Related Experiment Video
Updated: Mar 28, 2026
10:16
Protocol for Microplastics Sampling on the Sea Surface and Sample Analysis
Published on: December 16, 2016
51.3K
Shapley Additive exPlanations-Integrated Convolutional Neural Networks for Chemically Interpretable Fourier-Transform
Ahmad Cahyono Adi1, Romanus Hadyanto Ongan1, Humairah Humairah1
1Department of Computer Science and Electronics, Universitas Gadjah Mada, Yogyakarta 55281, Indonesia.
Analytical Chemistry
|March 26, 2026
Summary
This study introduces CNN-SHAP fusion, an explainable AI framework for identifying microplastics using Fourier-transform infrared (FTIR) spectroscopy. The method enhances transparency in automated analysis, achieving high accuracy by interpreting spectral data effectively.
Area of Science:
- Environmental Science
- Analytical Chemistry
- Data Science
Background:
- Fourier-transform infrared (FTIR) spectroscopy is crucial for microplastic identification due to its sensitivity and non-destructive nature.
- Current FTIR-machine learning (ML) models often lack transparency, acting as black boxes and limiting confidence in automated microplastic analysis.
- Interpreting which spectral regions drive ML model decisions is essential for validating automated FTIR analyses.
Purpose of the Study:
- To develop an explainable deep-learning framework, CNN-SHAP fusion, for transparent interpretation of FTIR spectra in microplastic identification.
- To integrate one-dimensional convolutional neural networks (1D-CNNs) with SHapley Additive exPlanations (SHAP) for attribution-based analysis of FTIR data.
- To enhance the reliability and reproducibility of automated microplastic classification using FTIR spectroscopy.
Main Methods:
- Developed a novel framework, CNN-SHAP fusion, combining 1D-CNNs with SHAP for post-acquisition spectral interpretation.
- Utilized CNN-derived spectral embeddings and SHAP-weighted wavenumber representations as meta-features in an ensemble learning architecture.
- Employed standardized preprocessing on a balanced dataset of six common polymers (HDPE, LDPE, PET, PP, PS, PVC) under controlled conditions.
Main Results:
- Achieved a mean cross-validated classification accuracy of 99.6% for microplastic identification.
- Attribution analysis confirmed that model predictions are driven by diagnostically relevant spectral regions for polymer identification.
- Identified key spectral features influencing predictions, including C-H stretching, carbonyl bands, aromatic features, and fingerprint-region patterns.
Conclusions:
- CNN-SHAP fusion offers a transparent and reproducible method for FTIR-based microplastic classification in laboratory settings.
- The framework links predictive performance with spectrally interpretable attributions, enabling informed evaluation of model behavior.
- This approach provides a foundation for validating FTIR-based microplastic analysis under more complex, environmentally relevant conditions.
