Research And Accuracy Evaluation Of Regional Precipitation Fusion Based On Machine Learning

Accurate precipitation estimation is very important for water resource management,ecological model simulation and natural disaster prevention.At present,improving the accuracy of precipitation data through data fusion has become one of research hotspots.Machine learning algorithms are gradually applied to precipitation data fusion.In this paper,multiple linear regression(MLR),feedforward neural network(FNN),random forest(RF)and long short-term memory neural network(LSTM)are used to fuse four kinds of satellite precipitation products and one reanalysis data in China.The correlation coefficient(CC),root mean square error(RMSE),relative deviation(RB),probability of detection(POD),false alarm ratio(FAR),and critical success index(CSI)are used as evaluation indexes.This paper analyzes the advantages and disadvantages of the four merging methods,and discusses the influence of data size on the fusion method.The main results and conclusions are as follows:(1)The fusion precipitation products(P1-MLR,P1-FNN,P1-RF,P1-LSTM,P5-MLR,P5-FNN,P5-RF and P5-LSTM)based on 1-year and 5-year historical precipitation data were developed.(2)Taking the rain gauge precipitation data(RGD)as the actual precipitation value,six indexes(POD,FAR,CSI,RMSE,CC and RB)were used to quantitatively analyze the error and daily precipitation event monitoring ability of the merged precipitation data(M-MLR,M-FNN,M-RF and M-LSTM)generated by the cross validation of RGD in the same target year.The results show that: compared with grid precipitation products,the errors of M-MLR,M-FNN,M-RF and M-LSTM are generally smaller,but only RF and LSTM can improve the daily precipitation event detection capacity.(3)M-MLR is the best fused precipitation data for monitoring daily precipitation events in the(sub)tropical monsoonal climate zone and highland mountain climate zone,as well as the best fused precipitation data for monitoring daily precipitation events during summer,and it is the best fused precipitation data for monitoring daily precipitation events at different elevations.In addition,MLR is the best performing fusion method for the moderate precipitation scenario.However,MLR is also the worst performing fusion method for the light precipitation scenario.M-FNN has the worst accuracy in all regions and the worst accuracy in spring,summer,and autumn.However,FNN is not the worst performing fusion method in all categories of rainfall.M-RF is the least error fused precipitation data in all regions,all seasons,and all elevations,and is the best fused precipitation data for monitoring precipitation events in temperate monsoonal and temperate continental climate zones,and is the best fused precipitation data for monitoring daily precipitation events in spring,autumn,and winter.In addition,RF is the best performing fusion method in heavy precipitation scenarios.M-LSTM severely underestimates precipitation in all seasons and at all elevations.However,LSTM is the best performing fusion method in the light precipitation scenario.(4)Training the model based on historical years of RGD and fusing the gridded precipitation products can also produce merged precipitation data with better overall accuracy than the gridded precipitation products.Compared with the merged precipitation data generated based on the RGD of the target year as training data,the error of the P1 precipitation product is larger,but the daily precipitation event monitoring capability of both P1-FNN and P1-LSTM is better.(5)For the MLR model,the daily precipitation event monitoring capability of the merged precipitation data does not change much when the data size used to calculate the weights are thousands,tens of thousands,hundreds of thousands,and millions,respectively.Moreover,the error of the merged precipitation data does not change much when the data size used to calculate the weights are tens of thousands,hundreds of thousands,and millions,respectively.For the FNN model,the accuracy of the merged precipitation data is the best when the data size of training data is hundreds of thousands.For the RF model,its fusion effect becomes better as the data size of the training data increases.For the LSTM model,the merged precipitation data has the best ability to monitor daily precipitation events when the data size of its training data is hundreds of thousands.The RMSE and CC of the merged precipitation data are improved with the increase of the data size of the training data.(6)Compared with the internationally popular precipitation dataset which named MSWEP V1(Multi-Source Weighted-Ensemble Precipitation),the accuracy of all P1 precipitation products and P5 precipitation products are generally better.