Comparative performance of regionalization methods for model parameterization in ungauged Himalayan watersheds

Study region: The study region is 23 different watersheds across Nepal.Study focus: This study aims at assessing the strengths and weaknesses of widely used regionalization methods for simulating daily hydrograph and flow duration curve in a comparatively large sample of 23 medium to small-sized watersheds across Nepal. We employed a deductive approach based on extractable watershed properties to test the performance of four regionalization methods: principal component regression (PCR), random forests (RF) under regression-based methods, spatial proximity (SP), and physical similarity (PS) under donor-based methods in a New hydrological insights for the region: The GR4J rainfall-runoff model coupled with Cemaneige snow module (GR4J-CN) could provide good simulation for majority of the watersheds with median NSE of 0.76 and 0.74 for calibration and validation periods respectively. Model simulation using parameter values predicted from different regionalization methods showed satisfactory results in majority of the watersheds for daily hydrograph simulation. While there wasn't a single method that performed well in all of the watersheds, the physical similarity methods was found to be the most robust. Visual comparison of errors in flow duration curve (FDC) also indicated physical similarity method as a better approach in ungauged watersheds of Nepal. Further experiment using multiple donors using the output averaging option was found to increase the performance of donor-based methods while the parameter averaging option resulted in a drop in performance. The study provides a comprehensive assessment of regionalization methods and advocates the use of hydrological model regionalization as a promising tool for streamflow prediction in ungauged Himalayan watersheds.