Integration of Adaptive Emulators and Sensitivity Analysis for Enhancement of Complex Hydrological Models

A hydrological system involves a significant number of parameters to describe the real-world phenomena which ultimately reflects into the computational burden during model fitting and simulation. In this study, the proposed Adaptive Emulator Modelling based Optimization (AEMO) framework is presented which minimizes the complexity and computational burden, and enhances the overall efficiency of the representation of a physical watershed model. Here, AEMO consists of machine learning, sensitivity analysis, adaptive modelling, and sampling. The efficacy of the proposed AEMO framework is carried out through the Soil and Water Assessment Tool (SWAT) hydrological model for Peachtree Creek Watershed, Atlanta, USA. This watershed is subjected to flash floods in case of heavy rains due to a number of narrow streams situated in urban areas, and the water levels can rise quickly within a few hours or minutes of a rainfall event. Hence, this type of watershed is more complicated during the simulation and it is necessary to analyze the sensitivity during calibration of each iteration to provide information for the next iteration. The results concluded that the AEMO achieved exceptional results in both performance and computational burden when compared to the existing method. The NSE performance metric for the default SWAT model and hybrid model (SWAT + AEMO) showed the values 0.49 and 0.81, respectively. Hence, it displayed that AEMO enhanced the SWAT conceptual model prediction. The proposed framework incorporating future climate data can provide accurate information on water and disaster management.