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

Hey guys, here’s this week’s edition of the Spatial Edge — your weekly map to the latest geospatial news. This newsletter is a bit more like RGB than SAR - we’re impacted by the weather, but we’re tried and true. The aim is to help you become a better geospatial data scientist in less than 5 minutes a week.

In today’s newsletter:

• AlphaEarth Insights: Embeddings reveal hidden land-cover structure.

• Disaster AI: Domain adaptation improves damage detection.

• Urban Planning: Deep learning unifies fragmented city data.

• 3D Remote Sensing: AI models finally learn elevation.

• Crop Adaptation: Climate-responsive planting strategies simulated globally.

Subscribe now

Research you should know about

1. Cracking open Google’s geospatial black box

I’ve previously covered Google’s AlphaEarth a few times on here. In short, it’s one of the new geospatial foundation models trying to turn satellite and Earth observation data into reusable “embeddings”. In simple terms, these embeddings are compressed numerical descriptions of places on Earth. The tricky part is that these numbers are usually hard to interpret. A new paper tries to open that black box by asking what the individual embedding dimensions seem to capture when applied to land cover classification.

The main finding is that the embedding space appears to have a kind of internal geography. Some dimensions act like specialists, strongly linked to one land cover type, such as permanent water, built-up areas, grassland, bare soil, or moss and lichen. These make intuitive sense because certain land covers have very distinctive spectral or structural signatures. Water, for example, has very low reflectance in near-infrared and shortwave infrared bands, while bare soil and sparse vegetation have a very different year-round pattern.

The more interesting part is that other dimensions are shared across multiple land cover types. These seem to capture messy transition zones: cropland and grassland, shrubland and tree cover, bare soil and wetlands, or snow and sparse vegetation. That’s useful because real landscapes rarely fit neatly into clean categories. The paper’s bigger point is that geospatial foundation models may be learning both the “core” of land cover classes and the fuzzy spaces between them. If that holds up across more tasks and regions, it could make these models easier to prune, explain, and use for things like change detection, land degradation monitoring, and environmental planning.

Another pretty interesting finding is that the embeddings do a lot of unnecessary heavy lifting. Because the researchers mapped out exactly what each dimension does, they found that you simply do not need to use all 64 of them. Depending on the specific terrain you are looking at, you can hit 98% accuracy using just two to twelve targeted dimensions. This means developers can prune away the dead weight and slash their computing and storage costs by up to 90%, making it far cheaper to run high-resolution global models without losing any meaningful detail.

2. Building AI that works across different disaster zones

When a major disaster strikes, emergency responders increasingly rely on AI to rapidly scan satellite imagery and assess building damage. This allows human operators to stop squinting at fuzzy photos and focus on deploying rescue teams. However, there is a fundamental flaw in how these systems are currently built. If an AI model is trained on satellite photos of an earthquake in Turkey, it could completely fail to detect damaged buildings from a hurricane in Louisiana. This phenomenon, known as ‘domain shift’, happens because the AI gets confused by different camera angles, weather conditions, or regional architecture, rendering the tool useless for new, unseen disasters.

To solve this, researchers experimented with a technique called Supervised Domain Adaptation (SDA). They took a top-tier AI model originally trained on a massive global dataset of various disasters and fine-tuned it using a small batch of specific data from Hurricane Ida. Rather than just relying on standard RGB satellite images, they also fed the AI a variety of mathematically enhanced pictures. They tested ‘unsharp masking’ to highlight tiny cracks and splintered roofs, ‘contrast enhancement’ to spot subtle water damage, and ‘edge detection’ to make the outlines of buildings stand out.

The researchers found that Domain Adaptation was absolutely essential; without it, the AI failed to identify severely destroyed buildings entirely. When paired with ‘unsharp masking’, the model achieved its absolute best results, boosting its ability to detect ‘major damage’ by 190% and ‘destroyed’ buildings by an incredible 341% compared to baseline models. Interestingly, the researchers found that combining all the image enhancements at once actually confused the AI, proving that when it comes to training robust disaster-response tools, less is often more.

3. A deep learning framework for holistic urban planning

As I’m sure you all know, OpenStreetMap is one of the most useful free geospatial datasets out there. And I’m also sure you know the main issue with it: quality varies a lot by place. Some cities have really detailed road networks, building footprints, land-use polygons and points of interest. Others are patchier, especially in fast-growing or less well-mapped areas. A new paper in Scientific Reports looks at whether deep learning can help turn OSM into something more useful for urban planning. The case study is Krasnodar, a rapidly growing city in southern Russia, where the authors combine OSM with Sentinel-2 imagery, demographic data, traffic sensors and air-quality measurements.

The main idea is to build a single workflow that handles several planning tasks at once. The authors use different models for different jobs: a CNN for land-use classification, U-Net for building footprint extraction, LSTM networks for traffic forecasting, and a hybrid CNN-RNN model for short-term air-quality prediction. The results are pretty strong. The land-use model reached 91.6% accuracy, the building footprint model reached 94.0% accuracy, the traffic model had an RMSE of 3.6 vehicles per hour, and the air-quality model had an RMSE of 2.3 µg/m³ for PM2.5 forecasting. The framework also showed that OSM performed best for roads and buildings, while points of interest were less reliable, likely because they depend more heavily on voluntary tagging.

The useful part of this paper is the shift away from treating urban datasets as separate layers. Building footprints help make sense of urban structure, land-use classification helps identify functional zones, traffic prediction captures mobility pressure, and air-quality modelling adds an environmental layer. Together, these outputs can support things like zoning updates, green-space planning, congestion management and pollution monitoring. There are still some clear limits: the study is based on one city, OSM quality is uneven, and deep learning models can be expensive for smaller planning departments to run. But the broader direction is interesting: free geospatial data, satellite imagery and relatively standard deep learning models can now be stitched together into a pretty practical urban intelligence system.

4. Teaching AI to see the third dimension in satellite imagery

When it comes to analysing satellite imagery, most modern Vision-Language Models (VLMs) have a massive blind spot: they can only see in 2D. While these models are fantastic at looking at a flat image and identifying forests, roads, and buildings, they have absolutely no concept of height or terrain. This is a pretty big problem if you are trying to use AI for things like predicting where a flood will hit or assessing landslide risks, where the physical slope and elevation of the land are far more important than what it looks like from above.

To fix this vertical blind spot, researchers built GeoHeight-Bench, a new testing framework specifically designed to see if AI can understand height. Because getting experts to manually annotate the elevation of thousands of buildings and hills is incredibly expensive, the team used an automated AI pipeline to generate the data instead. The resulting benchmark forces models to tackle a range of 3D problems, from simply retrieving the exact elevation of a specific pixel, to identifying the tallest building in a city, to predicting which areas of a map are most likely to flood based on the terrain.

To prove that AI can actually learn to see in 3D, the researchers also built their own model called GeoHeightChat. Instead of just feeding the AI raw optical images, they used a two-stage training process to teach it how to implicitly guess the geometric structure of a landscape just by looking at a standard 2D photo. When tested on the new benchmark, GeoHeightChat completely blew the competition out of the water, scoring 44.14% overall compared to GPT-4o’s 21.61%. This proves that by injecting a sense of vertical scale into these models, we can unlock a whole new level of practical, real-world problem-solving for remote sensing AI.

You can check out the project page here.

Geospatial Datasets

1. High-altitude wetlands map

This newly developed dataset presents the very first global, high-resolution (30 m x 30 m) map of high-altitude wetlands across the world’s major mountain regions. The researchers combined Sentinel imagery, elevation data, and global ecoregion maps with machine learning to identify more than 30,500 km² of wetland ecosystems. The dataset is useful because high-altitude wetlands remain poorly mapped despite their importance for water storage, runoff regulation, and ecosystem functioning, particularly across remote mountain regions such as the Himalayas and Andes. You can access the data here and the code here.

2. Watershed hydrometeorological dataset

The MACH dataset compiles hydrometeorological data across 1,014 US watersheds, combining daily meteorological forcings with streamflow observations over a 44-year period, with some catchments extending back as far as 75 years. It also includes watershed-level attributes such as geology, topography, land cover, and climate characteristics. You can access the data here and the code here.

3. Climate-responsive crop selection dataset

This dataset explores how changing crop varieties and planting schedules could help agriculture adapt to climate change by simulating crop yields and water use for maize, wheat, soybean, and rice between 1961 and 2100. One of the more interesting aspects is its focus on adaptation timing, showing how relatively small shifts in sowing dates can substantially affect water demand and crop performance under warming conditions. You can access the data here and the code here.

4. European Atlantic wave climate dataset

A new dataset provides high-resolution projections of future wave conditions along the European Atlantic coastline using regionally downscaled CMIP6 climate simulations. The researchers generated 3-hourly directional wave spectra for more than 1,000 offshore locations across both historical (1985 to 2014) and future (2030 to 2059) periods under two emissions scenarios, including both integrated wave parameters and full spectral energy information. You can access the data here and the code here.

Other useful bits

• An important new global assessment has applied the Red List of Ecosystems framework to mangroves for the first time, providing a detailed look at the pressures facing coastal ecosystems worldwide. The report finds that roughly half of global mangrove ecosystems are now considered at risk due to factors such as coastal development, sea-level rise, pollution, and changing sediment dynamics. Beyond the conservation implications, the assessment also highlights the growing role of geospatial monitoring in tracking ecosystem degradation and coastal resilience over time.

• Vantor is partnering with Google Earth AI to integrate more advanced geospatial reasoning capabilities into its Tensorglobe platform. The aim is to move beyond basic object detection towards understanding more complex spatial behaviour and relationships within imagery and geospatial data. One of the more notable aspects is that these capabilities are being designed for classified and air-gapped environments, where many large AI systems are still difficult to deploy securely.

• NOAA is using AI and cloud infrastructure to modernise weather forecasting workflows by migrating large volumes of historical climate and meteorological data into cloud-based systems. The broader goal is to speed up forecasting pipelines for extreme weather events and improve model efficiency, particularly as data volumes continue to grow. At the same time, NOAA officials have stressed that AI tools are intended to support rather than replace scientific expertise, especially for high-stakes forecasting decisions.

• Google Earth is expanding its professional geospatial toolkit by adding global data layers and native Earth Engine raster support directly into its no-code interface. The update also includes an AI-powered feature called “Ask Google Earth”, which allows users to search for and visualise datasets using natural language prompts. The broader direction is fairly clear: reducing the technical barriers required to work with large-scale geospatial datasets and remote sensing products.

Jobs

• 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.

• Stantec is looking for a GIS Analyst under their Energy and Resources budiness unit based in Toronto.

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

Just for Fun

USGS has brilliantly released a seamless, globally consistent geologic map of the entire Moon at an incredible 1:5,000,000 scale.

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 slightly cheaper than your ESRI subscription. 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:

• Nightlights Harmonisation: Deep learning links DMSP and VIIRS.

• Map Navigation: AI struggles with route planning.

• Earth World Models: RS-WorldModel predicts future satellite scenes.

• Satellite Image Editing: RSEdit reshapes remote sensing imagery.

• Google Urban Embeddings: S2Vec summarises the built environment.

Subscribe now

Research you should know about

1. Harmonising nighttime light data across decades

Anyone who’s worked with nightlights data knows the long-running headache: DMSP-OLS data goes back to 1992 but suffers from saturation in city centres, blooming, and no on-board calibration. VIIRS, which kicked in from 2012, fixes most of these issues with finer resolution and a much wider dynamic range. The problem is, if you want a long time series stretching back before 2012, you’re stuck stitching together two sensors that don’t really speak the same language. A new paper addresses this with a deep learning approach that translates DMSP imagery into something that looks like it came out of VIIRS.

The authors use a CUT-based generator (a type of image-to-image translation model) trained on overlapping DMSP and VIIRS data. The idea is to teach the model what a VIIRS scene “should” look like given a DMSP input, so you can then run it on pre-2012 DMSP data and get a harmonised series. On their held-out test set of 1,862 patches, they get an R² of 0.87 on radiance, an RMSE of 0.114, and an SSIM (structural similarity) of 0.79. Visually, around 75% of cases look pretty clean, with the model successfully recovering internal structure in cities that DMSP had flattened into uniform blobs.

It’s not flawless though. About 8% of patches throw up artefacts, and the failure modes are pretty instructive. In megacity cores where DMSP is fully saturated at DN=63, the model can’t hallucinate plausible internal structure and produces checkerboard patterns from the upsampling kernel. Around isolated bright points like ports and gas flares, you get halo effects where the model over-generates radial glow. So if you’re using harmonised data for analysis of dense urban cores or industrial sites, you’ll want to flag those areas. For everything else, this looks like a useful step towards a properly continuous nightlights record from 1992 to today.

2. Testing if AI can navigate a map

Multimodal large language models (MLLMs) are becoming pretty smart at analysing satellite images. If you show them a photo of a landscape, they can usually identify the buildings, spot the rivers, and classify the terrain. But there’s a massive difference between simply recognising objects and actually planning a route through them. If a rescue team needs to know the safest path for an off-road vehicle across a flooded disaster zone, simply identifying the water isn’t enough; the AI needs to understand the vehicle’s capabilities, map out the terrain, and plot a strict, step-by-step route. Until now, there hasn’t been a reliable way to test if AI models can actually do this kind of complex, constraint-based planning.

A new study introduces NeSy-Route, a massive new testing framework designed to evaluate how well AI models can navigate remote sensing imagery. Instead of just asking the AI to pick the best route from a multiple-choice list, NeSy-Route forces the model to complete three increasingly difficult tasks. First, the AI has to read a text prompt (like “find the fastest route for a pedestrian avoiding water”) and translate those rules into strict mathematical constraints. Next, it has to look at the satellite image and correctly identify which parts of the terrain match those constraints. Finally, it has to generate an actual, coordinate-by-coordinate path from start to finish.

The researchers tested 10,821 different scenarios across a variety of top-tier AI models, and the results were pretty revealing. While most advanced models were pretty good at understanding the text constraints, they completely fell apart when asked to actually plot the route on the map. Even the smartest proprietary models generated highly inefficient paths, while open-source models often plotted routes that drove straight through restricted obstacles. Ultimately, this benchmark proves that while AI is great at describing satellite images, it still lacks the deep, global reasoning required to reliably navigate the real world. I don’t think this is surprising though, as LLMs are not trained/fine-tuned for the task (yet…).

3. A unified world model for Earth observation

If 2025 was the year of ‘foundation models’, 2026 is the year of ‘world models’. Basically, the idea is that a World Model understands complex environments and predict future states for things like autonomous driving and robotics. However, applying this concept to Earth observation has been pretty difficult. Typically, remote sensing algorithms are split into two completely separate camps. Models either understand and describe changes between satellite images, or they synthesise new landscapes from scratch. Because these tasks are kept isolated, the algorithms miss out on shared geographic knowledge, leaving generative models hard to control and analytical models unable to grasp future physical constraints.

RS-WorldModel is a new system that bridges this gap by tackling both spatiotemporal change analysis and future scene forecasting simultaneously in three stages. These stages are worth understanding because each does something distinct. First, Geo-Aware Generative Pre-training teaches the model to predict future scenes conditioned only on geospatial metadata, with no text involved. Second, Synergistic Instruction Tuning jointly trains the model on both change description and text-guided forecasting, so the two tasks reinforce each other. Third, Verifiable Reinforcement Optimization uses GRPO with task-specific rewards: cosine similarity for forecasting, and an LLM judge for change descriptions that can detect physically implausible outputs (e.g. impossible illumination changes given the sun angle). The ablations show each stage matters. Stripping out the geographic metadata alone bumps FID from 50.28 to 53.72, and removing the pre-training stage entirely sends FID up to 73.55.

The numbers are pretty compelling. On change description, RS-WorldModel beats Qwen3-VL-32B on most metrics despite being 16x smaller, with ROUGE-L jumping 21% over the same-size base model. On forecasting, it cuts FID by 48% relative to CRS-Diff and even beats Gemini-2.5-Flash Image (43.13 vs 46.14). The qualitative examples are particularly nice: when prompted with a winter scene description, the model preserves the underlying settlement layout while adding realistic snow cover and long shadows from a low sun angle, whereas the baselines either miss the structure or generate geographically implausible textures. For anyone working on land-use change, climate adaptation, or any application where you’d want to simulate future satellite imagery under different scenarios, this looks like a pretty interesting tool.

4. RSEdit: giving AI the power to reshape satellite imagery

While text-to-image AI models can easily generate highly realistic pictures of cats or fantasy landscapes, they struggle when asked to edit remote sensing satellite imagery. If you prompt a standard AI editor to show what a village would look like after a flood, it will often hallucinate buildings that don’t exist, distort the strict top-down perspective, or break the laws of physics. This happens because general-purpose AI models lack a deep understanding of geography and are built using conditioning methods that don’t align with the rigid spatial rules of Earth observation data.

To fix this, researchers built RSEdit, a new framework specifically designed to edit remote sensing imagery. Instead of forcing satellite data to fit a generic AI model, RSEdit adapts the underlying architecture of top-tier diffusion models (like Stable Diffusion and PixArt-α) to respect the strict spatial layout of satellite images. For older models, it tightly binds the original image to the edit instructions pixel-by-pixel, and for newer models, it treats the edit as a continuous stream of data tokens. The researchers trained the system on over 60,000 before-and-after satellite images detailing everything from urban expansion to natural disasters.

The results are pretty impressive. When asked to simulate complex scenarios like a tsunami hitting a coastline or a wildfire damaging specific structures, RSEdit generated physically plausible, highly accurate edits without hallucinating extra details or warping the surrounding roads and vegetation. In direct comparisons using specialised building-damage assessment tools, RSEdit outperformed leading general-domain editors like InstructPix2Pix by a massive margin. Ultimately, this framework provides a powerful, on-demand way to simulate disaster impacts and urban growth, creating a valuable new source of training data for environmental analysis.

Geospatial Datasets

1. High-res in-season crop mapping dataset

The newly developed In-season Crop-type Data Layer (ICDL) provides 10-metre resolution crop planting maps across the contiguous United States. You can access the data here and the code here.

2. Ancient China population dataset

A pretty original new dataset reconstructs county-level population distributions in ancient China for the pivotal years of 2, 742, 1102, and 1820 AD. It estimates population figures for over 6,000 counties across four representative dynasties. You can access the data here.

3. Off-shore marine cadastre dataset

The US Bureau of Ocean Energy Management has recently released a new Offshore Marine Cadastre Data Collection, essentially consolidating various regionally focused datasets into one streamlined national resource. You can access the data here or here. You can also check out the navigator here.

4. Comprehensive crop yield dataset

The CYCleSS dataset offers a large-scale resource for predicting crop yields across England, blending precision agriculture with environmental data. It features anonymised yield data from 934 fields alongside 10-km resolution satellite, daily weather, and soil type information from 2015 to 2017. You access the data here and the code here.

Other useful bits

• Google Research introduced S2Vec, a new AI framework that learns to read the built environment of our cities just like a human would. By transforming complex geospatial data into simple mathematical summaries, it provides a way of predicting important socioeconomic and environmental patterns across the globe.

• The WMO has published its latest State of the Global Climate report.

  It brings together evidence on ocean heat, sea ice, temperatures, greenhouse gases, and other indicators of a changing climate. It is also a useful reminder of why long-term Earth observation matters for climate monitoring, risk management, and public policy.

• The World Bank has launched the Land Data Map, an interactive platform for exploring land-related data. The tool makes it easier to see how land systems connect with development outcomes such as economic growth, jobs, and investment. It is a useful resource for anyone working on land governance, urban development, or spatial planning.

Jobs

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

• UN DESA is looking for two interns in (1) Geodesy & Geospatial Research Support and (2) Digital Design, Data Visualization & Geospatial Communications.

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

• Stantec is looking for a GIS Analyst under their Energy and Resources budiness unit based in Toronto.

Just for Fun

NASA has released some images of the SWOT satellite to map the Earth’s ocean floor.

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. 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:

• Urban Physics: PDEs model city sprawl dynamics

• Map Training: OpenStreetMap improves satellite AI captions

• EO Embeddings: Blueprint for stronger geospatial representations

• City Graphs: New Python library for urban networks

• Aerosol Forecasts: AI delivers rapid global predictions

Subscribe now

Research you should know about

1. Why do cities sprawl? The physics of urban growth

Between 1985 and 2015, the total amount of land covered by cities doubled, eating into green spaces, worsening air quality and making infrastructure incredibly expensive to run. Despite this massive transformation, actually explaining how and why cities expand the way they do is pretty difficult. Traditional urban economics often uses models that treat cities as if they are in a state of perfect balance, assuming that land is instantly redeveloped to its optimal use like a neat, circular dartboard. The reality is that cities grow in messy, uneven and unpredictable ways, driven by everything from traffic congestion to developer whims.

A new study argues that we should look at cities not through the lens of traditional economics, but through the physics of ‘surface growth’. In the same way that physicists use Partial Differential Equations (PDEs) to model the jagged edges of a growing bacterial colony or the spread of a tumour, these same mathematical tools can be used to model the irregular, sprawling boundaries of a city. Unlike older economic models, PDEs can handle the chaotic, real-world feedbacks of urban growth, such as how the construction of a new railway line suddenly makes a distant suburb boom, which in turn demands even more roads.

Recent studies have shown that this isn’t just an abstract theory. When researchers applied these physics-based PDE models to historical data from London and Sydney, they successfully replicated the real-world shift from dense, packed city cores to sprawling commuter belts over the last 150 years. The models naturally captured the ‘angiogenic’ growth of the cities, where population density and transport networks feed off each other in a constant cycle. Ultimately, the researchers argue that by merging complexity science with satellite data, planners can finally get a mathematical grip on urban sprawl, allowing them to test the long-term impact of new infrastructure before pouring any concrete.

2. Using maps to teach AI about satellite imagery

When it comes to processing satellite imagery, vision-language models (VLMs) have historically relied on a technique called ‘pseudo-labelling’ to learn about the Earth. Because high-quality, human-annotated satellite data is incredibly rare and expensive, developers usually rely on massive, general-purpose frontier models like GPT-5 or Gemini to act as a ‘teacher’. These teacher models look at satellite images and generate captions, which are then used to train a smaller ‘student’ model. However, this approach is both expensive and fundamentally flawed. The student model can never be smarter than its teacher, meaning it inherits all of the teacher’s mistakes and hallucinations.

Researchers have introduced OSMDA (OpenStreetMap-based Domain Adaptation) to break this reliance on expensive external teachers. The core idea is pretty simple: instead of asking a massive AI to guess what is in a satellite photo, the researchers pair the satellite image with a rendered map tile from OpenStreetMap. Because modern VLMs are already excellent at reading charts and OCR, the model can simply read the map to identify the roads, buildings and land-use polygons, and then cross-reference this with the satellite image to write a highly accurate, geographically grounded caption. This allows a capable base model to act as its own annotation engine.

Using this self-contained pipeline, the team generated OSMDA-Captions, a massive dataset of over 200,000 highly detailed image-caption pairs, entirely for free. When they mixed this new dataset with existing remote sensing data to train their OSMDA-VLM model, it achieved state-of-the-art results across the majority of ten rigorous benchmarks. It completely outperformed models that were trained using expensive frontier teachers, proving that leveraging crowdsourced geographic data is a far more scalable and accurate way to adapt AI to the remote sensing domain.

3. The blueprint for EO embeddings

Earth observation satellites generate massive amounts of data every day. To make analyzing this imagery more efficient, developers use Geospatial Foundation Models to compress raw satellite images into compact mathematical representations called embeddings. This is something I’ve covered previously here. By reusing these embeddings, pipelines can run much faster without constantly reprocessing the raw images. However, the best way to extract and design these embeddings (e.g. choosing the right model architecture, the best layers, or the optimal pooling method) has remained unclear, which can lead to suboptimal performance in real-world applications.

A recent study introduces a systematic blueprint for designing these embeddings by rigorously evaluating various GeoFM configurations using the NeuCo-Bench framework. The research team tested how different architectural decisions, pretraining strategies, and data aggregation methods impact the AI’s ability to track diverse environmental factors, from crop coverage to biomass and urban heat islands.

The findings offer some useful guidelines for developers. The study shows that Vision Transformers consistently produce stronger, more versatile embeddings than Convolutional Neural Networks (CNNs) for these tasks. When compressing the spatial data, simple “mean pooling” (basically averaging out the features) proved to be the best strategy across the board (simple is best as they say…). Interestingly, the researchers discovered that pulling data from the final layer of a CNN often degrades accuracy for tracking continuous physical variables, meaning developers should extract data from intermediate layers instead. Finally, the study highlights that stitching together embeddings from models trained with different self-supervised learning objectives creates highly robust representations, effectively compressing satellite data by a factor of 500 while maintaining exceptional predictive power.

Geospatial Datasets

1. Cactus ecological and evolutionary dataset

CactEcoDB brings together spatial, environmental, trait, and phylogenetic data for over 1,000 cactus species. It combines occurrence records with a time-calibrated phylogeny, making it useful for studying biodiversity patterns, species evolution, and conservation risks. You can access the data here and the code here.

2. Pore-scale CO2-water interaction dataset

This new dataset captures how CO2 and water interact within porous media at the pore scale, with relevance for geoscience and carbon capture and storage. It includes 624 high-resolution 2D samples across 100 time steps under a constant injection rate, generated using numerical simulations. You can access the files here and the code here.

3. High-precision landslide event dataset

This dataset maps 1,582 landslide events across mainland China between 2008 and 2024. It uses LLMs to extract information from local news reports, capturing detailed timing and location data, often at village level. You can access the data here and the code here.

4. Forest disturbance deep learning dataset

The Deep4Dist dataset provides around 17,500 high-resolution image patches (0.2 m) from Rhineland-Palatinate, Germany. It combines RGB, near-infrared, and height data, with segmentation labels for disturbance types such as bark beetle damage, clear cuts, and windthrow. Access the data here and the code here.

Other useful bits

• A new Python library called City2Graph has just launched, allowing developers to transform complex urban datasets into seamless graph representations. It’s a pretty useful resource for anyone looking to build advanced spatial network analyses or power innovative artificial intelligence applications using city mobility and infrastructure data.

• The Planet and NVIDIA are teaming up to build a new AI-powered engine that processes planetary satellite imagery in seconds rather than hours. The system uses high-performance computing to generate higher-resolution outputs and enables a searchable, near real-time view of the Earth. This improves access to up-to-date geospatial data for monitoring and analysis.

• SpaceX has exceeded 10,000 active Starlink satellites in low Earth orbit following a double launch. This marks a significant expansion of its global broadband network, increasing coverage and capacity across regions.

• Researchers in China have introduced an AI-based system for global aerosol and weather forecasting. Trained on over 40 years of data, it produces detailed five-day forecasts in under a minute. The model supports faster environmental monitoring and improved understanding of atmospheric conditions.

Jobs

• UN DESA is looking for two interns in (1) Geodesy & Geospatial Research Support and (2) Digital Design, Data Visualization & Geospatial Communications.

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

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

• Stantec is looking for a GIS Analyst under their Energy and Resources budiness unit based in Toronto.

Just for Fun

In celebration of Earth Day, NASA has released a cool interactive tool that spells out any name using natural geographic features captured by Landsat satellites.

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 called your high school cover band Rasterheads, we have one thing to say to you. Welcome home… The aim of this newsletter is, as always, to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Map Compression: Optimising tiles without losing spatial detail

• Embedding Methods: Moving beyond averages improves model accuracy

• LLM Reliability: Reducing hallucinations in satellite image analysis

• 5G Prediction: Fixing spatial bias in network demand models

• Soil Dataset: High-resolution global soil carbon mapping

Subscribe now

Research you should know about

1. Shrinking massive map tiles without losing the details

If you’re reading this newsletter, chances are you love a good interactive map. And let’s face it, who doesn’t? Well, making them fast and customisable is actually pretty difficult. Right now, developers are forced to choose between two flawed options. Server-side rendering generates static image tiles that load quickly, but they lock users into a single visual style with no room for client-side tweaking. Conversely, client-side rendering sends raw vector data to the browser for fully flexible styling, but these vector files don’t have size limits. As datasets grow, vector tiles become bloated with overlapping geometries and endless attribute tables, which chokes network bandwidth and browser memory. Existing compression tools try to fix this by blindly dropping data, but this typically warps geographical boundaries and deletes the exact attributes users want to visualise. Not great…

A new study introduces HiFIVE, a framework that treats map tile reduction as a strict mathematical optimisation problem. Instead of indiscriminately deleting polygons, the system uses a two-stage approach to shrink tiles to a strict size limit while preserving their visual and semantic fidelity. First, a lightweight triage step makes rough cuts by discarding visually insignificant records. Next, a highly precise ‘sparsification’ algorithm steps in. This algorithm weighs the visual prominence of a shape (such as its pixel footprint) against the usefulness of its underlying data. It selectively nullifies specific attribute values that are unlikely to be used for styling, like random unique identifiers, while preserving categorical data like state names or road types.

When tested on massive terabyte-scale datasets, HiFIVE processed almost a billion records in under an hour. The resulting maps proved to be vastly superior to those produced by standard reduction tools. By intelligently choosing exactly which data points to keep and which to discard, the framework maintained crisp, undistorted boundaries and retained the necessary attributes for complex client-side styling. Ultimately, this approach provides a highly scalable way to explore massive spatial datasets directly in the browser without sacrificing rendering speed or visual accuracy.

You can access the code here.

2. Why averaging satellite data is a bad idea

Geospatial foundation models are brilliant at turning satellite images into highly detailed pixel-level data. However, many practical applications require labelling entire regions, such as a whole farm or a residential neighbourhood. To make this work, developers typically compress thousands of pixel vectors into a single representation using a technique called mean pooling, which is essentially just averaging everything out. The major flaw with this default approach is that it completely erases all the nuanced variations within a patch. When these models are deployed in new geographic areas, this loss of detail causes their accuracy to drop significantly.

A team of researchers built a new dataset called EuroSAT-Embed. They generated 81,000 embedding files using three different foundation models and benchmarked eleven training-free pooling methods. Instead of just averaging the pixels, they tested richer statistical summaries that capture the extremes and the variability within a specific region. For example, a method called stats pooling combines the minimum, maximum, average and standard deviation of the pixels, while covariance pooling looks at the complex relationships between all the different data points.

The results clearly show that moving beyond simple averages makes a massive difference for spatial generalisation. By using richer pooling schemes, the researchers managed to cut the geographic accuracy gap in half and boosted overall performance by up to 6 per cent on spatial splits. They recommend a straightforward three-tier strategy for anyone working with earth embeddings. Mean pooling is fine for a quick baseline, but stats pooling should absolutely be the default choice, while covariance pooling should be reserved for projects where peak accuracy is the absolute priority.

3. Stopping AI from hallucinating over satellite maps

Multimodal LLMs are great at many things, but they tend to make things up when looking at satellite imagery. This issue, known as hallucination, usually happens for two reasons. Either the models get distracted by the massive scale of remote sensing images and can’t find the right area, or the targets are simply too small and cluttered to see clearly. To figure out exactly what is going wrong, researchers built RSHBench, a new diagnostic tool. It proved that models mostly suffer from factual errors, like making up objects or misidentifying attributes, which then snowball into flawed logic.

A clever training-free framework called RADAR essentially forces the AI to use its own internal attention mechanisms to zoom in on the image in two distinct stages. First, it asks a location-oriented query to pinpoint exactly where the relevant visual evidence is hiding. Once it locks onto that specific region, it uses a content-oriented query to zoom in and extract the fine-grained details needed to answer the prompt accurately.

When put to the test across a variety of major open-source and proprietary models, the results were highly impressive. By forcing the AI to actively seek out and refine visual evidence before making a guess, the RADAR framework consistently boosted overall accuracy by 2 to 4 per cent. More importantly, it slashed both factual and logical hallucination rates by around 10 per cent. Ultimately, this gives developers a highly scalable, plug-and-play way to make foundation models much more reliable for analysing complex Earth observation data.

4. Smarter spatial predictions for 5G networks

Predicting exactly where mobile network traffic will spike is essential for planning 5G and 6G infrastructure. While machine learning is great at combining different types of data (like population density and local land use) to map out demand, it often falls into a trap called spatial autocorrelation. Because neighbouring city blocks tend to have very similar traffic patterns, standard AI training methods accidentally leak information between the training and testing datasets. As a result, network planners get a model that looks highly accurate on paper but fails to generalise when deployed in the real world, leading to poorly placed cell towers and wasted spectrum.

A team of researchers developed a new AI-driven framework that forces the model to actually learn rather than just memorise nearby locations. They introduced a two-stage splitting strategy. First, it divides the city into geographically separate clusters to cut off the spatial data leakage. Then, it refines these clusters based on local context, such as whether an area is residential or commercial, to make sure the AI is tested on a truly diverse set of environments. Finally, they apply a Spatial Error Model to scrub out any leftover geographical biases from the predictions.

They tested the framework across five major Canadian cities using a massive crowdsourced dataset of 15 million mobile usage measurements. The new method consistently slashed prediction errors compared to basic location clustering, giving engineers a much more reliable picture of true network demand. When applied to real-world scenarios like 5G spectrum planning, these tighter predictions successfully prevented the costly overestimation or underestimation of network congestion. Ultimately, this gives telecom operators a much sharper tool for figuring out exactly where to deploy their next wave of antennas.

Geospatial Datasets

1. Global soil carbon dynamics dataset

The OpenLandMap-soildb dataset provides the most detailed information to date on global soil carbon dynamics. It integrates over 600,000 field measurements with satellite data and machine learning, and maps soil organic carbon density and content at a 30-metre resolution across three soil depths from the year 2000 onwards. You can access the maps here, the files here, and the code here.

2. Inland water pH and alkalinity dataset

The newly compiled PHALK dataset provides a global assessment of pH and alkalinity across the world’s inland surface waters. It collates chemical data from 18 databases and 55 scientific publications, and maps over 1 million drainage basins. You can access the data here and the code here.

3. Global ocean boundary current dataset

GOXBT dataset provides high-resolution data on the dynamics of major ocean boundary currents. You can access the data here and the code here.

Other useful bits

• ESA’s CryoSat ice-monitoring satellite has received a new software upgrade that allows it to now measure Earth’s magnetic field. Originally designed simply to keep the spacecraft on track, its onboard magnetometer recently captured incredibly precise data during a massive solar storm, offering researchers a fantastic, unexpected bonus alongside dedicated magnetic missions.

• 25 new Open Buildings datasets are now available on HDX thanks to the team at Google Research.

  

• By measuring microscopic changes in gravity on the ocean's surface, a new satellite has helped researchers create the most accurate map of the seafloor to date. One finding is that small underwater hills cover a 70% of the ocean floor 🤯.

• Satellogic has just announced Merlin, a new satellite constellation designed to map the planet every day at 1m resolution. It’s expected to become operational by first half of 2027.

Jobs

• UN DESA is looking for two interns in (1) Geodesy & Geospatial Research Support and (2) Digital Design, Data Visualization & Geospatial Communications.

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

• FAO is looking for a Remote Sensing and GIS Specialist based in Rome.

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

• ERM is looking for a Spatial Analyst (Remote Sensing) based in Australia.

Just for Fun

The recent equinox saw the sun align perfectly over the equator, treating everyone across the globe to an equal measure of day and night as spring arrives in the north and autumn settles in the south.

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 has more rhythm than dembow… 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:

• Radar Model: Billion-scale AI improves SAR image understanding

• Water Resilience: New index captures true water scarcity risks

• Cloud-Free Imaging: AI merges cloud removal and sharpening

• Canopy Heights: Global 1m forest height dataset from Meta and WRI released

• Cropland Emissions: New dataset maps global agricultural emissions

Subscribe now

Research you should know about

1. A billion-scale model for understanding radar imagery

As long time readers know, Synthetic Aperture Radar (SAR) is an incredibly powerful tool for earth observation because it can see through clouds and operate in the pitch black. But teaching AI to understand radar images can be pretty difficult. Unlike normal optical photos, SAR images are plagued by granular “speckle” noise and severe geometric distortions caused by the radar’s side-looking angle. Even worse, SAR data is pretty fragmented. A model trained to identify buildings using one satellite’s specific frequency band or polarisation mode will often completely fail when looking at data from a different satellite. This extreme “domain gap” has prevented the creation of a universal AI model for radar imagery.

Researchers have built CrossEarth-SAR to solve this, the first billion-scale vision foundation model specifically designed to handle the massive diversity of radar data. To keep computational costs down while still using billions of parameters, they employed a “sparse Mixture-of-Experts” (MoE) architecture. Instead of every part of the neural network working on every image, the model uses a “router” to send specific types of radar images to specialised “expert” sub-networks. As part of this, the team built a novel “physics-guided” router that uses actual radar physics (e.g. speckle strength and directional entropy) to decide which expert should handle which image, ensuring the AI adapts to the specific characteristics of the satellite that took the picture.

To train this massive model, the team assembled a dataset of 200,000 radar images, using existing AI models to automatically generate “pseudo-labels” for unlabelled data. They then tested CrossEarth-SAR across 22 benchmarks designed to simulate real-world domain gaps. The model achieved state-of-the-art results on 20 of the 22 benchmarks, outperforming previous models by over 10 per cent in some of the hardest multi-gap scenarios. This suggests that by explicitly hardwiring radar physics into the AI’s routing mechanism, it’s possible to build a single, robust model capable of interpreting radar data from almost any satellite on Earth.

You can check out the project page here and the code here.

2. A smarter way to measure water scarcity

When we think of water scarcity, we usually just look at how much rain is falling or how full the reservoirs are. But a new paper from Scientific Reports argues that this is missing the bigger picture. Whether a community actually runs out of water has just as much to do with politics, infrastructure and economics as it does with hydrology. To fix this, researchers have proposed the Integrated Water Strategic Resilience Index (IWSRI). Instead of just measuring physical water, this new index acts as a broader, more holistic dashboard, combining data on climate variability, governance quality, political stability and economic adaptability to reveal a country’s true ‘water resilience’.

To test this new tool, the team applied it to the Middle East and North Africa, which is essentially ground zero for water stress. The results challenge some common assumptions. For example, Israel, Qatar and the UAE are practically deserts, yet they scored incredibly high on the resilience index. Why? Because at the time they had wealthy, stable governments that had invested heavily in desalination and wastewater recycling. They might be regretting that rating now… Anyway, on the flip side, countries like Yemen, Syria and Libya scored at the absolute bottom. It’s not just that they lack water; it’s that years of war, political instability and destroyed infrastructure have completely crippled their ability to manage the resources they do have.

The authors point out that traditional metrics like the Falkenmark Indicator would completely miss these nuances by focusing solely on per-capita water availability. By integrating sociopolitical factors, the IWSRI shows that water scarcity is often a manufactured crisis. While the researchers acknowledge that quantifying things like ‘governance’ can be tricky, this new index provides a much more realistic map of where the world’s water systems are actually breaking down.

3. Seeing through the clouds with a new all-in-one AI

Satellites monitoring the Earth face a frustrating compromise between detail and frequency. You can either get highly detailed images every few weeks or blurry images every few days. To make matters worse, optical sensors are frequently blocked by clouds, resulting in massive gaps in data. Until now, scientists have tried to fix this by using artificial intelligence to first remove the clouds and then artificially sharpen the images through a process called spatiotemporal fusion. However, treating these as two completely separate tasks often introduces compounding errors that ruin the final picture.

This is what SuperSTF tries to address. It’s essentially a framework that tackles cloud removal and image sharpening at the exact same time. Instead of operating directly on the raw pixels, the system uses a latent diffusion model paired with a Swin Transformer. In simple terms, it compresses the images into a more efficient hidden mathematical space, fuses the frequent blurry data with the rare sharp data, and then rebuilds a pristine image. The model is also fed specific data about cloud locations and the time of year to help it intelligently guess what is hiding beneath the cloud cover.

When tested across diverse regions like urban Washington and agricultural Texas, SuperSTF completely outperformed existing sequential methods. It successfully reconstructed sharp, cloud-free image sequences even when up to 70 per cent of the original high-resolution reference image was obscured by thick clouds. Because it operates in a compressed latent space rather than processing every single pixel during its diffusion phase, it’s also surprisingly fast and computationally efficient. Ultimately, this gives scientists a much more reliable tool for tracking rapid environmental changes, crop health, and urban development without having to constantly wait for a clear sky.

Geospatial Datasets

1. Meta & WRI’s high-res global canopy height map

The newly updated CHMv2 dataset, created by a collaboration between the World Resources Institute and Meta, provides a 1-metre resolution global canopy height map. You can access the data here and the code here.

2. Global cropland emissions dataset

A new dataset from Cornell University and the Land & Carbon Lab maps the greenhouse gas emissions generated by the world's croplands (rice, maize, wheat, and oil palm). You can access the data and code here.

3. Convective weather forecasting dataset

The newly introduced CIDS dataset is designed to train machine learning models for detecting and nowcasting severe convective weather. It covers southeastern China from 2018 to 2023, and captures intense events like short-duration heavy rainfall, thunderstorm winds, and hail at crisp 10-minute intervals. You can access the data here and the code here.

4. Global cropland and irrigation dataset

The MIRCA-OS dataset maps monthly irrigated and rainfed harvested areas between 2000 and 2015. It includes 5-arcminute gridded data for 23 major crop classes. You can access the data here and the code here.

Other useful bits

• ESA’s EarthCARE satellite has just delivered some new global maps of atmospheric dust trails. By using an advanced laser instrument to track how these tiny mineral particles travel across the globe, researchers are getting a wonderfully clear picture of how dust impacts our climate and weather systems.

• Google Earth’s 2026 roadmap has been unveiled, showcasing a shift towards professional workflows. We’re going to see everything from global elevation contours and seamless custom data imports to AI assistants that turn natural language into instant geospatial insights.

• A new initiative in Australia is using Google Earth AI and a bunch of spatial datasets to tackle heart disease in remote communities. By combining environmental data with health records, the approach identifies spatial patterns in risk exposure. The aim is to support earlier intervention by directing care to locations where underlying risk factors are highest.

Jobs

• The UN Department of Economic and Social Affairs is looking for two interns in (1) Geodesy & Geospatial Research Support and (2) Digital Design, Data Visualization & Geospatial Communications.

• FAO is looking for a Remote Sensing and GIS Specialist based in Rome.

• IOM is looking for a Displacement Tracking Matrix (DTM) Systems Developer and Architect based in Geneva.

• UNGSC is looking for a remote ArcGIS Enterprise Administrator from Europe, Middle East and North Africa.

• GHD is looking for a GIS Analyst based in Melbourne.

Just for Fun

This is pretty nuts… 143 million people who spent the last eight years catching Pokémon were actually building one of the largest real-world visual datasets in AI history. Niantic has just revealed that this staggering collection of over 30 billion crowdsourced images is now being cleverly used to train visual navigation AI for delivery robots.

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 frequent as a Marvel movie. 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:

• Spatial AI: LLMs struggle with pure GPS math.

• Change Detection: Foundation model tracks complex landscape changes.

• War Mapping: Satellites track destruction in conflict zones.

• Urban Shade: High-resolution dataset maps global pedestrian shade.

• Image Alignment: Massive dataset of aligned satellite pairs.

Subscribe now

Research you should know about

1. Do AI models actually understand GPS coordinates?

Large Language Models (LLMs) are increasingly being built into navigation apps and geographical tools, but their ability to actually understand and manipulate GPS coordinates has remained a bit of a mystery. Most spatial AI tests involve asking a model to describe where an object is in a photo of a living room, which doesn’t really translate to calculating the distance between Tokyo and Sydney. To bridge this gap, researchers built GPSBENCH, a new dataset designed to test whether LLMs genuinely understand the Earth’s coordinate system or if they are just bluffing.

The researchers tested 14 top-tier LLMs across two main tracks: ‘Pure GPS’ (geometric math like calculating distances or bearings) and ‘Applied’ (using world knowledge, like identifying which city sits at a specific coordinate). The results show a pretty stark divide. Most models are actually much better at the ‘Applied’ track, using their vast pre-trained world knowledge to guess locations, but they completely fall apart when asked to do the pure spherical geometry required for the ‘Pure GPS’ track. For instance, tasks like calculating the area of a polygon on the globe had a near 0 per cent success rate across the board, proving that the models haven’t actually learned geodetic formulas.

Interestingly, the study also revealed that LLMs have a highly hierarchical understanding of geography. They can almost always identify the correct country for a given GPS coordinate, but their accuracy drops off a cliff when asked to pinpoint the specific city. However, when the researchers added random ‘noise’ to the coordinates, the models’ performance didn’t drop, which proves they aren’t just memorising coordinates from databases but do have a genuine, albeit fuzzy, spatial understanding. Ultimately, this shows that if you want an AI to perform precise geographic reasoning, you still need to augment it with external calculators or specific GPS tools, because right now, their internal compass is a bit broken.

You can access the data and code here.

2. A new foundation model approach to tracking landscape changes

Tracking changes on the Earth’s surface usually relies on binary change detection, which simply spots where a landscape has been altered. Semantic change detection takes this a step further by identifying exactly what those changes are, such as a dense forest being replaced by a paved car park. While incredibly useful for urban planning and disaster relief, building models for this task is notoriously difficult. Most current systems rely on standard visual encoders trained on everyday photos, leaving them poorly equipped to understand the complex, top-down perspective of satellite imagery. Furthermore, because the vast majority of land remains unchanged between satellite passes, the models often get overwhelmed by the sheer imbalance of data and resort to lazy predictions.

A new paper has introduced PerASCD, a new framework driven by a massive remote sensing foundation model. By leveraging an AI that has been explicitly pre-trained on vast amounts of satellite data, the system possesses a much deeper, native understanding of geospatial features. The team paired this foundation model with a custom-built Cascaded Gated Decoder. Older methods typically forced the AI through highly convoluted, multi-branch pathways to separate the before and after semantics. This new decoder streamlines the entire process, handling feature compression, multi-level fusion, and upsampling in a single modular pipeline, which drastically simplifies the overall architecture.

The researchers also patched a frustrating mathematical glitch that has plagued previous models. Older systems used a rigid loss function to force the AI to separate changed and unchanged pixels, but the harsh mathematical boundaries of this rule would often cause the model’s accuracy to suddenly collapse during training. By introducing a softened version of this mathematical rule, the team smoothed out the learning gradients and eliminated the numerical instability. When tested on major benchmark datasets like SECOND and LandsatSCD, the framework achieved state-of-the-art performance, delivering significantly sharper boundaries and fewer false positives than previous methods.

3. Mapping war zones with combined satellite data

The ongoing war in Ukraine has caused immense damage to infrastructure, agriculture and the broader environment. Obviously, sending survey teams into active conflict zones to assess this wreckage is incredibly dangerous and often impossible. To get around this, a new study from Scientific Reports have developed a new method to automatically map the destruction from space using freely available satellite data. By combining radar data from Sentinel-1 with optical imagery from Sentinel-2, they have created a system that monitors conflict-related changes without putting anyone in harm’s way.

The interesting part of this approach is how it adapts to different types of terrain. The algorithm automatically classifies the land cover (separating urban environments from rural or agricultural areas) and then chooses the best analysis strategy for that specific zone. The researchers combined radar-based change detection with optical image classification to spot destruction that single-sensor methods might miss. They also added a ‘context-aware smoothing’ technique, which essentially filters out visual noise and local glitches to produce a much cleaner and more reliable map.

When put to the test, the results were pretty impressive. The team compared their automated findings against the official UNOSAT database and found that their tool successfully detected over 80 per cent of damaged buildings. Furthermore, their smoothing technique proved to be far more accurate at detecting built-up areas than existing global land-cover products like AlphaEarth. Ultimately, this gives humanitarian organisations a rapid and reliable way to pinpoint destruction in restricted regions.

You can access the data here and the code here.

Geospatial Datasets

1. Urban pedestrian shade mapping dataset

This dataset provides high-resolution simulations of pedestrian shade provision across nine geographically diverse global cities, including Amsterdam, Sydney, and Rio de Janeiro. You can access the data here and the code here.

2. SAR-Optical image alignment dataset

The SOMA-1M dataset offers a massive collection of over 1.3 million precisely aligned Synthetic Aperture Radar (SAR) and optical image pairs. The dataset will be publicly released here.

3. Multimodal environmental monitoring dataset

The newly introduced MMEARTH-BENCH is a collection of five environmental monitoring tasks designed to evaluate geospatial machine learning models on a global scale. It features 12 diverse data modalities and includes geographic test splits to specifically challenge how well algorithms adapt to unseen regions. You can access the dataset, code, and read more about the project here.

4. Global subnational farm size dataset

This dataset provides the most comprehensive and spatially detailed empirical assessment of global mean farm size to date, targeting the year 2000. Unlike previous products that rely on national averages or modelled data, this collection compiles census and survey information at the smallest available subnational administrative level for 101 countries, alongside national data for 99 others. You can access the data and code here.

Other useful bits

• The team behind AlphaEarth Foundations has just rolled out their 2025 Satellite Embedding dataset, cleverly packing an entire year's worth of satellite data into neat 10-metre pixels. This update makes it easier than ever to spot global shifts without the headache of processing raw imagery.

• Google DeepMind has showcased a new realistic city planner app powered by Gemini 3.1 Pro. It tackles complex terrain and simulates traffic to generate incredibly high-quality visualisations of future urban infrastructure.

• A new time series foundation model called TimesFM has just been released, which has been pre-trained with 100 billion data points. It’s showing pretty impressive performance across a whole host of domains, making it a useful resource for anyone keen to dive into machine learning and forecasting.

Jobs

• Mapbox is looking for a remote Software Development Engineer based in either Germany or Finland.

• The European Centre for Medium-Range Weather Forecasts (ECMWF) is looking for a (1) Data Management Engineer and (2) Climate Data and AI/Machine Learning Scientist, both based in either Bonn, Germany or Reading, UK.

• The European Organisation for Nuclear Research (CERN) is looking for a GIS Specialist and Developer based in Geneva.

Just for Fun

The Artemis II crew ran into a very Earthly problem whilst en route to the Moon. Just hours into their ten-day flyby, the astronauts had to radio mission control for IT support after their Microsoft Outlook crashed. It’s comforting to know that no one is safe from an email glitch.

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 Terra and exactextract enthusiasts... 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:

• Human Development: Satellites map local development levels.

• Flood Prediction: LLMs estimate damage without training.

• Ecological Mapping: Testing foundation models on ecosystems.

• Urban Inequality: Satellites reveal China’s built disparities.

• Hyperspectral Data: Planet expands its open dataset.

Subscribe now

Research you should know about

1. A much better way to map human development

For years, the UN’s Human Development Index (HDI) has been the gold standard for measuring a country’s wellbeing, factoring in health, education and income. The problem is that HDI data is generally only available at the national or, at best, provincial level. This makes it almost useless for local policy decisions, like deciding which specific neighbourhoods need a new clinic or school. A new paper from Nature Communications used satellite imagery and machine learning to produce the first high-resolution global map of HDI, estimating development levels for over 60,000 municipalities and a 10km-by-10km grid worldwide.

The team achieved this using an approach called MOSAIKS, which translates satellite imagery into a set of random convolutional features. Rather than trying to map irregularly shaped provinces directly, the model learns the relationship between the visual features of an area (like building density, road networks and nighttime lights) and its known HDI. Because this relationship is linear, they can train the model on coarse provincial data and then “downscale” it to predict HDI for much smaller areas based purely on what the satellites see. They validated this by testing it against ground-truth municipal census data in Mexico, Brazil and Indonesia, proving the model could accurately spot local variations.

The results expose a massive flaw in how we currently look at development. The researchers found that if you only use provincial data, more than half of the global population is assigned to the wrong HDI quintile within their country due to “aggregation bias”. It essentially lumps rich urban centres and poor rural outskirts into a single misleading average. By contrast, this new high-resolution data allows policymakers to target aid far more accurately. In a simulated test in Mexico, using municipal data rather than provincial data increased the accuracy of targeting aid to the poorest 10 per cent of the population by over 11 percentage points.

You can access the code here.

2. Predicting flood damage without training models from scratch

Predicting property damage immediately after a flood usually involves training complex machine learning models on highly specific regional data. This requires an immense amount of upstream effort to assemble datasets, engineer features and handle skewed data where extreme damage cases are rare. To bypass this bottleneck, a new study has introduced R2RAG-Flood, a new framework that uses LLMs to predict post-storm Property Damage Extent (PDE) without requiring any task-specific training or fine-tuning.

Instead of fitting new parameters, the framework uses Retrieval-Augmented Generation (RAG). It takes historical tabular data about flood events (like rainfall, elevation and building density) and converts it into concise text summaries paired with logical “reasoning trajectories.” When a new storm hits, the system looks at a specific target area and searches its knowledge base for similar historical cases nearby (geospatial neighbours) and typical examples of different damage levels (prototypes). It feeds these examples to an LLM, asking it to emulate the reasoning process to classify the new target into low, medium or high damage.

The team tested this approach on data from Hurricane Harvey in Harris County, Texas. While a traditional supervised model achieved 0.714 overall accuracy, the training-free R2RAG-Flood approach was highly competitive, hitting up to 0.668 accuracy depending on the LLM backbone used. More importantly, the RAG approach outputs a structured, human-readable rationale explaining exactly why it predicted a certain level of damage. When factoring in the cost of API calls versus the computing power needed to train a model from scratch, the lightweight LLM variants proved significantly more cost-efficient, offering a scalable way to rapidly assess flood impacts in new regions without heavy computational overhead.

3. Putting AI foundation models to the ultimate ecological test

Geospatial Foundation Models (GFMs) are, of course, all the rage for earth observation, but mapping complex ecosystems is a whole different beast compared to drawing basic land cover maps. To see if these models actually live up to the hype, researchers put two leading GFMs (Prithvi-EO-2.0 and TerraMind) to the test against a standard deep learning baseline. They wanted to find out if these systems could handle highly specialised jobs like estimating forest canopy density across the United States and tracking the extent of carbon-rich peatlands in Patagonia.

The results were pretty clear: both GFMs comfortably beat the standard baseline model across all the ecological tasks. By being pre-trained on massive amounts of satellite data, these models avoided the usual pitfalls of standard image classifiers and adapted much better to geospatial nuances. TerraMind slightly outperformed Prithvi on basic tasks, but it really pulled ahead when the researchers fed it multimodal inputs. By combining standard optical imagery with radar and elevation data, the model became much better at separating tricky vegetation types, like distinguishing vital peat moss from generic grass.

Despite the impressive performance, the study highlighted some pretty big roadblocks before we can completely hand ecological mapping over to AI. For one, these models are highly sensitive to the quality of the training labels. When the researchers used coarse, off-the-shelf datasets to teach the models, the AI struggled to draw accurate boundaries, proving that you still need high-fidelity, expert-verified field data. Furthermore, the standard 10-metre resolution of satellite inputs proved to be a major limitation for capturing fine-scale ecological changes. So while these foundation models are a massive leap forward, they still need sharper data and better ground truths to truly map out the complexities of the natural world.

4. Satellite mapping reveals deep built environment disparities in urban China

China’s cities have grown at a breakneck pace over the last few decades, but figuring out exactly how equitable this growth has been at the street level is notoriously difficult due to a lack of building-level data. To tackle this, a team of researchers just released SinoBF-1, a massive national map that categorises the specific function of 110 million individual buildings across 109 major Chinese cities. By combining 1-metre high-resolution satellite imagery with other multi-modal data, they essentially built a microscopic view of the country’s urban landscape to track exactly where resources are concentrated.

Using nine distinct indicators covering everything from facility accessibility to infrastructure sufficiency, the study reveals some pretty big disparities across different regions and city tiers. Unsurprisingly, top-tier megacities dominate when it comes to amenity diversity and access to services, with these perks dropping off a cliff as you move down to lower-tier cities. However, the data shows that mid-tier cities actually offer the most equitable housing allocation. Geographically, southern cities boast the best access to healthcare and education, but this comes at the cost of severe infrastructure overcrowding compared to other regions.

Zooming in on individual cities, the map also highlights a growing divide between old and new neighbourhoods. The researchers found that newer, recently expanded zones on the outskirts of cities suffer from much greater built environment disparities than the early-established urban cores. This spatial footprint serves as a physical ledger of national development policies over the past half century, offering a highly detailed framework for tracking urban inequality in rapidly developing areas.

Access the data here and the code here.

Geospatial Datasets

1. Open hyperspectral satellite dataset

Planet has recently expanded its Open STAC Catalog with nearly 100 new hyperspectral scenes, offering an impressive global footprint for geospatial exploration. You can access the data here.

2. Tile-drainage and Rotation-Enhanced Cropland (TREC) dataset

The newly developed Tile-drainage and Rotation-Enhanced Cropland (TREC) dataset significantly improves the spatial accuracy of ecohydrological models across the Conterminous United States. You can access the data here and the code here.

3. High-resolution global precipitation dataset

The newly updated PERSIANN-CCS-CDR Version 2.0 offers a remarkable high-resolution global precipitation climate record, boasting a 0.04° spatial and 3-hourly temporal resolution. You can access the data here and the code here.

4. Global water storage anomaly dataset

The newly released ML-TWiX dataset is a brilliant resource that extends our understanding of global total water storage anomalies back to 1980, well before the launch of the GRACE satellite missions. You can access the data and code here.

Other useful bits

• Vantor has cleverly teamed up with Google to integrate powerful Earth AI imagery models into its Tensorglobe platform, allowing these tools to be used in secure, air-gapped environments for the very first time. By combining cutting-edge AI with a massive archive of high-resolution satellite imagery, it offers a brilliant new way to effortlessly monitor infrastructure and track changes offline.

• To tackle AI's struggle with 'spatial grammar', researchers have built MapTrace, a new pipeline that generates millions of synthetic map-path pairs for training. This approach has already given models a massive boost in accurately tracing paths on real-world maps, proving just how useful synthetic data can be.

• A developer has cleverly forked Apple’s Embedding Atlas to create Geospatial Atlas, a new open-source tool for exploring massive datasets of up to 100 million points right in your web browser. It brings together tools like WebGPU, DuckDB, and MapLibre. This essentially provides a smooth and interactive way to visualise huge amounts of spatial data without needing heavyweight desktop software.

• A new AI methodology called Groundsource has just been unveiled, turning millions of public reports into a massive historical dataset for predicting urban flash floods. By mapping these events using Gemini and Google Maps, a new model can now forecast sudden flooding up to 24 hours in advance, offering a vital lifeline to communities worldwide.

Jobs

• ESA is looking for two EO Data Architect and AI Experts based in Frascati for their (1) Copernicus Ground Segment Systems Section and (2) Ecosystem Development Section.

• Development Seed is looking for a Technical Project Lead based in the US who will work across their NASA portfolio.

• The European Organization for Nuclear Research (CERN) is looking for GIS Specialist and Developer based in Geneva.

• Médecins Sans Frontières (Doctors Without Borders) is looking for a GIS Data Governance Specialist based in Geneva.

• The Center for International Forestry Research (CIFOR) is looking for a Spatial Data Scientist for Machine Learning & Remote Sensing who will be based in either Kenya, Nairobi or remote.

Just for Fun

Researchers at NOAA have shared a rather striking animation of high winds over the Great Lakes, created using their experimental RRFS weather model. It provides a brilliant demonstration of DESI, a clever tool designed to help forecasters effortlessly explore, compare, and share ensemble weather guidance.

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 just as rivetting as a Lex Fridman podcast. 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:

• Hidden Carbon: Earthquake repairs generate massive CO2 emissions.

• Spatial AI: Drones learn to measure 3D dimensions.

• Design Economics: Distinctive buildings boost surrounding property values.

• Location Data: Mobility patterns reveal true venue functions.

• Building Atlas: Open dataset provides global 3D models.

Subscribe now

Research you should know about

1. Earthquakes have a huge carbon footprint

When we talk about the environmental cost of buildings, we usually focus on the materials used to build them or the energy they consume. But a new study in Nature Communications shines a light on a hidden carbon cost: the emissions generated by repairing and rebuilding structures after earthquakes. By integrating seismic risk models with a massive database of construction materials, researchers have estimated that Europe’s building stock embodies nearly 14 billion tonnes of CO2 equivalent. More strikingly, the damage caused by earthquakes across the continent generates an average of 6.6 million tonnes of carbon emissions every single year— roughly the same as the annual emissions of 3.9 million cars.

The study introduces a new metric called ‘Average Annual Embodied Carbon’ (AAEC) to quantify this risk. Instead of just looking at financial losses, this measures the expected yearly carbon footprint of seismic damage. The results show a heavy concentration of this environmental burden in just a few countries. Türkiye tops the list, accounting for over half of the total annual emissions due to its high seismic activity and large number of vulnerable buildings. Italy, Romania, Greece, Spain and Bulgaria also feature prominently, meaning just six nations are responsible for about 90 per cent of Europe’s earthquake-related carbon emissions.

This research highlights a critical blind spot in our current sustainability strategies. While we are busy making buildings more energy-efficient, we often ignore the ‘embodied carbon’ that gets released when we have to fix or replace them after a disaster. The authors argue that seismic resilience is actually a form of climate action. By retrofitting buildings to withstand shocks, we aren’t just saving lives and money; we are also preventing millions of tonnes of CO2 from being pumped into the atmosphere during reconstruction. The team has released their datasets and models publicly, hoping to help policymakers integrate these hidden environmental costs into their disaster risk planning.

2. Giving drones a sense of depth

Remote sensing AI is getting sharper, yet it still struggles with a fundamental human skill: understanding 3D space. Most current models look at 2D pixels and can identify a ‘red car’, but they can’t reliably tell you its actual dimensions or how far away it is from the sensor. To bridge this gap, researchers have developed AirSpatial, a massive dataset derived from drone imagery that introduces 3D bounding boxes to the mix. By combining aerial photos with ground-level video, they have annotated thousands of vehicles with precise spatial data, allowing models to move beyond simple visual recognition to true spatial grounding where an AI can understand depth and scale.

The team used this data to train AirSpatialBot, an agent designed to recognise fine-grained vehicle attributes from the sky. The system uses a clever two-stage training strategy: it starts by learning standard 2D image understanding before fine-tuning on spatial tasks using techniques like Geometric Mapping Learning. Instead of just memorising what a specific car model looks like from above, the bot actually measures the vehicle’s length, width and height in the 3D scene. It then acts like a detective, cross-referencing these physical dimensions with a database to determine the exact make, model and even the price of the car, managing to distinguish between vehicles that look visually similar but differ in size.

In testing, AirSpatialBot significantly outperformed existing models like GeoChat, particularly in tasks requiring depth estimation and dimension measuring. The system excels at ‘target retrieval’, where a user can ask it to find a specific car (e.g. ‘white Mercedes-Benz C-Class’) and the bot locates it by calculating 3D coordinates. This approach is particularly flexible because it doesn’t require retraining when new car models come out; you simply update the reference table with the new dimensions. It’s a compelling step forward for drone-based embodied intelligence, suggesting that future aerial agents won’t just see the world as a flat image, but will navigate and understand it in full 3D.

3. The economics of good design

We all know that walking through a neighbourhood with stunning architecture feels better than walking through a sea of bland concrete, but quantifying that feeling has always been tricky. A new working paper takes a deep dive into the economics of architecture by combining a massive literature review with a new quantitative model. They found that distinctive buildings sell for about a 15 per cent premium compared to ordinary ones, confirming that people are willing to pay for good design. Perhaps more interestingly, they found a significant positive spillover effect where distinctive buildings boost the value of nearby properties by around 9 per cent. The catch is that building something unique is expensive, costing about 25 per cent more than a standard structure. This creates a classic coordination problem where developers under-invest in good design because they bear all the costs but only capture some of the value.

To understand how to fix this, the researchers modelled a neighbourhood where residents have different tastes for architecture. They ran a survey and discovered that people’s preferences for design are quite varied, roughly as diverse as their preferences for where to live within a city. This heterogeneity means that as you build more distinctive buildings, the premium people are willing to pay for them drops because you run out of ‘super-fans’ for that style. The model shows that without intervention, the market provides less distinctive architecture than is socially optimal. It turns out that the best way to solve this isn’t through complex zoning rules but through a direct Pigovian subsidy. Their simulations suggest that subsidising about 10 per cent of the construction costs for distinctive buildings creates the maximum welfare gain, balancing the extra cost against the aesthetic benefits for the community.

The study also throws cold water on some common planning tools. For instance, offering Floor Area Ratio (FAR) bonuses—letting developers build taller if they include good design—often backfires. Unless the city is already severely height-constrained, these bonuses tend to reduce welfare by distorting the housing supply. Similarly, relying on a ‘super-developer’ to build a whole district doesn’t necessarily help because they might restrict the supply of distinctive buildings to keep prices high. Designating small ‘distinctive districts’ (like conservation areas) can work well if they cover a small chunk of the neighbourhood, around 4 per cent, but expanding them too much leads to diminishing returns. The takeaway is that if we want better-looking cities, we should probably just pay for them directly rather than trying to engineer them through complex planning loop-holes.

4. Your location data knows more than you think

In the world of digital mapping, a point of interest (POI) is usually defined by static data: its name, category and coordinates. But as anyone who has ever visited a ‘quiet’ cafe that turned out to be a bustling co-working hub knows, a label doesn’t tell the whole story. To capture this nuance, researchers have introduced ME-POIs, a framework that enriches static POI data with patterns from large-scale human mobility. Instead of just knowing a place is a ‘restaurant’, the model learns its function by analysing visit sequences (whether it’s a quick lunch spot or a long-stay dinner venue).

The system uses a contrastive learning approach to align individual visit patterns with a global representation for each POI. One clever trick they employ is a ‘multi-scale distribution transfer’ mechanism. Since most places don’t get enough visits to generate robust data on their own, the model borrows insights from nearby, popular spots to fill in the gaps for less-frequented locations. This allows it to generate meaningful insights even for the ‘long tail’ of rarely visited businesses that usually get overlooked by data-hungry algorithms.

When tested on tasks like predicting opening hours, price levels and visit intent in Los Angeles and Houston, the mobility-augmented model consistently outperformed text-only baselines. For instance, adding mobility data boosted the accuracy of predicting a venue’s price level by over 75 per cent in some cases. It turns out that how people move through a city—when they arrive, how long they stay, and where they go next—reveals as much about a place’s identity as its official description does.

Geospatial Datasets

1. Global building dataset

GlobalBuildingAtlas presents the first open, globally complete dataset comprising building polygons, heights, and Level of Detail 1 (LoD1) 3D models. You can access the data here and the code here.

2. Semantic vehicle trajectory dataset

Captured via Unmanned Aerial Vehicles (UAVs) over the busy Guangzhou Bridge, the High-Density Semantic Vehicle Trajectory (HDSVT) dataset fills a gap in autonomous driving resources by providing long-duration tracking in high-density scenarios. You can access the data here.

3. Transport systems database

The African Transport Systems Database (AfTS-Db) stands as the first comprehensive, geolocated multi-modal transport database for the entire African continent. You can access the dataset here and the code here.

4. Global leaf chlorophyll content dataset

The MuSyQ Global LCC product revolutionises ecosystem monitoring by leveraging Sentinel-2 imagery to map leaf chlorophyll content at an impressive 10-m resolution. You can access the data here and the code here.

Other useful bits

• Europe’s Destination Earth project is taking climate modelling to the next level with a high-resolution ‘digital twin’ of our planet. This AI-powered simulation isn’t just about sharper weather forecasts; it’s already helping to optimise renewable energy and map urban heat, providing a crucial dashboard for understanding Earth’s complex systems.

• In its latest impact report, Planet Labs reveals how its daily Earth imagery is driving real-world change, from helping conservationists protect endangered species to spotting massive methane leaks in Texas.

• Albedo has shared a candid look at its Clarity-1 mission, proving that satellites can indeed thrive in the harsh environment of Very Low Earth Orbit (VLEO). Although the mission ended early due to hardware issues, it successfully validated the platform’s core tech, paving the way for ultra-high-resolution 10 cm imagery in the future.

• Data Center GeoSniffer is a fascinating new tool that merges AI permit scraping with satellite imagery to track critical infrastructure expansion across the US. By providing a ‘neural sniff’ of the grid, it allows users to look past the paper trail and spot physical changes like new substations and generator pads.

Jobs

• Center for International Forestry Research (CIFOR) is looking for a remote Spatial Data Scientist

• International Water Management Institute (IWMI) is looking for a Senior Geospatial Data Engineer based in Colombo, Sri Lanka

• Impact Initiatives is looking for a Senior Assessment officer based in Geneva, Switzerland.

• CERN is looking for a GIS Specialist based in Geneva, Switzerland

Just for Fun

New research reveals that space travel can actually cause an astronaut's brain to shift upwards in the skull, leading to disorientation that lingers long after they return to Earth. Understanding these ‘sensory conflicts’ is now seen as a vital step before we embark on longer missions to the Moon or Mars.

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 was named by the same folks who named GPT-5.4 (Thinking) High. 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:

• Data Centres: New maps show remaining US capacity.

• SAM3 Fixes: CONCEPTBANK improves segmentation under domain drift.

• Spatial AI: Why chatbots still struggle to drive.

• Health Mapping: AI links imagery and demographics.

• WorldPop 2: New global population grids at 100 metres.

Subscribe now

Research you should know about

1. We’re running out of room for data centres

The recent generative AI boom is driving a massive expansion of hyperscale data centres across the United States. While we have plenty of maps showing where these facilities are currently located, a new study takes a different approach by looking at where they can actually still be built. Rather than relying on future demand forecasts or basic zoning laws, the researcher built a constraint-first geospatial framework. This means they focused entirely on the strict physical, environmental and infrastructural limits that dictate where a massive server farm can realistically operate today.

To map this out, this new study divided the country into a uniform hexagonal grid and analysed the precise characteristics of existing hyperscale sites. The models layered in a huge amount of spatial data, including high-voltage power transmission lines, local substations, summer temperature peaks, flood risks and built-up land density. By using unsupervised machine learning approaches to compare these data layers against the environments of current facilities, the analysis identified remaining regions that share the specific structural DNA needed to support new development.

The findings suggest that the US is facing a pretty tight ceiling when it comes to future hosting capacity. Both models converged on a surprisingly limited envelope of feasible land, estimating that the total physically viable capacity left is somewhere in the tens of gigawatts (roughly 25 to 100 GW), rather than the hundreds of gigawatts that some optimistic industry forecasts assume. Ultimately, this shows that the tech industry is running out of easy real estate. Exceeding this physical limit will require a fundamental overhaul of the underlying electricity system and local regulatory landscapes.

2. Taming foundation models for the real world

The recently released SAM3 model has been pretty massive for text-to-segmentation. As with the other SAMs, it allows users to segment almost anything just by typing a text prompt. However, when you take it out of its comfortable pre-training environment and deploy it in the wild, things start to break down. The main culprits are data drift and concept drift. Data drift happens when the actual visual statistics shift, which is a pretty big issue in remote sensing where aerial viewpoints and spectral signatures look nothing like standard photos. Concept drift occurs when the meaning of a label changes depending on the context (like how the word “mouse” means an animal in a nature dataset but a clicking-device-thingy in an office dataset). When these shifts happen, the model’s static text prompts no longer align with the visual evidence, leading to missed objects and incorrect masks.

To fix this without the massive computational headache of retraining the whole model, researchers have proposed a parameter-free framework called CONCEPTBANK. Instead of relying on static text prompts, this system builds a dataset-specific dictionary derived directly from the target domain’s actual statistics. It works through a clever three-stage pipeline. First, it anchors target evidence by finding visual prototypes for each class. Second, it filters out noisy outliers to ensure it is only looking at representative examples. Finally, it uses a large language model to generate a variety of synonyms and attribute descriptions, fusing the best ones together to create a calibrated text anchor that perfectly matches the target dataset. It does all of this on the fly without updating a single weight in the SAM3 backbone.

The results are quite compelling, especially for those of us working with geospatial data (which I’m guessing is you). When tested across four different remote sensing benchmarks, CONCEPTBANK boosted SAM3’s average performance by 13 percentage points compared to the vanilla version. Datasets like LoveDA and iSAID, which usually trip up standard models due to their complex aerial perspectives, saw huge improvements. It also performed exceptionally well on standard natural scene datasets. Because the system bypasses the text encoder during actual inference and just retrieves the pre-calibrated anchors, it actually speeds up the processing time by 25 per cent. Ultimately, this approach provides a highly robust, plug-and-play way to optimise massive foundation models for messy, real-world satellite imagery.

3. Why chatbots make terrible drivers

LLMs are great. And I’m sure many of you are like me - our best friends are an LLM, like my OpenClaw Billy. But LLMs have historically failed when navigating the physical world. A model that can perfectly describe a photo of a kitchen cannot necessarily figure out how to move through it or pick up a mug. This happens because these models lack ‘spatial grounding’, which is the metric understanding of geometry and physics required to turn perception into action. To solve this disconnect, a team of researchers reviewed over 2,000 papers to create a massive new taxonomy that connects standard AI agent capabilities with spatial intelligence.

Their framework divides the landscape into three axes: the specific spatial task, the agent’s cognitive capabilities, and the physical scale. When mapping existing research onto this grid, they discovered a pretty big imbalance. Roughly 68 per cent of current methods focus on ‘meso-spatial’ tasks like room-scale navigation, leaving micro-scale robotic manipulation and macro-scale geospatial analysis completely underexplored. To bridge these gaps, the authors highlight some promising architectural patterns, specifically integrating Graph Neural Networks with language models. This helps agents maintain a structured, persistent understanding of spatial relationships over time rather than relying on text alone.

Another major takeaway is the absolute necessity of ‘world models’ for safe physical deployment. Since physical mistakes can cause irreversible damage, agents need these internal simulators to predict the consequences of their actions before actually making a move. Looking ahead, the researchers pinpoint several grand challenges that stand in the way of truly autonomous spatial systems. The most glaring issue is the lack of unified cross-scale representation, meaning an AI struggles to seamlessly transition from planning a city-wide route down to the centimetre-level precision needed to open a door. Solving these structural barriers will be essential for the next generation of robotics and autonomous vehicles.

4. Mapping public health from space

Predicting public health outcomes usually requires massive amounts of census and medical data. Satellite imagery can show us the physical environment, like green spaces or industrial zones, but it misses crucial demographic context. To bridge this gap, researchers have developed a new AI model called the Geo-Context Guided Visual Transformer (GCGVT). Instead of just merging data blindly, it uses a ‘Guided Attention’ mechanism. This allows the model to look at an aerial view of a neighbourhood while being explicitly guided by specific socioeconomic factors, pointing the AI to the most relevant physical features.

The team tested their model on predicting the prevalence of nine chronic diseases, including diabetes, obesity and depression, across 14,330 historical redlining zones in the US. They fed the AI aerial photos alongside 228 variables from the American Community Survey, covering details like age, income and race. The new model achieved an impressive average predictive score of 0.85, completely outperforming standard vision-language models like CLIP. Interestingly, the researchers found that they did not need all 228 variables. By combining the aerial imagery with just two demographic categories, such as Income and Race for predicting obesity, or Age and Housing for arthritis, the model still delivered highly accurate forecasts.

The real breakthrough here is the transparency of the system. Most deep learning models operate as a black box, making them hard to trust in healthcare settings. However, this framework assigns different AI attention heads to specific demographic categories, allowing users to see exactly how the model connects the dots. For instance, when predicting depression rates, the model visually highlighted how employment data correlated to specific housing blocks, and how education levels mapped onto nearby natural areas. This gives public health officials a highly interpretable tool to map out health disparities without relying on exhaustive, expensive on-the-ground surveys.

Geospatial Datasets

1. WorldPop Global 2 dataset

The newly released WorldPop Global 2 dataset provides highly detailed, annual geospatial population estimates at a 100-metre resolution for over 240 countries and territories, spanning the years 2015 to 2030. Built using recent census information and satellite-derived settlement modelling, this open-access resource also provides age and sex breakdowns. You can access the data here.

2. Long-term night-time light dataset

The Extended VIIRS-like Artificial Nighttime Light (EVAL) dataset provides a continuous, high-quality annual record of night-time illumination across China from 1986 to 2024. Using a two-stage deep learning framework, this 500-metre resolution map resolves common issues found in older satellite records, such as the underestimation of brilliant city centres and the blurring of intricate road networks. You can access the data here and the code here.

3. Impervious surface cover dataset

The China Impervious Surface Cover (CISC) dataset provides a 30-metre resolution map of artificial surfaces across mainland China for 2020 and 2022. It fuses together high-resolution 2-metre imagery from the Gaofen satellite series with 30-metre Landsat and SRTM data via a regionally adaptive deep learning model. You can access the data here and the code here.

4. Bioavailable strontium isotope data

If, unlike me, you’ve heard of bioavailable strontium isotopes, then you’re in luck. This compilation harmonises over 28,000 bioavailable strontium isotope (87Sr/86Sr) data points from more than 150 countries into one easily accessible resource. It draws from decades of research across soil, water, flora, and fauna samples. You can access the data here and here, and the code here.

Other useful bits

• The European Space Agency has just launched the EOPF Sentinel Zarr Explorer, a new platform that shifts massive Earth observation datasets into a new (and different) cloud-native format. This means we can now discover, visualise, and analyse satellite imagery directly in their web browsers without the headache of downloading colossal files.

• Researchers have just released PyGALAX, a new open-source Python toolkit designed to take the headache out of complex geospatial machine learning. By combining automated modelling with explainable AI, it gives users a much clearer picture of the 'why' behind spatial patterns, making it a fantastic resource for everything from urban planning to environmental monitoring.

• A new tool called the Interpretable Code-based assistant for Earth Observation (IC-EO) is making satellite data analysis accessible by translating simple, natural language questions directly into Python code. Developed by researchers alongside the European Space Agency, it allows even non-experts to tackle complex tasks like wildfire assessments and land-use mapping with impressive accuracy and transparency.

• Google Earth is tackling the headache of city gridlock by letting urban planners import their own custom traffic data directly onto the globe. By mixing this with built-in layers like vehicle trip counts and signal performance, it provides a new way of pinpointing bottlenecks.

Jobs

• Mapbox is looking for a Technical Support Engineer based in China

• UNEP - United Nations Environment Programme is looking for a Spatial Data Application Developer based in Nairobi, Kenya

• International Water Management Institute (IWMI) is looking for a Senior Geospatial Data Engineer based in Colombo, Sri Lanka

Just for Fun

This map outlines the European Union's envisaged transport and trade corridors across Africa.

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 more riveting than a Bad Bunny halftime performance. As usual, the aim is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Spatial Mapping: ANPs improve biomass uncertainty estimates.

• Image Sharpening: Synthetic data enhances hyperspectral resolution.

• Landslide Detection: Deep learning improves hazard monitoring.

• Malaria Threat: Climate change disrupts disease eradication.

• Building Heights: New deep learning estimation dataset.

Subscribe now

Research you should know about

1. Knowing when to doubt the data

A lot of folks use standard machine learning models like Random Forest or XGBoost to map environmental variables like aboveground biomass. These models are fantastic at making point predictions, but a new paper highlights that they are systematically overconfident when estimating uncertainty. When you use these standard baseline configurations, you end up with uncertainty estimates that fail to capture local spatial variability. For instance, if you’re mapping a landscape with sharp transitions between forests and cleared fields, traditional models generate the exact same narrow uncertainty interval they would anywhere else. They completely miss the local context. This basically means development organisations and carbon market verifiers could be making big conservation decisions using highly inaccurate confidence levels.

The researchers present Attentive Neural Processes (ANPs) as a much smarter alternative for spatial mapping. ANPs use an episodic training structure to learn spatial patterns directly, improving upon older methods that treat every observation in isolation. The model takes a sample area, splits the data, and learns to predict targets conditioned on the surrounding context. If the nearby data points are tightly clustered, the model assigns low uncertainty. If the context points are scattered and highly variable, it naturally expands its prediction intervals. The model essentially meta-learns how to adapt to local landscape heterogeneity on the fly.

The results show a massive difference between the two approaches. Across diverse forest types, the ANP consistently achieved well-calibrated uncertainty scores, whereas the traditional models fell way below expected coverage targets. When the models had to transfer their predictions to entirely new geographic regions, XGBoost coverage dropped to around 15 per cent because it simply could not adapt to the distributional shift. The ANP maintained strong coverage even as overall accuracy dipped. It is a pretty big reminder that having a model that knows when it is unsure is just as crucial as having one that makes accurate predictions.

2. Sharpening images with ‘dead leaves’

Hyperspectral imaging is awesome for identifying materials from space, but it usually suffers from a trade-off where you get rich spectral detail but poor spatial resolution. Deep learning can normally sharpen these images, but there’s a catch. There’s always a catch… These models typically need high-resolution ‘ground truth’ data to learn from, which simply doesn’t exist for most real-world scenarios. To get around this, researchers from Télécom Paris have come up with a clever unsupervised approach that relies on fully synthetic training data generated by something called the ‘dead leaves model.’

The main idea is to break the hyperspectral image down into ‘abundances’ (maps showing where specific materials are) and ‘endmembers’ (the spectral signature of those materials). The team then trains a super-resolution neural network solely on synthetic abundance maps created by simulating dead leaves falling on a canvas. While this sounds abstract, these random, overlapping shapes create statistical properties like edges and flat regions that mimic real-world abundance maps surprisingly well. This allows the network to learn how to sharpen features without ever seeing a real high-resolution image during training.

When tested on the standard ‘Urban’ dataset, this method didn’t just work; it actually outperformed state-of-the-art supervised methods like MCnet and HSISR. This is a pretty big deal because the competitors were trained on data with ground truth, while this model learned entirely from synthetic shapes. It demonstrates that we don’t always need expensive or impossible-to-acquire real-world datasets to train effective AI models. By using the dead leaves model to simulate texture and structure, the researchers provided a robust proof of concept for sharpening hyperspectral imagery in a completely unsupervised way.

3. Teaching AI to spot landslides

Landslides pose a massive threat to infrastructure in geologically unstable regions, yet our traditional methods for monitoring them (like ground surveys and visual inspections) are often too slow to be effective. To address this gap, a new paper in Scientific Reports looks into a new deep learning approach that employs enhanced Deep Convolutional Neural Networks (DCNNs) to detect landslides with high precision.

The researchers introduced several key modifications to make the system smarter. They integrated a spatial attention mechanism and attention-based Global Average Pooling, which essentially help the AI focus on the most critical parts of an image rather than getting distracted by irrelevant background noise. To further refine the training process, they utilised the Lookahead Adam optimizer and optimised learning rate schedules. These technical tweaks were designed to improve how the model extracts features and converges on a solution, ensuring it can generalise well across different types of terrain.

The performance metrics for this new approach are pretty impressive. In experiments using the Kaggle Landslide Dataset, the model achieved a testing accuracy of 95.8 per cent, with similar success observed when testing against the NASA Landslide Inventory. This level of accuracy suggests that attention-driven DCNNs could significantly outperform conventional assessment methods. By offering a reliable and efficient tool for automated monitoring, this technology could be crucial for early detection systems that help mitigate risks to communities living in danger zones. You can access the code here.

4. Climate chaos threatens malaria eradication

After decades of progress, the global fight against malaria has stalled, and climate change is emerging as a massive new hurdle. A new study in Nature has quantified this threat by looking at both the direct ecological impacts of a warmer world (like mosquito breeding rates) and the indirect, ‘disruptive’ impacts of extreme weather. By integrating 25 years of data on climate, disease burden and socioeconomic factors, the researchers project that climate change could lead to an additional 123 million malaria cases and 532,000 deaths in Africa between 2024 and 2050.

Interestingly, the study finds that it isn’t just about mosquitoes enjoying warmer weather. In fact, 79 per cent of the additional cases and 93 per cent of the extra deaths will likely be driven by extreme weather events like floods and cyclones. These disasters destroy housing, cut off access to healthcare and disrupt vector control programs, creating perfect conditions for outbreaks in areas that are already struggling. The research shows that while some regions might actually see a decrease in transmission due to excessive heat, this is far outweighed by the intensification of malaria in existing endemic zones, particularly in densely populated countries like Nigeria and Uganda.

This analysis challenges the traditional focus on ‘range expansion’, showing that 99 per cent of the additional burden will hit communities that already suffer from malaria, rather than spreading to new areas. The compounding effect of ecological changes and infrastructure disruption creates a ‘double whammy’ for vulnerable populations. To safeguard the goal of eradication, the authors argue we need to move beyond standard control measures and invest heavily in climate-resilient health systems, ensuring that clinics, supply chains and mosquito control efforts can withstand the increasingly volatile weather of the coming decades. You can access the code here.

Geospatial Datasets

1. Building height estimation dataset

M4Heights is a new benchmark dataset designed to advance building height estimation through deep learning, covering vast landscapes across Estonia, the Netherlands, and Switzerland. You can access the data here and the code here.

2. Carbon, water, and energy fluxes dataset

The Unified FLUXes (UFLUX) ensemble provides a globally consistent dataset of carbon, water, and energy fluxes, derived from a robust combination of eddy covariance data, satellite observations, and machine learning. You can check out the list of datasets here and the code here.

3. Automated crop monitoring dataset

This comprehensive dataset addresses the critical need for automation in crop monitoring by providing high-quality, diverse data for six crop species (sugar beet, spring wheat, sweet corn, soybean, potato, and wheat-faba bean intercrop). You can access the data here and the code here.

4. Root zone soil moisture dataset

TAMSAT-SM provides a reliable, long-term dataset of root zone soil moisture across Africa, designed to support operational agricultural drought monitoring. Spanning from 1983 to the present at a 0.25° spatial resolution, it uses the JULES land surface model forced with TAMSAT rainfall and NCEP reanalysis data, tuned to satellite observations. You can access the data here and the code here (registration required).

Other useful bits

• A new Earth Engine tutorial shows how to replicate this pretty awesome GOES GeoColor visualisation using Python. By blending daytime imagery with night-time city lights and infrared cloud data, this guide helps users create seamless animations that beautifully capture the transition from day to night across the winter solstice.

• Planet is helping to illuminate the world's ‘dark fleets’ with their new Maritime Domain Awareness solution. By combining daily satellite imagery with AI detection, this tool tracks ships even when they switch off their AIS signals.

• Google Earth is testing a new experimental feature that lets users bring their own KML, KMZ, and GeoJSON files directly onto the globe. This update makes it easier than ever to securely layer custom data over Google’s satellite imagery, helping users make smarter decisions, faster.

• A new tool called GeoSpy AI is turning heads with its ability to pinpoint the exact location of a social media photo in just two seconds. It’s pretty freaky… It maps the results in full 3D, offering a powerful (if slightly unnerving) demonstration of just how much information can be gleaned from a single image.

Jobs

• Development Seed is looking for a remote Technical Project Lead

• FAO - Food and Agriculture Organization of the United Nations is looking for a Geospatial Data Scientist based in Rome, Italy

• WHO - World Health Organization is looking for a remote Epidemic Intelligence and GIS Analyst: https://unjobs.org/vacancies/1772554525336

• WFP - World Food Programme is looking for a remote Earth Observation Analyst

Just for Fun

A stunning detailed map of the world’s 1,180 micro-tectonic plates.

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 safe space for the geospatially curious. The aim is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• US Migration: New granular data reveals neighborhood-level movement trends.

• 3D Cities: Deep learning creates urban models from radar.

• Solar Forecasting: Explainable AI improves renewable energy yield predictions.

• Agricultural AI: Google’s AlphaEarth model tested against traditional metrics.

• Renewable Scenarios: New dataset closes gaps in distributed energy studies.

Subscribe now

Research you should know about

1. A new high-resolution lens on US migration

US migration data has essentially been stuck in a ‘pick one’ scenario: reliable but clunky Census data that only goes down to the county level, or high-resolution proprietary data that is full of biases. To fix this, researchers have developed MIGRATE, a new dataset that fuses commercial address histories from Infutor with the reliability of Census constraints. By harmonising these sources, they created annual migration matrices between 47.4 billion pairs of Census Block Groups. That’s roughly 4,600 times more granular than the standard county-to-county data we’re used to seeing, effectively letting us zoom in from a satellite view right down to the neighbourhood level.

The team spent a lot of time cleaning up the messy commercial data, which tended to overcount wealthy and rural populations while missing others. Once corrected, the national trends revealed a stark picture of ‘homophily’, where people overwhelmingly move to areas demographically similar to where they left. The data shows that movers from plurality Black, Asian and Hispanic neighbourhoods are 5.5, 14.6 and 4.3 times more likely, respectively, to move to areas with the same plurality race group. Upward mobility is also uneven; while people generally aim for higher-income neighbourhoods, there is a clear racial gap. Even when starting from the same income level, movers from plurality Asian areas are more likely to upgrade to a wealthier block, whereas movers from plurality Black areas are less likely to do so.

This granular view proves its worth when looking at localised shocks like the California wildfires. Public county-level data completely missed the migration impact of the 2017 Tubbs and 2018 Camp fires because the displacement happened within county lines or was washed out by broader averages. In contrast, MIGRATE picked up a massive signal, showing that out-migration rates in the distinct burn zones spiked to over 37% and 46% respectively in the year following the fires. It turns out that a huge chunk of these displaced people didn’t leave the county, but just moved to a safer block nearby, a dynamic that remains invisible unless you have this kind of street-level resolution.

MIGRATE is available upon request here. You can also find the code here.

2. Building 3D cities from messy radar data

Getting accurate 3D models of cities is usually a bit of a headache involving expensive LiDAR flights or waiting for cloud-free satellite imagery. Synthetic Aperture Radar (SAR) offers an interesting alternative since it works day and night in any weather, and a technique called TomoSAR can even reconstruct 3D point clouds from space. The problem is that these point clouds are often pretty messy. They suffer from noise and gaps that make it tough to extract clean building heights. To fix this, researchers from the Technical University of Munich have developed a new deep learning framework that turns these raw, rough radar points into high-res height maps.

The main part of their approach is a ‘dual-topology’ network that toggles between two different ways of looking at the data. One branch processes the irregular 3D points to capture structural details, while a second branch projects everything onto a 2D grid to enforce spatial consistency. By constantly exchanging information between these two views, the model effectively learns to denoise the data and fill in the missing blanks where the radar signal didn’t bounce back. This allows it to generate continuous height estimates even from sparse inputs, without needing an external terrain model to tell it where the ground is.

The team tested this on TerraSAR-X data over Berlin and Munich, representing both high-quality and lower-quality datasets. The results showed a massive improvement over standard interpolation methods, successfully reconstructing building shapes that were previously lost in the noise. They also demonstrated that the framework is flexible enough to ingest standard optical satellite imagery alongside the radar data, which further sharpens the results. This is a pretty big step forward as it serves as a proof of concept for generating large-scale, weather-independent urban height maps directly from space.

You can access the code here.

3. Solar energy forecasting gets a transparent upgrade

Solar energy is great, but it can be a bit fickle. For example, clouds and local climate shifts can sometimes make it hard to integrate into the grid reliably. Instead of installing expensive sensors everywhere, researchers are turning to machine learning to predict solar irradiance based on weather data. A new study from Scientific Reports compared six different regression models to see which could best handle the job using data on temperature, humidity, wind and pressure.

The clear winner in their tests was the Random Forest Regressor, which edged out competitors like Gradient Boosting and Linear Regression with a regression score of roughly 0.90. But the researchers didn’t just want a number; they wanted to know why the model worked. To peek inside the ‘black box’ of the AI, they employed Explainable AI (XAI) techniques known as LIME and SHAP. These tools help interpret complex algorithms, breaking down exactly how much weight the model assigns to different variables when making a prediction.

The XAI analysis revealed that temperature is the biggest factor, showing the strongest positive correlation with solar radiation. In contrast, humidity and wind speed tended to have a negative impact. This insight is valuable because it allows engineers to forecast potential energy yields for new solar farm locations using readily available geospatial data, rather than deploying costly hardware first. By making these AI decisions transparent, grid operators can trust the forecasts more, leading to better stability and smarter investment decisions in renewable infrastructure.

You can access the data here.

4. Google’s giant AI model meets the farm

As I’ve previously covered on here, last year Google DeepMind released AlphaEarth Foundation (AEF), a massive model pre-trained on billions of images that harmonises optical, radar and climate data into ready-to-use ‘embeddings’. A new paper puts AEF to the test on specific agricultural tasks in the US, such as predicting crop yields and mapping tillage practices, to see if these pre-cooked features can actually beat traditional, hands-on remote sensing methods.

The results were pretty encouraging at a local level. When trained and tested in the same region, models built on AEF embeddings performed just as well as those built on carefully engineered remote sensing features. For tasks like estimating corn and soybean yields, the AI approach was competitive and offered a huge convenience advantage. Since AEF automatically handles data harmonization and fills in gaps caused by clouds, researchers didn’t have to spend weeks cleaning data to get a usable input.

However, the study uncovered some serious limitations when the models were pushed further. AEF struggled quite a bit with ‘spatial transferability’, which means a model trained in the eastern Corn Belt failed completely when tested in the west. Traditional satellite metrics handled this shift much better because they are based on consistent physical properties of plants. Additionally, AEF is a ‘black box’, so users have no idea what a feature like ‘A05’ actually represents physically, unlike an index like NDVI which directly measures vegetation vigour. This lack of interpretability suggests that while foundation models are powerful, we can’t quite ditch our physics-based indices just yet.

Geospatial Datasets

1. Distributed renewable energy scenarios dataset

The DDRE-33 dataset addresses the scarcity of public data for distributed renewable energy (DRE) studies within the standard IEEE-33 distribution network. This provide spatiotemporally correlated scenarios for both distributed wind and photovoltaic generation without the data gaps found in earlier collections. You can access the data here and the code here.

2. Riverine levee overtopping events dataset

The Levee Loading and Incident Dataset - Overtopping (LLID-OT) version 1.0 provides data on 487 riverine levee overtopping events across the US from the past 15 years. You can access the data here.

3. Electric distribution networks dataset

This harmonized geospatial dataset provides data on electric distribution networks across the US Northeast, covering seven states from Connecticut to Vermont. It features over 3.8 million line segments. You can access the data here and the code here.

4. LiDAR benchmark for railway corridors

SemanticRail3D and its enhanced V2 iteration provide a comprehensive mobile LiDAR benchmark designed specifically for railway infrastructure monitoring. It features 2 semantic classes, including rails, catenary systems, and masts and is comprised of approximately 2.8 billion points across 438 high-resolution segments. You can access SemanticRail3D here andSemanticRail3D-V2 here. You can access the code here.

Other useful bits

• Google DeepMind is powering a new agricultural layer in Google Earth that uses machine learning to map individual field boundaries across the Asia-Pacific region. This essentially allows users to pinpoint water resources and calculate acreage with precision, offering vital data for drought contingency planning.

• Researchers have introduced Embedded Seamless Data (ESD), an ultra-lightweight database that compresses twenty-five years of global 30-metre imagery into a surprisingly manageable size. With a massive 340x compression ratio, it allows the entire global land surface for a single year to fit on a standard workstation, democratising planetary-scale analysis. Check out the demo and the manuscript.

• Google has launched a handy new endpoint for its Weather API, giving developers a single, reliable source for global weather alerts. By unifying data from government agencies worldwide, it takes the hassle out of aggregating warnings and automatically surfaces local alerts based on simple coordinates.
  

Jobs

• FAO - Food and Agriculture Organization of the United Nations is looking for a Junior Agricultural Specialist based in Rome, Italy

• UNEP - United Nations Environment Programme is looking for a Information Management Assistant based in Nairobi, Kenya

• Zipline is looking for a Senior Analyst based in South San Francisco, USA

• FAO - Food and Agriculture Organization of the United Nations is looking for a remote Data Analyst

• OICT - Office of Information and Communications Technology is looking for a Chief of Service, Information Systems and Telecommunications based in New York, USA

Just for Fun

Those spectacular auroras seen in mid-January might have been a treat for stargazers, but they served as a stark warning for the upcoming Artemis II mission. With the launch just weeks away, the European Space Agency is using data from these severe solar storms to ensure the crew’s journey to the Moon remains safe from dangerous radiation.

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 weekly round-up of geospatial news. We’re a bit like Tensorflow. Some people still think we’re kinda cool, but nobody else is interested. In any case, the aim is to make you a better geospatial data scientist in less than 5 minutes a week.

In today’s newsletter:

• Urban Road Access: New metrics map road deprivation in Africa.

• AI Flood Mapping: Fine-tuned models improve disaster detection accuracy.

• Wildfire Forecasting: Global graph networks predict fires months ahead.

• Soil Health Sensing: Satellite data estimates carbon and nitrogen levels.

• Renewable Energy Data: Hourly climate projections for future power planning.

Subscribe now

Research you should know about

1. It’s not the distance to the road. It’s the dirt.

Access to motorable roads is a massive issue in rapidly urbanising regions like Sub-Saharan Africa, where informal settlements often get left off the map. A new study tries to address this by proposing a new way of modelling ‘road access deprivation’ in Nairobi, Lagos and Kano. Instead of just measuring distance, they combined two clever metrics. First, they developed a distance-agnostic accessibility measure that counts how many buildings obstruct the path between a structure and the nearest road. Second, they integrated data on road surface type to distinguish between paved and unpaved infrastructure. This allowed them to classify areas into low, medium and high deprivation levels based on whether residents have direct access to a paved road, a dirt track or are physically cut off from the network.

When they ran this model across the three cities using open geospatial data, the results showed that highly deprived areas are actually less common than you might expect. The majority of built-up zones fell into the low or medium deprivation categories, meaning most buildings do have direct access to some form of motorable road. However, the extent of severe disconnection varies quite a bit between locations. For instance, high deprivation areas made up only 11.8 per cent of Nairobi but climbed to 27.7 per cent in Kano. This suggests that while total physical isolation is a specific problem, a large chunk of the urban population is dealing with the ‘medium’ deprivation of having direct access but only to poor-quality, unpaved surfaces.

To see if these maps reflected reality, the team validated their model against data crowdsourced from local communities. The system proved highly effective at identifying well-connected, low-deprivation areas, achieving strong accuracy scores in Nairobi and Lagos. It struggled a bit more with the ‘medium’ category, often confusing it with low deprivation. This usually happened because the underlying data might misclassify a road’s surface type or because community members perceive access differently than the rigid definitions of the model. Despite these hiccups, the study shows that combining building obstruction metrics with road quality data is a promising, scalable way to identify disconnected neighbourhoods and help city planners prioritise upgrades in data-scarce environments.

2. Improving flood maps with heavy-duty AI

Flooding is causing havoc worldwide, with 2024 seeing devastating events from Valencia to Vietnam. While we have massive archives of Earth observation data, using it effectively is tough because standard AI models need huge amounts of labelled data. To tackle this, researchers took an existing Geospatial Foundation Model (GFM) called TerraMind and fine-tuned it specifically for flood detection. They used a new dataset called FloodsNet, which combines both Sentinel-1 radar and Sentinel-2 optical imagery for 85 global flood events. The goal was to see if retraining this pre-existing ‘brain’ with a rich, multimodal dataset could handle the messiness of flood data better than current methods.

The study ran experiments on four different versions of the model, tweaking the size and whether the core ‘backbone’ was frozen or allowed to update. They found that the ‘base unfrozen’ configuration offered the best balance, delivering strong accuracy without the heavy computational price tag of the larger models. While the massive 326-million-parameter version achieved the best recall among the GFM options, it didn’t offer enough of an overall performance boost to justify the extra cost. It seems that for practical flood mapping, a mid-sized model that can adapt its internal weights is the most efficient choice.

Things got interesting when they compared these GFMs against a standard U-Net, a conventional deep learning model that has been around for years. Surprisingly, the U-Net achieved better recall than every single TerraMind configuration, meaning it missed fewer flooded pixels overall. While the Foundation Models produced smoother and more coherent maps, the U-Net was often sharper at picking up fine details. This suggests that while multimodal GFMs are promising for unlocking vast data archives, traditional supervised models are still holding their ground when it comes to catching every last bit of water during a disaster.

3. Predicting wildfires with Earth-as-a-graph

Climate change is turning wildfire forecasting into a global priority. Traditional models often struggle to capture the complex, non-linear interactions between weather, vegetation, and human activity that drive fires over long periods. To address this, a team of researchers has developed FireCastNet, a new deep learning model that treats the Earth system as a giant interconnected graph. By combining 3D convolutional encoding with a Graph Neural Network (GNN) based on the GraphCast weather model, FireCastNet learns from both local fire dynamics and distant climate ‘teleconnections’, like how ocean temperatures in the Pacific might influence fire risk in Africa months later.

Trained on the comprehensive SeasFire datacube, which includes climate, vegetation, and human variables from 2001 to 2021, the model predicts burned area patterns up to six months in advance. When benchmarked against other state-of-the-art AI models, FireCastNet came out on top, particularly in fire-prone regions like Africa, South America, and Southeast Asia. The study found that giving the model longer time-series data significantly boosted its ability to make robust long-term predictions, confirming that fires are often the result of slowly accumulating conditions like drought and fuel buildup.

Interestingly, while a single global model performed well, the researchers also tested a ‘Local Area Modeling’ approach. This involved training specific models for different biomes, refining the graph’s mesh to be ultra-high resolution over a target region while keeping the rest of the globe coarser to capture those important long-range links. This local tuning improved accuracy in places like Europe and the Middle East but surprisingly didn’t help everywhere. In regions like Southeast Asia and Australia, the global model still reigned supreme, suggesting that for some parts of the world, those broad, planetary-scale connections are just as critical as local details.

4. Measuring soil from the sky

Precision agriculture relies heavily on accurate soil data, but traditional physical sampling is slow, expensive, and logistically difficult. A team from CO2 Angels and the European Space Agency has developed a scalable system to estimate soil properties like organic carbon and nitrogen directly from satellite data. They used a hybrid modelling approach that combines direct spectral observations with environmental drivers such as climate, terrain, and vegetation traits. By integrating physics-informed machine learning with foundation model embeddings, the system can infer underground soil health conditions even when the surface is covered by crops.

The researchers trained their model on a massive harmonised dataset of European croplands, incorporating data from Sentinel-2, Landsat, and MODIS. They applied a rigorous validation method known as spatial blocking, where test samples were separated from training data by blocks of approximately 100 kilometres. This strict separation ensures the model isn’t just memorising local geographic patterns but is genuinely learning to predict soil properties from environmental signals.

The performance varied significantly depending on the nutrient being analysed. The model proved highly effective for Soil Organic Carbon and Nitrogen, achieving a Concordance Correlation Coefficient of 0.77 for both, with a Mean Absolute Error of 5.12 g/kg for carbon and 0.44 g/kg for nitrogen. However, it found Phosphorus and Potassium much harder to pin down, with correlation coefficients dropping to 0.53 and 0.58 respectively. The system also successfully quantified its own uncertainty, achieving 90% coverage for its prediction intervals, which suggests it is a viable tool for carbon markets and broad-scale monitoring even if it isn’t quite ready for precise phosphorus management.

Geospatial Datasets

1. Climate projections and renewable energy generation dataset

This dataset addresses the critical need for high-resolution energy modelling by providing hourly climate projections and renewable energy generation data for mainland China. Spanning from 2021 to 2060, it uses analog-based temporal downscaling to refine daily Global Climate Model outputs into the hourly resolution required for accurate power system planning. You can access the data from the Zenodo projects cited (reference no. 41 to 60) and the code here.

2. Flood hazard mapping framework

This study presents a hybrid framework for flood hazard assessment in the Swat River Basin. It blends machine learning with coupled hydrologic–hydraulic modelling.

3. Soil Moisture Climatological Record

The SMCR provides the first global, satellite-derived soil moisture record that spans from 1980 to 2023 at a daily, 1 km resolution.

Access the dataset:

• 1980–1985

• 1986–1991

• 1992–1997

• 1998–2023

Access the code here.

4. Railway network dataset

The Italian Railway Network Dataset offers a comprehensive look at multi-type train operations across Italy, spanning from January to June 2024. It covers over 3,300 trains and nearly 1,400 stations. You can access the dataset here and the processing code here.

Other useful bits

• By analysing how ice flows over the bedrock, researchers have produced the most detailed map yet of Antarctica’s hidden ‘underbelly’, revealing thousands of previously unknown hills and valleys. This is a crucial leap for climate science, as knowing the shape of this buried landscape is essential for predicting exactly how the ice sheet will react to a warming world.

• Vantor has slashed satellite imagery delivery times from 90 minutes down to just 11. By processing data right at the edge rather than in central queues, their new approach ensures vital intelligence reaches the people who need it faster than ever before.

• ESRI founder Jack Dangermond envisions a ‘planetary nervous system’ where digital twins act as living, evolving entities rather than static snapshots. He argues that the industry’s next great leap depends less on new software and more on a cultural shift towards sharing data to solve global challenges.

• FAO’s Hand-in-Hand Initiative has released a new ‘suite’ of geospatial datasets aimed at transforming agriculture and rural development across more than 80 countries. By using advanced analytics to pinpoint high-impact investment opportunities, this project provides a vital roadmap for fighting global hunger through smarter, data-driven planning.

Jobs

• The World Bank is looking for a Senior Technical Lead for Geospatial Solutions based in Bulgaria.

• Genex Systems is looking for a GIS Analyst based in Mountain View.

• ESA is looking for a Graduate Trainee in Earth Observation Foundation Modelling based in Frascati.

• The Lincoln Institute of Land Policy is looking for a remote Geospatial Backend Software Developer based in the US.

• ETH Swiss GeoLab is looking for an Office Manager based in Zurich (and eventually, Lucerne).

Just for Fun

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 you can say things like ‘cloud native’ and people nod with agreement. The aim is to make you a better geospatial data scientist in just 5 minutes a week.

In today’s newsletter:

• Urban Access: Data maps the 15-minute city reality.

• Heat Resilience: Better health masks rising climate risks.

• Cloud Mapping: Ground cameras fix weather model blind spots.

• Ecosystem Resilience: ENSO weakens our greening world.

• Satellite Imaging: Sentinel-2A captures Earth in night mode.

Subscribe now

Research you should know about

1. Mapping the reality of the 15-minute city

The concept of the ‘15-minute city’ has become a bit of a buzzword in urban planning, yet we’ve had surprisingly little data on how many people actually live in one on a global scale. A new study published in Nature Communications tackles this data gap by mapping global accessibility to six essential categories, including healthcare, education and public transit. By combining Overture Maps POI data with high-resolution land cover and elevation models, the researchers created a ‘friction surface’ map. This basically measures the time it takes to travel across different types of terrain to reach these services.

To pull this off, the team integrated nearly 66 million points of interest from Overture Maps and Amap (for China) with population density data. They calculated travel times using a least-cost-path algorithm, which figures out the quickest route based on terrain and road networks. The results show that roughly 62.8% of the world’s urban population currently lives within a 15-minute walk of essential amenities. If you extend that radius to a 30-minute walk, the coverage jumps to 82.5%. While these numbers might sound high, they hide some pretty stark inequalities when you break the data down by region and income level.

The data reveals a significant divide between the Global North and South. Residents in high-income nations enjoy an average travel time of 15.5 minutes to essentials, whereas those in the Global South face an average trek of 30.8 minutes. The disparity in infrastructure density is even wider, with high-income countries having nearly nine times the number of amenities per capita compared to low-income nations. This confirms that wealth heavily dictates resource allocation, and it highlights just how much work remains to achieve the UN’s Sustainable Development Goals regarding equitable access to basic services.

2. Better health is hiding the impact of heat

Climate change is driving up temperatures, so I would’ve expected heat-related deaths to be skyrocketing. But a new study looking at 15 major German cities between 1993 and 2022 found something unexpected. Despite the summers getting hotter, the actual number of heat-related deaths has slightly declined. The researchers discovered that this is largely down to a drop in vulnerability. Basically, improvements in life expectancy and general health have acted as a buffer, protecting the population from the worst effects of rising temperatures.

To figure this out, the team used a clever method involving ‘counterfactual’ scenarios. They crunched the numbers to see what would have happened if life expectancy had stayed stuck at 1993 levels. The results were pretty interesting. Without those improvements in health and longevity, climate change would have driven a sharp rise in heat-related deaths. It’s a bit like running up a down escalator; we are getting healthier fast enough to outpace the rising heat, at least for now.

Of course, this doesn’t mean we are off the hook. The study found that the fingerprint of climate change on the deaths that do occur is getting stronger. The proportion of heat deaths directly attributable to climate change has been rising by about 5.6% every decade. In extreme years like the 2003 heatwave, climate change was responsible for a massive chunk of the mortality burden. This suggests that while better health is masking the problem, the underlying threat from a warming climate is intensifying.

3. Ground-based cameras are fixing the blind spot in weather models

We rely heavily on weather models for everything from routing flights to predicting renewable energy output, but these models have a significant blind spot: small, shallow cumulus clouds. These clouds cover a massive portion of the Earth’s surface, yet they are too small and short-lived for standard global models to simulate directly. The usual fix involves using approximations, but these introduce errors. Getting better data is tough because high-resolution satellites might only revisit a spot every few days, and radar systems only see a tiny slice of the sky. To solve this, researchers from Oxford and Imperial College London have developed Cloud4D, a framework that uses simple ground-based cameras to reconstruct clouds in four dimensions.

The system works by using a network of six synchronised cameras pointed skyward to capture the same cloud field from different angles. It applies a ‘homography-guided’ approach, which exploits the fact that clouds naturally form in layers, to map these 2D images into a 3D structure. A transformer model then refines this data to estimate the cloud’s liquid water content on a 25-metre grid. Unlike satellite data which can be days old, this system updates every five seconds 🤯. By tracking how these 3D reconstructions move over time, the model can even calculate horizontal wind vectors at different altitudes.

During a two-month real-world deployment, the system proved to be super effective. Cloud4D delivered an order-of-magnitude improvement in space-time resolution compared to state-of-the-art satellite products. Crucially, when the team cross-referenced their results with expensive collocated radar measurements, the camera-based system maintained a relative error of less than 10 per cent. This indicates that arrays of relatively cheap cameras could be a scalable way to gather the high-frequency data needed to validate and improve the next generation of climate and weather models.

4. The hidden fragility of a greener world

We often hear that the world is getting greener due to CO2 fertilisation, but a new study in Nature Communications suggests this greening hides a deeper problem. By analysing decades of satellite data, researchers found that while vegetation cover is increasing, its ‘resilience’ is actually dropping across over 40% of vegetated land. ‘Resilience’ is pretty much how quickly land can bounce back from a shock.

It turns out that a massive driver of this hidden instability is the El Niño-Southern Oscillation (ENSO), the major climate cycle that warms and cools the tropical Pacific. The researchers discovered that ENSO significantly impacts the resilience of 53% of the global vegetated area, meaning this single climate phenomenon is dictating the health of forests and grasslands on a massive scale.

The study breaks down exactly how ENSO messes with plants, identifying two main pathways:

1. ‘Atmospheric synchronisation’, where large-scale air circulation patterns directly link ENSO events to plant health.

2. The ‘climate-mediated’ pathway, where ENSO triggers local anomalies like heatwaves, droughts or floods that stress plants out.

The researchers found that El Niño events tend to cause sharp, immediate drops in resilience due to high heat and water stress, particularly in tropical forests. La Niña isn’t harmless either, though its impacts are sneaky; they often lag behind, hitting ecosystems hard up to three years after the event starts.

Looking ahead, the situation looks likely to intensify. Using climate models to project through to 2100, the team predicts that the footprint of ENSO-driven climate anomalies will expand by another 7 to 10%. This means regions that were previously relatively safe from these tropical climate swings, such as Eastern Siberia and the northern parts of North America, are going to start feeling the pressure. This expansion suggests we can’t just rely on planting trees to solve climate issues; we need to ensure the forests we have are actually tough enough to survive the increasingly volatile pendulum of global climate cycles.

Geospatial Datasets

1. Downscaled CMIP6 simulations

The Ouranos CRCM5-CMIP6 ensemble contributes to the NA-CORDEX initiative by providing dynamically downscaled climate simulations over North America. Driven by four global models, this project provides high-resolution 0.11° (approx. 12 km) projections spanning from 1950 to 2100 across various emission scenarios. You can access the data here and the code here.

2. Marine heatwave dataset

Capturing the unprecedented 2023 Marine Heatwave in the North Atlantic, this dataset features high-resolution observations from a drifting Spotter buoy off the west coast of Ireland. You can access the data here, the spectral code here, and the wind estimation tools here.

3. Shoreline change drone data

This dataset offers a seasonally structured record of shoreline change at Sugarloaf Island, North Carolina, tracking the evolution of a hybrid living shoreline project from 2023 to 2025. You can access the data here.

4. High-res CO₂ and pollutant emission data

The CHETNA-Road dataset offers a vital resource for understanding urban air quality in India, delivering 500 m gridded data on CO₂ and pollutant emissions across 15 major cities. You can access the data here and the code here.

Other useful bits

• ESA has successfully switched the ten-year-old Sentinel-2A satellite into night mode, capturing unexpected views of wildfires and fishing boats. These tests are providing vital insights for the Sentinel-2 Next Generation mission, ensuring future satellites can monitor the Earth around the clock.

• A new case study from Zambia highlights the power of open-source geospatial tools and free satellite data in mapping complex drought risks. By combining hazard indicators with vulnerability metrics, this report offers a practical blueprint for using accessible technology to strengthen climate resilience and early warning systems.

• Globe Telecom is breaking new ground in Southeast Asia by rolling out Starlink’s Direct to Cell service for everyday mobile phones. By acting as a ‘cell tower in space’, this partnership will bring vital connectivity to remote Philippine islands and ensure a lifeline remains open when terrestrial networks fail.

• 2026 is shaping up to be a landmark year for space exploration, featuring everything from new flagship telescopes to historic lunar missions. With NASA’s Artemis II returning humans to the Moon’s orbit and international players like China, India, and Japan launching ambitious projects, the year promises a thrilling mix of global cooperation and competition in the cosmos.

Jobs

• ICARDA - International Center for Agricultural Research in the Dry Areas is looking for a Data Scientist based in Abu Dhabi, UAE.

• UNODC - Office on Drugs and Crime is looking for a Data Visualisation Consultant based in Vienna, Austria.

• ESCAP - Economic and Social Commission for Asia and the Pacific is looking for an Intern - Economic Affairs based in Bangkok, Thailand

Just for Fun

Open Infrastructure Map reveals the hidden skeleton of our world, visualising the power lines and telecom networks that are usually invisible on the standard OpenStreetMap. It’s a pretty interesting open-source project that brings these global layers to light for everyone to explore.

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 geospatial newsletter that’s hotter than a Nelly song… 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:

• Maritime Risks: Narrow lanes face $192 billion in trade risks.

• Extreme Rainfall: Local history drives risks for a billion people.

• Simple Models: RNNs outperform complex setups for African rainfall.

• Ecosystem Loss: Agriculture threatens vital grasslands and non-forest landscapes.

• Building Data: High-resolution building characteristics for the Global South.

Subscribe now

Research you should know about

1. The massive $192 billion trade risk hiding in our shipping lanes

Global trade relies on a handful of narrow passageways, known as maritime chokepoints, which are surprisingly vulnerable to a bunch of different hazards. A new study in Nature Communications finds that about $192 billion of trade is exposed to disruptions every year from things like cyclones, droughts, and geopolitical conflict. While we often think of these as local problems, the effects are actually systemic. Many countries are heavily dependent on multiple chokepoints without even realising it, particularly those in Central Asia or West Africa that might be thousands of kilometres away from the actual trouble.

To figure out the actual risk, the team calculated the Expected Value of Trade Disrupted by looking at how likely a hazard is, how long a closure would last, and how much trade would be blocked. They also validated their transport model against real-world vessel movements from the PortWatch platform, which uses AIS data to track ships. This approach accounts for whether there’s a way to go around a blockage or if ships are just stuck. They even looked at how joint disruptions might happen, such as when a single cyclone hits multiple chokepoints at once.

Altogether, the annual economic bill for these disruptions is roughly $10.7 billion, with an extra $3.4 billion tacked on from spiked freight rates. These price jumps happen because rerouting ships pulls capacity out of the global fleet, which makes shipping more expensive for everyone everywhere. Geopolitical risks at the Taiwan Strait and the Suez Canal are the primary drivers of these losses. Even if a country isn’t near a conflict zone, it might still see higher inflation and pricier goods because of the way these maritime links are all stitched together.

You can access the data and code here.

2. We’re underestimating the next record-shattering rain event

We often think of extreme rainfall through the lens of a steady climate trend, but a new study in Nature Communications shows that local history is a much bigger driver of near-term disasters. The researchers found that record-breaking potential depends on a bunch of factors, but the size of the standing record is a pretty big one. If a region has a low historical record just because of natural luck, it’s in a much more dangerous spot for a surprise jump in the near future. This path-dependency means that our local experience of the past is just as important as the global models we use to predict the future.

The way climate change interacts with these records is pretty surprising because it doesn’t hit every place the same way. On average, moderate warming is set to hike up the chance of a record-breaking event by about 40% by 2050. However, the most sensitive areas are actually the ones with high standing records. These regions see the steepest increase in risk, with a 75% jump in the likelihood of a record-shattering event. It turns out that while low-record areas are most at risk right now, the high-record areas are becoming less safe at the fastest rate.

There’s also a human element involved in how these records turn into full-blown disasters. When a place hasn’t seen a record-breaking storm in a long time, the local society often loses its sense of risk, which leads to a massive gap in preparedness. This is a major threat for about a billion people who are likely to face record-shattering rainfall by the end of the century. Many of these hotspots are in fast-growing urban areas where poverty and rapid urbanisation make the population even more vulnerable to the next big surprise.

You can access the code here.

3. Complex models aren’t always better for rainfall

I've been looking at how we predict rainfall in Africa lately, as it's a pretty big factor for food security across the continent. A new paper in Scientific Reports examines the use of deep learning for daily forecasts in cities like Nairobi and Abuja. The researchers were analysing over 40 years of NASA’s MERRA-2 data, which is a popular reanalysis dataset for this kind of work, to see if they could improve on traditional stats by comparing single models like Recurrent Neural Networks (RNNs) against more complex hybrid setups.

The results are a bit of a reality check for anyone who thinks more complex is always better. Even though hybrid models are theoretically more powerful, the study found that single models actually did a better job in most spots. Specifically, the RNN came out on top in Nairobi, Pretoria, and Rabat (this was news to me, since I thought the RNN was all but dead…). There are a bunch of reasons for this, but it’s mostly because simpler models are more reliable, have fewer parameters, and don’t overfit the noisy weather data as easily. The only place where a hybrid model took the lead was in Abuja, which suggests that extra complexity only pays off when the rainfall patterns are particularly wild.

There’s also a bunch of interesting spatial stuff happening here. The researchers found that rainfall patterns between these cities aren’t really correlated, meaning local atmospheric conditions are driving the bus rather than one big continental system. Relative humidity was the biggest predictor in most spots, which I guess makes sense for the tropics. However, in drier places like Rabat, the weather from the previous two days was the most important factor. This highlights how important it is to use location-specific modelling because what works for one African city can’t just be copy-pasted onto another (something we’ve all been guilty of)…

4. Why our focus on deforestation is missing half the picture

While the world’s eyes are usually glued to deforestation rates in the Amazon, a new study shows that agricultural expansion is impacting non-forest ecosystems just as much. Grasslands, shrublands, and wetlands are often overlooked in global conservation talks, yet they store huge amounts of carbon and host a massive chunk of the planet’s biodiversity. By analysing land cover changes from 2000 to 2020, researchers found that the conversion of these open landscapes into croplands and pastures is often more extensive than the loss of actual forests. This is a pretty big deal because these areas are typically less protected and are more vulnerable to being ploughed up for farming.

There’s a bunch of reasons why this study stands out, but the main one is how it handles satellite data. The team used three different high-resolution datasets to track where these changes are happening, which is important because different models have different ways of defining a forest or a wetland. For example, some datasets might count tree crops like oil palm as forest while others see them as agriculture, which can lead to messy results. By comparing the data they analysed, they identified a consistent group of global hotspots where non-forest land is disappearing fast. Brazil, the United States, China, Russia, and Argentina topped the list, making up over 60 per cent of all non-forest conversion worldwide.

The environmental cost of this shift is massive. The study found that habitats for over 5,000 threatened species have been affected by this expansion, and over half of these creatures rely specifically on non-forest ecosystems. Even within supposedly protected areas, conversion rates for grasslands and wetlands are actually increasing. This suggests that current conservation frameworks are failing to shield these vital landscapes, particularly in places where agricultural suitability makes them a prime target for new farms. As supply chains come under more scrutiny, focusing solely on trees will likely leave some of the world’s most carbon-rich and diverse landscapes at risk.

You can access the code here.

Geospatial Datasets

1. Building characteristics dataset

This dataset derives high-res (100 m) annual building characteristics for the Global South from 2016 to 2023, based on the Google Open Buildings Temporal dataset. It provides six spatial layers per year, including building count, area, volume, height variability, and proximity metrics. You can access the data here.

2. Global vegetation dataset

VegVault is a huge, versioned SQLite database (~110 GB) that brings together global vegetation data across both the present day and the last ~20,000 years. You can access the data here and the code here.

3. City and town boundaries dataset

A new Global City and Town Boundaries (GCTB) dataset maps annual urban settlement boundaries worldwide from 2000–2022 at 30 m resolution. You can access the data here and the code here.

4. Cartographic map benchmark dataset

CartoMapQA is a new benchmark dataset for testing how well vision–language models can actually understand cartographic maps, not just talk about them. It includes 2,251 QA samples across 853 OpenStreetMap-rendered maps, covering a neat hierarchy of tasks from symbol/feature semantics and feature counting/name extraction through to scale-bar distance estimation and turn-by-turn shortest-route navigation. You can access the data and code here.

Other useful bits

• This handy Python–Flask tool lets you draw a box on a map and download OpenStreetMap or satellite tiles for offline use, exporting them as a ZIP or MBTiles file. It’s ideal for personal or educational projects, with features like tile count previews, live map previews, and flexible zoom levels.

• The Library of Congress Geography and Map Division has released its 2025 digital scan highlights, featuring notable 18th-century Dutch wall maps and a rare hand-drawn Islamic world map. The collection also includes significant historical documents such as Albert Speer's 1942 Berlin redevelopment plan and various newly digitized panoramic "bird's-eye-view" city maps.

• As satellite launches are projected to reach 1.7 million by 2030, astronomers are urging UN bodies to establish international standards to mitigate light and radio interference. While some operators are testing non-reflective coatings and signal-steering, the sheer volume of planned megaconstellations threatens to permanently degrade both optical and radio observations.

• Scientists have used Parker Solar Probe data to map the Alfvén critical surface, the "spiky" boundary where solar plasma becomes solar wind. These maps show the surface expands and becomes more irregular during solar maximum, offering a new tool for predicting how solar storms might affect Earth.

Jobs

• The Alan Turing Institute is looking for a Senior Research Associate based in London under its Environment Forecasting Mission.

• The Gates Foundation is looking for a Senior Strategy Officer, Geospatial Insights (LTE) based in Seattle under its Global Development Division.

• UNOCC is looking for a Geographical Information Systems Intern under its Data Unit (offering a flexible work arrangement).

Just for Fun

Check out this "Earthrise" footage captured by the Japanese lunar orbiter, SELENE (Kaguya).

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 — think of us as a safe space for the EPSG:4326-curious. The aim is to make you a better geospatial data scientist in less than five minutes a week.

In today’s newsletter:

• Damage Mapping: Sentinel data maps disaster damage effectively.

• Road Safety: Probabilistic model predicts crash risk uncertainty.

• Human Mobility: Satellite imagery predicts urban movement flows.

• Renewable Energy: Microsoft maps global solar and wind sites.

• Fire Emissions: Global database tracks landscape fire emissions.

Subscribe now

Research you should know about

1. Predicting human mobility flows from satellite imagery

Understanding how people move around cities is obviously quite important for transport systems and urban planning. Traditionally, we’ve relied on expensive surveys or mobile phone data to track these mobility flows. However, surveys are costly to update and mobile data comes with privacy concerns. A new study in Nature Communications proposes a different approach called Imagery2Flow. It uses a deep learning model to predict human mobility flows directly from satellite imagery. The authors tested this on the top 10 largest metropolitan areas in the US to see if the built environment alone could reveal how populations move.

The researchers developed a framework that extracts visual features from medium-resolution satellite imagery like Sentinel-2 and Landsat-8. The model encodes these images into high-dimensional vectors and uses a Graph Attention Network (GAT) to understand the spatial interactions between different areas. Essentially, it learns to predict the number of people moving between locations by analysing the visual characteristics of the urban landscape. It looks at the texture and arrangement of urban features to infer movement patterns without needing direct tracking data of individuals. This allows it to generate origin-destination matrices purely from open-source imagery.

The results were pretty impressive. The model outperformed traditional gravity and radiation models and even beat other deep learning baselines. Specifically, using 10-metre resolution imagery reduced the Root Mean Square Error by a range of 4.5% to 13.9% compared to machine learning models that just use visual features. Interestingly, the study found a strong link between the predicted mobility patterns and urban morphology. The model also showed good transferability. A model trained on data from 2020 could accurately predict flows using imagery from 2010. This suggests that the physical structure of a city holds deep clues about how we move through it.

You can access the code here.

2. Some surprising findings on resolution and damage mapping

Accessing Very High Resolution (VHR) imagery immediately after a disaster is often difficult or prohibitively expensive for humanitarian agencies. A new study from ETH Zurich investigates whether free, medium-resolution data from the Copernicus programme can effectively fill this gap. The researchers introduced xBD-S12, a massive dataset that pairs over 10,000 pre- and post-disaster image pairs from Sentinel-1 and Sentinel-2 with the established xBD benchmark. Their goal was to determine if these 10-metre resolution images could detect building damage across various disaster scenarios, ranging from floods to wildfires, without relying on commercial providers.

The results indicate that while resolution is a constraint, it isn’t a dealbreaker for every scenario. The models performed surprisingly well for large-scale events like wildfires and floods where spectral signatures change dramatically over wide areas. However, the 10-metre ground sampling distance struggled with sparse damage events, such as weak earthquakes where only individual buildings were affected amidst a dense urban fabric. The authors also found that architectural sophistication provided little advantage here. Simple U-Net models actually outperformed complex geospatial foundation models like Prithvi or DOFA, particularly when the models had to generalise to unseen disasters.

To achieve these results, the team adopted a two-step approach that first localises buildings and then classifies the damage. This method proved more robust for unseen locations compared to trying to do everything in a single joint optimisation step. Ultimately, this research suggests that while Sentinel data can’t quite replace VHR imagery for detailed forensic analysis, it serves as a powerful and accessible complement for rapid, wide-area assessments. It allows humanitarian teams to get a pretty big head start on understanding the extent of destruction when aerial or high-resolution satellite data is unavailable.

3. Estimating road crash risk and uncertainty from satellite imagery

Most models trying to predict car crash risk from satellite imagery have a pretty big flaw. They tend to be overconfident binary classifiers that just spit out a single “crash” or “no crash” prediction. A new study that’ll be presented at AAAI’26 proposes a smarter way to handle this using something called Beta Distribution Learning. By moving away from static yes or no answers, their model generates a full probability distribution for every location. This allows the system to express uncertainty, meaning it can tell you if a road looks genuinely safe or if the model just isn’t sure because the visual data is ambiguous.

The researchers achieved this by using a clever procedural labelling technique during training. When the model looks at a satellite image crop, it calculates an “influence score” based on how close a crash site is to the centre of the image and the size of the crop. This teaches the neural network that visual evidence is stronger when hazards are front and centre. Consequently, the model outputs two parameters (alpha and beta) that define a curve. A sharp peak in the curve means the model is highly confident in its risk assessment, while a flat, wide curve signals that the model is uncertain. It’s a lot more nuanced than a simple binary label and helps distinguish between safe roads and confusing ones.

Testing this on a dataset of Texas roadways showed that this probabilistic approach is much more reliable than standard methods. The model improved recall by around 20 per cent compared to state-of-the-art baselines, meaning it successfully identified many dangerous locations that other models missed. In a case study of the San Antonio River Walk, it correctly flagged a hazardous intersection near a parking garage that earlier models ignored. Even when the model made “mistakes” (like flagging a safe but complex highway interchange as risky), it did so with low confidence. This provides a helpful signal for human planners to investigate further rather than blindly trusting the machine.

Geospatial Datasets

1. Microsoft global renewables watch

Microsoft has released two global geospatial datasets mapping solar and wind energy installations worldwide, covering the period 2017 to Q2 2024. Together, they identify 375,197 individual wind turbines and 86,410 solar PV installations, each with estimated construction dates and associated land-use context.

2. Global fire emissions database

GFED5 is the latest version of the Global Fire Emissions Database, providing monthly global estimates of landscape fire emissions at 0.25° resolution from 2002–2022, with extensions back to 1997. You can access the data and code here.

3. Large reservoir operational variables dataset

GDROM v2 is a comprehensive open dataset covering daily inflow, release, storage time series, and operational rules for 2,017 large reservoirs across the contiguous United States. You can access the data and code here.

4. Potential light availability dataset

The SwissRad10 dataset provides nationwide, hourly maps of potential light availability across Switzerland at a 10 m resolution, covering a full annual solar cycle. You can access the data here and the code here.

Other useful bits

• Apple’s SHARP shows just how fast things are moving in 3D vision: generating sharp, photorealistic new views from a single image in under a second. It delivers big quality gains across benchmarks while dramatically cutting compute time, opening the door to more lightweight and scalable 3D scene reconstruction.

• Meta has released SAM Audio, a unified model that can isolate any sound from complex audio using text, visual, or time-span prompts— a big step forward for flexible audio understanding. The model, along with benchmarks and research resources, is now open to the community, opening up new creative and analytical possibilities across audio applications.

• Copernicus Sentinel-2 satellites are teaming up in early 2026, with Sentinel-2A and Sentinel-2B flying in tandem to image the same locations just 30 seconds apart, helping fine-tune radiometric calibration across the constellation. This short but carefully planned campaign will further boost data consistency and quality for everything from precision agriculture to methane monitoring.

• InSAR.dev has launched a new Sentinel-1 workflow on Google Colab, making burst-level InSAR preprocessing and analysis fast, portable, and genuinely reusable across platforms. With cloud-ready data, a GPU-accelerated phase unwrapper, and dramatic speed-ups (minutes instead of hours), it’s a big step toward making high-resolution elevation and deformation mapping accessible to everyone.

Jobs

• Department of Political and Peacebuilding Affairs-Department of Peace Operations-Shared Structure is looking for a Geographical Information Systems Intern based in New York, USA

• UNEP - United Nations Environment Programme is looking for an Intern for Policy and Analysis Unit based in Geneva, Switzerland

• Resources for the Future (RFF) is looking for a Summer Research Internship based in Washington, USA

• First Street is looking for a Lead Product Manager based in New York, USA

Just for Fun

This video shows how Earth rotates in space by stabilizing the Milky Way galaxy.

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