Mapping the mean annual precipitation of China using local interpolation techniques

Spatially explicit precipitation data are required in the research of hydrology, agriculture, ecology, and environmental sciences. In this study, two established techniques of local ordinary linear regression (OLR) and geographically weighted regression (GWR) and two new local hybrid interpolation techniques of local regression-kriging (LRK) and geographically weighted regression kriging (GWRK) were compared to predict the spatial distribution of mean annual precipitation of China. Precipitation data from 684 meteorological stations were used in the analysis, and a stepwise regression analysis was used to select six covariates, including longitude, latitude, elevation, slope, surface roughness, and river density. The four spatial prediction methods (OLR, GWR, LRK, and GWRK) were implemented with local regression techniques with different number of neighbors (50, 100, 150, and 200). The prediction accuracy was assessed at validation sites with the root mean squared deviation, mean estimation error, and R-square values. The results showed that LRK outperforms OLR and GWRK outperforms GWR, indicating that adding the kriging of regression residuals can help improve the prediction performance. GWRK gives the best prediction but the accuracy of estimation varies with the number of neighborhood points used for modeling. Although LRK is outperformed by GWRK, LRK is still recommended as a powerful and practical interpolation method given its computation efficiency. However, if LRK and GWRK are used to extrapolate prediction values, post-processing in the areal interpolation will be needed.