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

To address the low accuracy of spectral identification of mixed microplastic samples and the "black box" nature of deep learning models, this paper proposes an optimization method based on Gradient Weighted Class Activation Mapping (Grad-CAM). Model interpretability analysis using Grad-CAM reveals an imbalance in the feature extraction process-insufficient attention to the features of minor components in mixed samples. Based on these findings, multi-head attention and multi-region attention mechanisms are introduced to enhance the model's ability to extract global and local features. Considering the complex and diverse ATR-FTIR spectral features of microplastics and the gradient vanishing problem in deep learning models, Resblock residual connections are introduced on top of CNN, and a ResM2A-CNN model is designed. Experimental results show that the overall accuracy of the optimized model on the test set is improved from 95.67% to 99.92%. Notably, the ResM2A-CNN model exhibits strong classification ability for PA and its mixtures, outperforming other traditional models in various performance metrics.This study highlights the effectiveness of using interpretability analysis to guide model optimization, providing a breakthrough method for the rapid and accurate detection of microplastics. Furthermore, this research offers practical insights into modeling complex spectral data and optimizing deep learning models.