Streamflow regionalisation of an ungauged catchment with machine learning approaches

Several tools are available to simulate catchment streamflow with a higher level of accuracy. Validation of simulated streamflow of ungauged catchments is a challenge in hydrological science due to the non-availability of gauging data. Generally, complex linear and nonlinear mathematical approaches are used to generate regionalized streamflow for ungauged catchments with available observed hydrological data from a neighbouring catchment. Machine learning (ML) is broadly used to model complex nonlinear relationships between different variables. This study demonstrates how novel ML approaches such as support vector machine (SVM) and extreme gradient boosting (XGB) can be applied to generate regionalized streamflow to calibrate ungauged simulated flow from the existing hydrological model. This study was performed on two study areas and four catchments located in different climate zones. The Soil and Water Assessment Tool (SWAT) model was used for ungauged flow simulation, and ML was used for regionalization.