Examination and evaluation of multi-source monthly scale fusion precipitation product in China based on machine learning algorithm

Abstract

Grid format precipitation products have better spatial monitoring capabilities compared to ground meteorological station observations, but there are significant differences in performance among different products. This article evaluates the accuracy of nine monthly scale precipitation products TRMM, GPM, CMORPH, CHIRPS, ERA5, ERA5 Land, PERSIANN, PERSIANN-CDR, PERSIANN-CCS in China, and selects five better precipitation products from them. XGBoost is used to select the best precipitation products Three machine learning algorithms, random forest and multiple linear regression, were used for data fusion. Research has found that TRMM, GPM, CMORPH, CHIRPS, and PERSIANN-CDR products have relatively good accuracy. In high altitude and arid regions, the error of precipitation products significantly increases. After machine learning algorithm fusion, the optimal XGBoost algorithm model significantly improves product correlation coefficient, and significantly reduces root mean square error and bias. The three algorithms have shown good accuracy in each month, with XGBoost algorithm model products performing better in summer and random forest algorithm model products performing better in winter. Moreover, the three algorithm model products have shown high accuracy in different regions. Compared with the five original products before the fusion, the accuracy of the three algorithm model products has improved. The product fused with XGBoost algorithm has more variation and local precipitation details compared to the optimal original GPM product and meteorological station interpolation product in space.

Keywords

• monthly scale precipitation products
• machine learning
• evaluation indicators
• data fusion