Data and code for: Benchmarking predictive-performance loss from compressed urban cooling representations

This archive contains processed data, analysis code, benchmark outputs and figure source materials supporting the manuscript “Benchmarking predictive-performance loss from compressed urban cooling representations”. The study evaluates whether compressed urban-cooling representations, including k-means labels, score quartiles and an equal-weight score, retain held-out predictive performance and reconstruct the original continuous indicator space. The analysis uses 1,106 global urban centres and four continuous indicators representing surface thermal conditions, baseline water stress, vegetation and urban morphology. The archive includes country/ISO3-grouped fold assignments, grouped and random benchmark outputs, out-of-fold predictions, signed typology compression loss tables, country-block bootstrap summaries, large-country holdout sensitivity outputs, compression-curve results for k = 2–20, axis-reconstruction results, water-axis quantization diagnostics and publication/diagnostic figures. Typology compression loss is provided as a signed held-out R² difference between continuous and compressed representations. The archive should not be interpreted as providing intervention-performance estimates, vulnerability scores, risk rankings or urban-cooling priority maps. Vegetation-family outputs are included as internal-consistency diagnostics because of potential provenance overlap with vegetation indicators. Thermal-family outputs are provided as provisional benchmark outcomes subject to the provenance documentation included in the archive. The repository is intended to support transparency and reproducibility for the manuscript. Users should also cite and follow the licensing terms of the original upstream data providers documented in the provenance and source-data notes.