LAUSR

Abstract The use of remote sensing observation to estimate subsurface oceanic variables, including subsurface temperature anomaly (STA), is essential for the study of ocean dynamics and climate change. Here we report a new method that combines a pre-clustering process and a neural network (NN) approach to determine the STA using ocean surface temperature, surface height, and surface wind observation data at the global scale. Gridded monthly Argo data were used in the training and validation procedures of the method. Results show that the pre-clustered NN method was better than the same method without clustering, while also outperforming a clustered linear regressor and the random forest method recently reported. The new method was tested over a wide range of time (all months from 2004 to 2010) and depth (down to 1900 m). Overall, our best estimation resulted in an overall root-mean-squared error of 0.41 °C and a determination coefficient (R2) of 0.91 at the 50 m level for all months. The R2 decreased to 0.51 at 300 m but was still better than the calculation without pre-clustering. This method can be expanded to estimate other key oceanic variables and provide new insights in understanding the climate system.