Context-Aware Concept Distillation for Trustworthy Flood Prediction

A new framework called Context-Aware Concept Distillation (CACD) aims to make deep learning flood prediction models interpretable for disaster response authorities. Developed with domain experts, CACD distills opaque LSTMs into hydrology-aware surrogate models using an unsupervised pipeline that discovers a 'Hydrological Language' and a Residual Hypernetwork modulating concepts based on static basin characteristics. Tested on 5,203 basins globally, the model achieves high fidelity with a median Nash-Sutcliffe Efficiency (NSE) of 0.70, outperforming black-box baselines. The work addresses the societal challenge of trust and accountability in high-stakes public safety decisions.