A Deep-Neural-Network Based Early-Warning System for Blue-Green Algae Blooms from Sentinel-3 Satellite Imagery

Blue-green algal blooms are a major problem in bodies of water which can lead to deterioration of drinking water quality and to human health issues. Cyanobacteria, which have high concentrations in blue-green algal blooms, can produce harmful intracellular toxins (cyanotoxins, such as microcystins) which can cause skin and eye irritation, nausea, or more severe health issues upon ingestion. Therefore, it is important to be able to detect high cyanobacteria levels in bodies of water, especially when they are used as sources of drinking water. Traditional methods use in-situ sensors to measure levels of chlorophyll-a and phycocyanin, to determine the presence of such blooms. However, the deployment of such sensors is costly and also sometimes cumbersome. Recently, the use of satellite imagery has become another tool to determine the presence of algal blooms. The advantage of using such imagery is the large coverage area when compared to in-situ sensors, as well as their free availability. In this paper, we describe a study at Forest Pond, Newfoundland, Canada, where prediction of chlorophyll-a and phycocyanin levels is made and compared to measured levels from in-situ sensors. A machine learning method based on deep neural networks is introduced which is shown to determine such levels with fair accuracy as well as reduce the risk of false negatives when predicting the presence of blue-green algal blooms. This work has the potential to significantly reduce water testing costs by providing an alarm system for the presence of blue-green algae in bodies of water without the need for widespread in-situ testing, instead allowing testing to be performed only in a targeted manner where algal levels have been flagged by the alarm system as potentially being high.