FTIR-Based Textile Fiber Classification

Machine learning classification of textile fibers using FTIR spectroscopy and chemometric preprocessing

---

Overview

This repository contains the complete workflow for automated textile fiber classification using Fourier-Transform Infrared (FTIR) spectroscopy. The project implements multiple preprocessing pipelines and evaluates classification performance using Support Vector Machines (SVM) and Random Forest algorithms.

Key Features:

• Chemometric preprocessing (ALS baseline correction, SNV normalization, Savitzky-Golay derivatives)
• Multi-pipeline comparison (P1: ALS+SNV, P2: ALS+SNV+D1, P3: ALS+SNV+D2)
• Binary classification (Natural vs Man-made) and multiclass classification (fiber types/subtypes)
• Automated batch processing script for new samples
• Comprehensive visualization and analysis notebooks

---

Repository Structure

Experiments/
├── initial_analysis.ipynb                    # Exploratory data analysis and experimentation
├── initial_preprocessing_experiments.ipynb   # Preprocessing method comparison
├── initial_ml_classification.ipynb           # ML model development and testing
├── data_analysis_and_figures.ipynb           # Final dataset analysis checks and visualisations
├── pretreatments_and_classification.ipynb    # Final preprocessing pipeline evaluation
├── spectral_visualisation.ipynb              # Spectral plotting utilities
│
├── Preprocessing Pipeline - Script/          # Standalone batch processing tool
│   ├── process_spectra_pipeline.py           # Main script
│   ├── README.md                             # Script documentation and instructions
│   ├── exported_csvs/                        # Raw FTIR CSV input
│   ├── metadata/                             # Sample metadata
│   └── ml_datasets/                          # Generated feature matrices
│
├── raw_csv_data/                             # Raw FTIR spectra (CSV format) from PerkinElmer Spectrum exports
├── ml_datasets/                              # Preprocessed datasets for ML
├── images/                                   # Output figures
└── requirements.txt                          # Python dependencies

---

Datasets

Fiber Collections:

• Microtrace Forensic Fiber Reference Collection (synthetic fibers)
• Microtrace Arbidar Natural Fibre Collection (natural fibers)
• Bio-Couture & UNUSUWUL, associated with IMPACT+ (assorted fibers)

Sample Coverage:

• Natural fibers: Cotton, Linen, Jute, Silk, Wool
• Man-made fibers: Polyester, Nylon, Acrylic, Modacrylic, Regenerated Cellulose

Spectral Data:

• Spectral range: 4000-400 cm⁻¹
• Resolution: 4 cm⁻¹
• Multiple replicas per sample (1-3 scans)

---

Preprocessing Pipelines

Pipeline	Methods	Output Format	Use Case
P1	ALS + SNV	Absorbance	Baseline classification
P2	ALS + SNV + D1	1st derivative	SOTA
P3	ALS + SNV + D2	2nd derivative	SOTA, Comparative analysis

Preprocessing Details:

• ALS (Asymmetric Least Squares): Baseline correction (λ=1e6, p=0.001)
• SNV (Standard Normal Variate): Scatter normalization
• Savitzky-Golay: Derivative computation (window=15, polynomial=3)

---

Classification Performance

Binary Classification (Natural vs Man-made):

• Cross-validation accuracy: 98.5% ± 1.2%
• Test accuracy: 100% (15/15 samples)
• Algorithm: SVM with RBF kernel

Multiclass Classification (12 fiber subtypes):

• Cross-validation accuracy: 95.8% ± 2.1%
• Test accuracy: 93.3% (14/15 samples)
• Algorithm: Random Forest (200 trees)

---

Quick Start

Installation

# Clone repository
git clone https://github.com/reeha-parkar/FasTEX-KCL.git
cd FasTEX-KCL/

# Install dependencies
pip install -r requirements.txt

Workflow

1. Exploratory Analysis: Run initial_analysis.ipynb (For experimentations on different types of chemometric preprocessing techniques)
2. Preprocessing Evaluation and Classification: Run pretreatments_and_classification.ipynb
3. Results Visualization: Run data_analysis_and_figures.ipynb

Batch Processing New Samples

cd "Preprocessing Pipeline - Script"
python process_spectra_pipeline.py

See Preprocessing Pipeline - Script/README.md for detailed instructions.

---

Requirements

• Python 3.12+
• NumPy, Pandas, SciPy
• Matplotlib, Seaborn
• scikit-learn

See requirements.txt for complete list with versions.

---

Citation

If you use this work in your research, please cite:

Dataset:

@dataset{parkar2025dataset,
  author = {Parkar, Reeha and Jain, Angelica and Prendergast-Miller, Miranda and Stanton, Thomas and Sheridan, Kelly and Gallidabino, Matteo},
  title = {A dataset of infrared (ATR-FTIR) spectra for textile fibres of natural and man-made origin},
  year = {2025},
  publisher = {Mendeley Data},
  version = {V1},
  doi = {10.17632/rx3fjgz96x.1},
  url = {https://doi.org/10.17632/rx3fjgz96x.1}
}

Note: A data descriptor paper for this dataset is currently under review at Data in Brief journal. This citation will be updated upon publication.

Code Repository:

@misc{parkar2025ftir,
  author = {Parkar, Reeha},
  title = {FTIR-Based Textile Fiber Classification using Machine Learning},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/reeha-parkar/ftir-fiber-classification}
}

---

Author

Reeha Karim Parkar
ML Research Inern, King's College London
📧 reeha_karim.parkar@kcl.ac.uk | reehaparkar@gmail.com
🔗 GitHub

Supervisor/PI: Dr. Matteo Gallidabino, Department of Forensic Science, King's College London

Funding: IMPACT+

---

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Acknowledgments

• Microtrace LLC for the Forensic and Natural Fibre Collection
• IMPACT+
• King's College London, Department of Forensic Science

---

Last Updated: December 16, 2025