Interpreting Deep Machine Learning for Streamflow Modeling Across Glacial, Nival, and Pluvial Regimes in Southwestern Canada

The interpretation of deep learning (DL) hydrological models is a key challenge in data-driven modeling streamflow, as the DL are often seen “black box” despite outperforming process-based streamflow prediction. Here we explore interpretability convolutional long short-term memory network (CNN-LSTM) previously trained to successfully predict at 226 stream gauge stations across southwestern Canada. To this end, develop set sensitivity experiments characterize how CNN-LSTM model learns map spatiotemporal fields temperature and precipitation three regimes (glacial, nival, pluvial) region, uncover patterns learning. results reveal that has learned basic physically-consistent principles behind runoff generation for each regime, without being given any information other than temperature, precipitation, data. In particular, during periods dynamic more sensitive perturbations within/nearby basin where modeled, far away from basins. modeled magnitude timing perturbations, well day-to-day increases daily weather anomalies, found be specific regime. For example, summer months glacial increasingly generated by warm anomalies basins with larger fraction glacier coverage. This model's “glacier runoff” contributions explicit about coverage, enabled cell states strongly only glacierized summer. Our demonstrate decision making, when mapping consistent physical understanding system.