Comparison of univariate and multivariate geographically weighted regression for estimating air temperature over Iran

Station recording air temperature (Ta) has limited spatial coverage, especially in unpopulated areas. Since temperature can change greatly both spatially and temporally, stations data are often inadequate for meteorology and subsequently climatology studies. Time series of moderate-resolution imaging spectroradiometer (MODIS) land surface temperature (Ts) and normalized difference vegetation index (NDVI) products, combined with digital elevation model (DEM), albedo from Era-Interim and meteorological data from 2006 to 2015, were used to estimate daily mean air temperature over Iran. Geographically weighted regression was applied to compare univariate and multivariate model accuracy. In the first model, which only interfered with land surface temperature (LST), the results indicate a weak performance with coefficient of determination up to 91% and RMSE of 1.08 to 2.9 °C. The mean accuracy of a four-variable model (which used LST, elevation, slope, NDVI) slightly increased (6.6% of the univariate model accuracy) when compared to univariate model. RMSE dropped by 19% of the first model. By addition albedo in the third model, the coefficient of determination increased significantly. This increase was 32% of the univariate model and 23.75% of the 4-variable model accuracy. The statistical comparison between the three models revealed that there is significant improvement in air estimation by applying the geographically weighted regression (GWR) method with interfering LST, NDVI, elevation, slope, and albedo with mean absolute RMSE of 0.62 °C and mean absolute R2 of 0.99. In order to better illustrate the third model, t values were spatially mapped at 0.05 level.