torchvision for satellite imagery, minus the RGB assumptions

What it does

TorchGeo is a PyTorch domain library for geospatial and remote-sensing data, positioned as a companion to torchvision. It provides datasets, samplers, transforms, and pre-trained models that handle the usual satellite-imagery headaches: multispectral bands beyond RGB, varying spatial resolutions, and mismatched coordinate reference systems. You can compose datasets from multiple satellites or sources using set-style union and intersection operators, and the library automatically aligns projections and resolutions.

The interesting bit

It is the first library to support pre-trained weights for multispectral sensors like Sentinel-2, adopting torchvision’s multi-weight API so you aren’t stuck with RGB ImageNet baselines. It also bundles PyTorch Lightning datamodules and trainers with fixed train-val-test splits to reduce boilerplate and improve reproducibility.

Key highlights

• Geospatial datasets automatically reproject to a common CRS and resolution when composed
• Set operators (UnionDataset / IntersectionDataset) let you merge or overlap satellite sources without manual alignment
• Benchmark datasets for classification, segmentation, object detection, and change detection that return dictionary-style samples
• Pre-trained weights for non-RGB sensors via torchvision’s multi-weight API
• Lightning datamodules and trainers for reproducible experiments

Caveats

• Some datasets require you to bring your own imagery (e.g., Landsat), while others auto-download and checksum

Verdict

Remote sensing researchers and ML practitioners who want to skip geospatial data-wrangling should start here. If your data is already tidy RGB images, plain torchvision is still simpler.