Theory-Guided Deep Learning with AlphaEarth Embeddings for Flash Flood Prediction in Data-Scarce Regions

Flash floods are increasing in frequency and intensity due to climate change, yet reliable prediction remains difficult in regions with sparse hydrometeorological observations. Traditional hydrological models struggle without dense gauge networks, while purely data-driven approaches often produce implausible outputs. In this work, we introduce a theory-guided deep learning framework that integrates physics-inspired constraints with AlphaEarth satellite embeddings, a newly released global representation of multi-sensor Earth observation data available in Google Earth Engine. Our model combines dynamic drivers (rainfall, antecedent soil moisture) with static context (topography, land cover, and AlphaEarth embeddings) while enforcing monotonicity with rainfall, topographic consistency, and a rainfall–runoff budget. Using Sentinel-1 SAR flood masks from Pakistan as ground truth, we demonstrate that AlphaEarth embeddings improve spatial detail, and physics constraints enhance both accuracy and calibration. Our results highlight the potential of embedding-driven, physics-consistent ML to support climate adaptation by enabling trustworthy flood prediction in data-scarce regions.