A data fusion approach to enhancing runoff simulation in a semi-arid river basin

Abstract

Accurate streamflow modeling is crucial for water resource management in dry and semi-arid regions. This study proposes a novel approach combining machine learning (ML) with conceptual and physically-based models to address of traditional model limitations in Iran's semi-arid Jazmourian River Basin. The HBV and SWAT hydrological models are used for conceptual and physically-based simulations, respectively, while Support vector regression (SVR) and multilayer perceptron (MLP) integrate hydrological model outputs with hydro-meteorological variables. Using hydroclimatic data from two periods-1963-1989 (dry phase) and 1993–2019 (wetter phase)-the study evaluates model performance under contrasting conditions. The proposed "fusion SVR" and "hybrid SVR with whale optimization algorithm" (SVR-WOA) models demonstrate improved accuracy in simulating runoff peaks. The SVR-WOA model achieves a 26.17 % performance improvement over SWAT for 1993–2019 and 25.36 % for 1963–1989, with RMSE values of 9.90 m³/s and 10.33 m³/s, respectively. This highlights hybrid modeling's potential for diverse hydrological challenges.