A physics guided hybrid modeling framework for rainfall runoff forecasting in data scarce regions

Rainfall-runoff (R-R) forecasting in data-scarce regions remains challenging due to limited hydrometeorological observations, uncertainties in satellite-based precipitation products (SPPs), and the inability of individual modeling approaches to fully represent complex watershed dynamics. To address these challenges, this study develops a physics guided hybrid modeling framework that integrates bias-corrected satellite precipitation products, signal decomposition techniques, process-based hydrological modeling, and deep learning (DL) for improved runoff simulation and forecasting. In the proposed framework, physical knowledge is incorporated through the soil and water assessment tool (SWAT), which provides physically meaningful representations of watershed processes, while a long short-term memory (LSTM) network learns residual nonlinear relationships between physically simulated and observed runoff responses. Three widely used SPPs, including CHIRPS, IMERG, and PERSIANN-CDR, were evaluated against available rainfall observations, and a PDF-based additive bias correction approach was applied to improve precipitation estimates. Wavelet transform (WT) and variational mode decomposition (VMD) were further investigated within a time consistent feature construction framework to extract hydrologically relevant temporal patterns from corrected precipitation inputs and SWAT-simulated runoff outputs. The framework was evaluated in two hydroclimatically contrasting watersheds in Iran, namely the semi-arid Tabriz Basin and the humid Ghaem Shahr Basin. Results indicated that CHIRPS provided the most reliable precipitation estimates among the tested SPPs, while VMD outperformed WT in extracting temporal characteristics relevant to runoff generation. The proposed VMD-SWAT-LSTM framework generally achieved superior or comparable performance compared with standalone SWAT, LSTM, and conventional SWAT-LSTM models during training, validation, and testing periods under a temporally consistent evaluation framework. During the training phase, the proposed framework achieved CC = 0.99, NSE = 0.96, and RMSE = 2.61 m 3 . s − 1 in Tabriz Basin, and CC = 0.98, NSE = 0.97, and RMSE = 0.58 m 3 . s − 1 in the Ghaem Shahr Basin. The proposed approach also demonstrated improved robustness in multi-step-ahead forecasting (up to 7 days), although predictive performance decreased with increasing lead time, and maintained robust performance under uncertainty according to Monte Carlo (MC) analysis. The findings demonstrate that combining physically based watershed representation with decomposition assisted data-driven (DD) error learning within a physics-guided framework can improve R-R forecasting accuracy and robustness in data-scarce regions. The proposed methodology showed robust performance in the two investigated hydroclimatically contrasting basins, suggesting its potential for hydrological prediction in data-scarce regions. However, its transferability to other basins and broader hydrological and climatic conditions requires further evaluation.