Advancing marine debris counting during extreme weather events: Deep learning applications in Typhoons Saola and Haikui.

Abstract

Traditional sampling methods have been limited by the weather condition. If the a typhoon occurs in the study area, researchers can only collect samples before and after the event, as it is not possible to obtain data during the typhoon weather. In this study, we proposed the method noted as "Smart Debris Counting"(SDC), which integrated the deep learning and the shore-based fixed camera to investigate marine debris in the midst of a typhoon. With this approach, we collected the marine debris data and ten different algorithms was trained for it. The best-performing algorithm, which was evaluated on the dataset using mean Average Precision (mAP) and processing time, was selected for the continuous debris monitoring in the Dongshan Sea area during the typhoon event. The main results were as follows. (1) A new artificial intelligence algorithm was developed to effectively identify debris during extreme weather, which could achieve the mAP of 84.48 % and processing time of 0.2153 s/image. (2) This algorithm could realize the 8-days continuous collection of uninterrupted data, which collected 2080 images in total from 20 stations during the period of Typhoons Saola and Haikui. (3) Based on the SDC monitoring, the debris was increased by 8.3 % and 37 % respectively after Typhoon Saola and Haikui. Hence, using deep learning method to monitor marine debris is more efficient to acquire continuous-uninterrupted data, compared to some traditional sampling surveys. This is significantly valuable for understanding the spatiotemporal dynamics of debris distribution, clustering trends, and types within the region.