A XGBoost-Based Downscaling-Calibration Scheme for Extreme Precipitation Events

Extreme precipitation events have caused severe societal, economic, and environmental impacts through the disasters of floods, flash floods, and landslides. The coarse-resolution of satellite-derived precipitation data, however, makes it difficult to quantitatively capture certain fine-scale heavy rainfall processes. Therefore, to improve the spatial resolution and accuracy of satellite-based precipitation extremes, a downscaling-calibration scheme based on eXtreme Gradient Boosting (XGBoost_DC) was proposed in this study, where the XGBoost algorithm was applied in both downscaling and calibration procedures. The performance of XGBoost_DC was evaluated with two other comparative methods, in which XGBoost was only used in either downscaling (XGBoost_Spline) or calibration (Spline_XGBoost) process. The results showed that: 1) XGBoost_DC achieved the best performance, as it obtained the highest accuracy and well reproduced the occurrence and the spatial distribution of precipitation during typhoon events; 2) XGBoost_DC could capture the spatial variations of the precipitation. Although Spline_XGBoost obtained results only slightly worse than the XGBoost_DC, it significantly underestimated the spatial variability; and 3) the model assessment between the XGBoost_DC and Spline_XGBoost illustrated the essential contribution of XGBoost algorithm in downscaling process, and improved our understanding of the capability of machine learning (ML) algorithm in reproducing spatial variance of precipitation. These findings imply that our proposed downscaling-calibration scheme can be applied for generating high-resolution and high-quality precipitation extremes during typhoon events, which would benefit water and flood management, as well as other various applications in hydrological and meteorological modeling.