Super-resolution of subsurface temperature field from remote sensing observations based on machine learning

Abstract

Subsurface ocean observations are sparse and insufficient, significantly constraining studies of ocean processes. Retrieving high-resolution subsurface dynamic parameters from remote sensing observations using specific inversion models is possible but challenging. This study proposed two kinds of machine learning algorithms, namely, Convolutional Neural Network (CNN) and Light Gradient Boosting Machine (LightGBM), to reconstruct the subsurface temperature (ST) of the ocean’s upper 1000 m with a high resolution of 0.25° based on the satellite-based sea surface parameters combined with Argo float and EN4 profile data. We managed to improve the spatial resolution of ST from 1° to 0.25°. We employed two machine learning algorithms to set up monotemporal models of the four seasons and time-series models and adopted the determination coefficient (R2) and Root Mean Squared Error (RMSE) to evaluate the models’ prediction accuracy. The results show that LightGBM outperformed CNN in the case of small training samples. By contrast, in the case of big training samples, CNN outperformed LightGBM. Meanwhile, the ST with a high resolution of 0.25° predicted by the time-series CNN model can better observe mesoscale phenomena. This study provides more useful and higher-resolution data support for further studies on the warming and variability of the ocean interior under global warming.