Spectroscopic and machine learning strategies for automated classification and quantification of textile blends: Enabling material recovery and circular economy in the textile sector

Accurate identification and classification of textile materials are vital for improving recycling efficiency, ensuring quality control, and advancing circular economy practices in the textile industry. Traditional sorting technologies struggle to differentiate between complex textile blends, particularly when compositions contain low-percentage additives or chemically similar fibers. This study investigates the integration of Near-Infrared Spectroscopy (NIRS) with machine learning techniques to classify and quantify both pure and blended textile fibers across diverse composition ratios. The dataset encompasses pure fibers including Cotton, Polyester, Polyamide, Modal, Lyocell, Viscose, Wool, Silk, Linen, Polypropylene, Acrylic, and Elastane alongside binary blends (e.g., Cotton–Elastane, Cotton–Modal, Cotton–Polyester, Polyester–Elastane, and Polyamide–Elastane). Support Vector Machines (SVM), Random Forests (RF), and k-Nearest Neighbors (k-NN) were evaluated for classification, while SVR, RF Regression, and k-NN Regression were applied to predict blend ratios, supported by Principal Component Analysis (PCA) for spectral feature exploration. Comparative evaluations revealed complementary model strengths rather than a single dominant architecture: while SVM achieved superior performance in primary fiber classification tasks (reaching up to 97% accuracy under external validation and 96% under Leave-One-Out Cross-Validation), ensemble-based RF and distance-driven k-NN exhibited marked advantages in handling non-linear blend dynamics, imbalanced minority classes, and challenging low-concentration additives (2–10% elastane). For quantitative composition prediction, Random Forest Regression achieved exceptional precision with R 2 values exceeding 0.99 (e.g., for Polyester–Elastane blends), confirming its efficacy for compositional quantification. These findings demonstrate that integrating NIRS with tailored machine learning architectures enables both reliable automated identification and precise quantitative sorting, offering a robust framework for high-throughput sensor-based sorting systems essential for circular material recovery.