A spatial data mining algorithm for downscaling TMPA 3B43 V7 data over the Qinghai-Tibet Plateau with the effects of systematic anomalies removed

Precipitation plays an important role in the water cycle and in matter and energy exchanges. Acquiring accurate information on precipitation over the Qinghai-Tibet Plateau, which has a limited rain gauge network, has been a great challenge. Downscaling the TRMM Multisatellite Precipitation Analysis (TMPA) 3B43 Version 7 dataset (0.25 resolution) provided an optimal approach to estimating precipitation at 1 km resolution over this highland plateau. Our downscaling assumptions were that non-stationary relationships between precipitation and land surface characteristics occur and have varying two-dimensional scale effects, and that the relationships vary in different sub-regions having differing combinations of land surface characteristics, including vegetation index, topographical factors, and land surface temperatures. We used Cubist (a spatial data mining algorithm) to implement our assumption. Cubist separated the Qinghai-Tibet Plateau into sub-regions according to geographical similarities, and selected the most effective variables over each sub-region to build models. We found that: (1) the downscaled results using this algorithm were more accurate and precise than other commonly used algorithms (e.g., geographically weighted regression) and the original TMPA data at 0.25 resolution; (2) DEM showed limited correlation with precipitation over the Qinghai-Tibet Plateau; and (3) the effects of systematic anomalies in the original TMPA data were removed in the downscaled results based on Cubist. We conclude that Cubist is a promising algorithm able to take hundreds of variables into consideration, and in this study was used to retrieve precipitation estimates at approximately 1 km resolution.