• School of Sciences and Engineering
  • September 2023
    166 p.
    • The increased desire of society for plastic products has led to plastic becoming omnipresent in the marine environment. Due to its toxicity, abundance, and persistence at the sea, plastic debris is of particular concern to marine life mainly owing to the elevated risk of entanglement or ingestion which may prove lethal. In this study, various machine learning (ML) and deep learning (DL) tools were implemented for classifying, detecting and counting plastic marine debris in images and video recordings. Primarily, the importance of this research lies in the intelligent and swift detection and counting of plastic litter which can facilitate litter monitoring surveys around the world and improve estimations of plastic marine debris density in remote geographical areas. Initially, the Bag of Features (BoF) method was used to construct a plastic debris image classifier, which attained a classification accuracy of 62.5% when applied to images demonstrating plastic debris and sea life from 8 classes. However, motivated by the need to enhance the sophistication of the proposed classifier, the DL tool of the Bottleneck method (BM) was employed. By expanding the number of marine debris object classes that the classifier can recognise from 3 to 8, the BM attained a 4% improvement in the validation accuracy, which topped 90%. Interestingly, when the resolution of the examined images was lowered by 75% of their original size, the accuracy of the BM remained unchanged. Beyond the accurate classification of marine debris, detecting multiple objects in images and videos is of paramount importance. Charged with this task, the YOLOv5 tool proved the most successful, among the YOLO family, as it attained real-time object detection of 34 frames per second on video footage and realised the highest mean average precision of 92.4% on images. Coupled with the counting tools of the region of interest (ROI) line and the centroid tracking, YOLOv5 proved competent in counting marine debris items from video footage. Particularly, the centroid tracking tool realised an accuracy of ≈80% when it processed a video illustrating plastic litter from 9 classes. Lastly, for estimating the plastic litter density across the Cypriot coastlines and the litter’s physical dimensions, the YOLACT++ tool was utilised as it applies a mask on each detected litter item. Inspecting images depicting plastic litter from six beaches in Cyprus, this pertinent detector deduced a plastic litter density of 0.035 items/m2. Extrapolating to the entire shorelines of Cyprus, the YOLACT++ tool estimated about 66,000 plastic items weighing a total of about 1,000 kg as explained in the pertinent video https://bit.ly/35TQVVE. Concluding, the dominant length of all documented plastic litter ranged from 10 to 30 cm.

    Utilising Artificial Intelligent (AI) Tools to Discern Plastic Debris at the Marine Environment

    1. PhD thesis
    2. english
      1. Artificial Intelligence -- Environment -- Plastic Litter