Chilean groundwater-level benchmark: analysis-ready monthly depth-to-water table, split assignments and predictions

Analysis-ready benchmark for machine-learning prediction of groundwater levels in Chile. The record contains a quality-controlled table of 92,691 monthly depth-to-water values from 644 wells of the Chilean national monitoring network (1982–2021) with 37 declared predictors (topography, 19 long-term bioclimatic normals, ten monthly TerraClimate variables re-extracted for each record's own calendar month, land cover and coordinates); the train/test assignment of every row under five splitting designs (random, well-based, spatial block, chronological per well and chronological with a global origin) and five seeds; the predictions of Decision Tree, Random Forest and Extra Trees for every test well-month under the full predictor set and three ablations; and the metrics, sensitivity, tuning and importance results of every experiment. Any new method can be scored on exactly these partitions and compared with the same simple baselines (training mean, well training mean, temporal interpolation, neighbor interpolation and last observation). The source file is the national depth-to-water compilation of Venegas-Quiñones et al. (2024, Scientific Data) with environmental values extracted through Google Earth Engine; its SHA-256 checksum is given in SHA256SUMS.txt. File contents and column definitions are described in README.md. The code that produced every file is archived in a companion software record and at https://github.com/venegasquinones/Chile_groundwater_ML.