COMPARISON OF REGRESSION MODELS FOR SPATIAL DOWNSCALING OF COARSE SCALE SATELLITE-BASED PRECIPITATION PRODUCTS

This paper compared and evaluated the effects of explanatory power of regression models on predictive performance in component decomposition-based downscaling of coarse scale precipitation products. The regression models applied in this paper include (1) multiple linear regression (MLR), (2) geographically weighted regression (GWR), and (3) random forest (RF). From a case study of spatial downscaling of TRMM monthly precipitation products in South Korea, it was observed that GWR showed the highest explanatory power, followed by RF and MLR. From evaluation with independent rain gauge data, GWR-based downscaling outperformed other regression models. However, MLR-based downscaling with the lowest explanatory power showed better predictive performance than RF-based downscaling. Furthermore, the RF-based downscaling results could not preserve the overall patterns of original TRMM products. The GWR-based downscaling with the superior predictive performance included noisy artifacts in the downscaling result, which may be explained by over-fitting to the original coarse scale data. Thus, high explanatory power of regression models does not always improve predictive performance and it is suggested that other measures such as the preservation of spatial patterns of original coarse scale data should be considered for evaluation of downscaling results.