torchgeo

Why we included this project

Anyone who trains vision models on satellite or aerial imagery knows the pain of wrestling geospatial data into a form PyTorch will accept. TorchGeo exists to remove that friction: it is a PyTorch domain library in the spirit of torchvision, with CRS-aware loaders for over a hundred geospatial datasets, samplers that select geographically meaningful windows, multispectral transforms, and pretrained weights for backbones like SeCo, SatMAE, DOFA, and Prithvi. A pipeline can read 13-band Sentinel-2 input directly instead of hacking an RGB workaround, and the pieces drop into standard PyTorch and Lightning training loops. That makes it a practical starting point for land-cover classification, semantic segmentation of building footprints, or object detection, since the data plumbing and checkpoints are already in place. It is just as useful for ML engineers new to geospatial formats as for remote sensing experts who want standard deep learning workflows without rebuilding their tooling.