A Clustering-Assisted Regression (CAR) approach for developing spatial climate data sets in China

Abstract

There is an increasing demand for improving spatial resolution of climate data. However, an increase in resolution does not necessarily mean an increase in realism and accuracy if local spatial features, such as elevational effects, cannot be considered in developing higher-resolution climate data. The Gradient plus Inverse Distance Squared (GIDS) is a broadly accepted elevation-dependent method for spatial interpolation of climate data but is relatively less effective in predicting climate variables in mountainous regions. We developed a new method called Clustering-Assisted Regression (CAR). Instead of using a fixed number of neighboring observed stations within a moving window, we repeated cluster analysis to derive a new subset of stations for each estimated site in CAR. We used both GIDS and CAR to estimate monthly mean temperature and monthly precipitation across mainland China based on observation data from 719 national meteorological stations. Both GIDS and CAR interpolation methods behaved reasonably well in developing 1 km resolution spatial data of monthly mean temperature and monthly precipitation at a national scale in mainland China. The accuracy of monthly mean temperature in summer was higher than in winter whereas that of monthly precipitation in winter was better than in summer. Overall comparisons indicate that CAR was slightly more accurate than GIDS, especially for predicting local climate patterns.