// the find
torchgeo/terratorch
A Python toolkit for fine-tuning Geospatial Foundation Models (GFMs).
TerraTorch is a PyTorch Lightning wrapper for fine-tuning geospatial foundation models (Prithvi, TerraMind, SatMAE, Clay, DOFA, and others) on segmentation, classification, pixel regression, and object detection tasks, using TorchGeo and GEO-Bench data plumbing underneath. It's for people already doing earth-observation or remote-sensing ML who want to swap backbones/decoders via config instead of writing custom training loops.
It genuinely removes boilerplate: you point a YAML config at a GEO-Bench or TorchGeo dataset, pick a backbone (Prithvi, timm, Clay, DOFA...) and a decoder (SMP or mmsegmentation), and get a working fine-tune without writing dataset or model glue code. It covers both ends of the workflow, not just training — there's vLLM serving support and ONNX export examples, so you can actually get a model into production instead of stopping at a checkpoint. Institutional backing (IBM/NASA-affiliated, arXiv paper, active CI, biweekly community calls) means it's maintained rather than a one-off academic drop.
GDAL is a hard dependency, and the README itself admits install is complex and just punts to conda — anyone on a constrained pip-only environment will lose time here. The optional-extras list is sprawling (vllm, wxc, peft, visualize, geobenchv2, mmseg, surya, tortilla, rfdetr), which is flexible on paper but a real risk for dependency conflicts once you combine more than one or two. Object detection with Deformable DETR needs manually cloning a separate repo and compiling CUDA ops, Linux-only with specific CUDA/GCC versions — that part of the toolkit isn't really "pip install and go." It's also a fairly narrow audience tool (863 stars) — the abstractions are built around satellite/EO data specifics, so it won't generalize if your foundation-model fine-tuning isn't geospatial.