We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Microplastic Identification via Raman Spectroscopy: A Three‐Channel Feature Fusion Convolutional Neural Network
Original title: Microplastic Identification via Raman Spectroscopy: A Three‐Channel Feature Fusion Convolutional Neural Network
Summary
Scientists have developed a smarter AI tool that can identify tiny plastic particles with 99% accuracy, even in messy real-world samples like water or soil where other detection methods often fail. This matters because microplastics are increasingly found in our food, water, and even our bodies, and better detection tools help researchers track pollution levels and understand our actual exposure—an important step toward figuring out what these particles might mean for our health.
ABSTRACT The rapid and accurate identification of microplastics is critical for environmental monitoring and pollution control. While Raman spectroscopy is widely utilized for microplastic detection due to its nondestructive nature and high specificity, traditional classifiers and standard convolutional neural networks often struggle with insufficient feature representation and limited generalization when faced with complex background interference and spectral similarity. To address these challenges, we propose a novel Three‐Channel Feature Fusion Convolutional Neural Network (TC‐FFCNN) integrating attention mechanisms. The model constructs multidimensional feature inputs by fusing raw spectra with their first‐ and second‐order derivatives. Concurrently, it combines a multiscale convolutional architecture with attention mechanisms to achieve the adaptive enhancement of key spectral bands. Under standardized experimental conditions, the proposed model was systematically compared against other models, including KNN‐PCA, SVM‐PCA, random forest, PLS‐DA, and ResNet. Results demonstrate that TC‐FFCNN outperforms all comparison models, achieving an average precision, recall, and accuracy of 0.990 on the standard dataset. Furthermore, ablation studies validate the critical roles of both the multichannel inputs and the attention mechanisms in enhancing overall model performance. When evaluated on more challenging environmental samples, the performance of the above mentioned other models has significantly declined; however, TC‐FFCNN still maintained a high classification accuracy of 0.989, demonstrating superior robustness and generalization capabilities. Ultimately, this study establishes that the proposed TC‐FFCNN effectively improves the recognition accuracy of microplastic Raman spectra and maintains stable performance under complex conditions, providing a reliable technical framework for the rapid detection of microplastics in practical applications.