Multi-Source Rainfall Data Assimilation based on Broad Learning System over Yunnan Province

The accurate estimation of rainfall is always a topic of concern， given its pivotal role in accurately predicting rainfall-related disasters.This study proposed a multi-source rainfall assimilation technology based on a broad learning system （BLS） to improve the accuracy of rainfall estimation.Yunnan Province， located in China's low-latitude plateau， was chosen as the geographical area of interest to establish a multi-source rainfall assimilation model within this region.In particular， the model utilizes five satellite-derived rainfall datasets （3B42V7， IMERG， GSMaP， CMORPH， PERSIANN） and the latitude and longitude information as the source data， and the ground-based rainfall gauge data serves as the reference data.The time span of all the datasets is from April 2014 to December 2017.A leave-one-year-out cross-validation （LOYOCV） method was applied to verify the performance of the established assimilation model， where statistical indicators including Pearson’s correlation coefficient （CC）， root-mean square error （RMSE）， mean absolute error （MAE）， Nash efficiency coefficient （NSE） and Kling-Gupta efficiency （KGE） were used to quantify the accuracy of assimilation rainfall at different spatiotemporal scales.Concurrently， assimilation models based on support vector machine （SVM） and deep neural network （DNN） were established to highlight the accuracy and efficiency of the BLS， respectively.Additionally， the effectiveness of the latitude and longitude information within the proposed assimilation model was examined.The results show that the daily average statistical index of assimilation rainfall based on BLS is better than that of the other five satellite-based products in LOYOCV.At the temporal scale， the proposed assimilation technique effectively reflects the temporal variations observed in gauge-recorded rainfall.Moreover， it can accurately estimate the rainfall amounts during rainstorms in Yunnan Province throughout 2017.It is worth noting that the rainfall data generated through the BLS method outperforms the CMORPH product （the most accurate one among the five satellite-derived rainfall products） in both rainy and dry seasons （May to October and November to April of next year， respectively）.At the spatial scale， BLS-based rainfall results in most areas of Yunnan Province showed higher CC and NSE as well as smaller RMSE and MAE than the satellite-based products.The evaluation of the assimilation models based on BLS， SVM， and DNN highlights that the BLS exhibits superior functional mapping capabilities compared to SVM and demands fewer computational resources than DNN.It is reasonable to conclude that the multi-source rainfall assimilation approach utilizing the BLS while incorporating latitude and longitude information can enhance the precision of rainfall estimates in Yunnan Province.The proposed method presents practical significance in multi-source rainfall data assimilation.