A hybrid deep learning scheme with high accuracy for tracking aeolian saltating particles in high-speed pictures

Traditional saltating particle tracking algorithms (SPTs) for revealing the evolution of aeolian saltation may have lower ensemble recall rates (∼30 %) under moderate particle concentrations (∼100 particles per frame). For this reason, we propose a hybrid deep learning approach for tracking saltating particles, termed YOLOv8-IKFH, which integrates YOLOv8 for particle recognition with an improved Kalman filter and Hungarian algorithm (IKFH) for particle association in high-speed pictures. In our validation experiment, it shows that YOLOv8-IKFH can achieve the highest accuracy (∼93 %), recall rate (∼95 %), ensemble recall rate (∼50 %), and the longest average track length (35 particle locations per track) compared with the published SPTs due to optimal integration of particle recognition and association strategies like two-parameter constraint. In summary, the present work provides valuable insights for the development of future SPTs and a potential tool for deeply understanding the saltation process under strong winds.