| dc.description.abstract | Humans’ awareness of the environment is increasing nowadays; however, bottle marine debris (BMD) remains one of the most pressing global issues. Fields surveys of marine debris based on manpower is less efficient; therefore, this study proposes an automatic detection method on BMD using unmanned aerial vehicles and machine learning techniques.
The study sites are located on sandy beaches in Dayuan District, Taoyuan City. We first set three designed sites to create training datasets and test the detecting algorithm, and performances. Two real sites were then surveyed to evaluate our method in such a sandy complex beach that was intended to be used for long-term researches. The UAVs were operated at different fly heights to capture images with resolutions from 0.12 to 1.54 cm/pixel. The object detection algorithm You Only Look Once version 2 (YOLO v2) was trained to identify BMD and we added an image processing skill to remove image background noises. Data augmentation was used in training process to increase training data, and intersection over union (IoU) was adopted to evaluate the training efficiency. The results reveal that the skill of data augmentation helps IoU reaches over 0.68; and the skill of background removal has an advantage to reduce the processing time, as well as reducing noise resulting in much greater precision in real sites. From testing on both the designed and real sites with different image resolutions and processing skills, we found that approximately 0.5 cm/pixel could be the optimal resolution for aerial surveys on BMD. When operating the UAV with an image resolution of 0.5 cm/pixel, the performance indexes of mean precision, recall rate, and F1-score are respectively, 0.94, 0.97 and 0.95 at designed sites and are 0.61, 0.86, and 0.72 at real sites.
Our work contributes to advances in beach debris surveys, optimizes the automatic detection on machine learning approach, especially with the role of data augmentation step in training data and background removing procedure. | en_US |