Precipitation scale effect of the TRMM satellite in Tianshan,China

High-resolution precipitation data is conducive to objectively describe the spatial-temporal variability of regional precipitation,and the study of downscaling techniques and spatial scale effects can provide technical and theoretical support to improve the spatial resolution and accuracy of satellite precipitation data.In this study,we used a machine learning algorithm combined with a regression algorithm RF-PLS(Random Forest-Partial Least Squares)to construct a downscaling model to obtain three types of high-resolution TRMM(Tropical Rainfall Measuring Mission)downscaled precipitation data for the years 2000-2017 at 250 m,500 m,and 1km.The scale effects with topographic and geomorphological features in the study area were analysed.Finally,we described the spatial and temporal variation of precipitation based on the optimal TRMM downscaled precipitation data.The results showed that:1)The linear relationships between the TRMM downscaled precipitation data obtained by each of the three downscaled models(PLS,RF,and RF-PLS)and the precipitation at the observation stations were improved compared to the linear relationships between the original TRMM data and the precipitation at the observation stations.The accuracy of the RF-PLS model was better than the other two models.2)Based on the RF-PLS model,the resolution of the TRMM data was increased to three different scales(250 m,500 m,and 1 km),considering the scale effects with topographic and geomorphological features.The precipitation simulation effect with a spatial resolution of 500 m was better than the other two scales.3)The annual precipitation was the highest in the areas with extremely high mountains,followed by the mediumhigh mountain,high mountain,medium mountain,medium-low mountain,plain,low mountain,and basin.