Mapping Water Stress and Agricultural Adaptation Potential Using Multi-Source 16-Band Attention U-Net (Case of Taroudant, Morocco)

Water scarcity represents a major and growing constraint on agricultural production across arid and semi-arid regions, a challenge significantly intensified by global climate change and the resulting increase in drought frequency. This study addresses the critical need for high-resolution, actionable water stress mapping in the Taroudant region of Morocco, a known water stress hotspot within the Souss-Massa basin. We propose a novel, weakly supervised framework utilizing an Attention U-Net deep learning architecture. The model is trained on a unique 16-band multi-source remote sensing data stack, fusing optical, radar, soil moisture, climate, and socio- economic variables. Crucially, this framework produces detailed maps of agricultural water stress at a 3 km resolution, providing a regional overview essential for large-scale adaptation planning. A Smart Health Index (SHI), derived from satellite indicators, serves as the training label, effectively overcoming the logistical and financial barriers associated with extensive field data collec- tion. The resulting maps classify stress into Severe, Moderate, and Healthy categories, offering critical insights into key drivers such as root-zone soil moisture and plant water content. This work significantly contributes to data-driven decision-making for regional water resource management, targeted irrigation scheduling, and the implementation of effective, drought resilient agricultural practices. While high-resolution inputs are processed at 3 km to capture field-level spatial variability, the resulting stress maps are interpreted at a coarser effective scale, consistent with the spatial support of root-zone soil moisture observations, enabling robust regional-scale analysis.