Comparison of False Positive Case in Coastal Debris Using Deep Learning-Based Object Detection Models

Deep learning-based object detection models, such as YOLO and DETR, have been actively studied for monitoring coastal debris. While recent models exhibit minimal differences in quantitative accuracy and performance, the underlying algorithms and methodologies for object detection vary across models. Consequently, detection outcomes can differ based on the type of the debris and the characteristics of the coastal environment. Nonetheless, there is a notable lack of studies that provide a quantitative analysis of these findings. Therefore, this study analyzed the false positives of coastal debris using the YOLOv10 and RT-DETR models to identify the detection characteristics of each model. To ensure comparable performance between the two models, hyperparameters were fine-tuned to achieve a mean Average Precision (mAP) exceeding 0.9. A dataset of approximately 350,000 coastal debris images (sourced from https://www.aihub.or.kr/) was utilized to train both models, with an 8:2 split between training and validation sets. Coastal debris was classified into 11 categories: Glass, Metal, Net, PET Bottle, Plastic Buoy, Plastic ETC, Plastic Buoy of China, Rope, Styrofoam Box, Styrofoam Buoy, and Styrofoam Piece. To analyze the detection characteristics of the trained models, images of coastal with various types of debris were collected using UAVs. False positive objects were classified and systematically analyzed based on the detection results of the collected coastal debris images using the two model. The analysis of false positives revealed that the YOLOv10 model exhibited a 72% false positive rate for Styrofoam buoys, attributed primarily to the significant impact of object color and shape. In the RT-DETR model, false positive rates were observed at 22% for seaweed and 20% for Styrofoam buoys, with object color and surface composition as key contributing factors. Based on these findings, it is recommended to consider the characteristics of the coastal and the distributed debris when selecting a deep learning model for coastal debris detection. Future studies on precise classification of coastal debris and diverse environmental data will facilitate the selection of optimal deep learning models for specific field conditions.