Large-area automatic detection of shoreline stranded marine debris using deep learning

Marine debris is a global crisis impacting human health, wildlife, and coastal economies. Remote sensing and geospatial technologies are an efficient way to document marine debris across large areas, but the high costs of fieldwork and manual interpretation of debris imagery are major barriers to broader use of these methods. Recent advances in machine learning (ML) have brought rapid automation to many remote sensing domains, including the detection and classification of marine debris. This study evaluates the ability to detect and classify coastline marine debris objects in a real-world setting across a large geographic extent. Three leading ML object detection models were trained to detect and classify large shoreline stranded marine debris from a set of aerial images collected over 1,900 km of Hawaiian coastline. The three models evaluated 1,587 image chips containing 10,703 individual debris labels from 8 debris classes. The SS-MN was both the fastest model and provided the best percentage of accurate predictions, achieving an average precision of 72 %. Nonetheless, this performance was achieved at the expense of missing numerous debris objects, with an average recall of 40 %. The other models provided distinct advantages for certain object classes and use cases. The results show that ML techniques show strong potential to automatically detect and classify certain types of marine debris from aerial surveys. However, there are existing methodological and technical challenges left to overcome before ML methods can outperform human observers in manual interpretation of marine debris from imagery.