Detecting ephemeral inland water in hyper-arid coastal environments

Mapping inland water in hyper-arid coastal regions is challenging: ephemeral water, smooth surfaces, and moist soils are spectrally confused, while SAR also mistakes smooth surfaces for water. No prior study combines optical-SAR fusion for post-rainfall ephemeral detection in this setting. This paper presents a framework for Bahrain, with governmental collaboration, fusing Sentinel-2 optical and Sentinel-1 SAR in a U-Net model, pairing acquisitions under rainfall, cloud, and evaporation constraints. An RF baseline establishes per-sensor contributions; an ablation quantifies fusion benefit. Under LOSO evaluation over seven scenes, the fused U-Net reaches a mean water-class IoU of 0.573: about 96% above the RF baseline at the default 0.5 threshold (IoU 0.292), and 63% above it at the tuned 0.9 threshold (IoU 0.351). RF precision reached only 0.27-0.69 versus 0.62-0.90 for the U-Net, with improved recall, resolving the baseline's over-prediction spatially. This establishes a foundation for future smooth-surface discrimination, permanence, and uncertainty products.