Statement of Novelty

The proposed approach can be profitably used to control the quality of the textile recycling process thanks to its easiness and detection rapidity. The possibility to use low cost and automatic systems for the identification and sorting of End-Of-Life (EoL) textile products plays an important role into the feasibility of an efficient recycling process to apply to these materials since at moment waste textile sorting mainly takes place through time consuming manual operations.

Moreover, the European Commission has identified textiles (i.e., apparel and fabrics) as “priority product category for the circular economy”, encouraging the development of technological applications for textile waste reuse and recycling. Therefore, this approach can represent a big step forward toward the fulfillment of good recycling targets.

Introduction

The textile and apparel sector is crucial to the economies of all nations in terms of trade, employment, investment, and revenue [1]. On the other hand, supply chain operations are characterized by fragmented and low-tech systems having significant environmental implications such as the consumption of large volumes of water and chemicals, massive emissions of greenhouse gases and waste generation [2]. After food, housing and transportation, textiles are the fourth-highest pressure category for primary raw materials and water usage [3]. Moreover, the textile industry is the second most pressing in terms of land usage and the fifth for greenhouse gas (GHG) emissions. Indeed, the fashion industry (i.e. apparel and footwear) accounts for more than 8% of worldwide GHG emissions and 20% of annual wastewater production [4].

The global textile and apparel consumption has increased to an estimated 62 million tons per year, with a projected increase to 102 million tons by 2030 [5].

Each year, over 150 million tons of textile wastes are produced worldwide [6, 7]. According to projections, consumers in the European Union (EU) discard over 5.8 million tons (11.3 kg per person) of textiles annually [8].

Globally, about 75% of total textile waste is disposed of in landfills, 25% is reused or recycled and less than 1% is reprocessed back into clothes. In this context, developing reuse and recycling technologies for textile waste is critical for diverting it from landfills. As a result, addressing the negative consequences of this economic sector through the benefits of circular economy initiatives is critical for long-term sustainability [9]. Garment recycling has to be widely implemented in order to close the material loop and build an effective recycling system for all textile waste. In fact, recycling and reusing are the most effective textile waste disposal strategies that have less environmental impact. The European Waste Framework Directive (2008/98/EC) and following revisions establishes the basic waste management principle and mandates EU member states to implement a waste management hierarchy (prevention, reuse, recycling, and disposal) in waste management plans and waste prevention initiatives [10]. Moreover, the European Council (EC) supported sustainability by replacing the Waste Framework Regulation with a Circular Economy Package, which established a 70% goal for municipal solid waste (MSW) recovery and a 10% limit for landfilling by 2030 [11]. The Extended Producer Responsibility (EPR) policy was crucial in meeting such goals, leading to an average yearly increase of 13% in post-consumer textile collection [12]. Nevertheless, nowadays less than the 1% of all textiles in worldwide are recycled into new textiles and the textile recycling rate is about 25% in Europe and about 16% in the United States [13].

According to the United States Environmental Protection Agency (USEPA), textile waste accounts for roughly 5% of landfill area [14]. Therefore, millions of tons of textiles go to landfills around the world instead of being recovered and recycled [15]. Increasing textile recycling rates would help to reduce the negative environmental effects of the textile industry [16].

Textiles (i.e. garments and fabrics) were classified as a "priority product category for the circular economy" by the European Commission, stimulating the development of technological applications to improve the textile waste reuse and recycling. Moreover EU policies also mandate Member States to achieve high levels of textile waste separation by 2025, ensuring that waste collected separately is not incinerated or landfilled [13].

Textile recycling technologies are recently emerged in order to classify/sort the textile fibers typology [2]. Identification of textile fibers can be critical in recycling because identical chemical composition generally indicates comparable physical characteristics. As a result, classifying textile waste by material is essential for the proper implementation of a recycling system [17].

Textile fibers are usually defined as natural or man-made/synthetic (i.e. artificial) [18, 19]. More in detail, natural fibers can be of:

  1. i)

    vegetable-based origin (i.e. cellulosic fibers obtained from different part of plants such as leaves and seeds), constituting the so-called plant derived textiles,

  2. ii)

    animal-based origin (i.e. protein fibers) producing the animal derived textiles and

  3. iii)

    mineral-based origin (i.e. asbestos is the only naturally occurring).

Moreover, there are artificial textile coming from:

  • Manmade Cellulosic Fibers (MCF) that are regenerated materials usually made from the dissolved wood pulp or “cellulose” of trees (i.e. viscose deriving from cellulose) and

  • synthetic fibers coming from fossil-fuel-derived resources including petroleum-based chemicals and petrochemicals.

Textile waste requires several types of recycling procedures based on the fiber type. Therefore, the potential to use low-cost and automatic techniques for identifying and sorting EoL textile items is critical to define the feasibility of an efficient recycling process for these materials [20]. At the moment, waste textile sorting is mostly done manually, and the expense for this procedure accounts for 30% of the total costs of recycled textile: developing an automation sorting system could minimize process costs and time [21].

The present work is addressed to explore the possibility to use Short-Wave InfraRed (SWIR: 1000–2500 nm) spectroscopy to extract information regarding the fabric composition of different EoL textiles, useful to recognize plant derived textile such as cotton (i.e., vegetable-based origin fibers), animal derived textile as silk (i.e., animal-based origin fibers) and artificial textile (i.e., synthetic fibers or MCF) such as polyester or viscose.

Spectroscopy is a powerful tool for chemical analysis as, starting from the characteristic spectrum of each substance, it is possible to provide detailed and precise information on its physical–chemical features (i.e., structure or composition). Therefore, since spectral attributes are linked to the sample chemical composition, adopting this investigation methodology, it possible to characterize, classify and control the quality of different materials in several industrial sectors.

In order to perform systematic environmental remote and proximal sensing, SWIR spectroscopy techniques are widely used in the primary and secondary raw materials sector [22,23,24], in the pharmaceutical and chemical industry [25,26,27], in the cultural heritage sector, in the agricultural/food industry [28,29,30,31], in the medical field and for clinical applications [32,33,34] and more commonly in analytical science [35]. SWIR spectroscopy offers several advantages. Firstly, it is non-destructive, allowing to perform repeated measurements without altering the sample, thus enabling real-time monitoring. Secondly, samples undergoing SWIR spectroscopy analysis do not require complex sample preparation procedures. Moreover, this technique is high scalable in different industrial applications at various level (i.e., on-line, at-line and off-line). The main drawback of the utilization of such technique is that the material is analyzed only at surface-level. For this reason, studying samples with coatings and with multi-layered materials can be challenging.

However, the utilization of Near-InfraRed (NIR: 1000–1700 nm) represents a novel approach to be applied for textiles recognition. Such topic has been studied by various researchers and continues to be under investigation. NIR spectroscopy and Machine Learning (ML) techniques, i.e. Soft Independent Modeling of Class Analogy (SIMCA) and Linear Discriminant Analysis (LDA), were adopted by [20] to identify seven different textile fibers (i.e. cotton, cashmere, fine wool, tencel, polylactic acid–PLA, polypropylene–PP and polyethylene terephthalate–PET). Different authors adopted a Deep Learning (DL) approach, i.e. Convolutional Neural Network (CNN), to classify fiber samples from a NIR/SWIR spectral data [21, 36, 37]. In literature there are also different case studies on the application of HyperSpectral Imaging (HSI) in the Visible/NIR/SWIR range to identify and classify textile material for sorting purposes [17, 21, 38].

In a previous study an NIR-HSI system and a portable spectrophotoradiometer were adopted along with ML techniques, i.e. Partial Least Squares—Discriminant Analysis (PLS-DA), to classify cotton, silk, viscose, and some of their blends [39].

The main goal of this preliminary study is to further explore the possibilities of the utilization of portable spectrophotoradiometer for the fast recognition of textiles. For this reason, in the present work, a portable single spot spectroradiometer equipped with a contact probe was used to collect reflectance spectra in the SWIR range (1000–2500 nm). Finally, ML techniques were adopted to analyze and classify the spectral data. In more detail, Principal Component Analysis (PCA) was adopted to explore the decomposed spectral data. While, PLS-DA—based hierarchical modelling was adopted to firstly classify the fiber category (i.e., vegetable-based origin fibers, animal-based origin fibers and synthetic fibers / MCF) and then the material constituting the analyzed textiles.

Materials and Methods

Textile Samples

The analyzed textile samples were provided by an Italian tailor and fashion designer (Cinzia Tiso Handmade, San Lorenzo del Vallo, Cosenza, Italy). In more detail, 36 samples were selected from EoL textile apparels, as reported in Table 1, according to the material reported by producers on the apparel, and they were grouped according to the fiber origin in animal-derived textile, plant-derived textile and artificial textile, considering in the latter category synthetic fiber and MCF. Based on this classification, 12 different textile materials or material blends were identified in the sample set, 3 for animal-derived, 4 for plant-derived and 5 for artificial textiles.

Table 1 Analyzed fabric samples categorized according to material and fiber origin classes
Full size table

Methods

Portable Spectrophotoradiometer

In order to collect spectra in reflectance mode, a FieldSpec® 4 Standard-Res field portable spectrophotoradiometer (ASD Inc., Boulder, Colorado, United States) was employed. This portable instrument has a spectral resolution of 3 nm at 700 nm and 10 nm at 1400/2100 nm and it can acquire spectra in the Vis-SWIR ranges (350—2500 nm) [40].

The spectroradiometer is composed of a detector unit and a fiber optics cable that connects to a contact probe that is controlled by a computer. The detectors unit is realized by coupling three separate detectors with different holographic diffraction gratings. A VNIR detector (512 element silicon array: 350–1000 nm), a SWIR 1 detector (Graded Index InGaAs. Photodiode, Two Stage TE Cooled; 1001–1800 nm), and a SWIR 2 detector (Graded Index InGaAs. Photodiode, Two Stage TE Cooled; 1801–2500 nm) comprise the detector architecture.

The light source is a halogen lamp with a color temperature of 2901 ± 10% ºK. The contact probe spot is 10 mm in size. The ASD RS3 software was used for data collecting and calibration procedures [41]. The spectroradiometer was calibrated by referencing the dark current calibration file and acquiring a standardized white Spectralon® ceramic material via a white reference measurement. Following the calibration stage, the spectrum is acquired, and reflectance is computed for each sample.

The ASD ViewSpec Pro (Ver. 6.2.0) was used to stack the spectroradiometer spectra “.asd” data files into an ASCII text file. An ad hoc written routine was used to import the converted data into the MATLAB® environment (MATLAB R2019a; The Mathworks, Inc., Natick, Massachusetts, United States). Imported data files were analyzed in the MATLAB® environment using PLS_toolbox (Ver. 8.2., Eigenvector Research, Inc., Manson, Washington, United States). Datasets were created, objects were created, and classes were assigned.

Spectral Data Collection, Processing and Exploratory Analysis

For each fabric sample, 5 spectra were acquired in reflectance mode with the ASD portable spectrophotoradiometer for a total number of 180 analyzed spectra. The acquired spectral range was reduced to 1000–2500 nm. Spectral data were then subjected to spectral pre-processing to emphasize the differences between the classes that were considered, to improve sample spectral features and to minimize the effect of any external sources of variability.

The combination of pre-processing algorithms used in this study are: Standard Normal Variate (SNV), followed by Savitzky-Golay (SG) first derivative and Mean Centering (MC). SNV is a common pre-processing algorithm used in spectroscopy, that has the purpose to remove baseline variations and scaling differences between spectra [42].

The SG filter is one of the most used filter in chemometrics for differentiation [43, 44]. SG first derivative is useful to remove baseline variations and reduce data noise. Finally, MC was employed to remove constant offset, which was not relevant to the data variance interpretation.

Following the pre-processing procedure, Principal Component Analysis (PCA) was performed. PCA is an unsupervised method for reducing data dimensionality by projecting samples onto a low dimensional subspace with axes referred to Principal Components (PCs) [45]. PCA was used to investigate the variability of spectral data and to define classification setups.

Hierarchical Classification

To reach the classification target, a hierarchical classification procedure based on Partial Least squares—Discriminant Analysis (PLS-DA) was applied. Classification methods aim to build a model capable of identifying the class to which objects belong, on the basis of a certain number of descriptors. In other words, a classification algorithm tries to predict a property y of the sample, starting from a matrix of descriptors X. Among the classification algorithms, PLS-DA is a supervised classification method, combining the properties of partial least squares regression with the discriminating ability of a classification technique [46].

Models were calibrated by using the spectral information contained in the training sets and cross-validated using the Venetian Blinds (VBs) algorithm to assess the optimal complexity of the models.

The first classification rule in the hierarchy procedure set-up is represented by a PLS-DA classifier, built in order to discriminate the textiles according to the three different origins of its fibers: (i) animal-derived textile, (ii) plant-derived textile and (iii) artificial textile. In the second step, three PLS-DA were performed: one PLS-DA for each branch of fiber category was performed in order to discriminate the material of the analyzed textile. The spectra not classified by the hierarchical classification method were defined as “Not classified”. In Fig. 1, the dendrogram of the developed hierarchical procedure for classification is shown.

Fig. 1
figure 1

Dendrogram of the hierarchy of PLS-DA classifier and the obtained outputs. Each rule node represents a PLS-DA classifier

Full size image

To evaluate the classifier performance, Sensitivity, Specificity, Precision, Accuracy and Error Rate statistical parameters were calculated from the confusion matrix [47, 48] according to the following equations:

$${\text{Sensitivity}}=\mathrm{ TP}/({\text{TP}}+{\text{FN}})$$
(1)
$${\text{Specificity}}=\mathrm{ TN}/({\text{FP}}+{\text{TN}})$$
(2)
$${\text{Precision}}={\text{TP}}/({\text{TP}}+{\text{FP}})$$
(3)
$${\text{Accuracy}}=({\text{TP}}+{\text{TN}})/({\text{TP}}+{\text{TN}}+{\text{FP}}+{\text{FN}})$$
(4)

where TP (True Positive) is a positive instance that is classified as positive; FN (False Negative) is a positive instance that is classified as negative; TN (True Negative) is a negative instance that is a classified as negative and FP (False Positive) is a negative instance that is classified as positive.

The overall Precision and Accuracy are calculated averaging the single class statistical parameter weighted by the number of class spectra for calibration and cross-validation phases.

Results and Discussion

Exploratory Analysis

The raw reflectance spectra of the analyzed samples are reported in Fig. 2, whereas the raw reflectance and pre-processed spectra averaged according to the fiber origin and material classes are reported in Figs. 3 and 4, respectively. The stretching vibration of hydrogen groups such as N–H, O–H, and C-H provides the absorption bands in the NIR spectral range [49]. Plant-derived fibers are mostly composed of cellulose, which includes C-H, C–C, and O–H groups. The absorption at 1480 nm corresponds to the initial overtone of the O–H stretching from semicrystalline cellulose, which characterizes this type of fibers [50]. Regarding animal-derived textiles, interesting bands are visible between 1540 and 1580 nm, related to the NH groups [51]. Absorption in the spectral region between 1000 and 1400 nm can be due to the eumelanin pigment presence. The absorption in the region around 1400 nm represents the first overtone of O–H stretching vibration of water and around 1900 nm a combination of the O–H stretches and H–O-H bending vibration of the hydroxyl group from water is shown. The overtone of the C-H stretch of protein side chains and lipids is at around 1730 nm, while between 2000 and 2500 nm information about amino acid composition and animal keratin fibers are given. In terms of chemical composition, artificial textile (i.e. polyester, acetate, viscose, nylon, etc.) combination bands can be linked to the C–H, methyl (CH3) group and ester bond (CO–O) [20].

Fig. 2
figure 2

Average raw reflectance spectra of the analyzed textile samples

Full size image
Fig. 3
figure 3

Raw reflectance (a) and pre-processed (b) spectra averaged according to the fiber classes

Full size image
Fig. 4
figure 4

Raw reflectance (a) and pre-processed (b) spectra averaged according to the material classes

Full size image

The results of PCA performed on all the samples is reported in Fig. 5, in terms of PC1-PC2 score plot.PC1 and PC2explained 66.42% and 16.24% of the total variance, respectively. The scores of “Animal-derived textile” class are mainly clustered in the first quadrant, whereas the “Plant-derived textile” category scores are mostly located in the space of positive PC1 and negative PC2. Finally, the “Artificial textile” scores are scattered in the second, third and fourth quadrants of the plot. The partial overlapping in PCA of plant-based textile/artificial textile scores and animal-derived textiles/artificial textile scores in SWIR range can be attributed to the similarities in the spectral features of these materials.

Fig. 5
figure 5

PCA scores plot of the first two PCs for the fiber classes. The 95% confidence ellipse for each group (red for “Animal-derived textile”, green for “Plant-derived textile” and blue for “Artificial textile”) are reported in the plot, representing the region in which the scores of a group of samples are expected to fall with a 95% probability

Full size image

The score plot of the first two PCs of the investigated animal-derived, plant-derived and artificial textiles are reported in Fig. 6a, b and c, respectively. The PCA performed on spectra of animal-derived textile (Fig. 6a) explained a total variance of 66.48% with PC1 and 22.43% with PC2. “Wool” scores are located in the positive space of PC2, while silk-based fabric (i.e. “Silk” and “Silk blend”, being a blend of different type of silk fabrics) scores are clustered in the negative space of PC2 due to the different spectral behavior of the two textile classes.

Fig. 6
figure 6

PCA scores plot of the first two PCs for the material classes of a animal-derived, b plant-derived and c artificial textiles. The 95% confidence ellipse for each analyzed group is reported in the plots. (Color figure online)

Full size image

The PCA performed on spectra of plant-derived textiles (Fig. 6b) explained a total variance of 71.65% with PC1 and 14.74% with PC2. Linen-based fabric scores are grouped in the negative space of PC2 and positive space of PC1. Jute fabric scores are concentrated in the positive space of PC1 and PC2, whereas cotton fabric scores are scattered around the first, third and fourth quadrants. The scattered distribution of “Cotton” scores into the PCA plot and the partial overlapping of “Linen” and “Linen blend” with “Cotton” class are related to the similar spectral behavior of plant-derived textile (i.e. cotton, jute, linen and linen blend) composed mainly by cellulose-based fibers.

Finally, the PCA performed on spectra of artificial textiles (Fig. 6c) explained a total variance of 83.58% with PC1 and 11.78% with PC2. In this case, all the viscose blend textile scores are mainly in the negative space of PC2. With reference to the “Polyester” and the “Nylon-Polyester” classes, the scores of the first group are completely in the positive space of PC1 and negative space of PC2, while the second ones are in the negative space of PC1 and in the positive space of PC2. The separation of these two groups is probably due to differences in the spectral signatures around the OH combination bands region (i.e. 2000–2100 nm).

Classification

Classification First Rule: Fiber Origin Classes

Statistical parameters of the first PLS-DA classification set-up are reported in Table 2. Overall Precision and Accuracy (i.e. measured as spectra weighted average values) in calibration are 0.557 and 0.977 respectively, while in cross-validation Precision is 0.56 and Accuracy is 0.952. Lower values of Precision and Accuracy were obtained for “Artificial textile”, probably due to data variability being greater than that of the other categories.

Table 2 Statistical parameters of the first PLS-DA classification set-up
Full size table

Classification Second Rule: Material Types

Animal-Derived Textile Classes

Statistical parameters of the PLS-DA classification for animal-derived textile classes set-up are reported in Table 3. Overall Precision and Accuracy are 1.000 and 1.000 respectively, both in calibration and in cross-validation. In this case all the spectra are correctly classified.

Table 3 Statistical parameters of the second PLS-DA classification set-up: animal-derived textile classes
Full size table
Plant-Derived Textile Classes

Statistical parameters of the PLS-DA classification for plant-derived textile classes set-up are reported in Table 4.

Table 4 Statistical parameters of the second PLS-DA classification set-up: plant-derived textile classes
Full size table

Overall Precision in calibration is 1.000 and Accuracy is 1.000, whereas in cross-validation Precision is 0.839 and Accuracy is 0.919. In this case the material classes that reached the smallest values of Sensitivity in cross-validation (around 0.600) were “Jute”, “Linen” and “Linen blend”. This is probably due to the spectral similarities of the three textile classes linked to their comparable chemical compositions (i.e. cellulose based fibers).

Artificial Textile Classes

Statistical parameters of the PLS-DA classification for artificial textile classes set-up are reported in Table 5.

Table 5 Statistical parameters of the second PLS-DA classification set-up: artificial textile classes
Full size table

Overall Precision in calibration is 1.000 and Accuracy is 1.000, while in cross-validation Precision is 0.884 and Accuracy is 0.925. In this case, the smallest value of Sensitivity in cross-validation was reached by “Viscose-Polyester blend” (0.400) followed by “Viscose” (0.867) and “Polyester” (0.900). In this classification step, blends and the composing textile fibers (i.e. linen, jute and linen blend or viscose, polyester and viscose-polyester) achieved the worse performance parameters. This fact could be linked to the fiber percentage composition (i.e. a textile blend that has a considerably higher content of one fiber than another), which makes the material more similar in terms of spectral signature to that one.

In order to understand the weight of those errors, information about the fiber percentage composition of each textile blend could be very useful. Therefore, these aspects have to be further investigated in future studies.

Overall Classification Results

The statistical parameters of the hierarchical classification procedure based on PLS-DA classifiers is reported in Table 6. These statistical parameters are referred to the application of the classifier hierarchy to the whole dataset.

Table 6 Statistical parameters of the hierarchical classification model based on PLS-DA classification
Full size table

The overall Accuracy of the performed classification procedure is 0.984, while the Overall Precision is 0.933, not considering the “Linen” and “Linen blend” classes that are not correctly classified (i.e. Sensitivity equal to 0). The smallest values of Sensitivity were reached by “Viscose” (0.067), followed by “Silk blend” and “Nylon–polyester blend” (0.400). The highest values of Sensitivity (1.000) and Specificity (1.000) were achieved by the following classes: “Acetate-viscose blend”, “Jute”, “Silk”, “Wool” and “Viscose–polyester blend”. Even in this experimental set-up, the presence of textile blends reduces the classification statistical parameter values. Furthermore, it must be considered that the balance of the number of spectra in each product class is highly uneven, which can influence the predictive ability of the model.

Conclusions and Future Perspectives

The present study was carried out with the aim of recognizing EOL textile fabrics, using SWIR spectroscopy techniques, into a circular economy perspective.

Present findings show as the applied technique enables the recognition of textile fabric both according to the origin (i.e., plant-derived textile, animal-derived textile and artificial textiles) and the material type (i.e., cotton, jute, linen, silk, wool, viscose, nylon, polyester, and blends).

However, the obtained results showed some misclassifications probably due to the presence of textile blends containing an unknown percentage of different fibers, creating problems to the trained classification model. Based on the current findings, additional studies will be carried out to investigate a greater number of textile typologies (by balancing also the number of spectra of each class of product) and different textile blend samples, not only to identify them but also to evaluate the content in terms of fabric type percentage. Anyway, it's essential to highlight that the model's testing on novel samples, crucial for validation and robustness assessment, will be an integral part of our future work. This preliminary study sets the groundwork for forthcoming research aimed at expanding the dataset and rigorously testing the model’s predictive capacity on previously unseen fabric samples.

Furthermore, the next steps will include the set up of a hierarchical classification technique based on more complex models (i.e. Support Vector Machines, Convolutional Neural Networks, Random Forests) in order to achieve improved accuracy and precision for all selected classes.

In this context, the developed technique can be viewed as a methodological procedure to be systematically implemented at the scale of a recycling plant. A portable device could be used to perform a fast test on samples taken at the feed of a recycling plant, on recovered final products, and/or on materials from different process stream phases to ensure that the system is operating properly.

The recycling process could be efficiently automated by using a hierarchical classification approach, supplementing, or replacing human operations/sorting steps. Recycling operations might be accelerated and enhanced in this manner, resulting in significant cost savings.

Furthermore, a more circular and sustainable system would help to achieve many of the UN Sustainable Development Goals (SDGs). Likewise, "SDG 12: Responsible Consumption and Production" is a doorway to many of the other SDGs for the textile industry, including Goal 6: Clean Water and Sanitation, Goal 7: Affordable and Clean Energy, Goal 12: Responsible Consumption and Production, and Goal 13: Climate Action.