Comprehensive Raman spectroscopy analysis for differentiating toxic cyanobacteria through multichannel 1D-CNNs and...

Area of Science:

• Environmental Science
• Biotechnology
• Spectroscopy

Background:

• Cyanobacterial blooms pose risks to water quality and public health due to toxin production.
• Accurate identification of cyanobacterial species is essential for effective monitoring and management of harmful algal blooms (HABs).

Purpose of the Study:

• To develop and evaluate a Raman spectroscopy-based deep learning method for classifying four toxic cyanobacterial species.
• To improve the accuracy and interpretability of cyanobacterial species identification for HAB monitoring.

Main Methods:

• Acquisition of Raman spectra from four toxic cyanobacterial species using confocal Raman microscopy (532 nm excitation).
• Application of a multichannel one-dimensional convolutional neural network (1D-CNN) incorporating raw, baseline, and preprocessed spectral data.
• Utilizing Shapley Additive exPlanations (SHAP) for spectral region interpretability.

Main Results:

• The multichannel 1D-CNN achieved 86% classification accuracy, outperforming a single-channel 1D-CNN (74%).
• The multichannel approach demonstrated reduced overfitting compared to traditional methods.
• SHAP analysis identified key spectral regions crucial for accurate species classification.

Conclusions:

• Combining Raman spectroscopy with explainable deep learning offers a powerful tool for water quality monitoring.
• This approach facilitates early detection and identification of harmful algal blooms.
• The developed method enhances the accuracy and interpretability of cyanobacterial species classification.