Research On Multi-spatial Scale Ocean Sound Speed Prediction Method Based On Deep Learning

Ocean sound speed is a major environmental factor that affects the effectiveness of sonar systems and even determines the success or failure of underwater information warfare.Therefore,ocean sound speed prediction is an important aspect of oceanographic research and military ocean environment protection.With the continuous improvement of computer processing power and the rapid development of artificial intelligence,unprecedented opportunities and challenges have emerged for spatiotemporal prediction of ocean sound speed.The spatiotemporal continuity and heterogeneity of ocean sound speed exhibit significant uncertainty and complex nonlinear characteristics.Although numerous spatiotemporal sequence prediction algorithms have emerged,their prediction accuracy still needs to be further improved.Moreover,in the context of ocean sound speed prediction,most existing methods focus on sound speed profile prediction,while prediction methods for multiscale ocean sound speed are scarce.In light of this,this thesis proposes a deep learning-based approach for multiscale ocean sound speed prediction,specifically targeting the prediction of sound speed at different spatial scales.The specific research contents are as follows:Firstly,based on the BOA＿Argo gridded dataset,the main environmental factors affecting ocean sound speed,namely ocean temperature,salinity,and depth,are obtained.Due to the uneven vertical distribution,low horizontal resolution,and missing values in the original dataset,the modeling of ocean sound speed may lack precision.To address these issues,this thesis proposes a neural network interpolation model with multiple time spans,which can effectively enhance the spatial resolution of the data.Compared to single time span interpolation models,this approach is more efficient and lays the foundation for finer sound speed prediction.Secondly,this thesis constructs and improves the ocean sound speed prediction model.The influence of different spatial scale divisions,the network structure of the Conv LSTM model,and the prediction method on prediction accuracy are investigated.By comparative analysis,the optimal hyperparameters of the Conv LSTM model are determined.Furthermore,a spatiotemporal attention-based STA-Conv LSTM model is proposed,which allocates different weights to different time steps and spatial positions,replacing the assumption of equal importance for each time step and spatial position in normal circumstances.This approach further improves the prediction accuracy of ocean sound speed.Finally,the STA-Conv LSTM model’s predictive performance is validated by predicting ocean sound speed at different regions and multiple spatial scales.Experimental results show that the improved STA-Conv LSTM model achieves good prediction performance in the prediction of ocean sound speed at multiple spatial scales,demonstrating excellent generalization capability.The model shows improvements in prediction accuracy and interpretability.