融合星载GNSS-R数据和多变量参数全球海洋有效波高深度学习反演法

Deep learning retrieval method for global ocean significant wave height by integrating spaceborne GNSS-R data and multivariable parameters

星载GNSS-R作为一种新兴的观测方法,最近被应用于有效波高反演。现有研究通常使用从延迟多普勒图中提取的特征值以构建经验地球物理模型函数反演SWH。然而,使用多个变量参数作为模型输入具有很大挑战。为此,本文提出了一个融合星载GNSS-R数据和多变量参数反演全球海面SWH的深度学习网络模型(GloWH-Net)。为了验证本文模型的性能,ERA5、WaveWatchⅢ和AVISO SWH数据被用作广泛测试的参考数据,以评估GloWH-Net模型和先前模型(即经验模型和机器学习模型)的SWH反演性能。结果表明,当分别使用ERA5、WaveWatchⅢ和AVISO SWH作为参考值时,所提的GloWH-Net模型反演SWH的均方根误差分别为0.330、0.393和0.433 m,相关系数分别为0.91、0.89和0.84。相比基于最小方差估计器的经验组合模型反演SWH的均方根误差分别降低了53.45%、48.06%和40.63%;相比袋装树机器学习模型反演SWH的均方根误差分别降低了21.92%、18.72%和4.47%。表明了本文方法在反演全球海面SWH方面具有显著优势。

Global navigation satellite system-reflectometry(GNSS-R),as an emerging observation method,has recently been applied to the retrieval of significant wave height(SWH).Existing studies typically use extracting features from delay Doppler maps(DDMs)to construct empirical geophysical model functions(GMFs)for SWH retrieval.However,using multiple variable parameters as model inputs poses significant challenges.Therefore,this article proposes a deep learning network model(named GloWH-Net)that integrates spaceborne GNSS-R data and multivariate parameters to invert global sea surface SWH.To verify the performance of the proposed model,ERA5,WavewatchⅢ(WW3),and AVISO SWH data were used as reference data for extensive testing to evaluate the SWH retrieval performance of the GloWH-Net model and previous models(i.e.empirical and machine learning models).The results showed that when ERA5,WW3,and AVISO SWH were used as reference data respectively,the root mean square error(RMSE)of the proposed GloWH-Net model for retrieving SWH were 0.330 m,0.393 m,and 0.433 m,respectively,the correlation coefficients(CC)were 0.91,0.89,and 0.84,respectively.Compared with the empirical combination model based on the minimum variance estimator(MVE),the RMSE of SWH retrieval is reduced by 53.45%,48.06%,and 40.63%,respectively;Compared to bagging tree(BT)machine learning model,the RMSE of SWH retrieval decreased by 21.92%,18.72%,and 4.47%,respectively.This indicates that the deep learning method proposed in this article has significant advantages in retrieving global sea surface SWH.