🌐 Humans vs machines: poverty mapping edition
PLUS: measuring the impact of cool roofs on heat-related deaths, predicting forest fires, estimating building damages in conflict zones, and more.
Hey guys, here’s this week’s edition of the Spatial Edge — a weekly round-up of geospatial news that you can digest in less than 5 minutes. You can think of us as a safe space for map fetishists.
In today’s newsletter:
Poverty Mapping: CNNs outperform human experts.
Cool Roofs and Solar Panels: Significantly reduce urban heat deaths.
Forest Fire Severity: Remote sensing and ML enhance forest fire prediction.
Building Damage Assessment: Deep learning in conflict zones estimate damages.
CloudSEN12+: New dataset for cloud detection.
Research you should know about
1. Humans vs machines: poverty mapping edition
A new study in Scientific Reports explores how convolutional neural networks (CNNs) outperform human experts in estimating poverty levels from satellite imagery. I’ve previously covered some of the ways we can use deep learning and satellite images to estimate poverty. However, this paper examines whether CNNs are better at estimating poverty than humans when looking at satellite images.
Researchers compared methods of poverty estimation using satellite images from Tanzania. Human experts ranked 608 clusters from the Tanzania Demographic and Health Survey (DHS) based on high-resolution images and identified features they thought indicated wealth, like building materials and road quality. These features were used in machine learning models such as logistic regression and random forests. Separately, a CNN was trained on medium-resolution Sentinel-2 images to estimate poverty without human-defined features.
The CNN achieved higher accuracy in predicting wealth levels with less bias. Human evaluations tended to cluster around average wealth levels, missing the poorest and wealthiest areas. The study also shows that CNNs can detect more subtle patterns in satellite imagery that us mere mortals may overlook.
How much more evidence do we need that machines are taking over?
2. Addressing heat-related deaths with cool roofs and solar panels
A new study published in Nature Cities looks at how ‘cool roofs’ and rooftop solar panels can lower urban temperatures and reduce heat-related deaths in London.
Researchers used urban climate models (WRF BEP-BEM) to simulate how installing reflective "cool roofs" and solar panels on all buildings would affect London's air temperature during the ‘hot summer’ of 2018. They used a bunch of geospatial data like:
meteorological data
detailed population figures
mortality records from the UK's Office for National Statistics
The models compared scenarios with all buildings having cool roofs, all with solar panels, and a baseline with standard roofs, assessing changes in near-surface air temperature and heat-related mortality.
The findings showed that cool roofs could reduce average urban temperatures by 0.8°C, potentially reducing heat-related deaths by 32%. Rooftop solar panels could lower temperatures by 0.3°C, reducing deaths by 12%, while generating up to 20 TWh of electricity.
All in all, these are pretty compelling results. Of course, we’ll need to understand the costs of installing this type of infrastructure.
3. Using remote sensing and ML for bushfire severity
A new study uses remote sensing data and machine learning to better understand bushfire risks in Australia and improve how we manage them. This is pretty important given these disasters are happening more often and with greater intensity.
The team analysed 12 years of satellite data from NASA's FIRMS and Landsat programs to assess fire severity. They calculated indexes like NDVI and NBR (which assesses burn severity), and incorporated topographical and climate factors such as elevation, slope, temperature, and precipitation. Using this data, they developed a predictive model (using XG Boost), achieving an accuracy of 86.13% in forecasting fire severity.
Some key takeaways include identifying regions that are especially at risk for severe bushfires, which can help focus firefighting efforts. The team also suggest creating a UAV (i.e. drone) coordination model to improve real-time fire prediction and response.
4. Assessing building damage using deep learning
A new study looks at using deep learning to evaluate building damage in conflict zones, focussing on Mariupol, Ukraine. Quick damage assessments are pretty important during humanitarian crises, especially in conflict zones where traditional methods take too long and are clearly too dangerous.
The team put together a unique dataset of images showing Mariupol before and after the conflict, using high-res images. They tested a few different CNNs that were originally made for assessing damage from natural disasters. The study looked at how well these models could adapt to war damage using techniques like zero-shot learning (i.e. models without extra training) and fine-tuning.
The results showed that the CNN models could effectively assess war damage, achieving up to a 69% F1 score and 86% balanced accuracy. Zero-shot models performed better than random guessing but struggled particularly with distinguishing between different levels of damage. However, fine-tuning pre-trained models on conflict-specific data significantly enhances their accuracy.
Geospatial datasets
1. CloudSEN12+ dataset
CloudSEN12+ is an expanded dataset designed for cloud and cloud shadow detection in Sentinel-2 satellite imagery. It includes 849 new high-resolution, manually labelled images, making it the largest dataset of its kind. Previous annotations have also been reviewed to improve accuracy.
P.S. You can access the project site here and the dataset here.
2. World Bank road crash dataset
The World Bank released the Nairobi Road Safety Data 2023 which maps 30,000 road crashes across Nairobi. This dataset integrates data from police reports, hospital records, and spatial mapping. It includes attributes like crash type (e.g., vehicle-to-vehicle, pedestrian incidents), and severity levels (fatal, serious, minor).
3. Dynamic World+ dataset
This paper introduces Dynamic World+, a dataset that combines radar and colour data to improve land use mapping. It combines SAR data from Sentinel-1, with optical imagery from Sentinel-2.
P.S. You can read more about Dynamic World here.
4. GRID3 settlement extents data
The third version of the GRID3 settlement extents dataset for sub-Saharan Africa is now available. It covers 52 countries and more than 9 million settlements, providing info on where people live across the region.
Other useful bits
The Relief Visualization Toolbox (RVT) is a free tool and QGIS plugin for visualising digital elevation models (DEMs). RVT offers algorithms like hill-shading, sky-view factor, and openness. It makes it easy to highlight fine-scale terrain features. The guide can be accessed here.
The WMO’s new report on global water resources highlights an urgent water crisis. 2023 was supposedly the driest year for rivers in over three decades and saw the greatest glacier melt in 50 years.
Cycleosm, a new Python library, has been released to make it easy to extract bicycle infrastructure data from OpenStreetMap.
A recent study in Scientific Data examines the environmental impact of satellite mega constellations like Starlink and OneWeb. These satellites contributed 37-41% of emissions like black carbon and CO2 by 2022, primarily from rocket launches and re-entries. Using data from the ESA, the study creates a detailed emissions map to help inform sustainable space practices.
Jobs
Earth Pulse is looking for (i) a Geospatial developer intern to help with GEO data projects and (ii) an Earth Observation Data Analyst.
The University of Surrey is looking for a Research Assistant in Earth Observation and Machine Learning for Habitat Mapping.
National Geographic is looking for a Senior Interactive Cartographer with a strong background in interactive or animated maps.
UNDP is looking for a Geospatial Sciences Intern to assist its GIS team.
The United Nations University is looking for a Researcher for its geospatial and climate analytics initiatives.
UC Santa Barbara is looking for an Assistant Professor in Remote Sensing under the Department of Geography.
ESA is looking for (i) a Science Operations Scientist under its Science Operations Department, (ii) a Division Head for Earth Observation Mission Management and Product Quality, and (iii) a Forest Applications Specialist with EO experience.
The University of Liverpool is looking for a Postdoctoral Research Associate in Geographic Data Science.
Just for fun
NVIDIA delivered one of the first Blackwell B200 high-performance GPUs to OpenAI. It’s a lot more energy efficient than H100 GPUs and has a much bigger memory bandwidth.
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