Yuji HARUKI

Study on Rip Current Detection under Low Wave Conditions Using Deep Learning-Based Image Recognition

Naoyuki Inukai

 This study investigates a rip current detection method using deep learning, focusing on rip currents as one of the major causes of drowning accidents in coastal areas. Rip currents are strong offshore-directed flows generated by the interaction between waves and coastal topography near the shoreline. Identifying their occurrence locations is difficult for general beach users, and the recent decrease in beach lifeguards has made continuous monitoring more challenging. Although previous studies have attempted to detect rip currents using image recognition techniques, many of them mainly focus on relatively high-wave conditions. However, drowning accidents caused by rip currents also occur under low-wave conditions where swimming is generally considered safe. Therefore, developing a method for identifying rip current locations under such conditions is important.
 In this study, a deep learning-based detection model was constructed using the object detection algorithm YOLOv8. Video images recorded at Fujitsukahama and Ajirohama beaches were used to create training datasets. Several datasets were prepared by focusing on characteristic features of rip current occurrence, including the intersection of wave crests and disturbances on the water surface. Models trained with single features and models trained with combinations of multiple features were constructed to examine the influence of training data composition on detection performance.
 The performance of the models was evaluated in two stages. First, accuracy verification was conducted using test datasets with labeled rip current locations. The results showed that relatively high detection performance was obtained under conditions similar to those used in training, while performance decreased under different camera angles or coastal conditions. Models trained with multiple features tended to show improved recall, suggesting a reduction in missed detections. Next, detection rates were evaluated using video sequences. Although frame-based detection rates were not high, the detection rate improved when evaluated over time intervals corresponding to wave periods.
 These results suggest that constructing training datasets based on characteristic features of rip currents can contribute to detecting rip currents under low-wave conditions.