20 JUL 2026 (MON) 16:05 - 16:35

Improving Marine Plastic Debris Extraction Accuracy from Sentinel-2 Imagery Using an Object-Based Random Forest Classifier

Mr. YANG Chenjing    

( Supervisor: Prof Jimmy Li )

Abstract:

Marine plastic pollution has become a pressing global environmental issue, necessitating effective monitoring methods. Remote sensing, particularly Sentinel-2 imagery, offers a promising avenue for detecting floating debris. However, distinguishing plastic from spectrally similar features like waves, seaweed, and organic material remains challenging. This study develops an object-based Random Forest (RF) classification method using the Marine Debris Archive (MARIDA) dataset to improve the extraction of floating marine plastics from Sentinel-2 imagery. We compared the performance of this object-based approach against a traditional pixel-based classification method. Following image segmentation and feature extraction—including spectral indices, band statistics, and geometric properties—an RF model was trained and validated on MARIDA dataset. Results demonstrate that the object-based model significantly outperforms the pixel-based model, achieving a Producer’s Accuracy (PA) of 0.83 and User’s Accuracy (UA) of 0.94 for Marine Debris on the final test dataset, compared to a PA of only 0.57 for the pixel-based method. While misclassifications with waves and coastal waters persist, the object-based approach effectively reduces false negatives by integrating spatial and contextual information. Furthermore, a novel algorithm utilizing relative positional features is proposed to potentially differentiate wave objects from plastic debris based on their distinctive distribution patterns. This study highlights the advantage of object-based machine learning frameworks for remote sensing of marine debris and suggests pathways for further refining classification accuracy.