Frozen geospatial foundation model embeddings and handcrafted features for patch-level land surface temperature: reproduction notebook and cached data

Reproduction notebook and cached data for a benchmark of frozen geospatial foundation model embeddings against handcrafted Earth-observation features for patch-level summer land surface temperature (LST) prediction at 480 m in Flanders, Belgium, with an out-of-region test in Wallonia. Eleven feature representations of Harmonized Landsat and Sentinel-2 (HLS S30 v2.0) reflectance are compared as predictors of Landsat 8 and 9 Collection 2 surface temperature. They comprise three spectral indices, patch-mean reflectance, within-patch texture statistics and frozen embeddings from Prithvi-EO-2.0-300M-TL and TerraMind-1.0-base. One Random Forest is used throughout, with tile-grouped cross-validation, a paired tile bootstrap, a ridge-regression probe, a transfer test, a post hoc masked-pixel analysis and a post hoc analysis along the urban–rural gradient defined by the global map of local climate zones. The record contains the Jupyter notebook with the outputs of the reported run and the cache archive paper_cache_v1.1.0.zip, which holds the extracted reflectance and LST arrays, the local climate zone and ESA WorldCover built-up maps on the tile grid, the patch features, the embeddings, the model predictions and a manifest of SHA-256 checksums. With the cache, the notebook reproduces every reported number, table and figure on a CPU in a few minutes, without Earth Engine or a GPU, and every section ends with a verification step. README.md describes the files, the changes in this version and how to run the notebook. The notebook code is released under the MIT License and the cached data under CC BY 4.0. Version v1.1.0 adds the urban–rural analysis and leaves every result of v1.0.0 unchanged.