Detection of Harmful Algal Blooms in Lakes at the Regional Scale Using Satellite Remote Sensing and Machine Learning Techniques

Harmful algal blooms (HABs) in freshwater systems pose significant ecological and human health risks, necessitating effective monitoring and management strategies. These toxic blooms are being exacerbated by human impact due to climate change (global warming) and landcover changes resulting in worsening water quality. Traditional ground-based monitoring methods face limitations, including high costs, labor intensiveness, site-specificity, and disruptions due to unforeseen circumstances such as the recent pandemic. In this study, we propose a remote sensing approach to address these challenges, enabling cost-effective, widespread, and continuous monitoring of HABs in more than 35 lakes around New York State leveraging the existing in-situ data from various monitoring programs. In addition to a literature review covering different algorithms that are derived from simple band relationships to obtain chlorophyll-A (Chl-a) as a proxy for HABs presence, we test four such formulas on Sentinel-2 Multispectral Instrument (S2-MSI) Surface Reflectance (SR) data. There was a large range in accuracy as some formulas performed significantly better than others for different lakes. However, all exhibited ample error, leading us to incorporate different machine learning techniques including boosting and bagging models into an expanded study of several large lakes in NY State, including Lake Champlain. Various S2-SR bands were examined as input for the models and their importance was estimated using the in-situ data. The best-performing algorithm for Lake Champlain, our test region, Gradient Boosting, performed better than others with R2=0.85. This approach addresses the need for effective regional HABs detection and monitoring in freshwater systems through the incorporation of remote sensing and machine learning techniques.