AI4WaterGuard Dataset and Reproducibility Outputs: Harmonised Weekly Earth Observation Data for Cyprus (2019–2025)

This record provides the harmonised dataset and associated reproducibility outputs supporting the manuscript “A Multi-Source Earth Observation and Artificial Intelligence Framework for Multi-Model Anomaly Detection in Mediterranean Water Resources Monitoring,” currently under consideration at Scientific Reports. The AI4WaterGuard dataset contains 617,862 weekly observations from 1,974 approximately 2 km grid-cell centroids across Cyprus. It covers 313 weeks and six complete hydrological years, from 1 October 2019 to 30 September 2025, with 16 harmonised Earth Observation and meteorological features derived from Sentinel-2, MODIS, ERA5-Land, and SMAP L4 products. ESA WorldCover 2021 was used separately as a land-cover eligibility mask. The accompanying reproducibility outputs include fold-level and aggregate evaluation metrics for seven unsupervised anomaly-detection models evaluated under nested Leave-One-Hydrological-Year-Out cross-validation, the persistent-hotspot catalogue comprising 705 locations, the spatiotemporal event catalogue comprising 2,788 reportable events, cross-model and cross-year event-linkage summaries, and fold-level model-attribution summaries. A data dictionary and README are included. Source code: https://github.com/elenakakoulli/AI4WaterGuard (MIT licence). The curated dataset and author-generated reproducibility outputs are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) licence. Upstream Earth Observation and reanalysis products remain subject to their original providers’ terms, licences, and attribution requirements.