A statistical method based on remote sensing for the estimation of air temperature in China

Environmental applications require accurate air temperature (T-air) datasets with different temporal and spatial resolutions. Existing methods generally improve the estimation accuracy of T-air using environmental variables as auxiliary data to overcome problems related to sparse metrological stations. However, these data are always fixed and do not comprehensively explain the variations in T-air values at all temporal and spatial scales. Moreover, these methods seldom consider the spatial heterogeneity of relationships between T-air and auxiliary data. This heterogeneity is often caused by several factors, such as land type, topography, and climate. This study proposes an estimation method to produce maximum, minimum, and mean T-air (T-max, T-min, and T-mean) datasets at different temporal and spatial resolutions using satellite-derived digital elevation model data and both nighttime and daytime land surface temperature data as auxiliary data. The method is based on the assumption that the relationships between T-air and the chosen auxiliary data vary spatially. These relationships were further explored using geographically weighted regression with adaptive bi-square kernel function. The derived relationships were used to construct a T-air estimation model. Monthly T-air data with 5-km resolution and 8-day T-air data with 1-km resolution were produced for 2010. The results show that the proposed method can accurately represent the variations in T-air; the R-2 values were in the range of 0.95-0.99 for the monthly T-air data and 0.93-0.99 for the 8-day T-air data. The root mean square errors (RMSEs) for the monthly and 8-day T-max, T-min, and T-mean data of the year 2010 were 1.29 and 1.45 degrees C, 1.24 and 1.29 degrees C, and 0.8 and 1.2 degrees C, respectively. These results were compared with those from other estimation methods, specifically the estimation of T-air based on multiple linear regression (EATMLR) and regression kriging (EATRK). The proposed method was found to produce RMSEs that were 25-26% smaller than EATMLR and 34-42% smaller than EATRK.