Accurate discrimination and quantification of cellulosic fibers using NIR hyperspectral imaging and optimised machine learning

Near-infrared hyperspectral imaging (NIR-HSI) enables non-destructive acquisition of spectral signatures that reflect physicochemical properties of textile fibers. However, for cellulose-based fibers, a major category of textile materials, the high similarity in their spectral response poses significant challenges for reliable identification and quantitative analysis. Here, we present specifically designed compact machine learning (ML) models, including the emerging Kolmogorov-Arnold network (KAN), to detect subtle variations in NIR spectra for two central tasks: (i) determining the blend ratio of cellulosic blend and (ii) classifying eight pure cellulosic fibers, including ecological materials. The spectral range 1410–1610 nm was identified as region of interest (ROI). A comprehensive analysis of classical and neural network-based ML models was conducted, with model size evaluated as an important factor for hardware constrained, in-situ NIR-HSI applications. Interestingly, tailored compact models ranked among the top performers. The KAN, one of the most compact architectures studied, achieved highest performance in both tasks: a mean absolute error (MAE) of 0.90 % for blend ratios and an accuracy of 93.8% for fiber classification, outperforming a 1D convolutional neural network. Our results highlight the potential of tailored ML architectures integrated with NIR-HSI for accurate identification and quantification of cellulosic fibers, advancing the precision and practical applicability of chemometric textile analysis. • NIR hyperspectral imaging was applied for non-destructive discrimination and quantification of cellulosic fibers and blends. • Specifically designed and size-optimized classical and neural network machine learning models, including the emerging KAN, were developed and compared. • Determination of the blending ratio of cotton/viscose blends and discrimination of cellulosic fibers, including cotton, viscose and six ecological plant fibers. • Interestingly, compact models ranked among the top performing models. • KAN achieved best performance: MAE < 0.90 % in cotton/viscose blend ratio determination and 93.8 % accuracy in classifying cellulosic fibers.