Harmful Cyanobacterial Bloom Detection Using Deep Learning and Sentinel-2 Imagery

Harmful cyanobacterial blooms (CyanoHABs) pose a growing threat to human health and aquatic ecosystems, making early detection and monitoring critical. Current satellite-based approaches, such as the Cyanobacteria Assessment Network (CyAN) using the Sentinel-3 OLCI sensor, are optimized for chlorophyll <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a</i> but are limited by a coarse 300 m pixel size. While Sentinel-2 lacks the spectral specialization of Sentinel-3, it provides finer sampling (10–60 m), enabling monitoring of smaller lakes and rivers if suitable models are applied. Here, we present a method to generate a large, geographically diverse CyanoHAB dataset by aligning CyAN data with Sentinel-2 imagery, resulting in 938,607 images and 1.86 billion labeled pixels across the continental United States. Using this dataset, we train a deep learning model that predicts CyanoHAB risk levels with 86.2% accuracy. Bands 4 (665 nm) and 5 (705 nm) are most influential, but randomizing any of 11 bands reduces accuracy by at least 10% in one or more classes, demonstrating the importance of the full spectral range. Both the dataset and code are publicly available (<uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://zenodo.org/records/14230064</uri>; <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/cschloer/hab_detection_s2</uri>).