Optimization of a hybrid microplastic model method based on grad-CAM combined with convolutional neural networks

Area of Science:

• Analytical Chemistry
• Environmental Science
• Computer Science

Background:

• Microplastic identification accuracy is limited, especially for mixed samples.
• Deep learning models often lack transparency, hindering trust and optimization.
• Attenuated Total Reflectance Fourier-Transform Infrared (ATR-FTIR) spectroscopy presents complex spectral data challenges.

Purpose of the Study:

• To improve the accuracy and interpretability of deep learning models for mixed microplastic identification.
• To address feature extraction imbalances in deep learning models for complex spectral data.
• To develop a novel model for rapid and accurate microplastic detection.

Main Methods:

• Gradient Weighted Class Activation Mapping (Grad-CAM) was used for model interpretability analysis.
• Multi-head attention and multi-region attention mechanisms were incorporated to enhance feature extraction.
• A Convolutional Neural Network (CNN) with Resblock residual connections was designed, forming the ResM2A-CNN model.

Main Results:

• The optimized ResM2A-CNN model achieved an accuracy of 99.92%, a significant improvement from 95.67%.
• The model demonstrated superior classification performance for Polyamide (PA) and its mixtures compared to traditional models.
• Interpretability analysis successfully guided model optimization, revealing feature extraction imbalances.

Conclusions:

• Interpretability analysis is effective in guiding deep learning model optimization for spectral identification.
• The ResM2A-CNN model offers a breakthrough for rapid and accurate microplastic detection.
• This research provides valuable insights into modeling complex spectral data and optimizing deep learning approaches.