Comparing Meanshift/SVM and Mask-RCNN algorithms for beach litter detection on UAVs images

Novel methodologies for beach litter monitoring take advantage of machine learning techniques applied to highresolution aerial-photogrammetric images acquired by unmanned aerial vehicles (UAV). In this study, a first attempt to classify multiple classes of litter items from very highresolution images derived from an aerophotogrammetric survey is proposed. For this purpose, a comparison between segmentation and classification algorithms is proposed for the automatic identification of beach macro-litter items (>2.5 cm). In particular, the Meanshift segmentation algorithm joined with the Support Vector Machine (SVM) classification algorithm were compared to a Convolutional Neural Network (Mask- RCNN). Two beach sectors, one on the Atlantic coast of Portugal and the other on the Adriatic coast of the Apulia region (Italy) were used as training site while test cases were conducted on the Italian site. Segmentation and classification phases were conducted considering three labels associated with three litter types. Results show a better performance of the Mask-RCNN algorithm with a precision value of 0.178 for 3 classes. The proposed comparison shows that algorithms based on Convolutional Neural Networks provide a useful tool for beach litter analysis programs and highlights how their training phase requires a specific survey plan to be followed during UAV image acquisitions.