Weekly dispatches on geospatial machine learning — remote sensing, earth observation, and everything in between.
Sentinel-2 captures an image of the same patch of Earth around every 5 days, but with different sub-pixel offsets. Can we use that jitter to super-resolve the imagery? We try with a gaussian splat approach!
Say you have a GPU and want to run a deep learning model on all incoming Sentinel-2 imagery — the model you pick is the difference between ~10 GPU-hours and ~2,000 GPU-hours per year on H100. This post is about benchmarking the throughput of these deep learning models.
DeltaBit is a web app for training change detection models entirely in the browser on per-pixel AEF embedding differences (served as XYZ GeoTIFF tiles).
How many bits do you need for planetary-scale earth embedding retrieval? Binary-quantized embedding retrieval over 50M Sentinel-2 patches — entirely in the browser, no backend or server beyond cloud-native formats and storage.
A single year of DINOv3 embeddings for Earth costs 6.1 PB — more than the Sentinel-2 archive that produced them. How much can you compress without losing accuracy?
Chesapeake RSC is a benchmark that tests whether segmentation models use long-range spatial context by asking them to label road pixels hidden under tree canopy. Spoiler: they mostly don’t.
Using Alpha Earth Foundation embeddings to predict urban/rural classification and population density at the census block level — with a simple linear model hitting 92.5% accuracy.
An end-to-end walkthrough of our ICLR 2026 ML4RS tutorial: train a DeepLabV3 model on the Earth Surface Water dataset and run inference on arbitrary Sentinel-2 scenes with TorchGeo.
A new blog for geospatial machine learning explorations, experiments, and weekly finds.