Comparison of Different Methods for Spatial Downscaling of GPM IMERG V06B Satellite Precipitation Product Over a Typical Arid to Semi-Arid Area

Abstract

Spatial downscaling is an effective way to obtain precipitation with sufficient spatial details. The performance of downscaling is typically determined by the empirical statistical relationships between precipitation and the used auxiliary variables. In this study, we conducted a comprehensive comparison of five empirical statistical methods for spatial downscaling of GPM IMERG V06B monthly and annual precipitation with a relatively long time series from 2001 to 2015 over a typical semi-arid to arid area (Gansu province, China). These methods included two parametric regression methods (univariate regression, or UR; multivariate regression, or MR) and three machine learning methods (artificial neural network, or ANN; support vector machine, or SVM; random forests, or RF), which were used to downscale the satellite precipitation from 0.1° (∼10 km) to 1 km spatial resolution. Five commonly used indices which were normalized differential vegetation index (NDVI), elevation, land surface temperature (LST), and latitude and longitude were selected as auxiliary variables. The downscaled results were validated using a total of 80 rain gauge station data during 2001–2015. Results showed that latitude had the overall largest correlation with IMERG annual precipitation, also evidenced by feature importance measurements in RF. The downscaled results at monthly scale were overall consistent with the results at annual scale. The machine learning-based methods had better predictive ability of the original IMERG precipitation than parametric regression methods, with larger coefficient of determination (R2) and smaller root-mean-square error (RMSE) as well as relative root-mean-square error (RRMSE). The downscaled 1 km IMERG precipitation by parametric regression methods had obvious underestimations (positive residual errors) in the south and east of Gansu province and overestimations (negative residual errors) in the west. In addition, the validation results of parametric regression downscaling methods showed large improvements after residual correction, while the improvements were small in the machine learning-based methods. However, the interpolation algorithm included in residual correction can cause certain errors in the downscaled results due to the ignorance of precipitation spatial heterogeneity. The machine learning-based RF downscaling had the smallest residual errors and the overall best validation results, showing great potentials to provide accurate precipitation with high spatial resolution.