Abstract
Morocco’s oases, critical agroecological systems in arid regions, face escalating water scarcity due to climate variability, groundwater depletion, and a historic decline in palm groves from 15 million to 4 million trees over the past century. This study introduces a machine learning (ML)-based precipitation forecasting framework to enhance water resource management in four semi-arid Moroccan regions: Errachidia, Figuig, Tata, and Zagora. Leveraging a 1981–2025 dekadal rainfall dataset from the Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS v2), we developed and compared four ML models: XGBoost, CatBoost, Long Short-Term Memory (LSTM), and Temporal Convolutional Network (TCN). CatBoost emerged as the most effective, achieving a testing of 0.9818 and a mean squared error (MSE) of 0.5430 mm on historical data (2019–2025), and a fine-tuned 90-day forecast (March 5–June 3, 2025) with of and MSE of 0.3364 mm. Historical trends revealed declining precipitation post-2015, underscoring the need for predictive tools. These findings demonstrate CatBoost’s superior ability to capture nonlinear rainfall dynamics, offering a scalable solution for climate-resilient water management in water-scarce regions. However, challenges such as data sparsity and model interpretability highlight the need for enhanced observational networks and explainable AI approaches to maximize practical adoption.
