Satellite segmentation with a transformer upgrade

What it does

GeoSeg is a semantic segmentation toolbox built specifically for remote sensing—urban aerial shots, drone footage, and satellite passes. It wraps a unified training pipeline around PyTorch Lightning and timm, letting you benchmark CNNs like MANet against Vision Transformers such as UNetFormer or DC-Swin without rewriting boilerplate. The repo includes dataset adapters for ISPRS Vaihingen, Potsdam, UAVid, LoveDA, and OpenEarthMap, plus sliding-window inference for tiles too large to load into GPU memory at once.

The interesting bit

The project’s own UNetFormer (published in ISPRS) keeps a UNet-like encoder-decoder but swaps in transformer blocks, and the recent addition of PyramidMamba shows the maintainers are tracking the state-space model wave. Most academic code stops at a test script; this one actually ships tooling to chop up and reason over “huge” remote sensing images, which is where real-world deployment lives.

Key highlights

• Ships UNetFormer, DC-Swin, BANet, and CNNs (MANet, ABCNet, A2FPN) under one benchmark script
• Built on PyTorch Lightning 2.0 and timm, so multi-GPU training and backbone swaps are mostly config-driven
• Includes patch-splitting utilities and sliding-window inference for oversized aerial tiles
• Provides pretrained backbone weights via Google Drive and Baidu Disk
• Reproduction tables for standard remote-sensing benchmarks are included

Caveats

• Dataset preparation is strictly DIY: you must download raw imagery from official portals and run the provided patch-splitting scripts yourself
• Architectural intuition is left to the linked papers; the README is heavy on execution examples and light on model explanations
• Mamba support depends on additional packages not covered by the base requirements file

Verdict

Grab it if you need a unified remote-sensing segmentation benchmark that covers both CNNs and transformers. Look elsewhere if you want a batteries-included dataset or a general-purpose segmentation framework.