*5.1. Limitations of GWR Downscaling*

Previous studies have shown that, in the case that an established model can adequately estimate precipitation, the quality of the initial satellite precipitation data becomes a critical factor in determining the quality of the downscaling results [14,17]. In this study, the proposed two-step OI-GWR method combining data merging and downscaling successfully solved this problem. Nevertheless, some limitations of the GWR method were also found during this research. First, during the process of variable selection for GWR, when the variables with the best global correlation with precipitation were applied for GWR downscaling, the obtained results were not optimal. In addition, the inclusion of more explanatory variables (e.g., DEM, SLP, ASP, CVT, HSHD, and/or TWI) did not yield better results than those achieved using the single NDVI variable. Second, unlike the stepwise regression method, GWR cannot automatically identify and eliminate variables that are not significantly related to the dependent variables, which can be considered a shortcoming of the GWR method itself. Even after stepwise regression was performed for variable screening and the selected variables were input into the GWR model, the desired results were not obtained. Presumably, this can be explained by the fact that stepwise regression is a global regression method, which therefore selects variables from the perspective of global correlation, whereas the GWR method relies on point-by-point regression and therefore focuses on local modeling. For this reason, variables selected on the basis of stepwise regression or correlation analysis may not be suitable for GWR. Accordingly, further investigation will be necessary to improve the process of variable selection for GWR.
