基于机器学习的SMAP卫星海表盐度反演

SMAP Satellite Sea Surface Salinity Inversion Model Based on Machine Learning

针对传统海表盐度的物理机制反演模型拟合过程复杂且反演精度不高等问题,借助大范围、全天时、L波段探测的SMAP卫星微波海洋遥感产品,以北太平洋(135°~165°E,15°~45°N)范围为研究海域,利用深层神经网络(Deep Neural Network,DNN)和支持向量机(Support Vector Machine,SVM)建立海表盐度(Sea Surface Salinity,SSS)遥感反演模型。验证结果表明:DNN与SVM模型测试集反演SSS与Argo(Array for Real-time Geostrophic Oceanography)实测SSS的均方根误差(Root Mean Square Error,RMSE)分别为0.1790和0.2570,平均绝对误差(Mean Absolute Error,MAE)为0.1293和0.1821,最小绝对误差为0.6426和2.0380,最大绝对误差为1.3241和2.3732,反演模型数据与实测Argo数据拟合后的的相关系数分别为0.89和0.84。总体来看,DNN模型比SVM模型的反演精度更高,但两者均显著提高了SMOS盐度产品精度,能够为相关研究提供数据支撑。

Sea surface salinity(SSS)is an import feature in the global ocean circulation and climate change.In order to solve the problem of complicated inversion process and poor accuracy of retrieved SSS,this paper uses the SMAP mission products of large-scale,all-day and L-band observation,takes(135°—165°E,15°—45°N)in the North Pacific as the research area,and establishes two SSS remote sensing models by Deep Neural Network(DNN)and Support Vector Machine(SVM).The experiment results show that the root mean square error of the DNN and the SVM model relative to Argo measured SSS are 0.1790 and 0.2570,respectively.The average absolute error are 0.1293 and 0.1821,respectively.The minimum absolute error are 0.6426 and 2.0380,and the maximum absolute error are 1.3241 and 2.3732,respectively.After fitting the inversion model data with the measured Argo data,the correlation coefficients are 0.89 and 0.84 respectively.In general,the inversion accuracy of DNN model is higher than that of SVM model,but both of them significantly improve the accuracy of SMOS salinity products,which can provide data support for relevant research.