Reconstructing subsurface temperature profiles with sea surface data worldwide through deep evidential regression methods

In this paper, the combination of the deep evidential regression network (DERN) and empirical orthogonal function (EOF) methods, namely, the DERN-EOF method, was proposed to reconstruct the subsurface temperature profiles. Firstly, the DERN method was compared with the previous machine learning methods used in the subsurface temperature field reconstruction, for example the Self-organizing map method. The new network was shown to have certain advantages of higher accuracy and the capability to provide uncertainty estimation, which enables the reconstruction of the subsurface field with higher belief. Analysis of the method performance suggests that the DERN could improve reconstruction accuracy by almost 20% and that the uncertainty statistics agrees well with the data distribution. Then the DERN method combined with the EOF method was used to reconstruct subsurface temperature profiles at a global scale. Results suggest that the reconstruction error in highly dynamic regions, like the Kuroshio Extension and Gulf Stream regions, is significantly higher than that in other regions, and the uncertainties increase with the reconstruction error, suggesting that the proposed method can provide a relatively believable uncertainty estimation.