Onboard Near-Real-Time Detection of Harmful Algal Blooms Using a Convolutional Neural Network and Multispectral Imagery from a Satellite in Low Earth Orbit

Freshwater harmful algal blooms (HABs) pose significant ecological and public health risks worldwide. Detecting HABs soon after they form is critical to managing the damage they cause. While in-situ measurements are more accurate at detecting and measuring their toxicity levels, satellite imagery is more adept at capturing the spatial and temporal dynamics of these blooms over large geographic regions. Satellites can also more persistently monitor for HABs. In the past, empirical methods and machine learning methods have used multispectral satellite imagery to estimate HAB biomass. To build upon the current body of research, this paper investigates an approach to expedite HAB detection by utilizing a convolutional neural network (CNN) deployed onboard a CubeSat in low Earth orbit to detect HABs in near-real-time. The CNN is trained with multispectral imagery from the Sentinel-2 satellite constellation aggregated with in-situ cyanobacteria cell counts from the Seabass CAML dataset. The results successfully demonstrated the capability of a CNN to detect cyanobacterial blooms using multispectral imagery. After classifying HAB predictions into 5 severity classes, the best performing model achieved a RMSE of 1.33 between HAB severity levels. Training the CNN on 30m GSD imagery with RGB and red edge (B05) bands achieved a RMSE of 1.83 between HAB severity levels, which was inadequate for detecting HABs in small inland water bodies. Improved performance was observed with 10m ground sample distance (GSD) band combinations. The best performing networks utilized all of Sentinel-2's 10m and 20m spectral bands.