.

3. Climate Data Store expanded dataset

The Copernicus Climate Change Service (C3S) has expanded its Climate Data Store (CDS) with a bunch of new open-access climate records and sectoral indicators spanning the atmosphere, land, ocean, cryosphere, and water cycle. You can access CDS here.

4. Cross-platform flood detection dataset

The Sen2GF3Floods dataset provides a multi-source flood mapping resource consisting of 21,483 standardised 10-m resolution image patches (256 × 256 pixels) covering nine major global flood events. You can access the data here and the code here.

Other useful bits

• Google Research found that in 10 major US cities, systematically rerouting under 2% of navigation app users away from bottlenecks can measurably improve traffic speeds and reduce carbon emissions for all road users. By dispersing traffic across underutilised peripheral roads, this is a scalable approach to enhancing network-wide efficiency and urban sustainability.

• Planet Labs is advancing its "Planetary Intelligence" initiative by embedding AI across its satellite hardware, data processing, and user applications to deliver near-real-time Earth observation insights. By integrating edge compute into satellites, sharpening image resolution via AI, and testing a natural-language mapping agent, the company aims to make complex global monitoring accessible to non-technical users.

• European Southern Observatory says that surpassing 100,000 satellites in orbit could render ground-based telescopes obsolete by causing irreversible light pollution and image corruption. The research notes that proposed megaconstellations, including thousands of highly reflective orbital mirrors, threaten to brighten the night sky by up to 300% and permanently disrupt global astronomical observations.

Jobs

• Esri Netherlands is looking for an Analytics Consultant based in Rotterdam.

• UNOPS is looking for an ICT Officer based in Valencia.

• UNDP is looking for a Spatial Planning and GIS Expert based in Malaysia.

• ESA is accepting applications for their mid-year internship opportunities under:

  • Spaceship EAC Team

  • Spaceship ECSAT Team

  • LUNA Analog Facility Team

  • Vulcan Team

Just for Fun

ESA has marked the 1000th image of its "Earth from Space" series with a Copernicus Sentinel-3 satellite capture of Florida, Cuba, and the Bahamas, returning to the exact region featured in the programme's inaugural 2004 edition.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge — the GPT-5.6 Sol of geospatial news. We’re not as impressive as some other news outlets, but we’ll get you where you need to go… In any case, the aim is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• US Census: Harmonising 30 years of housing data.

• Green Growth: Satellites track cleaner economic growth.

• Africa’s Power: Mapping future electricity expansion.

• Methane Trends: Wetlands mask emission reductions.

• Earthquake Damage: Satellite dataset for building damage.

Subscribe now

Research you should know about

1. Fixing the fragmented history of US census data

For researchers trying to track long-term socioeconomic trends in the United States, the decennial census can be a pain... Every ten years, the census redrawing process updates the boundaries of ‘block groups’ to reflect population shifts and urban development. While this keeps current data accurate, it basically fractures the historical record, creating geographic gaps and mismatches that make it almost impossible to directly compare modern housing statistics against those from past decades. A new study published in Scientific Data tries to address this exact problem, presenting a geospatial framework that successfully reconstructs lost or misaligned historical housing data and harmonises it to modern census boundaries.

Using California as a primary test case, the researchers pulled block-group level housing data from the 1990, 2000, 2010, and 2020 censuses. To fix the thousands of missing or incompatible data points across the older datasets, they employed Ordinary Kriging (an interpolation method that essentially estimates missing values by analysing the spatial patterns of the surrounding data). Once these historical gaps were filled, the team used zonal statistics to map the repaired historical data onto the most recent 2020 census block group boundaries. The end result is a single, seamless dataset that allows for direct, apples-to-apples comparisons of California’s median housing values across a 30-year span.

This framework resolves a major blind spot in longitudinal research, providing urban planners and policymakers with a reliable, continuous view of historical housing trends without the noise of shifting administrative borders.

2. Tracking green growth from space

Cities generate over 60% of the world’s GDP and roughly 70% of its greenhouse gas emissions, making them the ultimate battleground for the concept of “green growth”. The problem is that measuring this at a local level has always been a nightmare. Relying on self-reported city emission inventories is messy, and global models are too broad to capture what’s actually happening block by block. A new paper from Nature Cities tackles this by turning to space, using satellites to directly monitor nitrogen dioxide (NO₂) pollution across 5,435 cities worldwide between 2019 and 2024.

The researchers combined this high-resolution satellite pollution data with fine-grained local GDP estimates, and fed both into an advanced statistical model that stripped away the “noise” of local weather patterns. This allowed them to map the “decoupling” of economic growth from fossil fuels. The results show a massive shift: roughly 80% of the 2,475 cities with significant trends successfully achieved relative decoupling, acting as “cleaner + richer” hubs. This trend was heavily concentrated in China, Europe, and North America, where cities managed to grow their economies while driving down their NO₂ footprints, likely through a combination of strict regulations, electrification, and a shift away from heavy industry.

However, the data also highlighted a concerning counter-trend. Around 16% of cities (mostly located in India, Iran, and Russia) fell into the “dirtier + richer” category, locking themselves into a trajectory where economic expansion remains fundamentally tied to fossil-fuel-burning infrastructure and unregulated industrial growth. This new satellite-based framework provides a highly objective, globally consistent tool for policymakers to monitor the real-time effectiveness of green interventions.

3. A look at Africa's power future

Africa’s electricity sector is growing rapidly to meet massive demand, but keeping track of exactly where and how this power is being generated has been surprisingly difficult. A new study from Nature Communications changes that by introducing a detailed spatial database of African power plants. The researchers spent huge amounts of time data mining open access sources to map out 3,139 operational, under construction, and planned power facilities across the continent. By tracking this expansion up to 2030, they have given us a pretty incredible look at how the continent’s energy landscape is shifting over the next few years.

The numbers point to a massive boom in energy generation. The researchers forecast a 57% increase in electricity production between 2023 and 2030. While fossil fuels are still set to dominate, renewables are picking up serious momentum and are expected to grow from 19% to 34% of the total energy mix. However, this growth comes with some pretty big trade-offs. The surge in new hydropower dams will require a massive amount of water, which is going to put intense pressure on already stressed river basins like the Zambezi, Nile, and Congo. At the same time, the ongoing reliance on coal and natural gas means carbon emissions will continue to climb, even though the continent’s current build rate is actually falling slightly short of its international climate commitments.

4. Natural wetlands drive global methane rebound

Methane is a greenhouse gas that has been rapidly accumulating in our atmosphere, with growth rates hitting record highs in 2020 and 2021 before slowing down in 2023. While international agreements, like the Global Methane Pledge, aim to slash anthropogenic methane emissions by 30% by 2030, one of the big challenges is untangling human-made emissions from natural ones. A new paper from Nature Communications addresses this by using a blend of satellite data to track global methane budgets from 2019 to 2024, revealing that massive fluctuations in the Earth’s natural wetlands and atmospheric chemistry are actively masking human mitigation efforts.

The team used an advanced atmospheric inversion system called GONGGA, feeding it high-resolution satellite imagery from TROPOMI and GOSAT. They found that during the 2020–2022 La Niña period, the “methane sink” (i.e. the chemical process where hydroxyl radicals naturally scrub methane from the atmosphere) weakened significantly, driving up global methane levels. However, as the El Niño weather pattern took hold in 2023, the sink recovered and massive droughts across South America suppressed wetland emissions, bringing the atmospheric methane growth rate back down to 2019 levels. But in 2024, as La Niña returned and the droughts broke, wetland emissions rebounded violently. This massive natural release of methane effectively offset the recovering atmospheric sink, resulting in a near-baseline growth rate that hid the underlying surge in emissions.

The study highlights a pretty big problem for global climate policy. The natural variability of wetland emissions and the atmospheric sink combined to create an annual swing of 12.6 million tonnes of methane. This is essentially a figure large enough to entirely erase a full year of the emission cuts promised under the Global Methane Pledge. The researchers stress that as climate change makes weather patterns like La Niña more frequent, these natural methane releases will likely intensify. This means that tracking progress on global emission targets will require sophisticated satellite monitoring systems that can explicitly separate natural wetland surges from actual human emission reductions.

Geospatial Datasets

1. Paleoclimate reconstruction dataset

The DPastCliM-NA dataset provides a bias-corrected, high-resolution monthly reconstruction of temperature and precipitation for North America over the past 2,000 years, featuring explicit uncertainty estimates. You can access the data here and the code here.

2. Earthquake building damage dataset

The Turkey Earthquake change detection dataset (TUE-CD) comprises 1,656 pairs of bi-temporal satellite images captured by the WorldView-2 satellite to assess building damage after the February 2023 earthquake. You can access TUE-CD here and MSI-Net here.

3. Vectorised building footprint dataset

The PCFootprint dataset provides the first large-scale public baseline for extracting vectorised building footprints from airborne laser scanning point clouds. You can access the data here.

4. Disaster risk knowledge graph dataset

An AI-powered data pipeline transforms millions of global news articles from 2014 to 2024 into structured disaster risk knowledge graphs. You can access the data here, the code here, and the interactive dashboard here.

Other useful bits

• Google’s Gemini 3.1 model allows organisations to automate the transformation of unstructured street-level imagery into structured spatial databases through multimodal visual reasoning. By using bounding box detection and integrated code execution, this approach removes the need for training bespoke computer vision models and streamlines complex infrastructure analysis.

• Fine-scale vegetation mapping in West Kalimantan reveals that high-res data is critical for identifying fragmented forests and community agroforestry systems often obscured by conventional land-use planning. This mapping project provides local communities and policymakers with an accurate evidence base to protect biodiversity and prevent illegal plantation expansion into protected peatlands.

• Innovators are tackling the energy supply bottleneck created by AI power demands by developing space-based solar technologies that beam continuous energy directly to terrestrial grids. Rather than necessitating new infrastructure, this method uses near-infrared light to augment the output of existing solar farms, significantly increasing their operational value and reliability.

Jobs

• Vantor is looking for a (1) Geospatial Data Scientist and a (2) Senior Geospatial Data Engineer based in McLean.

• The University of Illinois is looking for a Research Specialist (GIS Analyst) under their Department of Paediatrics based in Chicago.

• Planet is looking for a Pre-Sales Solutions Architect based in Latin America (preferably Colombia).

• ESA is accepting applications for their mid-year internship opportunities under:

  • Spaceship EAC Team

  • Spaceship ECSAT Team

  • LUNA Analog Facility Team

  • Vulcan Team

Just for Fun

South Korean map service Kakao Map has gone viral for providing unexpectedly detailed geographical data on North Korea, including labels for mountains, universities, and railway stops that remain largely obscured on competing platforms.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge — your weekly map to the latest geospatial news. The aim, as always, is to make you a better geospatial data scientist in less than five minutes a week.

First up, I’ll be heading to the AI For Good Summit in Geneva next Tuesday, so hit me up if you’ll be around. I’m always up for a coffee and a geospatial chat.

In today’s newsletter:

• Rural Markets: Satellites track weekly market activity.

• Urban Hazards: Geospatial models flag danger zones.

• Groundwater Loss: Urban sprawl threatens aquifers.

• Urban Weather: Atlanta gets high-res meteorology data.

• AI Economy: Stanford tracks AI’s economic impacts.

• 5G Networks: Glasgow releases urban performance data.

Subscribe now

Research you should know about

1. A new way to measure rural economic activity

In many rural areas, weekly markets are where a lot of the local economy actually happens. Farmers sell crops, traders move goods around, and households buy what they need. But these markets are often informal, scattered across remote areas, and poorly recorded in official data. A new Nature Communications study shows that satellites can help fill this gap. The basic idea is that outdoor markets look different on market days than on other days, because of stalls, crowds, vehicles, and temporary structures. A single satellite image might barely show this, but a long time series of images can reveal a repeated weekly pattern.

The authors use PlanetScope imagery to compare each image with nearby non-market-day reference images, then look for places where brightness and colour change on a regular day-of-week cycle. They validate the approach using known marketplace data from Kenya, Malawi, and Mozambique, where the method achieves high precision, generally above 90% across a broad range of signal strengths. They then apply it across Ethiopia, excluding Addis Ababa, and map 1,776 periodic markets between 2017 and 2024. The resulting map is pretty revealing: around 70% of Ethiopia’s population lives within 10 km of one of these detected markets, and 63% of the markets sit in places with no measurable night-time light emissions. So this picks up a kind of rural economic activity that nightlights would mostly miss.

The activity data is also interesting. Market activity rises and falls with the agricultural calendar, increasing around harvest periods and responding to rainfall shocks. A one standard deviation increase in growing-season rainfall is associated with a roughly 4.2 point increase in harvest-season market activity. The authors also use the data to track disruption during Ethiopia’s recent conflict, finding that periods with more conflict events generally line up with lower market activity. The method has limits, especially where markets are covered, hidden by trees, very cloudy, or not operating around the satellite overpass time. But as a way to monitor rural economic life at high frequency, especially where surveys are sparse or unsafe, this is a pretty neat addition to the geospatial toolkit

2. Spotting urban danger zones with geospatial models

A new Scientific Reports paper looks at a pretty useful problem in urban remote sensing: how to monitor several hazards at the same time. Most satellite-based hazard studies focus on one thing, like floods, wildfire risk, or land movement. This paper tries to bring these together by combining AI with several sources of satellite imagery to map four hazards across two areas in Puglia, southern Italy: floods, land displacement, sediment connectivity, and urban wildfires.

The authors used a mix of Sentinel-1, Sentinel-2, COSMO-SkyMed, and SAOCOM data, alongside environmental and human-made features like elevation, slope, rainfall, soil stability, road distance, land cover, and population density. They trained separate random forest models for each hazard on 100 x 100 m grid cells, then used SHAP analysis to see which factors were actually driving the predictions. This is the interesting bit: the model shows why a place is being flagged. For example, flood risk was strongly linked to drainage position, elevation, soil thickness, and satellite-derived flood probability, while wildfire risk was mainly explained by fuel availability, land cover, and fire climate conditions.

The results were fairly strong, especially for land displacement, sediment connectivity, and wildfires. The multi-source approach generally beat single-hazard models, with big gains for land displacement in both study areas. The SHAP analysis also picked up some useful local patterns. In Gargano, steep slopes and post-wildfire conditions helped explain land instability and sediment movement. In Bari, land displacement was more closely linked to roads, pointing to the role of construction, altered drainage, and other human pressures. The broader takeaway is that multi-hazard mapping works best when it treats hazards as connected systems, rather than separate layers on a map.

3. The toll of urban sprawl on groundwater

As cities continue to expand rapidly around the globe, they transform the natural landscape by replacing soil and vegetation with concrete and asphalt. This unchecked urban sprawl poses a bit of a threat to underground aquifers, which rely on permeable land to soak up rainfall and replenish their stores. However, the exact impact of this urban growth on groundwater recharge and quality has been difficult to map and quantify on a city-wide scale. A new paper from Scientific Reports addresses this gap by focusing on Nnewi, a rapidly expanding industrial city in southeastern Nigeria, using a combination of remote sensing satellites and GIS to analyse changes over the past decade.

The researchers used Landsat imagery from 2013 and 2024 to map the changing landscape, feeding the data into a GIS model to calculate groundwater recharge potential. The results were pretty big. Over the eleven-year period, dense vegetation plummeted by 71%, while impervious, built-up surfaces expanded significantly to cover over 11% of the total area. Crucially, the model revealed that areas capable of recharging the aquifer now make up just 28% of the city. Furthermore, the expansion of concrete and loss of vegetation not only starved the aquifers of water but directly caused water quality to plummet, with industrial and high-density residential areas showing dangerous spikes in nitrates, heavy metals, and dissolved solids.

The study proves that as a city paves over its natural landscape, it effectively suffocates its own water supply while simultaneously trapping and funnelling pollutants directly into the remaining water table. These findings serve as a bit of a warning for rapidly urbanising cities, especially in Sub-Saharan Africa, where groundwater is often the primary source of drinking water. The authors stress that city planners must urgently incorporate “green infrastructure,” such as permeable pavements and urban parks, to restore the natural water cycle before the damage to local aquifers becomes irreversible.

Geospatial Datasets

1. High-res urban meteorology dataset

The HUMID-Atlanta dataset provides a 1-square-kilometre resolution hourly meteorological record covering the Atlanta metropolitan area from 2010 to 2023. Access the data here and the code here.

2. AI economic indicators

Stanford Digital Economy Lab’s AI Economic Indicators provide a comprehensive framework to monitor the macroeconomic and labour market impacts of AI across the US. It combines three measurement tools: the Canaries Dashboard for payroll employment tracking, the Takeoff Tracker for explosive economic growth indicators, and the Adoption Monitor for firm-level survey results. Access the indicators here.

3. Land-atmosphere coupling metrics dataset

The Global Observationally-Based Land-Atmosphere Coupling Metrics (GOLAM) dataset provides global fields of land-atmosphere interaction parameters at spatial resolutions of 0.1° and 0.25°. Access the data here and the code here.

4. Urban 5G network performance dataset

The Glasgow Public and Private 5G Performance Dataset provides 720 empirical outdoor network measurements collected across 15 Glasgow neighbourhoods over a three-day period in April 2025. Access the data and code here.

Other useful bits

• A comparative satellite analysis has highlighted 10 global water bodies suffering from severe "continental drying" caused by climate change and unsustainable land use. Long-term imagery reveals that critical freshwater sources, including Argentina's Parana River, Iran's Lake Urmia, and the South Aral Sea, have experienced drastic volume reductions that threaten local ecosystems and regional water security.

• Google Research has developed a high-res deep learning framework that converts pixel-based satellite imagery of agricultural landscapes into a vectorised dataset of fine-scale ecological features. This automated pipeline uses Google Earth AI and Google Earth Engine to programmatically classify and map millions of small-scale vegetation structures across England, providing a precise inventory to support nature restoration without compromising food security.

• The open-access Harvest2Market platform, developed by NASA Harvest, centralises satellite imagery, market indicators, shipping data, and global crop reports to bridge information silos in agricultural monitoring. This web-based system enables analysts to evaluate food supply chain risks by mapping how regional events, such as severe weather or trade disruptions, impact interconnected global markets.

• A satellite has successfully utilised a vision-language model in orbit to identify specific environmental features and infrastructure from natural language queries without ground-based human analysis. The demonstration, conducted on Loft Orbital's YAM-9 spacecraft using a streamlined version of Google DeepMind’s Gemma 3 model, proves the viability of on-orbit data triage and paves the way for autonomous, real-time monitoring constellations.

Jobs

• EIB is looking for (1) a Data Warehouse Officer and (2) an Associate DevSecOps Engineer, both based in Luxembourg.

• UNOOSA Germany is looking for a GIS / Remote Sensing Expert based in Bonn.

• Médecins Sans Frontières UK is looking for (1) a Missing Maps Coordinator and (2) a GIS Specialist (Maternity Cover) based in London

• IOM is looking for a Data and Research Intern based in Vienna

Just for Fun

Back-on-Track.eu has launched an interactive 2026 European night train map tracking 205 regular sleeper lines, highlighting five new routes alongside ten closures.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge — we’re sort of like the R of the geospatial world. Free, intuitive, but for some reason people still keep going back to Stat…In any case, the aim is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Earthquake Risk: Bayesian models predict building-level damage.

• Disaster Tracking: High-frequency satellites monitor fires and storms.

• Wind Siting: AI identifies optimal wind farm locations.

• Smart Farming: Decentralised AI improves pest management.

• Field Mapping: Three billion agricultural fields mapped globally.

Subscribe now

Research you should know about

1. Predicting earthquake damage with physics and statistics

Earthquakes pose a massive threat to dense urban areas, but predicting exactly how a city will hold up during a major tremor can be pretty complex. Standard risk assessments often rely on simplified charts or historical data, which can miss the nuanced structural differences between individual buildings. A new paper from Scientific Reports presents a large-scale seismic risk framework that combines rigorous physics-based structural modelling with the predictive power of Bayesian networks. By focusing on District 2 of Tehran as a case study, the team managed to evaluate thousands of buildings to map out precise zones of vulnerability.

The researchers categorised the district’s buildings based on their age, number of storeys, and structural materials (such as steel or reinforced concrete). They then ran intensive simulations to see how each building type would react to varying earthquake intensities. Because it’s obviously impossible to perfectly predict every variable in a sprawling city, the team fed their simulation results into a Bayesian probabilistic network, which is a statistical model that excels at handling uncertainty. The results revealed that newer buildings generally fare better, but specific vulnerabilities stood out. For instance, reinforced concrete structures were more prone to damage from horizontal shifting (drift), while steel buildings suffered more from raw shaking acceleration. Alarmingly, the model pinpointed two specific zones in the district facing up to a 59% chance of complete destruction in a severe earthquake.

This granular, data-driven approach gives city planners and emergency responders a bit more of a sharper tool for identifying weak points in urban infrastructure before disaster strikes.

2. Tracking fast-moving disasters from space

As climate-related disasters like wildfires and flash floods become more frequent and severe, the need for real-time monitoring from space has never been greater. Earth Observation (EO) models powered by artificial intelligence have made massive leaps recently, with developers building geospatial foundation models (more on that here and here), that are trained on vast amounts of unlabelled satellite imagery. However, nearly all of these existing models are trained on imagery from polar-orbiting satellites, which provide incredibly detailed, high-resolution pictures but only pass over the same spot on Earth every few days. This makes them entirely unsuitable for tracking fast-moving, dynamic events like a rapidly spreading forest fire or a developing storm system.

A new study introduces HighFM, a new foundation model explicitly designed for high-frequency environmental monitoring. Instead of using highly detailed but infrequent images, the team trained their model on more than 2 terabytes of imagery from the Meteosat Second Generation platform. These geostationary satellites provide lower-resolution images but update every 15 minutes, offering a near-constant view of the Earth. To handle this unique data, the researchers adapted a masked autoencoder architecture to retain fine-grained, minute-by-minute temporal information that other models typically discard.

To test HighFM, the researchers asked the model to identify active fires and map cloud cover. When benchmarked against traditional AI models and even other state-of-the-art EO foundation models, HighFM consistently proved superior. The model’s ability to understand the rapid, short-term variability between images allowed it to detect fires and clouds with greater accuracy and fewer false alarms. The researchers noted that while HighFM was specifically tested on fire and cloud detection, the architecture could easily be adapted for tracking other rapid phenomena, like storm development or solar energy forecasting, offering a powerful new tool for civil protection and disaster response agencies.

3. Optimising wind farm locations in data-scarce regions

Wind power is a pretty important part of the global transition to renewable energy, but building a wind farm is an incredibly expensive long-term investment. To ensure a wind farm is economically viable, planners need to carefully select sites based on a bunch of technical, environmental, and socio-economic factors. Typically, researchers use Multi-Criteria Decision-Making (MCDM) tools, which rely on expert opinions to weigh factors like wind speed, terrain slope, and proximity to roads. However, these methods are pretty static, pretty subjective, and unable to predict the suitability of locations that haven’t been directly sampled. A new paper from Scientific Reports introduces a new framework that uses expert knowledge, objective efficiency analysis, and predictive machine learning to identify the best wind farm locations in the data-scarce Amhara Region of Ethiopia.

The framework operates in three stages. First, a Fuzzy Analytic Hierarchy Process (FAHP) was used to gather and weight expert opinions, identifying wind speed and proximity to power grids as the most critical factors while smoothing out the vagueness inherent in human judgment. Second, Data Envelopment Analysis (DEA) was applied to objectively measure the efficiency of these candidate zones, pinpointing the North Shewa zone as the absolute best location for converting available land and wind into energy output. Finally, the researchers trained three machine learning models (Random Forest, SVM, and XGBoost) on the FAHP-DEA data. The Random Forest model proved to be the most accurate, learning the complex spatial patterns to predict wind farm suitability across a massive 31-million-pixel map of the region.

This integrated approach successfully delineated 1,698 square kilometres of high-priority land within North Shewa, divided into two distinct development corridors. These findings were strongly corroborated by independent engineering assessments and the Ethiopian government’s own development pipeline. By successfully combining subjective expert weighting, objective efficiency screening, and machine learning generalization, this framework provides a highly accurate, scalable blueprint for strategic wind energy planning that can be easily adapted for other developing regions worldwide.

4. A smart farming framework for pest management

Sustainable agriculture faces massive challenges from unpredictable climate changes, varying soil conditions, and sudden pest outbreaks. Precision agriculture aims to solve these problems by using sensors and AI to manage farms efficiently. However, models like Convolutional Neural Networks (CNNs) struggle to understand the complex, real-world relationships between weather, soil, and crops. Furthermore, centralising data from multiple farms onto a single server creates privacy and bandwidth issues. To tackle these challenges, a team of researchers developed GNN-ML-FRL, a highly advanced, decentralised AI framework designed to identify pests and model environmental factors across diverse farming regions.

This new framework combines three techniques. First, Graph Neural Networks (GNNs) map out the spatial relationships between different sensors and farm plots, allowing the system to understand how a pest outbreak in one area might affect a neighbouring field. Second, Federated Learning (FL) enables different farms to collaboratively train the AI without ever sharing their private, raw data with a central server. Finally, Meta-Learning (ML) and Reinforcement Learning (RL) allow the system to rapidly adapt to new, unseen environmental conditions and make real-time decisions, such as optimising irrigation or pesticide use.

When tested against a massive dataset containing over 52 million environmental sensor readings and 75,000 images of insect pests, the GNN-ML-FRL framework outperformed all standard AI baselines. It achieved an 89.3% accuracy rate in pest identification and demonstrated a significant reduction in prediction errors across geographically unseen farms. The researchers also used the framework to analyse genetic crop data, proving its versatility. Ultimately, this scalable, privacy-preserving AI system offers a robust blueprint for the future of sustainable, data-driven farming. You can access the code here.

Geospatial Datasets

1. Global agricultural monitoring and field boundary dataset

The collection integrates a multi-source framework that combines high-resolution spatial mapping with continuous biophysical tracking to capture field-level dynamics and macro-scale crop conditions. This maps over 3 billion agricultural field units at a 10-metre resolution across 241 countries and territories. You can access the FOTW project here and the GEOGLAM Crop Monitor here.

2. High-resolution canopy height dataset

The NAIP-CHM dataset provides a continuous 0.6-metre resolution canopy height and vertical structure model across the US. Derived from National Agriculture Imagery Program aerial imagery spanning 2012 to 2023, the dataset characterizes the full vertical profile of the landscape, mapping both natural vegetation heights and human-made infrastructure. You can access the data here and the code here.

3. Text-guided object counting dataset

The Cross-domain Large-scale Object Counting (CLOC) dataset is a text-guided object counting collection designed to evaluate generalist vision models across diverse imaging conditions and density distributions. It unifies multi-source heterogeneous annotations from public repositories into a standardized category-specified format, containing approximately 220,000 images, 619 categories, and 15 million instance point annotations. You can check out the project page here and the code here.

4. Earthquake high-frequency magnitude catalogue

The Italian Catalog of Earthquakes based on the high-frequency earthquake Magnitude m3Hz (ICEM) provides high-frequency magnitude records for 1,951 historical and recent Italian earthquakes spanning from 1117 to 2020. You can access the data here.

Other useful bits

• Google Earth has introduced new artificial intelligence-powered data layers that extract infrastructure assets like fire hydrants and storm drains directly from Street View imagery. Available for Professional and Professional Advanced users on web and Android, these layers combine with updated building footprints and land parcels to allow remote inspection of utility access and site constraints.

• An open-source tool called SplatAnalyzer utilises cloud inference to convert 3D Gaussian splats into structured data by generating 3D bounding boxes and object labels in under a minute. Designed for applications in VFX, gaming, digital twins, and physical AI, the interactive kit allows users to isolate, highlight, and manipulate specific assets within a scene without needing local compute.

• The Environmental Defenders Office has launched an interactive map on World Oceans Day to help the Australian public track offshore fossil fuel expansion in their waters. The tool outlines the environmental risks of deep-sea drilling and seismic blasting to marine biodiversity, while providing users with information on how to exercise their legal rights in environmental decision-making.

Jobs

• SEI is looking for a Mid–senior Researcher in data systems and digital governance based in Stockholm.

• EIB is looking for a Data Warehouse Officer based in Luxembourg.

• PAHO is looking for a remote GIS and Information Products Consultant.

• Vizzuality is looking for a remote Software Engineer based in either Spain, Portugal or UK.

• GiveDirectly is looking for a Remote Sensing Analyst based in New York.

Just for Fun

Hold on to your horses, NASA’s Voyager 1 will cross the 1 Light-Day threshold on November 18, 2026. That’s about 16 billion miles from Earth. If you use marathons as your unit of measurement, that’s like 610 million marathons.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge — we’re sort of like Playboy for geospatial nerds. In any case, the aim is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Rice Monitoring: Drones improve crop health estimation.

• Generative Earth Observation: COP-GEN captures real-world variability.

• Damage Assessment: 3D disaster dataset for structural damage.

• Levee Erosion: Multi-scale wave overtopping dataset.

• Ocean Nutrients: Warming seas increase nutrient stress.

Subscribe now

Research you should know about

1. Taking precision agriculture to new heights

If, like me, you live in Asia, then you know that rice is kinda a big deal. Monitoring the Leaf Area Index (LAI) of rice crops is an important part of understanding their health and water use efficiency. Farmers and researchers traditionally measure this by hand, which is incredibly slow, labour-intensive, and difficult to scale across massive agricultural operations. To speed things up, scientists are increasingly turning to drones equipped with advanced cameras. However, standard drone imagery often struggles with changes in lighting or gets confused by the soil background when the crop canopy is either too thin or too dense. A new paper from Scientific Reports tries to address this by blending multiple types of drone data to create a highly accurate monitoring framework specifically tailored for rice during its critical booting stage.

The research team flew drones equipped with both standard RGB and multispectral cameras over rice fields at altitudes of 30 and 60 metres. Instead of just looking at basic colours, they extracted a complex mix of colour indices, spectral vegetation indices, and spatial texture features. They also developed new texture feature indices that specifically couple spatial patterns with spectral data. When they fed this fused data into six different machine learning models, the results were highly impressive. They found that flying at a lower 30-metre altitude and combining colour, vegetation, and the new texture indices yielded the best outcomes. Using this specific combination, a Convolutional Neural Network and a Random Forest model achieved an accuracy score of over 90%, proving that blending spatial and spectral data easily overcomes the blind spots of using just one type of sensor.

These findings provide a pretty interesting solution for those looking to implement precision agriculture. By confirming the ideal flight altitude and the exact mix of data features needed, the researchers have created a reliable blueprint for tracking crop health from the sky. Interestingly, the study also noted that standard machine learning models like Random Forest performed incredibly well given the relatively small sample size, sometimes even outpacing more complex deep learning networks.

2. Cop-Gen: A generative approach to Earth observation

In the geospatial space (or ‘geospace’, if you will), satellites gather petabytes of data daily, capturing everything from optical imagery and radar to terrain elevation and land cover. The challenge lies in integrating these diverse “modalities” into a single, cohesive system. Traditionally, we’ve relied on deterministic models, which take a set of inputs and produce a single, averaged “best guess” output. However, the real world is obviously pretty ambiguous. A single type of terrain or land cover could look pretty different depending on the weather, the season, or the lighting. By forcing a single output, these models fail to capture the true variability of Earth’s physical systems.

Researchers built COP-GEN to address this: a new multimodal generative model designed to handle the massive influx of data from the Copernicus satellite program. COP-GEN uses a “latent diffusion transformer” architecture, which allows it to process all the different types of satellite data at their native resolutions simultaneously. Instead of producing one averaged image, COP-GEN acts as a stochastic model, meaning it generates multiple, physically plausible versions of a scene based on the input data. The model can even perform “zero-shot” translation, meaning it can fill in missing data (e.g. generating a realistic optical image using only radar and elevation data) without needing to be retrained for that specific task.

When tested against a massive global dataset, COP-GEN proved pretty good at capturing the natural variance of the Earth’s surface. In a head-to-head comparison with a leading deterministic model, COP-GEN was able to cover 90% of the possible real-world variations of a scene, whereas the competing model collapsed to just 2.8%, producing the same image over and over. You can access the code here.

Geospatial Datasets

1. Post-disaster structural damage 3D dataset

The 3DAeroRelief dataset provides the first 3D benchmark dataset specifically designed for large-scale outdoor post-disaster structural damage assessment. It comprises dense 3D point clouds reconstructed via Structure-from-Motion and Multi-View Stereo techniques using optical RGB imagery captured by low-cost unmanned aerial vehicles over hurricane-damaged areas. You can access the data here and the code here.

2. Levee erosion and wavefront velocity dataset

The collection provides a multi-scale dataset compiled from four experimental field campaigns and corresponding laboratory tests designed to evaluate the spatiotemporal erosion resistance of earthen levee cohesive clay covers under wave overtopping loads. You can access the data here and the code here.

3. Glacier inventory dataset

The Tajikistan Glacier Inventory (TGI) provides a national-scale glacier dataset for Tajikistan derived from 3-metre resolution PlanetScope satellite imagery acquired in 2024. The collection maps 12,325 glaciers covering a total area of 7,720.26 square kilometres. You can access the data here and the code here.

Other useful bits

• A new study combining two decades of NASA Aqua satellite data with global ocean surveys has mapped nutrient stress in marine microorganisms caused by warming waters. By tracking shifts in the carbon-to-chlorophyll ratio of phytoplankton alongside genetic markers, researchers found that ocean stratification is increasingly limiting nutrient availability, particularly in subtropical gyres.

• The EU is investing billions of dollars in several large-scale rail tunnels to bypass natural geographic barriers and shift passenger and freight traffic from air and road to rail. While mega-projects like the Gotthard, Brenner, and Mont Cenis base tunnels promise to significantly reduce travel times across the continent, they face substantial construction delays, ballooning costs, and a lack of completed connecting lines from neighboring countries.

• A recent article addresses the challenge of building machine learning models in geospatial applications where satellite data is abundant but accurate field labels are rare and expensive. The author outlines a practical workflow for small environmental datasets, highlighting that rigorous spatial validation and transparent mapping of uncertainty are critical to avoid artificial inflation of model accuracy.

• ESA has awarded a contract to Thales Alenia Space and Airbus Defence and Space to build two Copernicus Sentinel-1 Next Generation satellites to ensure long-term continuity for Europe’s radar Earth observation capabilities. The new satellites will feature advanced C-band synthetic aperture radars that extend coverage to the polar regions and offer a fourfold improvement in geometric resolution compared to the first-generation mission.

Jobs

• PAHO is looking for a remote GIS and Information Products Consultant.

• Vizzuality is looking for a remote Software Engineer based in either Spain, Portugal or UK.

• IOM is looking for a GIS Manager based in Berlin.

• GiveDirectly is looking for a Remote Sensing Analyst based in New York.

• UNITAR is looking for (1) Geospatial Analysts - Satellite Imagery and GIS and (2) Geospatial Quality, Assurance, Training and Analytical Support Specialists based in Geneva.

• WFP is looking for a Geospatial Data Engineer based in Rome.

Just for Fun

A time-lapse captured from a SpaceX Dragon spacecraft shows a spectacular southern aurora caused by a recent solar event, with the phenomenon dancing directly below the vehicle.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge — a safe space for the cartographically curious. As usual, the aim is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Poverty Mapping: Tempov updates wealth maps using satellites.

• Thermal Imaging: Dual AI models sharpen infrared imagery.

• Satellite Training: Pixel diversity improves foundation model performance.

• Weather Extremes: AI forecasts struggle beyond past records.

• Nightlights: NASA reveals shifting global illumination patterns.

Subscribe now

Research you should know about

1. Stanford’s foundation model for predicting poverty

Traditional methods for measuring poverty in developing nations rely on household surveys and census data. These methods can be expensive, slow, and infrequent, which leaves policymakers working with outdated information when deciding where to allocate resources or build infrastructure. Over the last decade, researchers have started using daytime satellite imagery to try to measure economic activity, looking for visual proxies like new roads or agricultural expansion. While this works well for a static snapshot of an area, these models often struggle to track changes in wealth over time, confusing seasonal changes like fields turning brown in the winter for actual economic decline.

To fix this, a team of researchers built Tempov, a new geospatial foundation model specifically designed to track long-term economic changes (more info on foundation models here). The model was trained on a massive dataset of three million satellite images spanning two decades. Interestingly, the researchers intentionally fed the model ‘bi-temporal’ pairs, i.e. pictures of the exact same location taken during different seasons of the same year. This forced the model to ignore transient, seasonal changes like snow or dead crops and focus entirely on permanent, human-made changes like new buildings or roads.

Tempov proved to be pretty effective. The researchers demonstrated that if Tempov was given a decade-old census from a country, it could accurately predict the modern wealth distribution using only 5% of modern survey data. In resource-strapped nations, this means governments can achieve highly accurate poverty mapping at a fraction of the cost of a full census. To prove its scalability, the team deployed Tempov to map the entire continent of Africa, producing a high-resolution, 6km-by-6km grid showing the exact changes in wealth from 2015 to 2025. This open-source tool could dramatically change how we monitor and combat global poverty.

2. Sharpening the world of thermal imaging

Enhancing the resolution of remote sensing imagery can be pretty important for applications like autonomous driving and environmental monitoring. But Super-resolution for infrared or thermal images brings a specialised set of challenges. Unlike standard visible light photographs, thermal images lack rich textures and rely heavily on clear object outlines and stable regional temperatures. If a super-resolution system sharpens the image too aggressively, it can create fake hot spots or unstable temperature readings that ruin the utility of the data.

To tackle this, a team of researchers developed a dual-branch framework for the NTIRE 2026 Infrared Image Super-Resolution Challenge. Instead of forcing one neural network to do everything, they combined two different architectures to capture the best of both worlds. The first branch uses a transformer model called HAT-L to focus on sharpening local edges and fine contours. The second branch employs a state-space model named MambaIRv2-L to handle the big picture, ensuring long-range temperature consistency across the entire scene. The system processes the low-resolution images through both branches independently, applies specific test-time enhancements to each, and simply averages the two outputs together to create the final image.

This collaborative approach proved pretty successful in generating clear and accurate thermal imagery. By blending the aggressive edge recovery of the transformer with the global stability of the state-space model, the fused output comfortably beat either model working on its own. The researchers evaluated their system using the official NTIRE 2026 development and hidden test sets, alongside a custom evaluation batch derived from the Caltech Aerial RGB-Thermal dataset.

3. Mapping out the best data for satellite AI

As I mentioned above, in recent years, developers have built massive ‘foundation models’ designed to analyse satellite imagery for tasks like mapping populations, tracking floods, or estimating crop yields. To train these models, researchers use huge datasets of unlabelled satellite images, a process known as pretraining. However, there are no real standards for how this data is collected. Some models are trained exclusively on images of European cities, some are trained on a perfectly even grid of the entire Earth, and others focus heavily on specific biomes like deserts or forests. Until now, no one had rigorously tested how these different geographic sampling strategies actually affect the AI’s final performance.

To solve this, researchers built seven distinct pretraining datasets (six covering individual continents and one global mix) and used them to train identical foundation models. They then tested these models on a variety of global and local mapping tasks. The results were highly surprising. The model trained exclusively on images of Europe consistently outperformed all the others, including the globally trained model, across almost every single task. Even when the AI was asked to map farms in Africa or populations in South America, the Europe-trained model proved to be the most accurate.

To figure out why the European dataset performed so well, the researchers measured the ‘diversity’ of all the datasets across different metrics. They discovered that simply having a wide variety of continents, biomes, or landcover types in the training data didn’t actually boost performance. What mattered most was ‘spectral diversity’, which measures the sheer complexity and variety of colour and light values within each individual image. Because the European images happened to have the highest spectral complexity, they provided the best training ground for the AI, proving that when it comes to training satellite models, the richness of the pixels matters far more than the geographic coordinates they came from.

4. AI vs Physics: Forecasting extreme weather

Geospatial models have begun to target weather forecasting, often outperforming leading physics-based numerical models on standard benchmarks. Models like GraphCast and Pangu-Weather are faster, more energy-efficient, and generally provide more accurate daily forecasts. However, scientists have questioned whether these data-driven models can reliably predict unprecedented, record-breaking weather events. Unlike physics-based models that calculate the actual laws of thermodynamics and fluid dynamics, AI models rely entirely on finding patterns in historical training data. A new study from Science Advances put this to the test by evaluating how well leading AI models could predict record-breaking heatwaves, cold snaps, and windstorms compared to the gold-standard physics-based model, the European Centre for Medium-Range Weather Forecasts’ HRES.

The researchers built a massive dataset of record-breaking weather events from the years 2018 and 2020. These events were specifically chosen because they exceeded any measurement recorded in the AI models’ training data (which covered 1979-2017). When tested on these unprecedented extremes, the AI models consistently failed, significantly underperforming the physics-based HRES model. The AI models not only failed to predict the frequency of these record-breaking events, missing many of them entirely, but they also severely underestimated their intensity.

Importantly, the researchers found that the larger the extreme event was compared to historical records, the worse the AI’s prediction became. It was as if the AI models hit a soft “ceiling” based on their training data and simply could not imagine an event being any worse.

These findings highlight a dangerous “extrapolation” problem in current AI weather models. Because they are not grounded in the actual laws of physics, they struggle to predict events that fall outside of their historical training data. While AI offers incredible speed and general accuracy, this study proves that we cannot solely rely on them for early warning systems or disaster management, particularly in a rapidly warming climate where unprecedented extremes are becoming the new normal. For now, physics-based models remain essential for keeping communities safe from the worst that nature can throw at them.

Geospatial Datasets

1. European agricultural land use systems map

This collection provides a harmonised, Europe-wide agricultural land use systems map at a 10-m spatial resolution covering the period from 2017 to 2021. You can access the data and code here.

2. Global sea and sea-Ice surface temperature dataset

This dataset introduces a 43-year Climate Data Record providing daily, gap-free global fields of combined sea surface temperature and sea-ice surface temperature at a 0.05-degree spatial resolution from 1982 to 2024. You can access the data here and the code here.

3. Pan-Arctic super-resolved land surface temperature dataset

A new 42-year climate data record offers twice-daily, clear-sky land surface temperature (LST) data for the entire pan-Arctic region at a 1-km spatial resolution from 1982 to 2023. You can access the data here.

4. Global coastal soil carbon accumulation dataset

A new global dataset provides a comprehensive inventory of soil organic carbon accumulation rates within coastal wetlands. This open-access resource contains 8,998 standardised field observations derived from a systematic review of 247 peer-reviewed articles published between 1981 and 2025. You can access the data and code here.

Other useful bits

• New imagery from the NISAR satellite has revealed that Mexico City is sinking at an alarming rate of up to 0.8 inches a month due to the over-pumping of its underlying aquifer. It captures subtle ground movements in detail, providing some pretty shocking insights to help the city manage its infrastructure and pressing water crisis.

• As the number of active satellites skyrockets, the science academies of G7 countries have called for the creation of an Intergovernmental Panel on Space Sustainability (IPSS). This initiative aims to establish international standards and global thresholds to protect the night sky, manage orbital space debris, and balance commercial innovation with environmental safety before a critical tipping point is reached.

• By analysing a decade of data from its "Black Marble" program, NASA revealed that the Earth's nighttime lighting is highly dynamic and flickers with industrial shifts. The study shows how changes in artificial light trace real-world human stories, from economic booms and urban development to energy crises and green policy shifts. And if you’re a nightlights nerd, you might want to check this out.

Jobs

• WFP is looking for a Geospatial Data Engineer based in Rome.

• UNICEF is looking for a GIS Lead based in Florence.

• GiveDirectly is looking for a Remote Sensing Analyst based in New York.

• UNITAR is looking for (1) Geospatial Analysts - Satellite Imagery and GIS and (2) Geospatial Quality, Assurance, Training and Analytical Support Specialists based in Geneva.

• WWF is looking for a Senior Conservation Data Analyst/Engineer based in Washington, DC.

Just for Fun

Jupiter and Venus, on 9 June, had the appearance of being pretty close together. This is known as a ‘planetary conjunction.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge — your weekly map to the latest geospatial news. We buffer out the noise, snap to the most interesting features, and provide geospatial puns that are even worse than the GeoDA UI. In any case, the aim is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Trade Deaths: Global trade shifts pollution deaths abroad.

• Political Violence: Conflict scars the next generation.

• EO Embeddings: Compact vectors streamline satellite analysis.

• SAR Matching: Off-the-shelf AI aligns disaster images

• Climate Pledges: ClimActor tracks subnational climate action.

Subscribe now

Research you should know about

1. The hidden health costs of global trade

If you’ve ever been to a city, or read about cities in books, or seen a city in a movie, then you might have experienced the joys of air pollution from fine particulate matter (PM2.5). It causes around 5.1 million deaths a year. While it’s easy to blame the countries where this pollution is emitted, the reality of global economics paints a bit more of a complicated picture. When a wealthy country imports cheap, manufactured goods, it’s effectively outsourcing the dirty, dangerous industrial work required to make them. This is basically the idea of ‘imported emissions’. Until now, scientists have struggled to map exactly how this international trade shifts the health burden of air pollution from wealthy nations to poorer ones on a country-by-country basis.

A new study from Nature Communications quantifies this by combining global economic data, air pollution transport models, and epidemiological studies. The researchers discovered that roughly 40% to 48% of all economy-related PM2.5 deaths are caused by the production of goods meant for export. Most shockingly, 14% to 18% of all global air pollution deaths occur when a wealthy country (with a high GDP) imports goods from a poorer country (with a GDP at least 50% lower). This means that high-income nations like the US and the UK are effectively exporting thousands of pollution-related mortalities to countries like China, India, and Bangladesh in exchange for cheaper goods.

This massive imbalance exposes a major flaw in how governments currently calculate the ‘externality costs’ of pollution. In countries like the US, the economic cost of pollution is only calculated based on the lives of its own citizens. If a US policy causes pollution deaths in another country, those lives are valued at zero, essentially incentivising companies to offshore their most toxic practices to the poorest nations. The researchers propose a new ‘Fair Trade in Pollution’ model that forces wealthy nations to value the lives lost in producing countries as highly as their own citizens, a shift that would drastically alter the math behind international trade policies. You can access the code for this study here.

2. The generational toll of political violence

We all know that armed conflict and political violence cause massive, immediate suffering. However, much less is known about the ripple effects these events have on the daily lives of children and young adults long after the dust settles. When a community experiences political violence, the resulting trauma, economic instability, and breakdown of local services can create a culture of insecurity. In these high-stress environments, violence is often normalised as a way to resolve disputes, leading to an increase in interpersonal violence within families and romantic relationships.

To measure this phenomenon, researchers built a massive dataset combining the Violence Against Children and Youth Surveys (VACS) with the Armed Conflict Location and Event Data (ACLED) project. By cross-referencing these two databases, they were able to track the history of political violence across nine African nations and compare it directly to self-reported instances of interpersonal violence against adolescents and young adults. The results showed a stark, long-term connection. If a community experienced a significant increase in political violence over a 15-year period, the local youth were 5.5% more likely to experience emotional violence from family members and 3.0% more likely to suffer physical violence from an intimate partner. Notably, the study found no significant link between short-term political violence and increased youth violence, proving that this is a generational, slow-burning issue.

The study also highlighted how economic factors exacerbate these dangers, with young people in poorer households being 16.9% more likely to experience sexual violence following long-term political instability. These findings prove that the scars of political violence extend far beyond the battlefield, infecting the homes and relationships of the next generation. You can access the data and code here.

3. Optimising Earth observation with compact embeddings

Earth observation satellites capture petabytes of multispectral imagery every day, which researchers increasingly analyse using large Geospatial Foundation Models. Instead of constantly feeding raw data through these massive models for every new task, scientists are shifting towards an embedding-centric approach. In this setup, models process the imagery once to create compact, reusable vector representations known as embeddings. However, deciding exactly how to compress this data, such as which layers of the model to use or how to pool the information, significantly impacts performance. A new study from IBM systematically evaluates these design choices to help developers build highly efficient Earth observation pipelines without sacrificing accuracy.

Using an evaluation framework called NeuCo-Bench, the researchers tested various embedding configurations across multiple tasks, including land cover classification, biomass estimation, and cloud cover prediction. They discovered that Vision Transformers consistently outperformed traditional convolutional networks like ResNet when dealing with continuous environmental variables. Furthermore, simply taking the average of the spatial features, a technique known as mean pooling, proved to be the most robust aggregation strategy across the board. The team also noted that while transformer models perform best at their final layers, convolutional networks actually produce their most useful embeddings at intermediate stages, meaning the default practice of using the final layer is often a mistake.

The findings show that carefully designed, fixed-size embeddings can compress raw satellite data by over 500 times while remaining incredibly useful for a wide range of analytical tasks. The study also highlights that different training methods excel at different tasks, meaning that combining embeddings from multiple models can further boost overall reliability. By following these design guidelines, developers can create scalable, lightweight systems for monitoring our planet.

4. Can off-the-shelf AI match optical and SAR satellite imagery?

When a disaster like a hurricane strikes, thick cloud cover often makes standard optical satellite imagery useless. In these situations, responders rely on Synthetic Aperture Radar (more info on SAR here) which can see through clouds but produces grainy, distorted images that look radically different from regular photos. To quickly assess damage, analysts need to perfectly align, or “register,” these new SAR images with older optical basemaps. It’s a pretty notoriously difficult task because the two types of imagery are fundamentally different. While there are plenty of highly advanced AI image-matching tools available, almost all of them were trained on normal, ground-level photos. A new study tries to see whether these general-purpose AI matchers actually handle the extreme differences between overhead optical and SAR satellite imagery.

To find out, a new paper tested 24 different pretrained AI matchers on SpaceNet9, a dataset of paired optical and SAR satellite images. They evaluated the models in a “zero-shot” setting, meaning the AIs had to perform the task straight out of the box without any special training or fine-tuning for satellite data. The results were surprising. While models specifically trained for cross-modal tasks (like matching visible light to thermal images) performed exceptionally well, some general-purpose models performed almost identically. For example, a model called RoMa, which was only ever trained on normal photos, achieved the lowest error rate alongside the cross-modal specialist XoFTR. This suggests that the sheer size and diversity of data used to train modern foundation models gives them an unexpected ability to bridge the gap between radically different types of imagery.

However, the study also revealed that picking the “best” AI model is only half the battle. The researchers found that how the model is deployed (specifically the geometric math used to align the images and how the images are cropped into tiles) often mattered more than the choice of AI itself. E.g., simply forcing the models to use an “affine” geometry model (which assumes the ground is relatively flat) rather than a “homography” model (which accounts for complex 3D perspective) slashed the error rate by quite a lot. The key takeaway for disaster responders is that while off-the-shelf AI matchers are surprisingly capable of handling satellite data, they require careful, domain-specific configuration to be truly reliable in the field.

Geospatial Datasets

1. Global agricultural field boundary map

The PRUE global dataset delivers the very first wall-to-wall map of agricultural field boundaries at a pretty granular 10-m resolution for 2024 and 2025. You can access the data here and code here.

2. Spatiotemporal hyperspectral EO dataset

The ChronoEarth-492K dataset introduces the first large-scale, temporally calibrated hyperspectral self-supervised learning collection, built on NASA’s EO-1 Hyperion mission archive spanning from 2001 to 2017. You can access the data and code here.

3. Climate pledges and emissions dataset

The ClimActor 2.0 database provides an expanded, openly available geospatial resource that maps climate targets, action plans, and emission inventories from subnational governments worldwide. You can access the data here and the code here.

4. National-scale sandy beach dataset

This new dataset introduces the first comprehensive, high-resolution national map of sandy beaches along the extensive 11,098.81 km coastline of India.

Other useful bits

• The creators of the Global Automation Atlas have mapped how automation will unfold across 124 countries by comparing over 18,000 distinct tasks. This new tool provides an incredible 2.3 million task-country comparisons.

• The new GeoX framework is impacting geospatial reasoning by using a self-play mechanism that allows AI to teach itself complex spatial logic without the need for costly human data. By generating and solving its own programming problems, this system creates a virtuous cycle of autonomous improvement that pushes vision-language models to state-of-the-art performance.

• NASA and IBM’s open-source Prithvi model has become the very first AI geospatial foundation model to be successfully deployed in orbit. By running directly on the International Space Station and the Kanyini satellite, the model system can quickly analyse floods and wildfires from space, cleverly saving massive amounts of bandwidth by processing Earth observation data before it even hits the ground.

• In the wake of Hurricane Melissa, Jamaica's National Spatial Data Management Branch demonstrated the power of resilient geospatial infrastructure by mapping nearly 200,000 damaged buildings within days. The nation used its National Geospatial Repository and partnered with international tech organisations to create a shared real-time dashboard, allowing emergency teams to coordinate relief and restore essential services with incredible precision.

Jobs

• UNITAR is looking for (1) Geospatial Analysts - Satellite Imagery and GIS and (2) Geospatial Quality, Assurance, Training and Analytical Support Specialists based in Geneva.

• WWF is looking for a Senior Conservation Data Analyst/Engineer based in Washington, DC.

• Google is looking for a Technical Program Manager under their Geo team based in Singapore.

• UNODC is looking for a Geospatial Information Management Intern based in Vienna.

Just for Fun

The Sentinel-6 Michael Freilich satellite has detected a massive swell of warm water moving across the Pacific Ocean, which means that El Niño will likely arrive later this year.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge — a place where you aren’t judged for being a little sinusoidal... Anyway, the aim as usual is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• AI Image Search: ProVG improves satellite object detection.

• Farm Boundaries: PRUE maps fields at global scale.

• Car Dependency: New index maps urban transport reliance.

• Carbon Credits: REDD+ projects show major over-crediting.

• Ocean Data: WOD23 compiles centuries of ocean measurements.

• Photosynthesis Data: TRAX GPP maps global productivity.

Subscribe now

Research you should know about

1. Teaching AI to search satellite images like a human

A new study introduces ProVG, an AI framework designed to dramatically improve object detection in satellite imagery based on text descriptions. Spotting a specific target in a remote sensing image can be pretty difficult because the scenes are massive and packed with identical-looking objects. Usually, AI models try to find a target by processing the user’s entire text prompt as a single chunk. It turns out this approach isn’t great. When a model processes a sentence all at once, it doesn’t pick up on fine-grained clues like spatial relationships or specific physical attributes, which often leads to the model highlighting the wrong object entirely.

To fix this, the researchers looked at how human brains actually process visual searches and built a progressive framework that mimics our natural cognition. Instead of swallowing a sentence whole, ProVG breaks the text prompt down into three distinct pieces: (1) global context, (2) spatial relations, and (3) object attributes. The model then applies these clues in a strict sequence to filter the image. First, it surveys the global context to get a general understanding of the scene. Next, it uses spatial clues to narrow down the search to a specific neighbourhood. Finally, it verifies the exact target by matching the specific physical attributes.

Alongside this structured searching method, the framework also includes special modules to handle the extreme scale variations found in satellite photos, helping it fuse details from both zoomed-in and zoomed-out perspectives. The researchers tested ProVG against a huge lineup of existing models across two major remote sensing benchmarks. The results were pretty impressive. By simply decoupling the language and forcing the model to search step-by-step, ProVG consistently outperformed current top-tier systems, proving to be incredibly accurate at drawing both basic bounding boxes and precise pixel-level masks around targets.

2. Mapping the world’s farms from space

Knowing exactly where farm fields begin and end can be pretty important for tracking crop yields, managing water resources, and monitoring climate initiatives. There are too many approaches to name that uses AI to automatically draw these boundaries by scanning satellite images. However, when these models are taken out of the lab and deployed in the real world, there can be a lot of issues… A model that works perfectly in the sunny plains of Illinois might completely mess up when scanning the cloudy, mountainous terraces of Japan, drawing jittery lines or hallucinating extra fields.

To solve this, researchers built PRUE (Practical Recipe for Field Boundary Segmentation at Scale). They tested 18 different AI architectures against a massive global dataset of fields to see what actually works. This involved introducing new, real-world tests to measure how the AI handled changes in brightness, picture scaling, and the order of the images it was fed. They discovered that while some of the largest, most hyped ‘foundation models’ struggled, a specialized U-Net architecture crushed the competition. By feeding this model mathematically enhanced images and tweaking how it penalises mistakes during training, the researchers created an incredibly robust system.

The results speak for themselves. PRUE scored 6% higher in overall accuracy than previous top-tier models and proved remarkably stable when the lighting or image scale was tweaked. To prove it works at scale, the researchers used PRUE to map out every single agricultural field in Japan, Mexico, Rwanda, South Africa, and Switzerland, covering 4.7 million square kilometres in just a few hours.

3. Decoding car dependency in modern cities

To reach net-zero emissions, urban planners are pushing for car-free zones, better cycling infrastructure, and improved public transport. However, moving away from private vehicles is incredibly difficult because decades of car-centric urban growth have transformed the automobile from a convenience into an absolute necessity. For many residents, giving up their car means losing access to essential services, leisure activities, and jobs. A new study addresses this by quantifying exactly how reliant different neighbourhoods are on private vehicles compared to public transport.

A team of researchers developed the Car Dependency Index to map the accessibility gap between driving and taking public transport across 18 major cities in Europe and North America. By crunching high-resolution geospatial data, they calculated how many points of interest a resident could reach within a normal journey time using either a car or a bus and train network. The data revealed a lot of spatial inequalities, with peripheral districts almost entirely trapped in high-dependency cycles while central areas were far better equipped for car-free living. When looking closely at Vienna, the team also discovered that this neighbourhood dependency was a massive driver of car ownership, completely independent of a household’s income level.

To see if infrastructure upgrades could fix the problem, the researchers simulated the impact of a planned metro expansion in Rome. While the new transit line successfully removed an estimated 60,000 commuting vehicles from the roads, the benefits were extremely localised to the people living right next to the new stations. This proves that building a single new train line is simply not enough to break the cycle of car dependency across an entire city. Ultimately, if policymakers want to dismantle car-based systems and build truly sustainable urban spaces, they need to commit to massive, network-wide public transport expansions rather than isolated projects.

4. The truth behind forest carbon over-crediting

Forest carbon credits, particularly REDD+ projects, are meant to fund tropical forest conservation. However, recent controversies revealed that many early projects issued far more credits than they actually earned, causing voluntary carbon markets to lose over a billion dollars in value. Some industry figures pushed back against these claims, arguing that independent evaluations used flawed global satellite data that simply missed the real deforestation. To settle the debate, a new study from Nature Communications synthesised data from six independent evaluations covering 44 REDD+ projects to see exactly where the discrepancies came from.

The results showed that while most projects did successfully reduce deforestation, they still claimed to have saved nearly 11 times more forest than they actually did. Importantly, the researchers proved that the use of global satellite data was not the problem; these tools actually detected more deforestation than the custom layers used by the projects themselves. Instead, the massive over-crediting stemmed from how projects chose their baseline reference areas. Projects routinely selected highly accessible, heavily degraded comparison zones that were far more vulnerable to logging than the actual protected areas. When combined with overly pessimistic predictive models, this made the projects look vastly more effective than they really were.

To ensure carbon markets genuinely benefit the climate, the researchers argue that the industry must abandon flexible, prediction-based baselines. Future credits must rely on rigorous, after-the-fact evaluations using independent methods to measure real-world impact. You can access the data here and the code here.

Geospatial Datasets

1. World ocean database

The World Ocean Database 2023 (WOD23) is a large digital collection of oceanographic in situ profile measurements covering the instrumental record from 1772 to 2022, with around 18.6 million water column profiles and 3.6 billion measurements across 27 physical and chemical variables, including temperature, salinity, and oxygen. You can access the data here.

2. High-res global GPP dataset

The newly introduced TRAX GPP dataset provides a 0.05° global record of terrestrial photosynthesis from 2001 to 2024, using a hybrid Long Short-Term Memory framework with seasonal-trend decomposition and atmospheric CO₂ assimilation constraints to capture carbon cycle dynamics and CO₂ fertilisation effects. You can access the data here and the code here.

3. Long-term soil analysis dataset

The Sol_Run dataset provides a seventeen-year record of soil analysis data for Réunion Island from 2008 to 2024, with more than 266,000 records from nearly 25,000 unique samples covering physical and chemical properties such as pH, nutrient content, and cation exchange capacity. You can access the data here.

4. High-res Mediterranean wildfire dataset

The CalWildFire dataset brings together eleven years of daily in situ wildfire and meteorological data for the Mediterranean region of southern Italy, combining ground-validated meteorological measurements, field-based burned area records, and calculated fire weather indices. You can access the data here.

Other useful bits

• Planet has launched SuperRes technology, a new AI-based super-resolution dataset that sharpens its satellite imagery from 3 metres to 2-metre resolution. The update makes small-scale details, such as rooftop markings and individual trees, easier to inspect while retaining Planet’s daily global monitoring coverage.

• Earth Index is officially live offering free global access to its AI-powered satellite search tools. The platform is already being used to identify issues such as illegal deforestation and hazardous quarries through large-scale satellite image search.

• Render a House has launched a new 3D mode that allows users to place their own 3D models directly within the Google Earth environment. The update makes it easier to see how proposed projects fit into their real-world surroundings and create more consistent project images.

Jobs

• World Vision is looking for a WASH Data Fellow based in the US.

• Google is looking for a Technical Program Manager under their Geo team based in Singapore.

• UNODC is looking for a Geospatial Information Management Intern based in Vienna.

• Esri is looking for a GIS Solution Engineer – Global Alliances & Partners based in San Antonio.

Just for Fun

Saturn and Neptune in retrograde, traced across 34 nights from May 2025 to February 2026. This composite shows Saturn as the brighter foreground planet and Neptune as the dimmer object behind it, both appearing to move backwards across the sky as Earth overtook them on its faster inner orbit around the Sun. Saturn moved from Pisces into Aquarius and back again, while Neptune remained in Pisces, during their closest apparent pairing since the Saturn-Neptune conjunction of 1989.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now

Hey guys, here’s this week’s edition of the Spatial Edge — a safe space for map fetishists… The aim is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Disaster Damage: AI assesses buildings after natural disasters.

• Map Queries: LLMs make spatial databases conversational.

• Data Heat: AI infrastructure warms nearby landscapes.

• Satellite Labels: MAPLE teaches AI land-cover hierarchies.

• Urban Plastics: DURUM tracks microplastics in stormwater.

Subscribe now

Research you should know about

1. A sharper lens on disaster zones

A new study introduces a three-step AI framework designed to improve how we assess building damage from space. When a natural disaster strikes, quickly figuring out which buildings are destroyed is pretty important for sending rescue teams to the right places. However, the satellite imagery used for these assessments is often blurry or full of noise, which makes it pretty difficult for automated systems or even human experts to spot crucial details like collapsed roofs or cracked walls. Traditional object detection models can spot buildings but fail to interpret the severity of the damage, while newer Vision-Language Models (VLMs) can analyse scenes but often get confused and ‘hallucinate’ when looking at low-resolution photos.

To overcome these limitations, researchers from Dakota State University built a hybrid pipeline and tested it against the xBD dataset, a massive benchmark collection of satellite images covering various global disasters. The new framework tackles the problem in three steps. First, it uses a Video Restoration Transformer to artificially upscale the low-resolution satellite images, quadrupling their size to make structural details stand out. Next, it uses a YOLOv11 object detection model to draw precise boxes around every building in the image and crops them out. Finally, these high-resolution cropped images are fed into several different VLMs, which act as a ‘jury’ to assess the damage severity of each building on a four-point scale. By combining multiple models, the system drastically reduces the bias and hallucination risks of relying on just one AI.

When evaluated on imagery of Hurricane Matthew and the Moore Tornado from the xBD collection, the results showed that this multi-stage approach was pretty decent. By upscaling the images first, the VLMs were able to produce much more accurate and semantically rich descriptions of the damage. Furthermore, the team found that the Qwen3 family of models consistently outperformed Gemma3 models, proving to be far more reliable at identifying completely destroyed buildings. Ultimately, this pipeline could provide first responders with a faster way to prioritise rescue efforts without having to set foot in a hazardous area. However, I do remain sceptical about the super-resolution approach…

2. When maps learn to answer questions

Querying complex geographic or time-based databases usually requires some coding skills. If you want to find all the restaurants within a specific flood zone, you can’t just type that into a search bar and expect a database to understand you. Developers build Natural Language Interfaces to Databases so everyday users can bypass the complex code and just ask questions. A new paper maps out the entire landscape of this highly fragmented field, reviewing everything from the datasets used to the evolution of the algorithms themselves.

The paper breaks down how these systems have evolved across four main categories. Early models relied on strict, handcrafted grammar rules that worked well but struggled with casual or unexpected phrasing. Programmers then moved to hybrid neural networks that learned from data but still relied on rigid templates. Today, the focus has shifted entirely to LLMs and multi-agent systems, where different AI agents collaborate to interpret a query, check the database schema, and write the code. While these modern approaches can be incredibly flexible, they often suffer from geo-hallucination, where the AI invents imaginary spatial relationships or references tables that simply do not exist.

Beyond the models, the review highlights a pretty big issue with the data used to test them. Right now, there isn’t a single, unified benchmark that accurately reflects the messy and ambiguous way real people ask geographic questions. Instead, researchers rely on a mix of highly controlled synthetic datasets and real-world GIS databases that make it nearly impossible to compare different models fairly. Ultimately, the authors conclude that before we can build AI that truly understands space and time, the industry needs to ditch these fragmented tests and create a robust benchmark that accounts for real-world geographic ambiguity.

3. AI’s next environmental problem is heat

We hear a lot about the massive energy demands of artificial intelligence, but there’s a hidden side effect we rarely talk about: the heat. As tech companies scramble to build massive AI data centres across the globe, these power-hungry hubs are pumping enormous amounts of thermal energy into the atmosphere. To figure out the real-world impact, researchers combined two decades of satellite land surface temperature records with the exact locations of over 8,000 data centres built outside densely populated cities. This allowed them to isolate the specific temperature changes caused by these facilities without getting confused by typical urban warmth.

The data reveals a pretty crazy phenomenon that they call the ‘data heat island effect’. On average, the land surface temperature around a new AI hyperscaler spikes by 2°C right after it starts operating. This warming footprint isn’t just confined to the building itself. The heat ripples outward, raising local temperatures for up to 10 kilometres in every direction. When researchers overlaid this temperature data with global population maps, they discovered that over 340 million people currently live within these artificially warmed zones. It’s a pretty big deal, considering this extra heat places significant stress on local healthcare, energy grids, and water supplies.

Tackling this growing problem requires a mix of smart software and innovative hardware. On the coding side, developers can reduce compute loads by cleaning up messy data before processing or using stuff like carbon-aware token generation to improve efficiency. Physically, engineers are exploring some incredibly interesting tech to keep temperatures down. This includes adiabatic circuitry that actually recycles energy during computation, hybrid liquid cooling systems, and special passive radiative coatings that reflect solar heat away from the buildings. By combining these hardware and software upgrades, the industry has a real shot at keeping our environment cool(er) as AI continues to scale.

4. Teaching AI the family tree of satellite imagery

When AI models look at satellite imagery, it’s usually trained to spot flat, isolated labels like a ship, a tree, or a road. But the real world is pretty highly structured, and these labels actually belong to a complex hierarchy. E.g. a ship is part of a port area, which itself is a type of artificial surface. While some models try to learn these hierarchical relationships, they can get confused when a single image contains multiple objects from completely different branches of the taxonomic tree. Furthermore, these older methods are pretty computationally heavy and tend to fall apart when they do not have massive amounts of training data.

A recent paper introduces MAPLE, a new framework designed to solve this multi-path classification problem by explicitly teaching AI the hierarchical structure of remote sensing data. The system starts by feeding the model text-based descriptions of how different land cover types are related, creating a semantic map of the categories. It then uses a Vision Transformer to scan the satellite image and a Graph Convolutional Network to pass information along the branches of the label tree. Finally, it uses an adaptive fusion mechanism to dynamically weigh the visual evidence from the image against the structural rules, ensuring it makes logical predictions across every level of the hierarchy without any manual tuning.

The researchers tested MAPLE across several major remote sensing datasets that were mapped to standard land-cover classifications. The results were pretty impressive, particularly when the model was given very little data to learn from. In these restricted scenarios, MAPLE outperformed standard classification baselines by up to 42%, proving that understanding the relationships between categories gives models a massive head start. Interestingly, it achieved these gains while adding only a tiny 2.6% increase in computational overhead. The framework also proved quite effective at reducing dumb mistakes between tricky categories, making it a robust and highly efficient tool for Earth observation.

Geospatial Datasets

1. Wind tunnel gas dispersion dataset

The RED:VAPOR dataset captures gas sensor responses to vapour plumes in scale-model landscapes, using a wind tunnel fitted with both low-cost and advanced sensors to map synthetic gas dispersion across a dense 3D grid under controlled wind conditions. You can access the data here and the code here.

2. Urban runoff microplastics dataset

The Dataset of Urban RUnoff Microplastics (DURUM) brings together 180 urban stormwater microplastic sampling efforts across 15 countries from 2018 to 2024, covering runoff from streets, residential catchments, and industrial areas. You can access the data here and the code here. You can access the data here and the code here.

3. Multi-sensor EO dataset

The BigEarthNet.txt dataset provides more than 464,000 co-registered Sentinel-1 and Sentinel-2 satellite images paired with detailed textual descriptions, with nearly 9.6 million annotations covering land-use classes, spatial relationships, and visual question-answering pairs. You can access the data here.

4. Global soil water isotope dataset

This new dataset provides a harmonised global collection of soil-water stable isotope records from 1975 to 2024, bringing together more than 27,000 records from 463 observation sites across 37 countries and organising them into standard soil-depth intervals. You can access the data here.

Other useful bits

• While AI is often discussed in terms of its environmental footprint, DINOv3 is also being used for nature restoration finance. The model uses computer vision to count individual trees and measure canopy height from satellite imagery, helping investors verify local restoration projects at lower cost and direct funding towards farmers and communities.

• For the first time in history, agricultural field boundaries have been mapped at a global scale. The 10-metre resolution dataset is publicly available at no cost and can be explored through the FTW Explorer.

• Google Earth has just released an update that allows users to import Shapefiles and 3D models directly onto its real-world canvas. This improvement enables us to render industry-standard data as cloud-native layers, and they’ve even added elevation profiles to the measure tool to help you easily visualise detailed terrain data.

Jobs

• Yale University is looking for a GIS Research Specialist based in Connecticut.

• NTT Data is looking for a Esri/ArcGIS Specialist based in Arlington.

• BP is looking for a Senior GIS Specialist based in Denver.

• TomTom is looking for a Staff UX Designer under their Map Visualisation team based in Amsterdam.

• Esri is looking for a remote GIS Project Manager under their Assessment Department based in Canada.

Just for Fun

A research team aboard the Polarstern icebreaker has discovered a small, uncharted island in the Weddell Sea after sailing into a mysterious “danger zone” that was previously hidden from all nautical maps.

That’s it for this week.

I’m always keen to hear from you, so please let me know if you have:

• new geospatial datasets

• newly published papers

• geospatial job opportunities

and I’ll do my best to showcase them here.

Yohan

Subscribe now