Detecting harmful algal blooms from reconstructed hyperspectral satellite data

Harmful algal blooms (HABs) are increasingly affecting marine, coastal and inland water bodies due to excessive nutrient input, anthropogenic pollution,= and climate change. Thus, accurate detection and classification of HABs is crucial for monitoring their impact and implementing mitigation strategies. Remote sensing has proven to be an effective tool for continuous monitoring and detection of HABs. However, existing remote sensing techniques face significant challenges, particularly due to the limited spectral resolution of multispectral sensors. To address this, a novel two-stage deep learning framework based on a one-dimensional convolutional neural network (1D-CNN) is proposed in this study. The first 1D-CNN model reconstructs hyperspectral reflectance spectra from Sentinel-2 multispectral data. This model was trained using data from the Hyperspectral Imager for the Coastal Ocean (HICO) and validated against the GLObal Reflectance community dataset for Imaging and optical sensing of Aquatic environments (GLORIA). The model results showed consistently high PCC values (>0.98) and low RMSE and MAE, indicating strong agreement between reconstructed and in-situ spectra, and effective preservation of key spectral features. Then, the second 1D-CNN model utilizes the reconstructed hyperspectral data to classify three commonly occurring HABs, namely, Noctiluca scintillans, Trichodesmium erythraeum and Microcystis aeruginosa. The model achieved an overall accuracy of 97.3%, with a Macro-F1 score: 0.970 and a Weighted-F1 score: 0.973, demonstrating high classification precision across all classes. The low misclassification rate (2.7%) highlights the model’s robustness and the effectiveness of spectral features derived from reconstructed hyperspectral data in capturing class-specific variability. Furthermore, the model’s applicability was verified using Sentinel-2 data acquired over different marine environments around the world. This study demonstrates the potential of hyperspectral reconstruction to enhance HAB classification where hyperspectral sensors are unavailable, thereby supporting real-time monitoring and global efforts to mitigate the ecological and economic impacts of HABs.