Effects of surface contamination on automated textile sorting using NIR-spectroscopy

As of January 2025, all EU member states must implement separate textile collection, but contamination (e.g., dirt, moisture) remains a challenge for sorting. NIR spectroscopy is a promising technology for automated textile sorting, though issues like contaminated materials still hinder its performance. This study showcases how contamination affects the identification of textiles by analyzing samples from residual waste with varying contamination levels. The impact of contamination on the classification accuracy of cotton-rich, polyester-rich, and polycotton textiles was investigated, as well as the resulting spectral differences and the identification of contaminants. Results indicate that moisture presence significantly influences spectral behavior, affecting the classification models' performance. Excluding the most moisture-sensitive wavelengths (1368-1459 nm) improved classification accuracy for polyester-rich samples. However, this exclusion further complicated the already challenging classification of polycotton samples, while cotton-rich accuracy experienced only a slight decline. Consequently, the average accuracy dropped from 83.1% to 78.9%. These findings suggest that excluding moisture-sensitive wavelengths can improve classification accuracy for specific material types (e.g., polyester-rich) but may increase misclassification risk for others (e.g., polycotton and cotton-rich), highlighting the need to tailor models to target fractions. While other contaminants affected classification less strongly, heavier surface contamination may require lowering the classification threshold to limit misclassification risk. This work can aid in improving NIR classification of contaminated post-consumer textiles and thereby support the circular economy of one of the fastest growing waste streams by enabling the recovery of resources that would otherwise be lost.