A Study to Compare Three Different Spatial Downscaling Algorithms of Annual TRMM 3B43 Precipitation

High spatial resolution, high accuracy precipitation data is essential for understanding basin-scale hydrology and the spatiotemporal distribution of regional precipitation. Since satellite precipitation products are often too coarse for practical applications, it is necessary to develop spatial downscaling algorithms. In this study, we investigated three downscaling algorithms based on the Multiple Linear Regression (MLR), Random Forest (RF), and Geographic Weighted Regression (GWR), respectively. They were used to downscale annual precipitation from 2005 to 2016 from the Tropical Rainfall Measuring Mission (TRMM) from 25 km $\times 25$ km to 1 km $\times$ 1km. Ground observations were used to validate the accuracy of the downscaled precipitation. The results showed that (1) GWR can capture precipitation spatial distribution of the original TRMM but MLR and RF can only capture global trend without residual correction. While after residual correction, MLR and RF also can capture spatial distribution of the original TRMM. (2) Residual correction was indispensable for the MLR-based and RF-based downscaling algorithms but not recommend for the GWR-based algorithm. (3) GWR and MLR were easy to overfit while RF can avoid overfitting well. When no overfitting existed, the GWR-based algorithms had the best accuracy among three algorithms. But with the number of predictors increasing, the accuracy of MLR-based and GWR-based algorithms would decrease but the accuracy of RF-based algorithms would increase which would eventually make the RF-based algorithms have the best accuracy among three algorithms. (4) The MLR-based, RF-based, and GWR-based algorithms improved the resolution of the original TRMM 3B43 at cost of reducing its accuracy.