裂流為近岸碎波帶中向外海流動之高速水流通道，常具有低對比、邊界模糊、形態不固定與不易肉眼辨識等特性。相較於一般物件偵測任務中邊界明確、形體固定之目標，裂流並非實體物件，因此在影像辨識上具有較高難度。本研究以交通部中央氣象署設置於宜蘭外澳海灘之岸基攝影機影像作為資料來源，建立裂流影像資料集，並比較 YOLOv8、YOLOv12、RF-DETR 與 RT-DETRv4 四種深度學習物件偵測模型於裂流辨識任務中的表現。
研究中首先依據既有裂流影像判釋準則，從外澳海岸監測影像中篩選裂流案例，並加入數值模式結果作為輔助判定依據，使裂流案例之選取能同時具備影像特徵與水動力條件上的對應關係。影像資料經時間平均處理後，分別建立 10 分鐘與 1 小時時間平均影像資料集，並以相同資料切分方式供四種模型進行訓練、驗證與測試。模型評估指標採用精確率、召回率與 F1-score，並進一步透過實際推論案例比較不同模型於邊界模糊、低對比及背景紋理干擾條件下之辨識差異。
研究結果顯示，四種模型於 10 分鐘時間平均影像資料集中皆取得較佳 F1-score，表示 10 分鐘時間平均影像較能保留裂流辨識特徵。實際推論案例顯示，四種模型的差異主要出現在邊界模糊、白沫分布不連續或背景紋理干擾較強之影像中。YOLOv8 與 YOLOv12 較容易出現部分漏判，RT-DETRv4 則較容易將非裂流海面紋理誤判為裂流，而 RF-DETR 對裂流位置與範圍之掌握較完整。綜合量化指標與推論案例結果，本研究認為 RF-DETR 可作為後續近岸裂流影像辨識應用中優先考量之模型；若要進一步應用於即時監測，後續仍需補充推論速度、硬體部署與現場連續影像測試。

Rip currents are fast, offshore-directed flow channels that occur within the nearshore surf zone. Unlike common object detection targets with fixed shapes and clear boundaries, rip currents are not physical objects. They usually appear in coastal images as low-contrast, blurred, and irregular channel-like features formed by surf-zone gaps, low-foam regions, darker water channels, and sea-surface textures. These characteristics make rip current detection more challenging than ordinary object detection tasks.
This study used shore-based coastal monitoring images from Waiao Beach, Yilan, Taiwan, to establish rip current image datasets and compare four deep learning object detection models: YOLOv8, YOLOv12, RF-DETR, and RT-DETRv4. Rip current cases were selected based on visual interpretation criteria and further supported by numerical model results of nearshore flow fields. A total of 77 rip current cases were used to generate 10-minute and 1-hour time-averaged image datasets. The datasets were divided into training, validation, and testing sets using a case-based splitting strategy to reduce data leakage.
Model performance was evaluated using precision, recall, and F1-score. The results showed that all four models achieved higher F1-scores using the 10-minute time-averaged image dataset than using the 1-hour dataset, indicating that 10-minute averaged images can better preserve important rip current features while reducing short-term wave fluctuations. Among the four models, RF-DETR achieved the best overall performance on the 10-minute dataset, with a precision of 0.82, recall of 0.85, and F1-score of 0.83. YOLOv8 showed balanced performance and served as a stable baseline model, while YOLOv12 did not clearly outperform YOLOv8 in this study. RT-DETRv4 produced more false detections due to its lower precision.
The inference case analysis showed that differences among the models became more obvious under blurred boundaries, discontinuous foam patterns, and complex background textures. YOLOv8 and YOLOv12 were more likely to miss some rip current regions, whereas RT-DETRv4 tended to misclassify non-rip sea-surface textures as rip currents. RF-DETR generally provided more complete detection of rip current locations and visible image extents. However, RF-DETR could still produce missed detections when rip current features were weak, which is critical for coastal safety applications. Therefore, although RF-DETR is considered the most suitable candidate among the four tested architectures under the dataset and experimental settings of this study, future studies should further reduce missed detections, expand difficult samples, integrate numerical model information, and evaluate real-time deployment performance.

王敘民, 董東璟, 蔡政翰, & 林芳如. (2022). 海岸裂流觀測之研究.土木水利, 49(6). https://doi.org/10.6653/MoCICHE.202212_49(6).0005
傅仲偉. (2022). 利用深度學習之更快速區域卷積神經網路分析裂流影像/Rip Current Detection from Images by Faster R-CNN.
簡子傑. (2024). 基於 YOLO 深度學習演算法之裂流辨識模型研究/Study on the Rip Current Recognition Model based on YOLO Deep Learning Algorithm.
蔡政翰,董東璟. (2022). 海岸裂流監測與預警技術研究(2/3),交通部中央氣象局委託研究計畫(期末)成果報告。
蔡政翰. (2023). 海岸裂流監測與預警技術研究(3/3),交通部中央氣象署委託研究計畫(期末)報告。
蔡政翰,董東璟. (2024). 海岸裂流監測與作業化預警試用研究(1/3),交通部中央氣象署委託研究計畫(期末)報告。
蔡政翰,董東璟. (2025). 海岸裂流監測與作業化預警試用研究(2/3),交通部中央氣象署委託研究計畫(期末)報告。
Brighton, B., Sherker, S., Brander, R., Thompson, M., & Bradstreet, A. (2013). Rip current related drowning deaths and rescues in Australia 2004–2011. Natural Hazards and Earth System Sciences, 13(4), 1069-1075.
Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., & Zagoruyko, S. (2020, August). End-to-end object detection with transformers. In European conference on computer vision (pp. 213-229). Cham: Springer International Publishing.
Castelle, B., Scott, T., Brander, R. W., & McCarroll, R. J. (2016). Rip current types, circulation and hazard. Earth-Science Reviews, 163, 1-21.
Dalrymple, R. A., MacMahan, J. H., Reniers, A. J., & Nelko, V. (2011). Rip currents. Annual Review of Fluid Mechanics, 43(1), 551-581.
De Silva, A., Mori, I., Dusek, G., Davis, J., & Pang, A. (2021). Automated rip current detection with region based convolutional neural networks. Coastal Engineering, 166, 103859. https://doi.org/10.1016/j.coastaleng.2021.103859
Drozdzewski, D., Shaw, W., Dominey-Howes, D., Brander, R., Walton, T., Gero, A., ... & Edwick, B. (2012). Surveying rip current survivors: preliminary insights into the experiences of being caught in rip currents. Natural Hazards and Earth System Sciences, 12(4), 1201-1211.
Dumitriu, A., Tatui, F., Miron, F., Ralhan, A., Ionescu, R. T., & Timofte, R. (2025). Ripvis: Rip currents video instance segmentation benchmark for beach monitoring and safety. In Proceedings of the Computer Vision and Pattern Recognition Conference (pp. 3427-3437)..
Everingham, M., Eslami, S. M. A., Van Gool, L., Williams, C. K. I., Winn, J., & Zisserman, A. (2015). The Pascal Visual Object Classes Challenge: A Retrospective. International Journal of Computer Vision, 111(1), 98–136. https://doi.org/10.1007/s11263-014-0733-5
Haller, M. C., Honegger, D., & Catalan, P. A. (2014). Rip Current Observations via Marine Radar. Journal of Waterway, Port, Coastal, and Ocean Engineering, 140(2), 115–124. https://doi.org/10.1061/(ASCE)WW.1943-5460.0000229
Holland, K. T., Holman, R. A., Lippmann, T. C., Stanley, J., & Plant, N. (1997). Practical use of video imagery in nearshore oceanographic field studies. IEEE Journal of Oceanic Engineering, 22(1), 81–92. https://doi.org/10.1109/48.557542
Holman, R. A., & Stanley, J. (2007). The history and technical capabilities of Argus. Coastal engineering, 54(6-7), 477-491.
Jocher, G., Chaurasia, A., & Qiu, J. (2023). Ultralytics YOLOv8 (版 8.0.0) [Software]. https://github.com/ultralytics/ultralytics
Lascody, R. L. (1998). East central Florida rip current program. National Weather Digest, 22(2), 25-30.
Lin, T. Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., ... & Zitnick, C. L. (2014, September). Microsoft coco: Common objects in context. In European Conference on Computer Vision (pp. 740-755). Cham: Springer International Publishing.
Liao, Z., Zhao, Y., Shan, X., Yan, Y., Liu, C., Lu, L., ... & Chen, J. (2025). RT-DETRv4: Painlessly Furthering Real-Time Object Detection with Vision Foundation Models. arXiv preprint arXiv:2510.25257.
Longuet-Higgins, M. S., & Stewart, R. W. (1964). Radiation stresses in water waves; a physical discussion, with applications. Deep-Sea Research and Oceanographic Abstracts.
Ma, J., Ushiku, Y., & Sagara, M. (2022). The effect of improving annotation quality on object detection datasets: A preliminary study. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 4849–4858. https://doi.org/10.1109/CVPRW56347.2022.00532
MacMahan, J. H., Thornton, E. B., & Reniers, A. J. (2006). Rip current review. Coastal engineering, 53(2-3), 191-208.
Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., ... & Bojanowski, P. (2023). Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193.
Pitman, S., Gallop, S. L., Haigh, I. D., Mahmoodi, S., Masselink, G., & Ranasinghe, R. (2016). Synthetic imagery for the automated detection of rip currents. Journal of Coastal Research, 75(sp1), 912–916. https://doi.org/10.2112/SI75-183.1
Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 779-788).
Robinson, I., Robicheaux, P., Popov, M., Ramanan, D., & Peri, N. (2026). RF-DETR: neural architecture search for real-time detectiontransformers.
Sohan, M., Sai Ram, T., & Rami Reddy, C. V. (2024). A review on yolov8 and its advancements. In International conference on data intelligence and cognitive informatics (pp. 529-545). Springer, Singapore.
Tian, Y., Ye, Q., & Doermann, D. (2025). Yolov12: Attention-centric real-time object detectors. Advances in neural information processing systems, 38, 78433-78457.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30.
Wang, H.-M., Doong, D.-J., & Lai, J.-W. (2025). Rip Current Identification in Optical Images Using Wavelet Transform. Journal of Marine Science and Engineering, 13(4), 707. https://doi.org/10.3390/jmse13040707
Wright, L. D., & Short, A. D. (1984). Morphodynamic variability of surf zones and beaches: a synthesis. Marine geology, 56(1-4), 93-118.
Zhao, Y., Lv, W., Xu, S., Wei, J., Wang, G., Dang, Q., ... & Chen, J. (2024). Detrs beat yolos on real-time object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 16965-16974).
Zhu, D., Qi, R., Hu, P., Su, Q., Qin, X., & Li, Z. (2022). YOLO-Rip: A modified lightweight network for Rip currents detection. Frontiers in Marine Science, 9, 930478. https://doi.org/10.3389/fmars.2022.930478