Assessing Fourier and Latin hypercube sampling methods as new multi-model methods for hydrological simulations

The selection of a hydrological model plays a crucial role in simulating different hydrological processes and the water balance of a watershed. Different hydrological models differ in their structure, algorithms, and governing equations to solve the hydrological processes, which causes uncertainty in the simulation results. A multi-model approach minimizes the model uncertainty by combining two or more hydrological models. Existing multi-model methods have their advantages and limitations. This study introduces two simple and well-known sampling methods for assigning weights (SAW): Fourier and Latin hypercube sampling (SAW_FAST and SAW_LHC). We used four hydrological models: two lumped ( Identification of unit hydrographs and component flows from rainfall, evaporation, and streamflow data , and Modèle à Réservoirs de Détermination Objective du Ruissellement ), and two semi-distributed ( Soil and Water Assessment Tool , and Variable Infiltration Capacity ). We assess five different well-established multi-model methods: Bates–Granger averaging (BGA), Granger Ramanathan analysis method (GRA), multi-model super ensemble, equal weighted averaging method, and weighted averaging method along with the proposed methods (SAW_FAST and SAW_LHC). Our results show that SAW_FAST and SAW_LHC methods are sometimes more accurate than the other multi-model methods. BGA (KGE (0.744)) and GRA (NSE (0.528)) show high accuracy in combining the hydrological model outputs. SAW_FAST shows the highest NSE (0.708), CC (0.847), and R 2 (0.718) values while combining simulated results. Moreover, the SAW_LHC shows the highest KGE (0.782) and R 2 (0.840) values in all model combination results. Also, the 1:1 combination of lumped and semi-distributed hydrological models leads to more reliable results than the combination of similar structured hydrological models.