Spatiotemporal regression Kriging to predict precipitation using time-series MODIS data

Accurate prediction of the spatiotemporal distribution of precipitation is an important guide for more efficient agricultural production. However in the Xinjiang Uygur Autonomous Region, China, it is difficult to ensure accuracy due to sparse and unevenly distributed precipitation monitoring stations. The precipitation on raster grids must be predicted from point data. The combined China monthly mean meteorological data sets were used to build spatiotemporal geostatistical models to predict mean monthly precipitation in Xinjiang. Predictions in space and time were made for precipitation using spatiotemporal regression Kriging with some covariates. The Moderate Resolution Imaging Spectroradiometer 1-month images time series, topographic layers representing the normalized difference vegetation index and digital elevation model, and a temporal index to adjust for yearly periodic were used as covariates. The optimal covariates were selected by the all subset regression method to determine which predictor variables to be included in the multiple regression models. The accuracy of our mean monthly precipitation predictions was assessed by leave-one-out cross-validation. The prediction accuracy of the proposed spatiotemporal regression Kriging approach was compared with spatiotemporal multi-linear regression and spatiotemporal Kriging. These experimental results show that the normalized difference vegetation index, latitude, longitude, elevation, and time periodicity index are the optimal covariates for mean monthly precipitation prediction. The spatiotemporal distribution of precipitation in Xinjiang exhibits a distinctive pattern; in the north and west, there is more precipitation than in the south and east, respectively and more precipitation in the mountains than on the plains. Precipitation is closely related to topography. Annually, summer precipitation is the highest, followed by, spring and autumn, with winter the driest season in Xinjiang. Considering regression residuals the product-sum model was found to be suitable to fit the spatiotemporal variogram at higher accuracy. Precipitation maps generated by spatiotemporal regression Kriging were more accurate than those produced by the other tested methods, with lower MAE, RMSE, and BIAS values, and higher COR values for validation sampling sites. Spatiotemporal regression Kriging is an efficient method for accurate spatiotemporal prediction of precipitation in Xinjiang.