AlphaCAMELS: Dynamic Yearly AlphaEarth Embeddings for the 671 CAMELS-US Basins (2017-2025)

Per-basin, per-year AlphaEarth Foundations embeddings for all 671 CAMELS-US basins (Newman et al., 2015; Addor et al., 2017), covering 2017 to 2025 (9 calendar years). For each basin and year, the 64-dimensional AlphaEarth embedding is averaged over the pixels whose centers fall inside the basin polygon, using a cosine-latitude-weighted spatial mean computed from the public AlphaEarth annual GeoZarr mosaic. The data ship as one Parquet file per basin (by_basin/<gauge_id>.parquet; 671 files, 9 rows each), with columns gauge_id, year, n_valid_pixels, and the 64 embedding bands A00 through A63. Basin polygons follow the canonical CAMELS HCDN set, keyed by zero-padded 8-character gauge IDs. This release preserves the time axis so the embeddings can be used as dynamic, time-varying model inputs rather than a single static per-basin average. The accompanying Python package (alphacamels) reads the data offline and reconstructs the polygon-clipped pixel grids on demand. Embeddings are derived from AlphaEarth Foundations (Google DeepMind, 2025) and mirrored by Source Cooperative and The Earth Genome. Released under CC-BY-4.0.