Spatial Downscaling of Precipitation Data in Arid Regions Based on the XGBoost-MGWR Model: A Case Study of the Turpan–Hami Region

Abstract

Accurate and reliable precipitation data are important for analyzing regional precipitation distribution, water resource management, and ecological environment construction. Due to the scarcity of meteorological stations in the Turpan–Hami region, precipitation observation conditions are limited, and it is difficult to obtain precipitation data. Firstly, the applicability of TRMM 3B43v7, GPM_3IMERGM 06, and CMORPH CDR satellite precipitation data for the Turpan–Hami Region was evaluated, and the products with better applicability were selected. Next, the Extreme Gradient Boosting Algorithm (XGBoost) and the Shapley Additive Explanations for Machine Learning (SHAP) model were combined to carry out a feature importance analysis on the climate factors affecting precipitation (mean temperature, actual evapotranspiration, wind speed, cloud cover), from which climate factors with a greater influence on precipitation were selected. Combined with climate factors, normalized difference vegetation index (NDVI), slope, aspect, and elevation as explanatory variables, a Multi-Scale Geographically Weighted Regression (MGWR) model was constructed to obtain the monthly precipitation data of 1 km spatial resolution in the Turpan–Hami area from 2001 to 2020. Finally, the spatiotemporal distribution characteristics and changing trend of precipitation in the Turpan–Hami region from 2001 to 2020 were analyzed. The results show that (1) GPM_3IMERGM 06 satellite precipitation data exhibits good applicability in the Turpan–Hami region. (2) The precision verification of the downscaling results from a monthly scale and an annual scale shows that the accuracy and spatial resolution of the data are improved after downscaling. (3) From 2001 to 2020, the precipitation in the Turpan–Hami region showed an insignificantly increasing trend.