Torchgeo

TorchGeo is a PyTorch domain library for deep learning on satellite and aerial imagery. It provides datasets, samplers, transforms and pretrained models for geospatial machine learning, with interfaces designed to compose with PyTorch and Lightning. It is built for researchers and engineers who train models on remote sensing data. Datasets and samplers are aware of coordinate reference systems, so sampling windows align to the geographic projection of the data rather than to array indices. Models accept arbitrary band counts, including 13 band Sentinel 2 imagery, and pretrained weights include SeCo, MoCo, MAE, DOFA and DINO v2 variants. Training tasks cover classification, semantic segmentation, object detection and instance segmentation, with datamodules for sets such as EuroSAT. The site also links to TorchGeo Bench for reproducible model evaluation and to TerraTorch for foundation model fine tuning, all projects of the independent TorchGeo Organization. TorchGeo is a PyTorch ecosystem library and an OSGeo project, released under the MIT license and installed with pip. The home page lists version 0.9.0 and requires Python 3.12 or newer and PyTorch 2.2 or newer. It carries no price, and the site publishes no paid plans. It also links to documentation, GitHub, the SIGSPATIAL 2022 paper and recorded talks and tutorials.