Mapping and quantifying pelagic Sargassum in the Atlantic Ocean using multi-band medium-resolution satellite data and deep learning

Abstract

Pelagic Sargassum in the Atlantic Ocean is an important habitat for marine animals, yet frequent and massive beaching events around the Caribbean Sea have caused many problems. Currently, both retrospective assessment and near real-time monitoring over synoptic scales rely on medium-resolution multi-band satellite imagery and algorithms based on the alternative floating algae index (AFAI) or its alternatives. Despite the success in assessing long-term bloom patterns and providing synoptic information in near real-time, two weaknesses still exist: 1) due to spatial resolution limitations and optical complexity from the ocean, there are often false positive detection in nearshore waters (up to 30 km from shoreline). These waters are currently masked to avoid providing false information, thus leading to data gaps in nearshore waters; 2) due to perturbations by cloud-adjacent straylight and cloud shadows, both false positive and false negative detections are often encountered. The current solution is to dilate the cloudmask to make the adjacent pixels invalid, thus leading to data loss. In this work, a computer deep-learning model based on the Res-UNet architecture, namely U-net, was developed to overcome these difficulties. The model was trained and validated using a total of 8518 medium-resolution “ground truth” images carefully prepared from data collected by the Moderate Resolution Imaging Spectroradiometer (MODIS) onboard the Terra and Aqua satellites. After accounting for subpixel Sargassum coverage, the over accuracy of the DL model, as measured by its F1 score, is ∼92.5% (without masking nearshore waters or cloud-adjacent pixels), which is higher than that of the current AFAI model (86.2% after masking nearshore waters and cloud-adjacent pixels). Application of the DL model to the 2021 MODIS data (a major Sargassum year) showed substantially improved valid data coverage (monthly increases of 20–34%) in both nearshore waters and offshore waters, and increased Sargassum biomass during the summer months (by 10–20%) from the entire Sargassum belt extending from the west Africa to the Gulf of Mexico. In the minor Sargassum year of 2016, the DL model led to an average increase of about 7% in Sargassum biomass. The DL model also appears to be applicable to the Visible Infrared Imaging Radiometer Suite (VIIRS) from the Suomi National Polar-orbiting Partnership (SNPP) satellite and the Ocean Land Colour Instrument (OLCI) onboard the Sentinel 3A and 3B satellites. Merging multi-sensor observations (including MODIS) resulted in substantially improved valid data coverage in daily to weekly composites, thus improving near real-time monitoring.