Improving runoff simulation using soil moisture guided hybrid modeling: A MISDc-FFNN framework applied to data-sparse basins

Study region The approach is evaluated across three contrasting basins: the semi-arid Rheraya basin in Morocco and two temperate basins, Colorso (Italy) and Bibeschbach (Luxembourg). Study focus This study introduces a new approach that uses soil moisture as the main calibration variable in a hybrid modeling framework to improve runoff simulation rather than using discharge. This framework combines a two-layer version of the daily lumped MISDc hydrological model (Modello Idrologico Semi-Distribuito in continuo) with a compact Feedforward Neural Network (FFNN). The FFNN is integrated to improve model calibration and optimize parameter selection based on soil moisture dynamics. New hydrological insights for the region Across the three basins and using two soil moisture datasets, in-situ and ERA5-Land, the results show that the hybrid framework yields comparable performance to the traditional MISDc approach during calibration periods for the benchmark model and consistently improves robustness during independent evaluation periods. Absolute gains in correlation and efficiency are observed across all basins, with median increases of approximately 0.16–0.18 in temperate basins and 0.05–0.08 in the semi-arid basin, alongside moderate reductions in simulation error. In several evaluation periods, the hybrid approach maintains superior performance, while the traditional model exhibits limited predictive capability. Overall, the proposed hybrid framework enhances runoff estimation using both observed and ERA5-Land soil moisture data, offering a reliable solution in data-scarce environments and supporting water-resources management.