Machine learning–based feature prediction of convergence zones in ocean front environments

Abstract

The convergence zone holds significant importance in deep-sea underwater acoustic propagation, playing a pivotal role in remote underwater acoustic detection and communication. Despite the adaptability and predictive power of machine learning, its practical application in predicting the convergence zone remains largely unexplored. This study aimed to address this gap by developing a high-resolution ocean front-based model for convergence zone prediction. Out of 24 machine learning algorithms tested through K-fold cross-validation, the multilayer perceptron–random forest hybrid demonstrated the highest accuracy, showing its superiority in predicting the convergence zone within a complex ocean front environment. The research findings emphasized the substantial impact of ocean fronts on the convergence zone’s location concerning the sound source. Specifically, they highlighted that in relatively cold (or warm) water, the intensity of the ocean front significantly influences the proximity (or distance) of the convergence zone to the sound source. Furthermore, among the input features, the turning depth emerged as a crucial determinant, contributing more than 25% to the model’s effectiveness in predicting the convergence zone’s distance. The model achieved an accuracy of 82.43% in predicting the convergence zone’s distance with an error of less than 1 km. Additionally, it attained a 77.1% accuracy in predicting the convergence zone’s width within a similar error range. Notably, this prediction model exhibits strong performance and generalizability, capable of discerning evolving trends in new datasets when cross-validated using in situ observation data and information from diverse sea areas.