Detecting stranded macro-litter categories on drone orthophoto by a multi-class Neural Network.

The use of Unmanned Aerial Systems (UAS, aka drones) images for mapping macro-litter in the environment have been exponentially increasing in the recent years. In this work, we developed a multi-class Neural Network (NN) to automatically identify stranded plastic litter categories on an UAS-derived orthophoto. The best results were assessed for items that did not have substantial intra-class colour variability, such as octopus pots and fishing ropes (F-score = 61%, on average). Instead, performance was poor (37%) for plastic bottles and fragments, due to their changing intra-class colours. On average, the performance improved 24% when the binary detection (litter/non-litter, F-Score = 73%) was considered, however this approach did not discriminate the litter categories. This work gives a new perspective for the automated litter detection on drone images, suggesting that colour-based approach can be used to improve the categorization of stranded litter on UAS orthophoto.