A data-driven approach for predicting depth-averaged velocities in the early stages of underwater glider navigation

• Published in: Ocean engineering 2024-05, Vol.299, p.117417, Article 117417
• Main Authors: Li, Hualing, Zhou, Yaojian, Zhao, Yuning, Wang, Meishu, Wang, Zijian
• Format: Article
• Language: English
• Subjects: Depth-averaged velocity; Forecast; Underwater glider
• ISSN: 0029-8018; 1873-5258
• DOI: 10.1016/j.oceaneng.2024.117417

Predicting the depth-averaged velocities of underwater gliders in the early stages of navigation is crucial for task performance optimization. However, due to the insufficient number of data samples for the depth-averaged velocities of underwater gliders in the early stages of navigation and the complex mechanisms of depth-averaged flow, task performance optimization remains challenging. This study presents a data-driven approach, the CutMix-augmented Extreme Gradient Boosting (CM_XGB) method, to effectively address this issue. Initially, the CutMix method is employed to augment the depth-averaged velocity samples. Then, the XGB method is applied for prediction. The proposed CM_XGB method is compared to five different prediction methods using three real-world depth-averaged velocity datasets and three error evaluation metrics, and its effectiveness in accurately predicting depth-averaged velocities under the same parameter conditions during the early stages of observation missions is demonstrated. This study shows that the CM_XGB method accurately predicts depth-averaged velocities even with limited data. This method is especially useful for underwater glider missions with data challenges in the initial stages. The success of the CM_XGB method highlights its potential for broader applications in oceanographic research and related fields, providing a valuable tool for scientists and researchers working with limited datasets. •A method called CutMix-augmented Extreme Gradient Boosting (CM_XGB) is proposed for predicting depth-averaged velocities of underwater gliders in the early stages of observation missions. This method effectively overcomes the challenge of limited data samples and complex flow mechanisms.•We validated the effectiveness of the CM_XGB method using three sets of depth-averaged velocity data from underwater gliders, three error evaluation metrics, and five comparison models under different sample sizes. These included the initial 10, 20, 30, 50, 100, 300 samples, and the entire dataset for constructing various sample size depth-averaged velocity data.•For three sets of underwater glider data, CM_XGB excelled with samples under 30, showing its efficacy in early mission predictions. Beyond this, random forest was superior, underscoring tree ensemble methods’ effectiveness in forecasting gliders’ depth-averaged velocities.