Hydroclimatic-State Organization and Machine-Learning Assessment of Extreme Regional Rainfall in the Semi-Arid Peruvian Andes

Extreme rainfall in semi-arid mountain regions can occur within broader hydroclimatic states involving persistent precipitation, atmospheric moisture, and terrestrial wetness, yet their respective associations and prospective information remain poorly resolved in data-sparse environments. Here, we investigated the hydroclimatic organization of extreme regional rainfall in the semi-arid Peruvian Andes using satellite-derived and reanalysis observations spanning 2001–2025. Extreme rainfall was defined using the global 99th percentile of three-day accumulated precipitation, yielding 92 threshold-exceedance days that were declustered into 41 independent episodes. These episodes occurred preferentially under wet regional conditions, with the strongest separation from true non-extreme background days observed for multiday precipitation accumulation, followed by atmospheric and soil-moisture anomalies. Thirty-seven of the 41 independent episodes (90.2%) occurred when both total column water vapor (TCWV) and shallow soil-moisture anomalies were positive, corresponding to an odds ratio of 19.86 (year-cluster bootstrap 95% CI: 10.24–83.44). However, adjusted nested models provided no robust evidence for an independent shallow-soil-moisture contribution or a TCWV × soil-moisture interaction, indicating that the compound-wet state is better interpreted as a marker of organized regional wetness than as evidence of atmosphere–land synergy. Under a strictly non-overlapping rolling-origin architecture, hydroclimatic information available before the prediction date provided modest prospective discrimination of future three-day rainfall extremes. The complete hierarchical logistic model achieved a pooled PR-AUC of 0.126 compared with 0.094 for seasonality alone, whereas Random Forest and XGBoost showed no robust improvement over logistic regression. Native-resolution MODIS NDVI analyses likewise provided no robust evidence of post-event vegetation greening, although positive median responses occurred after aggregated ecological rainfall exposures. Overall, independent extreme regional rainfall episodes in the semi-arid Peruvian Andes were strongly organized within wet atmospheric and terrestrial hydroclimatic states, while evidence for atmosphere–soil interaction, strong predictive skill, and subsequent ecological response was substantially weaker. These findings demonstrate the value of integrated Earth-observation records for distinguishing hydroclimatic-state association from prospective prediction and ecological response in data-sparse mountain environments.