ucam-eo/geotessera

What it solves

GeoTessera is a Python library designed to simplify access to and manipulation of embeddings generated by the Tessera geospatial foundation model. It allows users to retrieve dense, 128-channel representation maps of the Earth's surface (derived from Sentinel-1 and Sentinel-2 satellite imagery) for specific regions and years, removing the need to manually handle raw satellite data or the complex underlying model inference.

How it works

The library uses a Parquet-based registry system to track available embeddings. Users specify a geographic bounding box, region file, or point, and the library identifies the necessary 0.1-degree tiles. It then downloads quantized embeddings and scale files via HTTPS from a public repository, dequantizes them into raw NumPy arrays, and can optionally export them as georeferenced GeoTIFF files for use in GIS software.

Who it’s for

It is intended for geospatial analysts, researchers, and data scientists working with satellite imagery who want to use precomputed foundation model embeddings for downstream analysis tasks without running the model themselves.

Highlights

• Flexible Export: Supports exporting data as raw NumPy arrays or georeferenced GeoTIFFs.
• Comprehensive Tooling: Includes a CLI for downloading data, generating coverage maps, and creating PCA-based visualizations.
• Efficient Data Handling: Uses quantized storage and a manifest-based registry to minimize bandwidth and local storage.
• Cloud-Native Access: Provides a preview API for streaming data directly from Zarr stores without full downloads.
• Interactive Visualization: Includes tools to create web maps and serve them locally for an interactive viewer.