Next-generation runoff prediction: Merging RFE, SHAP insights, and satellite data with innovative deep learning techniques

The study focuses on the Oued Ouahrane Ras in the Cheliff Basin, located in north-central Algeria. Accurate prediction of daily runoff is essential for effective water resource management, flood control, and agricultural planning. This study evaluates the performance of three advanced deep learning models—(a) Recursive Feature Elimination with Gated Recurrent Unit–Bidirectional Long Short-Term Memory (RFE-GRU-BiLSTM), (b) RFE with Gated Recurrent Unit–Convolutional Neural Network (RFE-GRU-CNN), and (c) RFE with Convolutional Neural Network–GRU–BiLSTM (RFE-CNN-GRU-BiLSTM)—for forecasting daily runoff in the study region. These models incorporate Recursive Feature Elimination (RFE) for dimensionality reduction and SHapley Additive exPlanations (SHAP) for post hoc feature importance analysis. Two distinct datasets were employed for model training and evaluation: satellite-based precipitation data from the Tropical Rainfall Measuring Mission (TRMM) and ground-based in-situ hydrological observations, covering the period from 1998 to 2012. The results indicate that the RFE-GRU-CNN model achieved the best performance on the TRMM dataset, yielding a minimum RMSE of 3.61 in model M8. Conversely, the RFE-CNN-GRU-BiLSTM model produced superior outcomes for the in-situ dataset, with an RMSE of 3.815. SHAP analysis identified the lagged discharge input (Q t − 2 ) as a key predictor across models, highlighting the catchment’s short-term memory, reflecting rapid runoff persistence controlled by active storage components. Although the implementation of RFE increased the average training time by approximately 43.37 s, the additional computational cost remained within acceptable limits. For example, training durations ranged from 1106.54 to 2425.52 s for the TRMM dataset and from 1224.71 to 1955.10 s for the in-situ dataset. These findings underscore the effectiveness of hybrid deep learning architectures in daily runoff prediction and emphasize the critical roles of both feature selection and dataset type in optimizing model performance. • Hybrid deep learning enhanced runoff prediction using RFE and SHAP insights. • RFE-GRU-CNN achieved RMSE = 3.61 and NSE = 0.893 on TRMM rainfall data. • RFE-CNN-GRU-BiLSTM attained RMSE = 3.79 and R = 0.94 with in-situ data. • SHAP identified lagged discharge (Qt−2) as the dominant runoff predictor. • RFE improved model accuracy with minimal increase in training time.