terratorch

Why we included this project

TerraTorch is a PyTorch Lightning toolkit that makes fine-tuning geospatial foundation models feel less like an exercise in plumbing. It pairs any supported backbone (Prithvi, TerraMind, Granite, plus models from TorchGeo and timm) with pluggable decoders, so you can define semantic segmentation, regression, or classification through Python or YAML instead of hand-writing training loops. Because it sits on TorchGeo, the hard parts of geospatial data are already taken care of: band handling, multi-temporal inputs, and loading. That same foundation also reaches into solar physics and weather modeling, and in practice an applied ML engineer can go from a pretrained checkpoint to a fine-tuned map of burn scars or crop types with a single config file. It is a framework to build on rather than a model zoo, which suits teams that want control over training and inference while keeping the repetitive parts concise.