基于多源水文数据融合的GRACE水储量精度校正

Accuracy Correction of GRACE Water Storage based on Multi-source Hydrological Data

全球水储量的时空变化研究,对理解全球水循环过程、安排农业生产、防治洪涝灾害等具有重要意义。GRACE重力反演与气象实验卫星为获取全球水储量变化提供了直接观测手段,但不同解算模型和方法导致最终水储量产品之间存在差异。针对现阶段GRACE水储量数据产品中所存在的不确定性,拟结合全球水文模型以及陆面模型数据,采用三角帽方法对现有水储量数据产品进行不确定性分析;选取不确定性最低的GRACE和全球水文模型数据产品,引入点面融合思想,利用卫星观测和模型模拟数据互相约束来选取训练点,基于BP神经网络、深度置信网络建立GRACE精度校正模型,对GRACE月尺度水储量数据进行校正。以美国加利福尼亚州为例,从产品不确定性定量评价、水储量异常长时序变化及空间分布定性分析和地下水井站点实测验证等方面对校正结果进行分析验证。结果显示:①校正结果相比原始卫星观测数据产品及模型模拟产品在不确定分析中具有更低的不确定性(GRACE CSR:25.32 cm,PCR-GLOBWB:33.10 cm,DBN:13.85 cm);②在长时序变化和空间可视化分析中,校正结果相较于原始数据减少了异常的波动;③地下水位监测水井站点验证中,校正结果在相关性、均方根误差以及平均绝对误差上均有提升。

Study on the spatiotemporal changes of global water storage is of great significance for understanding the global water cycle,arranging agricultural production,and preventing natural disasters. GRACE satellite provides direct observation for obtaining changes in water storage on a global scale. However,different solution models and methods lead to differences between the water storage products. To reduce the uncertainty in GRACE data,this paper combines the Global Hydrological Models(GHMs) and Land Surface Models(LSMs)data,and uses the Three-Cornered Hat(TCH)method to analyze the uncertainty of the existing products. On this basis,the idea of point surface fusion is introduced,using GRACE and model simulation data to select training points,and using machine learning methods to correct the accuracy of GRACE satellite data.This article takes California,USA as an example,and uses TCH,long-term quantitative analysis and groundwater well site data to verify the accuracy of the result. The results show that:(1)The correction results have lower uncertainty than the original GRACE data in the uncertainty analysis(GRACE CSR:25.32 cm,PCRGLOBWB:33.10 cm,DBN:13.85 cm);(2)In long time series analysis,the correction results are smoother than the original data,reducing abnormal fluctuations;(3)In the verification of the well site,the results show that results have improved in correlation,root mean square error,and average absolute error.