Deep Learning–Based Streamflow Forecasting in a Snowmelt-Dominated Basin: A Regime-Aware Evaluation

Artificial intelligence, particularly deep learning, is transforming discharge forecasting to address climate challenges and enhance water management by capturing complex, nonlinear, long-term dependencies. This study provides a process-based perspective on how sequence models interact with snowmelt-driven hydrologic variability and characterizes their behavior across monthly and seasonal regimes relative to operational flow thresholds. Modeling and forecasting of discharge in the Henry’s Fork basin are conducted using recurrent neural networks, specifically gated recurrent unit (GRU) and long short-term memory (LSTM). Input data were filtered using HUC-7 spatial codes. snow water equivalent (SWE) from SNODAS and temperature and precipitation from Daymet were extracted via the Google Earth Engine. Model training used grid search to optimize hidden units, learning rate, and batch size. Evaluations were conducted at two levels: overall performance and a monthly regime-disaggregated analysis under flood-threshold conditions. Incorporating time-lagged SWE substantially improved model performance. Attribution analysis using integrated gradients reveals model-dependent driver importance, with SWE and temperature more influential in GRU and precipitation stronger in LSTM. At the overall performance level, GRU outperformed LSTM (NSE 0.87 versus 0.81). Monthly regime-disaggregated analysis confirmed GRU’s advantage in most months and its stability during prolonged extremes and recession regimes, while LSTM retained an advantage in early melt-onset detection. These findings highlight regime-aware model selection: GRU’s stability across floods and recessions suits snowmelt-dominated, volatile basins, while LSTM offers niche value for short-duration, onset-focused regimes. Embedding these models within threshold-linked operational frameworks can strengthen flood-warning and water management under hydroclimatic variability.