Automatic detection of seafloor marine litter using towed camera images and deep learning. (March 2021)

• Record Type: Journal Article
• Authors: Politikos, Dimitris V.; Fakiris, Elias; Davvetas, Athanasios; Klampanos, Iraklis A.; Papatheodorou, George
• Is Part Of: Marine pollution bulletin. Volume 164(2021)
• Publication Date: 2021-03
• Subjects: Seafloor marine litter -- Object detection -- Mask R-CNN -- Deep learning -- Aegean Sea -- Mediterranean Sea; Marine pollution -- Periodicals; Marine Biology -- Periodicals; Water Pollution -- Periodicals; Mer -- Pollution -- Périodiques; Publications périodiques; Pollution des mers; Lutte antipollution; Electronic journals; 363.7394
• DOI: 10.1016/j.marpolbul.2021.111974
• Languages: English
• ISSNs: 0025-326X

Abstract

Abstract: Aerial and underwater imaging is being widely used for monitoring litter objects found at the sea surface, beaches and seafloor. However, litter monitoring requires a considerable amount of human effort, indicating the need for automatic and cost-effective approaches. Here we present an object detection approach that automatically detects seafloor marine litter in a real-world environment using a Region-based Convolution Neural Network. The neural network is trained on an imagery with 11 manually annotated litter categories and then evaluated on an independent part of the dataset, attaining a mean average precision score of 62%. The presence of other background features in the imagery (e.g., algae, seagrass, scattered boulders) resulted to higher number of predicted litter items compare to the observed ones. The results of the study are encouraging and suggest that deep learning has the potential to become a significant tool for automatically recognizing seafloor litter in surveys, accomplishing continuous and precise litter monitoring. Highlights: Object detection was used to automatically detect seafloor marine litter. The litter imagery was acquired from Greek waters. Background features in the litter imagery (seagrass, rocks, shadings) affected performance. Results show a mean average precision of 62%.