TorchGeo is a PyTorch domain library, similar to torchvision, specifically designed for geospatial data. It provides datasets, samplers, transforms, and pretrained models to help machine learning practitioners work with geospatial data and enable remote sensing experts to explore machine learning solutions.
Core Features
Geospatial Datasets and Samplers
- Data loaders for over 20 satellite imagery and mask datasets (Full list)
- Automatic reprojection and resampling of coordinate reference systems (CRS)
- Support for multimodal learning and data fusion
Benchmark Datasets
- Loaders for 50+ benchmark datasets (List)
- Enables rapid experimentation with various models
- Automatic download for most datasets
Pretrained Weights
- 40+ pretrained models on geospatial data (Models)
- Native support for multispectral imagery
- Enables transfer learning on smaller datasets
Reproducibility with Lightning
- Built-in train/validation/test splits and data augmentation
- Support for classification, regression, detection, and segmentation
- Command-line interface with YAML configuration support
Official Resources
- GitHub Repository:
https://github.com/microsoft/torchgeo - Documentation:
https://torchgeo.readthedocs.io/en/stable/
Note: Chinese documentation is not currently available.