Merge of multi-source precipitation products in Yalong River Basin of southwestern China based on LSTM

Objective With the intensification of global warming, the spatiotemporal pattern of precipitation has changed, leading to an increase in the frequency and intensity of extreme precipitation events. As an important basis for the study of watershed hydrological processes, regional precipitation is constrained by certain geographical environments such as alpine mountainous areas and permafrost glaciers, which hinders the comprehensive implementation of related detection efforts. Therefore, this paper aims to propose a framework for optimizing the accuracy of multi-source precipitation, which provides a reference for precipitation and hydrological simulation in areas with little or no data.

Method This study took the Yalong River Basin in the Tibetan Plateau of southwestern China as the research object. Based on the measured precipitation of 19 meteorological stations and three sets of precipitation data sets such as ECMWF re-analysis 5 (ERA5), climate hazards center infraRed precipitation with station data (CHIRPS) and the high asia refined analysis (HAR), the topographic factors and seasonal characteristics of precipitation were introduced. Using the long short-term memory (LSTM) method, a multi-source merged precipitation machine learning model was constructed, and a set of gridded precipitation data with a resolution of 0.25° × 0.25° from 1981 to 2017 was outputted.

Result The Pearson correlation coefficient between final output multi-source merged precipitation product and actual precipitation during the study period increased from 0.43 before merging to 0.82, and the root mean square error decreased from an average of 6.12 to 5.45 mm/d. The average critical success index (C_SI), probability of detetion (P_OD), and false alarm ratio (F_AR) were 0.64, 0.92, and 0.33, respectively. At the same time, based on the final output multi-source merged precipitation product, the VIC hydrological model was used for hydrological effect analysis, and the N_ash, RE, and C during the study period were 0.80, 1.39%, and 0.89, respectively.

Conclusion The above results indicate that the multi-source merged precipitation machine learning model constructed in this study can improve the spatiotemporal quality of precipitation products in the Yalong River Basin to a certain extent, and provide a certain reference for obtaining high-precision precipitation products and conducting hydrological simulation in areas with little or no data.