Reconstructing Sound Speed Profile From Remote Sensing Data: Nonlinear Inversion Based on Self-Organizing Map

Abstract

By establishing a linear regression relationship between the projection coefficient of the empirical orthogonal function (EOF) of the sound speed profile (SSP) and remote sensing parameters of the sea surface, the single empirical orthogonal function regression (sEOF-r) method was used to reconstruct the underwater SSP from satellite remote sensing data. However, because the ocean is a complex dynamical system, the parameters of the surface and the subsurface did not conform to the linear regression model in strict sense. This paper proposes a self-organizing map (SOM)-based nonlinear inversion method that used satellite observations to obtain anomalies in data on the sea surface temperature and height, and combined them with the EOF coefficient from an Argo buoy to train and generate a map. The SSP was then reconstructed by obtaining the best matching neuron. The results of SSP reconstruction in the northern part of the South China Sea showed that the relationship between the parameters of the sea surface and the subsurface could be adequately expressed by the nonlinear neuronal topology. The SOM algorithm generated a smaller inversion error than linear inversion and had better robustness. It improved the average accuracy of reconstruction by 0.88 m/s and reduced the mean-squared reconstruction error to less than 1.19 m/s. It thus offered significant promise for acoustic applications.INDEX TERMS Sound speed profile, Empirical orthogonal function, Self-organizing map, Nonlinear inversion.

Wide-area sound speed profile estimation based on a pre-classification scheme for sound speed perturbation modes

“…The SOM inversion process is briefly described as follows (Chapman and Charantonis, 2017;Li et al, 2021).…”

“…If the time series and spatial orientation were expanded to introduce more samples, this would lead to inconsistent basis functions for the inversions and introduce more significant errors due to differences in the perturbation mechanisms. To maintain the consistency of the basic functions, the traditional treatment is to divide the sea area into individual 1°x 1°or 2°x 2°latitude and longitude grid cells and then extract the Empirical orthogonal function (EOF) vector basis functions in each grid (Chen et al, 2018;Li et al, 2021). When the sample data is insufficient, the whole sea area is treated as a large grid in a unified manner.…”

Estimation of Sound Speed Profiles Using a Random Forest Model with Satellite Surface Observations

“…Subsequently, Su proposed an integrated learning algorithm, known as the extreme gradient boosting (XGBoost), which could predict the thermohaline profile of the global ocean above 2000 meters [19]. Chen successfully reconstructed the SSP over 1000 meters in the northwestern Pacific Ocean using the SOM method [20,21], and Li et al further improved the SOM neural network and successfully reconstructed the SSP in the northern South China Sea [22]. It can be concluded that the nonlinear model based on machine learning performs well in terms of solving problems related to nonlinear ocean dynamics.…”

Estimation of sound speed profiles based on remote sensing parameters using a scalable end-to-end tree boosting model

“…By combining satellite remote sensing data and machine learning, Su et al and Li et al used classical machine learning methods and support vector regression to predict global ocean temperature profiles above 1000 m (Su et al, 2015;Li et al, 2017). Li et al successfully inverted the SSP in the South China Sea through a neural network based on a selforganizing map (Li et al, 2021).…”