Filtering out false Sargassum detections using context features

Since 2011, the distribution extent of pelagic Sargassum algae has substantially increased and now covers whole Tropical North Atlantic Ocean, with significant inter-annual variability. The ocean colour imagery been used as only way to monitor regularly such a vast area. However, detection is hampered by cloud masking, sunglint, coastal contamination other phenomena. All together, they lead false detections that can hardly be discriminated classic radiometric analysis, but may overcome considering shape context detections. Here, we built machine learning model base exclusively on spatial features filter out after process. Moderate-Resolution Imaging Spectroradiometer (MODIS, 1 km) data from Aqua Terra satellites were generate daily map Alternative Floating Algae Index (AFAI). Based this index, presence in Ocean was inferred. For every aggregations, five contextual indices extracted (number neighbours, surface temporal persistence, distance coast aggregation texture) then random forest binary classifier. Contextual at large-scale most important Trained multi-annual (2016-2020) set, performs filtering an accuracy ~ 90%. This leads reduction detected pixels 50% over domain. method provides reliable while preserving high resolutions (1 km, daily). resulting consistent literature for seasonal fluctuations, maximum coverage 2018 minimum 2016. dataset will useful understanding drivers dynamics fine large scale validate future models. methodology here demonstrates usefulness complementing classical remote sensing approaches. Our could easily adapted datasets containing erroneous