TorchGeo is a PyTorch library designed for deep learning on geospatial data, providing a framework for building and training neural networks for tasks such as semantic segmentation, object detection, and change detection. It serves as a comprehensive pipeline for remote sensing, featuring specialized dataset loaders and multispectral image preprocessing tools.
The library is distinguished by a dedicated remote sensing model zoo and extensive support for transfer learning, allowing users to integrate pre-trained weights optimized for specific satellite sensors. It also includes support for self-supervised learning on unlabeled geospatial data and provides utilities for mapping multispectral weights to custom input channels.
The project covers broad capability areas including geospatial dataset management—such as coordinate reference system reprojection, spatial sampling, and dataset splitting—as well as data preprocessing and the implementation of model architectures for various data dimensions. It further provides command-line interfaces for executing training workflows and tools for visualizing geospatial datasets, including imagery and model predictions.