Forecasting groundwater drought in data-scarce regions using a machine learning approach and Med-CORDEX climate projections: the case of the Haouz aquifer (Morocco)

Abstract

Groundwater is a critical resource in semi-arid regions, particularly in the Haouz plain of central Morocco, where climatic variability and growing anthropogenic pressures are causing increased stress on aquifer systems. This study aims to assess future groundwater drought in the Haouz aquifer under conditions of data scarcity by integrating regional climate projections from the Med-CORDEX initiative with advanced machine learning techniques. The research is driven by the need for reliable, spatially resolved forecasts in regions where hydrological and groundwater data are limited or unavailable. The core methodology involves the use of meteorological drought indices to quantify drought events based on climate variables. These indices were calculated using historical and projected climate data derived from Med-CORDEX simulations under two Representative Concentration Pathways: RCP 4.5 and RCP 8.5. In the absence of dense ground-based monitoring networks, the study relies on ERA5 reanalysis data and virtual station datasets to create an input matrix suitable for predictive modeling. Machine learning models were trained to estimate groundwater drought conditions using climate predictors and geographical variables. Among the models tested, Random Forest exhibited superior performance, capturing non-linear interactions and delivering high predictive accuracy (R² > 0.9). The results reveal a significant intensification of drought conditions over time, particularly in the long term under the RCP 8.5 scenario, with increased occurrence and severity of extreme drought events projected in the latter half of the 21st century. The western part of the aquifer is identified as highly vulnerable, experiencing the most pronounced drought intensification. In contrast, the eastern portion shows a degree of resilience, maintaining near-normal drought conditions even under severe climate scenarios. This spatial variability underscores the importance of localized groundwater management strategies. The study concludes that coupling regional climate projections with machine learning offers a promising approach for groundwater drought forecasting in data-scarce environments. The modeling framework developed is scalable and adaptable to similar hydrological systems facing data limitations.