ADDRESSING THE CLASS IMBALANCE PROBLEM IN THE AUTOMATIC IMAGE CLASSIFICATION OF COASTAL LITTER FROM ORTHOPHOTOS DERIVED FROM UAS IMAGERY
Keywords: garbage, mapping, marine, convolutional neural networks, oversampling, class weighting, classifier thresholding
Abstract. Unmanned Aerial Systems (UAS) has been recently used for mapping marine litter on beach-dune environment. Machine learning algorithms have been applied on UAS-derived images and orthophotos for automated marine litter items detection. As sand and vegetation are much predominant on the orthophoto, marine litter items constitute a small set of data, thus a class much less represented on the image scene. This communication aims to analyse the class imbalance issue on orthophotos for automated marine litter items detection. In the used dataset, the percentage of patches containing marine litter is close to 1% of the total amount of patches, hence representing a clear class imbalance issue. This problem has been previously indicated as detrimental for machine learning frameworks. Three different approaches were tested to address this imbalance, namely class weighting, oversampling and classifier thresholding. Oversampling had the best performance with a f1-score of 0.68, while the other methods had f1-score value of 0.56 on average. The results indicate that future works devoted to UAS-based automated marine litter detection should take in consideration the use of the oversampling method, which helped to improve the results of about 7% in the specific case shown in this paper.