Automatic identification of saltating tracks driven by strong wind in high-speed video using multiple statistical quantities of instant particle velocity

• Published in: Aeolian research 2024-12, Vol.70-71, p.100940, Article 100940
• Main Authors: Zhou, Hongji, Mei, Fanmin, Lin, Chuan, Pu, Mengjie, Xi, Aiguo, Chen, Jinguang, Su, Jin, Dong, Zhibao
• Format: Article
• Language: English
• Subjects: Aeolian saltating track; Categorical boosting; data collection; Machine learning; Particle velocity; sand; Tree model; Tree-structured Parzen Estimator; trees; wind speed
• ISSN: 1875-9637
• DOI: 10.1016/j.aeolia.2024.100940

•Automatically identifying saltating tracks driven by strong wind in high-speed video.•Instant saltating velocity and the TPE method have optimized the tree models.•The model CatBoost with the D3 dataset is the best classifier for saltating tracks. The evolution of saltating tracks driven by strong wind remains unknown due to the low accuracy or recall rates of saltating particle tracking algorithms (SPTs). Manual identification of saltating tracks becomes a primary bottleneck because of low efficiency, restricting the development of new SPTs with high accuracy. Herein, we proposed an optimized tree model for automatically identifying saltating tracks in the high-speed video under strong wind through establishing the dataset with multiple statistical quantities of instant saltating velocity (MSQV) and the workflow embracing the Tree-structured Parzen Estimator (TPE). The optimized Categorical Boosting model by the D3 dataset (CatBoost-D3) could be considered the best classifier among the tree models, owning the higher accuracy (0.9352), precision (0.9348), recall (0.9352), F1-score (0.9350) and area under an receiver operating characteristics curve (AUC, 0.9730), and lower time cost. The best performances were associated with the ensemble effect of critical and secondary features, distinct from the previous finding which revealed only the effect of critical features on enhancing AUC value. Additionally, one observed that the present model was comparable to other optimized tree model by the dataset with double-class and outperformed the other tree model by the dataset with multi-class. The present work offers a new avenue for identifying hop trajectories and tracking sand particle flow via machine learning in the future, and a new channel for reunderstanding the relationship between midair collision and saltation under strong wind through automatic identification of saltating tracks.