Sargassum detection and path estimation using neural networks

Abstract

Sargassum has affected the Mexican Caribbean coasts since 2015 in atypical amounts, causing economic and ecological problems. Removal once it reaches the coast is complex since it is not easily separated from the sand, damaging dune vegetation, heavy transport compacts the sand and further deteriorates the coastline. Therefore, it is important to detect and estimate the sargassum mats path to optimize the collection efforts in the water. There have been some improvements in systems that rely on satellite images to determine areas and possible paths of sargassum, but these methods do not solve the problems near the coastline where the big mats observed in deep sea end up segregating in little mats which often do not show up in the satellite images. Besides, the temporal scales of nearshore sargassum dynamics are characterized by finer temporal resolution. This paper focuses on cameras located near the coast of Puerto Morelos reef lagoon where images are recorded of both beach and near-coastal sea. First, we apply preprocessing techniques based on time that allows us to discriminate the moving sargassum mats from the static sea bottom, then, using classic image processing techniques and neural networks we detect, trace, and estimate the path of the mat towards the place of arrival on the beach. We compared classic algorithms with neural networks. Some of the algorithms we tested are k-means and random forest for segmentation and dense optical flow to follow and estimate the path. This new methodology allows to supervise in real time the demeanor of sargassum close to shore without complex technical support.