Ella
Mahlamäki
*a,
Inge
Schlapp-Hackl
b,
Tharindu
Koralage
b,
Michael
Hummel
b and
Mikko
Mäkelä
a
aVTT Technical Research Centre of Finland Ltd, PO Box 1000, 02044 VTT Espoo, Finland. E-mail: ella.mahlamaki@vtt.fi
bAalto University, School of Chemical Engineering, Department of Bioproducts and Biosystems, PO Box 16300, 00076 Aalto, Finland
First published on 22nd April 2025
Elastane detection is important for textile recycling as elastane fibers can hamper mechanical and chemical fiber recycling. Here, we report the use of near-infrared imaging spectroscopy and class modelling to detect 2–6% elastane in consumer cotton fabrics to provide alternatives to current detection methods, which are invasive and time-consuming. Our method automatically identified outlier fabrics and measurements with class-specific clustering and showed higher classification accuracies by averaging across individual pixel spectra to reduce sampling uncertainty. The final classification results showed median test set true positive and true negative rates of 89–97% based on randomized resampling. Class modelling offers clear benefits compared to commonly used discriminant classifiers as it allows modelling new classes using only a set of target samples without requiring representative training objects from all the other classes. Overall, these results open the possibility for fast non-invasive detection of small amounts of elastane in cotton, taking us a step closer to a circular economy of textiles.
We report a novel method to detect elastane in cotton fabrics based on near infrared (NIR) imaging spectroscopy and class modelling. Cotton, together with regenerated cellulosic fibers viscose and Lyocell, covers 32% of global textile fiber production,5 and is often combined with elastane for clothing applications. The same elastane properties that provide clear benefits in clothing, however, complicate several unit operations during textile recycling. For example, end-of-life textiles are usually disintegrated by shredding and highly elastic polyurethane fibers can block and clump the shredders.6 Subsequent dissolution and spinning of cellulosic textile fibers are also hampered by trace quantities of elastane.7 Elastane can be found on the surface or in the core of a yarn, which makes its detection challenging. Recent academic publications and patents in elastane fiber identification have mainly focused on non-invasive methods using traditional NIR spectrometers or hyperspectral cameras.8–12 Most of these methods did not specifically focus on detecting elastane in cotton or elastane as a minor component but focused on classifying broader textile material categories. Studies including elastane blends as one of the textile categories utilized mainly linear and non-linear discrimination methods.10,11 Hohmann et al. and Langeron et al. mentioned poor separability of elastane using traditional point spectrometers and linear class models.8,9 In addition, Cura et al. found that elastane content had a larger influence on the detectability of elastane in cotton fabrics than its location in the fabric.12
Our method offers three distinct aspects of novelty compared with the current alternatives in the fiber identification field. First, we focus on cotton where elastane is used as a minority component and show how fabrics with 2–6% elastane were reliably identified with median true positive and true negative rates of 89–97% based on randomized resampling. Second, we illustrate how outlier fabrics and measurements can be automatically identified and how averaging across individual pixel spectra improved classification accuracy by reducing sampling uncertainty. Outliers are important as fabrics with incorrect labels can distort the chemical subspace described by the class model. In addition, averaging across individual pixel spectra provides a sampling advantage compared to traditional point spectrometers, which determine spectra from a single localized area within a fabric, because both chemical and physical properties are known to influence NIR measurements.13,14 Third, we used class modelling which offers clear benefits for model training compared with commonly used discriminant classifiers. One-class classifiers enable modelling a new class by training a separate class model to describe the features of that class without representative training objects from the other classes. Overall, we showcase a novel non-invasive alternative to time-consuming chemical methods for elastane identification that can take us a step closer to a circular economy of textiles.
Class | Number of fabric objects | Composition reported in the fabric label | |
---|---|---|---|
Cotton (%) | Elastane (%) | ||
Cotton | 100 | 100 | 0 |
Cotton–elastane | 101 | 94–98 | 2–6 |
The images were initially determined in raw signal intensity counts and later converted to reflectance values using a two-point linear reflectance transformation. This transformation was based on measurements taken from a reflectance target and dark current readings.19 The noise at extreme wavelengths was eliminated by excluding variables outside the range 1000–2500 nm. Consequently, the number of spectral variables was reduced to 270.
![]() | (1) |
![]() | (2) |
A supervised class model was then determined for the cotton–elastane class (Table 1) using soft independent modelling class analogy, SIMCA.24 The class spectra were first decomposed with PCA and two distance metrics determined to assign a class boundary. These distances were the squared Mahalanobis distance of a score object to the center of the model subspace and the squared Euclidean distance to the model subspace, eqn (3) and (4), respectively:
![]() | (3) |
![]() | (4) |
![]() | (5) |
c ≤ clim | (6) |
Training and test sets were generated randomly by selecting one third of the cotton–elastane objects to the test set with all the cotton objects. A suitable number of PCs for the SIMCA classifier was identified through Monte Carlo cross-validation.26 Two thirds of the training objects were randomly selected to a cross-validation training set and the true positive rate (TPR) of the remaining cross-validation test set objects was determined based on the number of PCs used for class description. This cross-validation procedure was repeated 1000 times and the number of PCs with a median true positive rate closest to the confidence level 95% was selected for the final classifier. Additional resampling was then performed by randomly assigning the objects to training and test sets during 1000 resampling iterations and the distributions of the TPR and true negative rate (TNR) were determined. The analysis workflow is visualized in Fig. 1.
The effect of sampling was evaluated by varying the number of pixels used for calculating the average spectrum of each fabric object. Squared regions of interest of 1–1024 pixels were selected from the center of each image and the median true positive rates were determined with a similar Monte Carlo cross-validation procedure as described above. The analyses were performed using in-house scripts developed in Matlab® (The MathWorks, Inc.) including functions from the PLS toolbox (Eigenvector Research, Inc.).
![]() | ||
Fig. 2 An illustration of the imaging principle modified from Mahlamäki et al.17 in (a), conjoint principal component (PC) scores across both classes (CO denoting cotton and EL elastane) and respective principal component loadings with the most important wavelength differences in (b), disjoint class specific principal component scores with identified outliers colored in (c), and revised conjoint principal component scores after outlier removal in (d). The wavelengths marked in the loading plot in (b) are discussed in text. |
Scores of the first two PCs, however, showed an overlap in the two fabric classes. Some cotton objects were located over the cotton–elastane class and some cotton–elastane objects were located over the cotton class in the PC score space, see Fig. 2b. This overlap suggested that some of the fabric labels were not in line with the information in the NIR spectra. We decomposed the two classes separately with disjoint PCA and separated the class scores of the first two PCs into two groups with k-means clustering to determine the overlapping objects. The results are shown in Fig. 2c. The majority groups in both classes were assumed to be true representatives of that class and the minority groups were deemed as potential outliers. There were altogether 38 potential outliers, 21 in the cotton class and 17 in the cotton–elastane class, which was 19% of our sample set. The disjoint PCA loadings are given in Fig. S3 in the ESI† which shows that the wavelength differences of the majority and minority groups were associated with the same wavelengths as shown in Fig. 2b. These results suggested that some cotton fabrics likely contained synthetic fibers and some cotton–elastane fabrics did not contain elastane, or that the elastane was not visible in the NIR measurements.
We manually unraveled the outlier fabrics and visually inspected the separated fibers with a microscope to confirm our observations on the potential outliers. The results showed that the outlier fabrics with a cotton label did contain synthetic fibers and four of the fabrics also contained elastane. The microscopic structural images of these findings are shown in Fig. S4.† Synthetic fibers and elastane were also found in most of the cotton–elastane outliers (Fig. S5†). According to European Union Regulation no. 1007/2011 fabrics are allowed to have up to 2% of extraneous fibers without mentioning it on the fabric label if they are added unintentionally during production.30 This regulation accounts for small changes in fiber content during a manufacturing process. The NIR measurements were influenced not just by the extraneous fibers, but also by the structure of the fabrics. Five of the cotton–elastane outliers had a knitted fabric structure, where elastane was in the core of the fiber surrounded by cotton yarn (Fig. S6†).31–33 This suggested that NIR light had not penetrated through the cotton yarn to detect the elastane fibers. Eleven of the outliers had a woven structure, where elastane yarn was found only in the weft direction, and one fabric contained synthetic fibers other than elastane. Moussa et al. (2006) studied the effects of the fabric structure on light scattering from fibers and found that the highest back-scattering occured when the direction of incident light was perpendicular to the incidence plane for yarn samples and to the weft direction in the woven fabric.34 In addition to the direction of the light, the penetration depth of NIR light into the sample must be sufficient to detect elastane fiber in the core of the yarn. The fabric structures of the potential outliers containing elastane suggested that elastane was not on the surface of the fabric or elastane fiber, in relation to the incidence plane, was in the direction of increased scattering, and was not visible in the NIR spectra. All potential outlier fabrics in the minority groups were deemed as outliers and removed from the final sample set. After outlier removal the conjoint PC scores showed two clusters, and the loadings showed that the wavelength differences of the cotton and cotton–elastane groups were associated with the same wavelengths as shown in Fig. 2b. The revised PC scores are shown in Fig. 2d. Overall, disjoint PCA and clustering provided a convenient way to automatically identify fabrics or NIR measurements which were not in line with the information reported in the fabric label.
We focused on training a SIMCA classifier specifically for the cotton–elastane class and divided the objects of that class into training and test sets. All cotton class spectra were designated to the test set. Significant features in the training set were extracted with a PCA model, which was used to determine the relevant model distances for the classifier. An illustration of the model distances and the class boundary in a one-dimensional principal component model is shown in Fig. 3a. An appropriate number of PCs was estimated by Monte Carlo cross-validation.26 The results showed that a median true positive rate closest to the predetermined 95% confidence level was achieved with only one PC based on 1000 random sampling iterations, see Fig. 3b. Cross-validation performance started to decrease with more than one PC, which indicated that a one component class model was sufficient to reliably describe the properties of the target class.
We also determined cross-validation performance with different preprocessing combinations based on first and second derivative Savitzky–Golay filtering and standard normal variate transformation.35 The results are shown in Fig. S7,† where we report the effect of derivative filtering on different filter window lengths. A wider filter window led to excessive smoothing and removed important features in the average spectra, while a narrower filter window likely generated excess noise to the signals. An appropriate processing method was important to enhance the chemical features of elastane. We evaluated the effect of pixel averaging by controlling the number of pixels used for determining the average fabric spectra. As shown in Fig. 3c, median true positive rates were closest to the predetermined confidence level 95% using 1024 or all fabric pixels. Cross-validation performance decreased with a decrease in the number of pixels used for averaging. Using fewer pixels for averaging increased sampling uncertainty and reduced our capability to extract chemically relevant features from the spectra.
We then determined the performance of the final SIMCA classifier with the randomly chosen test set. Classification of the test set spectra resulted in a true positive rate of 93% and a true negative rate of 97% with an overall accuracy of 96%. Fig. 3d shows the squared Mahalanobis distances within the model and the squared Euclidean distances to the model for the test set objects with the derived class boundary, and Fig. 3e reports the confusion matrix based on the test set. In Fig. 3d the axes are logarithmically transformed for visualization, which makes the linear class boundary appear non-linear. Two cotton–elastane fabrics and two cotton fabrics were misclassified. The misclassified fabrics were similarly manually unraveled as the outliers. The microscopy images of the two misclassified cotton fabrics showed synthetic fibers (Fig. S8†). The misclassified cotton–elastane fabrics contained 3% of elastane based on the fabric label. The elastane fiber was found in the core of the fibers from both misclassified fabrics, which suggested that elastane in these fabrics was not visible in the NIR measurements.
As shown in Fig. 3d and e, we tested the SIMCA classifier after we had randomly assigned one third of the cotton–elastane spectral objects in the remaining sample set to a separate test set with the cotton objects. There were thus 56 objects in the training set and 107 objects in the corresponding test set (Fig. 3e). The results suggested promising classifier performance based on determined TPR and TNR which, however, varied depending on which exact objects were assigned to the training and test sets. To account for this variation, we performed additional resampling by randomly assigning the objects to training and test sets during 1000 resampling iterations and determined distributions of TPR and TNR. The results are illustrated in Fig. 3f. The randomized resampling iterations led to median true positive and true negative rates of 89% and 97%, which indicated that our classifier correctly identified cotton fabrics with small amounts of elastane independent of how the training and test set objects were chosen.
NIR imaging spectroscopy provides an appealing alternative for fast and non-invasive identification of elastane in textiles. Imaging spectrometers determine material-specific chemical fingerprints from every pixel of an image scene, which provides large amounts of data. These data can then be used for training image classification or regression models as part of high-throughput and automated sorting systems, which are currently breaking ground in the textile field.37 However, scaling imaging spectrometers for industrial applications presents challenges such as increased costs, demands on processing speed and complexity. Instruments working on wider NIR wavelength regions, such as 1000–2500 nm, generally use expensive cryogenically cooled MCT detectors, whereas imaging in shorter regions up to 1700 nm can be performed with cheaper InGaAs sensors.38,39 The required wavelength region is determined by the application and the spectral features as different chemical bonds interact with light on specific wavelengths. The speed of the measurement is specified by spectral and spatial resolutions, the frame rate of the detector, and the speed of the conveyor belt. These determine how many pixels are imaged and how quickly they are captured, which then defines the time required to process the data.
Beyond costs and speed, a significant challenge for industrial applications is the large amount of variation and contaminants in post-consumer textile materials.12 Variations in fabric types, sources, quality and condition generate a lot of spectral variation. Contaminants, such as dirt, oil, or residues from previous processing steps, can change the spectral signature of the fabric as they may introduce additional absorption or reflection features that are not representative of the target fabric material. Detecting detailed properties or minor fiber components in fabrics, such as elastane, thus requires that the collected textile waste is first sorted.37
Here, we have focused on a small problematic category of textile waste and presented on lab-scale how 2–6% elastane can be detected in cotton fabrics, but future efforts should acknowledge these challenges when developing machine vision algorithms for industrial applications of textile material identification. Based on the principles of our classification model, we believe that our method is applicable also to fabrics with elastane content outside the range 2–6%, and generalizable to other elastane fabrics with similar characteristics as long as the classifier can be trained on correct labels and the elastic fiber is visible in the NIR spectra.
Footnote |
† Electronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d5an00107b |
This journal is © The Royal Society of Chemistry 2025 |