Two Algorithms for Sargassum Segmentation in Citizen Science Images

The arrival of large quantities of Sargassum to the Caribbean beaches, mainly since 2018, is an emerging problem that has generated ecological, economic, and health challenges in the coastal localities of the region. Thus Sargassum monitoring has increased on the coasts and beaches to study and measure this phenomenon for better management and exploitation. Since 2018, Machine Learning methodologies have been incorporated to support the traditional techniques used for Sargassum monitoring, aligned with the context of this study. Sargassum monitoring in situ is relevant for the improvement of satellite and proximal remote sensing techniques. A dataset of 938 images was constructed, showing Sargassum on the beach, using photographs from a crowdsourcing scheme. This dataset was used for training and testing the u-net and pix2pix architectures to identify similarities and differences in the results when evaluated with the same subset of test data. The results show that both architectures demonstrate acceptable performance in image segmentation despite the small and unbalanced dataset. The performance of both architectures was very similar; therefore, implementation choice should also depend on the final purpose. From the perspective of the dataset used, this analysis is one of the first of its kind.