LAUSR

The accurate determination of the ocean sound speed profile (SSP) is essential for oceanographic research and marine engineering. Traditional methods for acquiring SSP data are often time-consuming and costly. Machine learning techniques provide a more efficient alternative for SSP inversion, effectively addressing the limitations of conventional approaches.This study proposes a novel SSP inversion model based on a grouped dilated convolution (GDC) Informer architecture. By replacing the standard one-dimensional convolution in the Informer model with GDC, the proposed model expands its receptive field and improves computational efficiency. The model was trained using Argo profile data from 2008 to 2017, incorporating empirical orthogonal function (EOF) decomposition data, geographic location, temporal information, and historical SSP data, enabling SSP inversion across diverse regions and time periods.The model’s performance was evaluated using mean absolute error (MAE), root mean square error (RMSE) and mean absolute percentage error (MAPE) metrics. Experimental results demonstrate that the Informer-GDC model achieves evaluation metrics of 0.355 m/s and 0.611 m/s for MAE, 0.241 m/s and 0.394 m/s for RMSE, and 0.018% and 0.025% for MAPE compared with measured data from 2018.Compared to the LSTM and Informer models, the proposed model improves MAE, RMSE, and MAPE by 46.51% and 29.66%, 51.65% and 39.28%, and 51.25% and 37.08%, respectively. These findings highlight the superior accuracy, stability, and efficiency of the Informer-GDC model, marking a significant advancement in SSP inversion methodologies.