Harmful Algal Bloom (HAB) Detection Using Deep Convolutional Neural Networks and Remote Sensing Data

A total of 198 Sentinel-2 Level-2A datasets over a 13,000-acre reservoir in Virginia were used to determine if deep convolutional neural networks (CNNs) could detect the presence of harmful algal blooms (HABs) – an event during which cyanobacteria replicate at abnormal rates and create toxins harmful to living organisms. The water collection campaign focused on over 45 sites throughout the lake and produced the bulk of the ground-truthed data. The Sentinel-2 datasets provided the spectral data used to match to laboratory spectral measurements. Finally, CNNs were used to learn and detect the spectral characteristics of HABs on remote sensing data. The proposed approach to the development of the CNN and the curating of the Sentinel-2 data showed that CNNs could detect harmful algal blooms with a 93% accuracy after at least 50 epochs.