TorchGeo Dataset Overview

Dataset Type

Geospatial Datasets and Samplers

• Type: Geospatial dataset
• Description: Datasets that include geospatial metadata, supporting multiple satellite sources (e.g., Landsat 7 and 8) and agricultural data (e.g., Cropland Data Layer (CDL)).
• Functionality: Supports union and intersection operations, automatically handling differing coordinate reference systems (CRS) and resolutions.
• Sampler: Provides a RandomGeoSampler for sampling small patches from large geospatial images.

Benchmark Datasets

• Type: Benchmark dataset
• Description: Contains input images and target labels suitable for image classification, regression, semantic segmentation, and object detection tasks.
• Examples: Includes the Northwestern Polytechnical University (NWPU) VHR‑10 dataset.
• Functionality: Supports automatic download, verification, and extraction.

Pre‑trained Models

• Description: Models pre‑trained on multispectral sensor data using torchvision’s multi‑weight API.
• Example: Provides a ResNet‑18 model pre‑trained on Sentinel‑2 imagery.

Reproducibility

• Description: Utilises Lightning data modules and trainers to simplify experiment setup and result comparison.
• Example: Includes a training example for semantic segmentation on the Inria Aerial Image Labeling dataset.

Installation

• Method: Recommended installation via pip.
• Command: pip install torchgeo

Documentation

• Location: ReadTheDocs
• Content: API documentation, contribution guide, and tutorials.

Citation

• Paper: Provides a citation format for referencing the software in publications.