Coupling hyperspectral imaging and deep learning to detect bloom-forming toxic cyanobacteria in mixed assemblages

• Microcystis, Dolichospermum and Chrysosporum were mixed in unequal proportions. • A hyperspectral imagery library was created with the pure and mixed biomass. • Derived spectral mixtures had unique and distinguishable features in the VIS-NIR. • Artificial Neural Networks outperformed Random Forest in classifying cyanobacteria. • Detecting cyanobacteria even at low proportions (6%) still achieved 95% accuracy. Cyanobacterial blooms, particularly those dominated by toxin-producing species, pose critical risks to ecosystems and public health. Monitoring strategies have evolved significantly in the last decades, but still lack near real-time capabilities for effective risk estimations. This study explores the potential of combining hyperspectral imagery (HSI) with machine learning and deep learning to detect the presence of specific cyanobacterial taxa in complex mixtures, even when present at low proportions. This work simulates a real-world scenario in which different cyanobacterial taxa coexist in varying proportions. A HSI library was created from pure cyanobacterial cultures and combined in binary mixtures of three species representative of bloom-forming potentially toxic cyanobacterial genera: Microcystis aeruginosa, Chrysosporum ovalisporum , and Dolichospermum crassum . Reflectance spectra from the visible to near-infrared (VIS-NIR) range of each image were extracted, randomly grouped and filtered. The resulting spectra were preprocessed and used to develop Random Forest (RF) and Neural Network (NN) classification models intended to classify these mixtures. Hyperparameter optimization was performed to achieve the best configuration. Learning curves and classification metrics were utilized to monitor the overall training process as well as the model performance. The resulting models demonstrated remarkable performance, achieving 91-95% accuracy in classifying pure and mixed assemblages, and 85-90% in defining the proportion of each species, with NNs consistently outperforming RFs by 4-6% in both tasks. Even cyanobacteria present at low proportions (e.g., 6%) were effectively detected and quantified. These findings have important implications for improving and automating cyanobacterial monitoring and risk assessment in current water management strategies.