Skip to content

Embeddings

Summary: Notes on vector embeddings and representation learning across domains. Last updated: 2026-08-20

2026

  • Downscaling Weather Prediction using TESSERA Embeddings: Downscaling weather prediction using Tessera embeddings — replaces hand-crafted terrain descriptors with TESSERA satellite foundation model embeddings to downscale coarse 25km weather grids into finer-resolution predictions, validated against Microsoft's Aurora AI model and ERA5 reanalysis. Related: Climate, Deep_Learning, AI. Keywords: TESSERA, embeddings, weather, downscaling, foundation model, Aurora, ERA5

  • Butterfly Habitat Mapping with TESSERA Embeddings: Tests whether TESSERA satellite embeddings can distinguish quality butterfly habitat from random locations, using UK Butterfly Monitoring Scheme data across 2,921 sites to train a logistic regression classifier validated with permutation testing. Related: Python, Data. Keywords: TESSERA, embeddings, butterfly, habitat mapping, logistic regression, Python, scikit-learn

  • Building Classification with GeoAI - Google AlphaEarth in Python: Day 5 of the #30DayMapChallenge demonstrating building type classification on OSM building footprints using Google AlphaEarth embeddings with the GeoAI Python package. Related: Web_Mapping, Python. Keywords: GeoAI, AlphaEarth, embeddings, OSM, building classification, Python, 30DayMapChallenge

  • S2Vec: Mapping the Modern World: Google Research's S2Vec framework uses masked autoencoding on rasterized Google Maps tiles (S2 Geometry cells) to generate general-purpose embeddings of the built environment, enabling prediction of socioeconomic and environmental metrics like population density, income, and carbon emissions at global scale. Related: Urban_Planning. Keywords: S2Vec, embeddings, masked autoencoding, S2 Geometry, Google Maps, urban analytics, geospatial ML, self-supervised, socioeconomic

  • Geospatial Skills for Coding Agents: A catalog of installable geospatial skills for coding agents (including Claude), providing plug-in workflows for GDAL, GeoParquet validation, and TESSERA embeddings via the geotessera CLI. Related: Claude_Code, Remote_Sensing, CNG. Keywords: Claude Code, coding agents, GDAL, GeoParquet, TESSERA, geotessera, geospatial, remote sensing, Python

  • Streaming millions of TESSERA tiles over HTTP with Zarr v3: Describes how TESSERA's geospatial embedding system was restructured from millions of individual NumPy files into sharded Zarr v3 stores per year, enabling efficient HTTP range requests for single-pixel to regional data retrieval with xarray/dask compatibility. Related: CNG, XArray, Data. Keywords: TESSERA, Zarr, embeddings, HTTP, geospatial, xarray, dask, cloud native

  • The Technical Debt of Earth Embedding Products: Examines fragmentation and interoperability challenges in Earth embedding products, arguing that standardizing how embeddings are distributed, stored, and accessed is the real bottleneck for geospatial foundation models. Related: Deep_Learning. Keywords: embeddings, geospatial, foundation models, interoperability, technical debt, cloud native

2025

Earlier

  • AlphaEarthFire: AlphaEarth × MODIS burn dataset builder and model trainer using AEF embeddings to model slow fire variables and predict forest fires. Related: Remote_Sensing, Deep_Learning, Python. Keywords: embeddings, fire prediction, MODIS, AlphaEarth, foundation model, Python