🌐 Are geospatial foundation models all hype and no substance?
PLUS: new POI dataset from Foursquare, alarming news about the environment, and more.
Hey guys, here’s this week’s edition of the Spatial Edge — a weekly round-up of geospatial news, that’s a bit like a warm beer on a hot summer’s day. It’s a bit niche, but some British people are weirdly into it. 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:
Foundation models: Are they actually…useful?
Foursquare data: 100m points of interest made available.
EarthView: New large-scale geospatial dataset.
Global warming: new worrying info from ECMWF.
Research you should know about
1. Are geospatial foundation models all hype and no substance?

If I could sum up geospatial data science in 2024 in just two words, it’d be “foundation models”. Last year saw the release of numerous geospatial foundation models (GFMs) from NASA and IBM’s Prithvi, to Satlasnet, SpectralGPT and Clay.
You can think of these GFMs as trying to do for satellite imagery, what large language models did for text. Note the keyword ‘trying’, here.
The basics of GFMs
At their core, GFMs use similar deep learning methods to LLMs (notably vision transformers — or ‘ViTs’ for the cool kids), to process reams of satellite data. This process creates ‘embeddings’, which you can think of as a ‘compression’ of an image. They basically capture important parts of an image - e.g. whether an image patch is covered by buildings, whether it’s in the desert, whether it’s likely to be an Asian location, and so on.
A quick embedding example
Imagine dividing a satellite image into smaller “patches.” You take one patch—like the one in the orange square below:
We can then take this patch and create embeddings of it. In the graph below, we have two axes - “greenness” and “tree”.
We can see that our patch is both pretty green and pretty tree-like. So it’s located in the top right.
Now this is a simplistic representation with just two dimensions (”greenness” and “tree”). Real GFMs do this in hundreds of dimensions, but the concept is the same. Once the model is trained, if it sees another forest patch somewhere else, it recognises it by placing it in that same patch-of-green corner of its internal “map.”
Why finetuning matters
Now the big issue is that, on its own, these GFMs can’t do much. We can’t use GFM’s out-of-the-box like ChatGPT. Instead, GFMs need fine-tuning for specific tasks like flood mapping or deforestation detection.
Previously, we’d have to collect a huge dataset of labelled images (e.g. thousands of images of trees) and train a model from scratch. However, thanks to GFMs, that data requirement has shrunk dramatically.
You can start with a pre-trained GFM, add a smaller set of labelled examples (trees), and end up with a pretty decent model in less time and with less data. That’s the real draw of GFMs: they often need less data to be effective.
Is this just an old concept with a new name?
There’s a bit of a debate in the geospatial community about whether GFMs are simply a rebrand of older techniques: i.e. taking a pre-trained model and fine-tuning it.
And I get this take.
At the end of the day, GFMs are not a panacea. If you don’t have high-quality labelled data, then GFMs are not going to magically work.
However, they do help when you have less quality data (but still at least some).
And ViTs do technically work well in creating embeddings that understand the context and relationships between patches of satellite images. This is an advancement on some of the more traditional deep-learning based approaches to satellite imagery analysis.
The caveat
A ‘foundation model’ doesn’t automatically mean a quality model. Poorly trained GFMs could be rubbish, just like we have some bad quality large language models (BloombergGPT anyone?).
In fact, for certain tasks (such as Burn Intensity Mapping), the Prithvi 2.0-300m performed on par with a vanilla U-Net in terms of its IoU and F1 scores. Not too impressive…
Evaluating GFMs
So in order to assess the quality of these GFMs, we need:
Benchmark Datasets: Standardised frameworks like PANGAEA and Geo-bench are needed to measure how GFMs perform on specific tasks, such as crop monitoring or urban change detection.
Leaderboards: Public rankings are needed to show which GFMs deliver the goods. This will bring more transparency about which models are more likely to provide reliable results for specific tasks.
Introducing AnySat
Ok that was a long pre-amble. But I wanted to go through these points before I get to the key paper of the week: AnySat, a new GFM.
AnySat’s angle is that other GFMs, like Prithvi, only work with particular sensors (e.g. the Harmonised Landsat / Sentinel dataset). AnySat, however, works on diverse sensors. They train on aerial images at 0.2m resolution, monthly 250m satellite data, as well as SAR.
And supposedly on downstream tasks like land cover classification, tree species identification or flood segmentation, AnySat matches or outperforms current specialised models.
You can access the code here.
Geospatial datasets
1. Foursquare places 100M dataset
Foursquare Places 100M is a big open dataset containing info on over 100 million real-world points of interest from around the globe. It includes metadata such as coordinates, category labels, and addresses.
2. EarthView remote sensing dataset
EarthView is a new large-scale remote sensing dataset spanning 15 trillion pixels from NEON, Sentinel, and newly released 1 m data by Satellogic, covering 2017–2022. The authors also introduce EarthMAE, a Masked Autoencoder approach to handle this diverse multi-source imagery. You can access the dataset here.
3. CloudSEN12 plus dataset
CloudSEN12Plus extends the original CloudSEN12 dataset with additional Sentinel-1 and Sentinel-2 scenes from 2020–2021, which targets multi-temporal cloud detection.
Other useful bits
According to Copernicus Climate data, 2024 is projected to be the first year where global temperatures exceed 1.5°C above pre-industrial levels. This highlights the combined effects of El Niño and ongoing greenhouse gas emissions. I covered some of this here.
This R Cookbook for Geological Spatial Data is a handy resource for working with geological and geospatial data in R. It covers a bunch of techniques for data processing, analysis, and visualisation, tailored for geologists and spatial scientists.
Microsoft has launched an AI-powered platform to accelerate materials discovery, enabling faster simulation and prediction of new materials with transformative potential. Of course, you can’t really launch a platform these days without AI being a part of it. But in any case, by integrating AI with more advanced computational techniques, researchers aim to tackle challenges in energy, sustainability, and manufacturing.
ICEYE has launched two new SAR satellites, further expanding its EO capabilities. These satellites enhance global monitoring for applications like natural disaster response, environmental changes, and infrastructure monitoring.
Jobs
The International Atomic Energy Agency (IAEA) is looking for a Satellite Imagery Analyst under its State Infrastructure Analysis Section.
Conservation International is looking for a Senior Manager for Monitoring and Evaluation.
First Street is looking for a Senior Geospatial Data Engineer under its Data Science Department.
FAO is looking for a Fishery Officer (Data Systems) under its Fishery Information and Knowledge Management Team.
Sustainable Development Solutions Network is looking for a GIS Developer and Analyst.
Vizzuality is looking for a Data Engineer with Python and data processing experience.
Just for fun
Saturn as captured by NASA’s Cassini spacecraft. Sometimes you just need to stand back and admire the beauty of nature….
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