Detecting Harmful Algal Blooms in the Gulf of Maine using a Hybrid Model

Harmful algal blooms are a growing threat to marine ecosystems, aquaculture, public health and tourism industries. This study quantifies the value of augmenting simulated outputs of a regional hydrodynamic model with satellite data input to detect harmful algal blooms using machine learning model in Gulf of Maine. And evaluates performance using in-situ Imaging FlowCytobot observations spanning 2018–2025 for five toxin producing algal taxa. Outputs of a regional ocean hydrodynamic model, including sea surface temperature and salinity, were incorporated as predictors with spatio-temporal matched in-situ algal cell concentrations as target labels. Satellite-derived water-leaving reflectance were produced using three atmospheric correction algorithms, enabling systematic comparison of spectral features. ACOLITE atmospheric correction algorithm consistently outperformed OC-SAC and C2RCC in terms of average F1-score across all five taxa. Using hydrodynamic outputs improved detection across three atmospheric correction processes, with F1-score improvements ranging from 2.9% (ACOLITE) to 9.2% (C2RCC); where the best hybrid configuration (ACOLITE + ROMS) achieved average F1-score of 0.745 and provided highest detectability for Karenia spp. with F1-score of 0.871. This framework advances operational harmful algal bloom monitoring by demonstrating that atmospheric correction quality and physical oceanographic context act as independent, partially substitutable performance levers for coastal water quality assessment in optically complex nearshore environment.