Prediction of Precipitation using Multiscale Geographically Weighted Regression
Murat Taşyürek, Mete Çelik, Ali Ümran Kömüşcü, Filiz Dadaşer ÇelikPrediction of precipitation at locations that lack meteorological measurements is a challenging task in hydrological applications. In this study, we aimed to demonstrate the potential use of multiscale geographically weighted regression (MGWR) method used to predict precipitation based on relevant meteorological parameters. Geographically weighted regression (GWR) is a regression technique proposed to explore spatial non-stationary relationships. Compared to the linear regression technique, GWR considers the dynamics of local behaviour and therefore provides an improved representation of spatial variations in relationships. Multiscale geographically weighted regression (MGWR) is a modified version of GWR that examines multiscale processes by providing a scalable and flexible framework. In this study, the MGWR method was used to predict precipitation, which is an essential problem not only in meteorology and climatology but also in many other disciplines, such as geography and ecology. A meteorological dataset including elevation, precipitation, air temperature, air pressure, relative humidity, and cloud cover data from 184 stations in Türkiye was used, and the performance of the MGWR was assessed in comparison with that of global regression and classical GWR based on root mean square error (RMSE), Nash-Sutcliffe efficiency (NSE), and correlation coefficient (R) calculated between measured and simulated precipitation. The RMSE values calculated for the global regression, GWR, and MGWR methods were 4.64 mm, 3.53 mm, 2.9 mm, respectively. NSE values and R values were -0.63, 0.03, 0.35 and 0.04, 0.42, 0.63, respectively. These results demonstrated that the MGWR model outperformed other approaches in precipitation prediction.