Evaluating the Functional Realism of Deep Learning Rainfall‐Runoff Models Using Catchment Hydrology Principles
Ara Bayati, Ali Ameli, Saman Razavi
Abstract
Abstract Deep learning (DL) models such as Long‐Short‐Term‐Memory (LSTM) networks have achieved exceptional predictive accuracy in rainfall–runoff modeling. Yet these models learn from statistical correlations rather than hydrologic insights, raising the question of whether their internal functional reasoning is physically reliable. Despite previous studies highlighting unexpected outcomes from LSTMs under long‐term climate shifts, functional realism—defined as the extent to which a model's internal functioning aligns with defensible mechanisms of streamflow generation—remains largely underexplored. We introduce a hydrology‐specific Explainable AI (XAI) framework that opens the black‐box of LSTM. It extracts nonlinear, lag‐dependent, and time‐varying Impulse Response Functions (IRFs) which quantify the functional relationships that LSTM uses to reflect the isolated influence of precipitation ( P ), temperature ( T ), and potential evapotranspiration ( PET ) on simulated streamflow. IRFs reveal how LSTMs internalize streamflow generation during events, offering a catchment hydrology perspective for evaluating model realism. Applying this framework to 672 North American catchments with strong LSTM predictive skill, we find that high accuracy often masks hydrologically implausible reasoning: in over 70% of rain‐dominated basins, short‐term temperature rises unexpectedly raise simulated streamflow and enhance celerity rate even without rainfall; in snow‐dominated regions, PET is misattributed as a driver of snowmelt‐related flow and enhances the catchment's celerity rate. We conclude that correlation‐driven learning can compromise the robustness of LSTM‐based forecasts under weather extremes and short‐term and long‐term climatic shifts. Our framework bridges deep learning with hydrologic understanding and offers a scalable diagnostic for assessing the functional realism of DL models across diverse catchment types.