Comprehensive Raman spectroscopy analysis for differentiating toxic cyanobacteria through multichannel 1D-CNNs and SHAP-based explainability

Abstract

Cyanobacterial blooms pose significant environmental and public health risks due to the production of toxins that contaminate water sources and disrupt aquatic ecosystems. Rapid and accurate identification of cyanobacterial species is crucial for effective monitoring and management strategies. In this study, we combined Raman spectroscopy with deep learning techniques to classify four toxic cyanobacterial species: Dolichospermum crassum, Aphanizomenon sp., Planktothrix agardhii and Microcystis aeruginosa. Spectral data were acquired using a confocal Raman microscope with a 532 nm excitation wavelength and subjected to preprocessing and filtering to enhance signal quality. We evaluated a multichannel one-dimensional convolutional neural network (1D-CNN) approach that incorporates raw spectra, baseline estimations, and preprocessed spectra. This multichannel approach improved overall classification accuracy, achieving 86% compared to 74% with a traditional single-channel 1D-CNN using only preprocessed spectra while maintaining low overfitting. Shapley Additive exPlanations (SHAP) were applied to identify critical spectral regions for classification to enhance interpretability. These findings highlight the potential of combining Raman spectroscopy with explainable deep learning methods as a powerful tool for water quality monitoring and the early detection of Harmful Algal Blooms (HABs).