Estimating ocean thermocline from satellite observations with a multi-head attention-based neural network

A thermocline inversion model based on multi-head attention mechanism within a neural network framework is developed to estimate and analyze the ocean thermocline features, including depth (updepth and base), thickness, and intensity, in the western Pacific Ocean. This model employs Argo-derived thermocline product alongside various satellite remote sensing observations of ocean surface parameters, such as sea surface height, salinity, temperature, and wind. Specifically, three independent inversion models are executed using a dataset spanning the previous five years for training purposes, with the resulting model parameters used to estimate thermocline features in March, June, September, and December of 2016. The analysis reveals that thermocline updepth mainly located in the east of the Kuroshio extension area occurring in winter and spring; the seasonal distribution of the thermocline base is characterized by deeper depths at higher latitudes in the northern hemisphere during winter and spring, and from summer to winter in the southern hemisphere; the thermocline intensity in tropical regions is observed to be shallower yet stronger, exhibiting significant variations along the latitude with distinct seasonal changes. The seasonal distribution characteristics and variation trends of the updepth, base and intensity of the thermocline calculated by the multi-head attention neural network are generally consistent with the referenced Argo-derived thermocline dataset. Notably, the proposed intelligent inversion model for thermocline could also be utilized under condition of certain position with high flexibility, and exhibits faster convergence and greater accuracy compared to the classic Bi-LSTM model under comparable experimental conditions.