Scholarlink

<p>Mesoscale eddies significantly affect sound propagation, yet their complex internal sound speed profiles suffer from substantial reconstruction errors, and only few studies specifically addressed this reconstruction. To bridge this gap, we used multi-source satellite data and Argo profiles to identify eddies and build a temperature-salinity-sound speed dataset. Sea surface parameters (temperature, height anomalies, salinity, density) and Argo density serve as inputs for a random forest (RF) algorithm to learn the surface-to-underwater sound speed mapping. Concurrently, a unified eddy dynamic model reconstructs the internal density field. By combining these environmental parameters with the reconstructed density generates eddy sound speed profiles, an integrated PIRF-DEN model that merges machine learning and physical modeling was established. Evaluations demonstrate the superiority of the model. By incorporating density input, the reconstruction accuracy was significantly improved, the mean absolute error (MAE) and root mean square error (RMSE) were reduced to 0.83 and 1.39 m/s, respectively, which is 87.3% and 83.7% less than that of the sEOF-r method. Integration of the eddy model effectively characterized the vertical density structure, whose constraint lowered the overfitting risk and enhanced the accuracy and stability over sEOF-r, sEOF-RF, and RF models. Propagation loss calculations using the reconstructed sound speed showed high correlation (coefficient: 0.77) with measured data, further confirming its reliability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Integrating machine learning and physical models for reconstructing sound speed profiles in mesoscale eddies

  • 期刊论文

  • 发表日期：

    2026年5月26日

查看全文
  • Hongchen Li,
  • Ming Li,
  • Penghao Wang,
  • Kefeng Mao,
  • Yuhang Zhu,
  • Yuhang Liu

摘要

Mesoscale eddies significantly affect sound propagation, yet their complex internal sound speed profiles suffer from substantial reconstruction errors, and only few studies specifically addressed this reconstruction. To bridge this gap, we used multi-source satellite data and Argo profiles to identify eddies and build a temperature-salinity-sound speed dataset. Sea surface parameters (temperature, height anomalies, salinity, density) and Argo density serve as inputs for a random forest (RF) algorithm to learn the surface-to-underwater sound speed mapping. Concurrently, a unified eddy dynamic model reconstructs the internal density field. By combining these environmental parameters with the reconstructed density generates eddy sound speed profiles, an integrated PIRF-DEN model that merges machine learning and physical modeling was established. Evaluations demonstrate the superiority of the model. By incorporating density input, the reconstruction accuracy was significantly improved, the mean absolute error (MAE) and root mean square error (RMSE) were reduced to 0.83 and 1.39 m/s, respectively, which is 87.3% and 83.7% less than that of the sEOF-r method. Integration of the eddy model effectively characterized the vertical density structure, whose constraint lowered the overfitting risk and enhanced the accuracy and stability over sEOF-r, sEOF-RF, and RF models. Propagation loss calculations using the reconstructed sound speed showed high correlation (coefficient: 0.77) with measured data, further confirming its reliability.