View in EDS

Prediction of in situ seafloor sediment sound speed with machine learning.

Saved in:

Bibliographic Details
Title:	Prediction of in situ seafloor sediment sound speed with machine learning.
Authors:	Chen, Mujun^1,2,3 (AUTHOR), Zhang, Linqing^2,3 (AUTHOR), Hu, Xinfeng^2,3,4 (AUTHOR), Meng, Xiangmei^2,3 (AUTHOR), Kan, Guangming^2,3,4,5 (AUTHOR) kgming135@fio.org.cn, Tong, Siyou¹ (AUTHOR) tsy@ouc.edu.cn, Wang, Jingqiang^2,3,5 (AUTHOR), Li, Guanbao^2,3,5 (AUTHOR), Liu, Baohua⁵ (AUTHOR), Hua, Qingfeng^2,3,5 (AUTHOR), Wu, Siqi^2,3,4 (AUTHOR), Zhang, Xiaobo⁶ (AUTHOR)
Source:	Journal of Oceanology & Limnology. May2026, Vol. 44 Issue 3, p963-977. 15p.
Subject Terms:	Machine learning, Composition of sediments, Ocean bottom, Ensemble learning, Optimization algorithms, Physical acoustics, *Marine sediments
Geographic Terms:	East China Sea
Abstract:	The acoustic properties of seafloor sediments are crucial for accurate acoustic field prediction, seafloor resource exploration, and marine disaster prevention. However, traditional prediction equations, often based on laboratory-measured sound speeds, suffer from low precision and discrepancies with in situ measurements. To address these issues, we employed eXtreme Gradient Boosting (XGBoost) machine learning algorithms to develop high-precision in situ sound speed prediction models for seafloor sediments. The models were constructed using in situ sound speed and sediment physical property data (density, water content, porosity, median grain size, and grain group content) from 48 sites in the East China Sea shelf. Through feature parameter reduction and hyperparameter optimization, the optimal XGBoost model achieves training and validation R2 values of 0.989 and 0.977, respectively, having hyperparameters set at n_estimators=49 and max_depth=6. Compared to other machine learning models and empirical equations, the XGBoost model based on density, water content, sand content, and median grain size exhibited the lowest mean absolute error (MAE) and mean absolute percentage error (MAPE) at 5.603 m/s and 0.366%, respectively. This represents significant improvements over existing models, with MAE reductions ranging 2.165–118.903 m/s and MAPE reductions 0.137%–7.657%. This study thus provides an innovative and highly accurate method for predicting the in situ sound speed of seafloor sediments. [ABSTRACT FROM AUTHOR]
Database:	Energy & Power Source

Description
Abstract:	The acoustic properties of seafloor sediments are crucial for accurate acoustic field prediction, seafloor resource exploration, and marine disaster prevention. However, traditional prediction equations, often based on laboratory-measured sound speeds, suffer from low precision and discrepancies with in situ measurements. To address these issues, we employed eXtreme Gradient Boosting (XGBoost) machine learning algorithms to develop high-precision in situ sound speed prediction models for seafloor sediments. The models were constructed using in situ sound speed and sediment physical property data (density, water content, porosity, median grain size, and grain group content) from 48 sites in the East China Sea shelf. Through feature parameter reduction and hyperparameter optimization, the optimal XGBoost model achieves training and validation R2 values of 0.989 and 0.977, respectively, having hyperparameters set at n_estimators=49 and max_depth=6. Compared to other machine learning models and empirical equations, the XGBoost model based on density, water content, sand content, and median grain size exhibited the lowest mean absolute error (MAE) and mean absolute percentage error (MAPE) at 5.603 m/s and 0.366%, respectively. This represents significant improvements over existing models, with MAE reductions ranging 2.165–118.903 m/s and MAPE reductions 0.137%–7.657%. This study thus provides an innovative and highly accurate method for predicting the in situ sound speed of seafloor sediments. [ABSTRACT FROM AUTHOR]
ISSN:	20965508
DOI:	10.1007/s00343-025-5119-8