联合GRACE与GRACE-FO反演2002～2020年长江流域陆地水储量变化

探讨采用不同激励函数的BP和RBF神经网络方法填补GRACE与GRACE-FO卫星空缺数据的精度及可行性，并基于最优方案对缺失数据进行填充；利用ITSG-Grace2018和ITSG-Grace_operational时变重力场模型反演2002～2020年长江流域陆地水储量变化，并结合GLDAS模型、降水、气温及长江流域水资源公报等数据对该区域的陆地水储量变化进行综合分析。结果表明：1）隐含层激励函数为线性整流函数（ReLU）的BP神经网络算法具有较好的拟合效果，可用于填充GRACE与GRACE-FO卫星任务间的数据空缺；2）长江流域的陆地水储量变化具有一定的区域差异性，主要表现为上游东部与中游大部分地区陆地水储量以5 mm/a左右的速率上升，上游中西部区域下降，下游基本保持不变；长时间序列的GRACE/GRACE-FO时变模型能够反映长江流域2019年的干旱与2017年、2019年的洪涝等灾害。

Combining GRACE and GRACE-FO to Derive Terrestrial Water Storage Changes in the Yangtze River Basin from 2002 to 2020

This paper first discusses the accuracy and feasibility of using different activation functions of BP and RBF neural network methods to fill the gap data of GRACE and GRACE-FO satellites, and fills in the missing data based on the optimal scheme. We use the ITSG-Grace2018 and ITSG-Grace operational time-varying gravity field models to derive the changes of TWS in the Yangtze river basin (YRB) from 2002 to 2020, and finally, combining with the GLDAS model, precipitation, temperature, the Yangtze River Water Resources Bulletin and other data, we comprehensively analyze the changes of TWS. The research results show that: 1) The BP neural network algorithm with the hidden layer activation function as the rectified linear unit (ReLU) is effective in filling the data gap between GRACE and GRACE-FO satellite missions; 2) TWS changes in the YRB have certain regional differences, which are mainly manifested in TWS increase in eastern part of the upstream and most part of the midstream at a rate of about 5 mm/a, and decline in the upstream mid-west, while the downstream is basically unchanged. The GRACE/GRACE-FO long-term series time-varying model can reflect the drought in 2019 and the floods in 2017 and 2019 in the YRB.