IMPACT ANALYSIS OF ERRORS IN COARSE SCALE SATELLITE DATA ON PREDICTIVE PERFORMANCE OF SPATIAL DOWNSCALING

摘要

The purpose of this study is to analyze the impact of error in coarse scale satellite data on predictive performance of spatial downscaling with fine scale auxiliary data. When downscaling coarse scale data with fine scale auxiliary data, the quality or error of the coarse scale data affects the downscaling results through decomposition into trend and residual components. Particularly, this study focuses on the impact of the magnitude of errors in the coarse scale data on trend component estimation. After generating the error-free true precipitation data (EO) at a coarse scale from rain gauge data, two synthetic precipitation data sets at a coarse scale such as data with small errors (El) and data with large errors (E2), which represent actual satellite precipitation products, are generated and used as input for spatial downscaling. Geographically weighted regression, which typically produces very high explanatory power, was selected as a trend estimation model, and area-to-point kriging was applied to residual correction. When downscaling the EO data, the predictive performance was not changed irrespective of residual correction. As the input coarse scale data included errors, however, residual correction led to an improvement of predictive performance, compared with the case of using trend estimates as the downscaling result. This result implies that the errors in the coarse scale data had a greater influence on trend estimates than residual estimates. Thus, residual correction should be applied for spatial downscaling of coarse scale data with errors. As the magnitude of errors in the coarse scale data increased, the predictive performance became worse, as expected. This poor predictive performance could not be adjusted through residual correction. Therefore, it is necessary to correct intrinsic errors in the coarse scale data during downscaling if a priori error information could be available.