Article

Using Infrared Raman Spectroscopy with Machine Learning and

Deep Learning as an Automatic Textile-Sorting Technology for
Waste Textiles
Pei-Fen Tsai and Shyan-Ming Yuan *

Department of Computer Science, National Yang Ming Chiao Tung University, ChiaoTung Campus,
Hsinchu 300093, Taiwan; [email protected]
* Correspondence: [email protected]; Tel.: +886-35712121 (ext. 56631)

Abstract: With the fast-fashion trend, an increasing number of discarded clothing items
are being eliminated at the stages of both pre-consumer and post-consumer each year. The
linear economy produces large volumes of waste, which harm environmental sustainability.
This study addresses the pressing need for efficient textile recycling in the circular economy
(CE). We developed a highly accurate Raman-spectroscopy-based textile-sorting technol-
ogy, which overcomes the challenge of diverse fiber combinations in waste textiles. By
categorizing textiles into six groups based on their fiber compositions, the sorter improves
the quality of recycled fibers. Our study demonstrates the potential of Raman spectroscopy
in providing detailed molecular compositional information, which is crucial for effective
textile sorting. Furthermore, AI technologies, including PCA, KNN, SVM, RF, ANN, and
CNN, are integrated into the sorting process, further enhancing the efficiency to 1 piece per
second with a precision of over 95% in grouping textiles based on the fiber compositional
analysis. This interdisciplinary approach offers a promising solution for sustainable textile
recycling, contributing to the objectives of the CE.

Keywords: circular economy; textile recycling; Raman spectroscopy; artificial intelligence;

machine learning; neural networks; textile sorting

Academic Editors: Stefan Oniga and

József Sütő

Received: 26 November 2024 1. Introduction

Revised: 18 December 2024
The projection of global plastic production from 2000 to 2100 shows that plastic waste
Accepted: 20 December 2024
generation will increase drastically, reaching over 15 billion metric tons by 2100, as shown
Published: 25 December 2024
in Figure 1a [1]. The textile sector is the second-largest contributor to this increase. With
Citation: Tsai, P.-F.; Yuan, S.-M.
the high demand for fast fashion and the lack of an effective solution to reduce, reuse, and
Using Infrared Raman Spectroscopy
with Machine Learning and Deep
recycle [2], textiles tend to have a short product life. As shown in Figure 1b, it is difficult to
Learning as an Automatic reduce product manufacturing effectively to meet consumer demand. Consequently, there
Textile-Sorting Technology for Waste has been an increasing number of studies on developing waste-textile-recycling technology
Textiles. Sensors 2025, 25, 57. to extend the textile life cycle, in recent years.
https://doi.org/10.3390/s25010057
In 2020, in the global textile industry report [3], there was a polyester (PES) fiber
Copyright: © 2024 by the authors. demand of 69 million metric tons, representing 60% of the market share, which highlights
Licensee MDPI, Basel, Switzerland. the highest textile demand and an increasing trend, as seen in Figure 2a,b. Consumers
This article is an open access article
frequently dispose of old clothing, leading to textile waste that often ends up in landfills
distributed under the terms and
or incinerators, causing land pollution and contributing to greenhouse gas emissions, as
conditions of the Creative Commons
Attribution (CC BY) license
shown in Figure 3 (purple flow). To achieve environmentally sustainable governance
(https://creativecommons.org/ (ESG), implementing the 3Rs—reuse, reduction, and recycling—is critical for extending the
licenses/by/4.0/). product life cycle. To shift the flow from purple (unrecycled) to blue and green (recycled)

Sensors 2025, 25, 57 https://doi.org/10.3390/s25010057

Sensors 2025, 25, 57
(ESG), implementing the 3Rs—reuse, reduction, and recycling—is critical for extendi
2 of 27
the product life cycle. To shift the flow from purple (unrecycled) to blue and green (re
cled) materials (Figure 3), an efficient and accurate AI-based sorting system is essential
materials (Figure 3), an efficient and accurate AI-based sorting system is essential for both
both open-loop and closed-loop recycling.
open-loop and closed-loop recycling.

Figure 1. Global plastic demand and life cycle by sector. (a) Global plastic production, waste, and
in-use volume projections
FigureFigure 1.from
1. GlobalGlobal2000
plastic
plastic
to
demand
2100and
demand (Stegmann
and life
lifecycle
etsector.
by by
cycle
al.,
sector. (a) 2022)
Global [1]. (b)
plastic
(a) Global
Product
production, lifeandcycle
waste,
plastic production,
by
waste, a
in-use volume projections from 2000 to 2100 (Stegmann et al., 2022) [1]. (b) Product life cycle by sector
sector (Geyer, R.in-use
et al., 2017)projections
volume [2]. from 2000 to 2100 (Stegmann et al., 2022) [1]. (b) Product life cycle
(Geyer, R. et al., 2017) [2].
sector (Geyer, R. et al., 2017) [2].

Figure 2. The trends and shares of the global textile demands. (a) Global textile demands for po
Figure 2. The trends Figure 2. The trends
and shares and shares
of the globalof the global textile
textile demands.(a)
demands. (a) Global
Global textile demands
textile for polyester,
demands for poly
ester, cotton, and others. (b) Global textile demand share percentages in 2020 (redrawn based
cotton, and others. (b) Global textile demand share percentages in 2020 (redrawn based on [3]).
ester, cotton, and others. (b) Global textile demand share percentages in 2020 (redrawn based on
[3]).
[3]).
sors 2025, 25, x FOR PEER REVIEW 3 of
Sensors 2025, 25, 57 3 of 27

Figure 3. The textile life-cycle flow includes un-recycling, reuse, and closed-loop and open-loop
cycling. The un-recycling of waste textiles is in the linear economy, while the 2Rs of reuse and rec
cling are in the circular economy (redrawn from (MDBC, 2021) [3]).

1.1. Textile Recycling and the Circular Economy

The textile life cycle can be classified into four modes, as defined by A. Payne [4] an
illustrated in Figure 4, as: un-recycling (Figure 4a), open-loop recycling (Figure 4b), an
closed-loop recycling (Figure 4c).
Open-loop recycling (OLP), also called “downcycling”, occurs at both the pre-co
sumer and post-consumer stages and is relatively easier to achieve compared to close
loop recycling. In contrast, closed-loop recycling involves reprocessing polyester textil
enabling them to reenter the same production cycle through chemical recycling. This pr
cess maintains high product value and aligns with circular economy (CE) principl
which emphasize
FigureFigure 3. textile
3. The the
The textile elimination
life-cycle
life-cycle flow of waste
flowincludes
includes and pollution,
un-recycling,
un-recycling,reuse, andthe
reuse, circulation
closed-loop
and of products
and open-loop
closed-loop and an
open-loop
recycling. The un-recycling of waste textiles is in the linear economy, while the 2Rs of reuse and
materials,
cycling. Theand the regeneration
un-recycling of nature.
of waste textiles is in the linear economy, while the 2Rs of reuse and re
recycling are in the circular economy (redrawn from (MDBC, 2021) [3]).
The transition toward a circular
cling are in the circular economy (redrawn from (MDBC,economy directly aligns
2021) [3]). with Sustainable Develo
1.1. Textile Recycling and the Circular Economy
ment Goal (SDG) 12: Responsible Consumption and Production. By implementing close
1.1. TextileThe textile lifeand
Recycling cycle can
the be classified
Circular into four modes, as defined by A. Payne [4] and
Economy
loop recycling and fostering innovative textile-sorting technologies, the environmen
illustrated in Figure 4, as: un-recycling (Figure 4a), open-loop recycling (Figure 4b), and
impact The oftextile
PET textiles
closed-loop life cycle
recycling
can be minimized.
can
(Figure be
4c).classified into four modes, as defined by A. Payne [4] a
illustrated in Figure 4, as: un-recycling (Figure 4a), open-loop recycling (Figure 4b), a
closed-loop recycling (Figure 4c).
Open-loop recycling (OLP), also called “downcycling”, occurs at both the pre-co
sumer and post-consumer stages and is relatively easier to achieve compared to close
loop recycling. In contrast, closed-loop recycling involves reprocessing polyester textil
enabling them to reenter the same production cycle through chemical recycling. This p
cess maintains high product value and aligns with circular economy (CE) principl
which emphasize the elimination of waste and pollution, the circulation of products a
materials, and the regeneration of nature.
The transition toward a circular economy directly aligns with Sustainable Develo
ment Goal (SDG) 12: Responsible Consumption and Production. By implementing close
loop recycling and fostering innovative textile-sorting technologies, the environmen
impact of PET textiles can be minimized.

Figure 4. Cont.
ensors 2025, 25, x FOR PEER REVIEW 4 of 28
Sensors 2025, 25, 57 4 of 27

Figure 4. A4.PET
Figure textile’s
A PET life
textile’s lifecycle
cycle (Redrawn from
(Redrawn from (A.(A. Payne
Payne et 2015))
et al., al., 2015))
[4]. (a)[4]. (a) Un-recycling
Un-recycling a PET a PET
garment
garment fromfrom cradle
cradle totograve.
grave.(b)
(b) Open-loop
Open-loop recycling
recyclingin ainPET garment’s
a PET life cycle.
garment’s (c) Closed-loop
life cycle. (c) Closed-loop
recycling in a PET garment’s life cycle.
recycling in a PET garment’s life cycle.
Open-loop recycling (OLP), also called “downcycling”, occurs at both the pre-
1.2. Challenges in Achieving
consumer and a Circular
post-consumer Economy
stages and in theeasier
is relatively Textile Industry
to achieve compared to closed-
loop
Onerecycling. In contrast,
of the primary closed-loop
challenges recycling involves
for achieving reprocessing
a circular economypolyester textiles,
in the textile industry
enabling them to reenter the same production cycle through chemical recycling. This
is the lack of efficient sorting methods because of complex fiber combinations in waste
process maintains high product value and aligns with circular economy (CE) principles,
textiles [5].emphasize
which Unlike plastic packaging
the elimination waste,
of waste and textiles
pollution,consist of varying
the circulation material
of products and blends,
complicating the the
materials, and recycling process.
regeneration In addressing this issue, our study focuses on SDG 9:
of nature.
Industry,The
Innovation, and Infrastructure,
transition toward by developing
a circular economy directly alignsadvanced artificial-intelligence-
with Sustainable Develop-
basedment Goal (SDG)
sorting models.12: At
Responsible
Taiwan’sConsumption
Institute ofand Production.
Industrial By implementing
Technology Researchclosed-
(ITRI), we
successfully created a highly accurate Raman-spectroscopy-based textile sorter. This in-
novation enables the precise sorting of textile materials, facilitating the transformation of
PET textile waste from a linear economy to a sustainable circular economy. This approach
can mitigate water pollution, minimize greenhouse gas emissions, and extend the product
 Sensors 2025, 25, 57 5 of 27

loop recycling and fostering innovative textile-sorting technologies, the environmental

impact of PET textiles can be minimized.

1.2. Challenges in Achieving a Circular Economy in the Textile Industry

One of the primary challenges for achieving a circular economy in the textile industry
is the lack of efficient sorting methods because of complex fiber combinations in waste
textiles [5]. Unlike plastic packaging waste, textiles consist of varying material blends,
25, x FOR PEER REVIEW complicating the recycling process. In addressing this issue, our study focuses 5 of 28on SDG 9:
Industry, Innovation, and Infrastructure, by developing advanced artificial-intelligence-
based sorting models. At Taiwan’s Institute of Industrial Technology Research (ITRI),
we successfully created a highly accurate Raman-spectroscopy-based textile sorter. This
innovation and fostering a global commitment to sustainable textile management. By in-
innovation enables the precise sorting of textile materials, facilitating the transformation of
tegrating advanced technologies, like AI-based sorting, industries can collaboratively pro-
PET textile waste from a linear economy to a sustainable circular economy. This approach
mote a circular can
economy
mitigatewhile
waterprotecting the environment
pollution, minimize greenhouseand meeting the
gas emissions, andrising
extendde-
the product
mand for sustainable materials.
life cycle, as shown in Figure 5.

Figure 5. Adding waste

Figure textile collection
5. Adding and sorting
waste textile to and
collection the production
sorting to theisproduction
critical forismoving from
critical for a from a
moving
linear economy tolinear
a circular economy.
economy to a circular economy.

Achieving these goals requires collaboration across sectors, countries, and organiza-
1.3. Vibrational Spectroscopy and AI for Textile Sorting
tions, emphasizing SDG 17: Partnerships for the Goals. Partnerships among policymakers,
Vibrationalresearchers,
spectroscopy [6], including
industries, near-infrared
and environmental (NIR) [7]
organizations areand Raman
essential tech- innova-
for driving
niques [8–10], is tion
widely used to analyze molecular compositions. NIR spectroscopy relies
and fostering a global commitment to sustainable textile management. By integrating
on vibrational bond absorption
advanced (from 780
technologies, likenm to 2526sorting,
AI-based nm), while Raman
industries canspectroscopy
collaborativelyde-
promote a
tects vibrational circular economyinelastic
states through while protecting theWe
scattering. environment and meeting
utilize informative the rising
Raman demand for
spectra
sustainable materials.
as a vibrational spectrum for analyzing textiles’ compositions, which serve as chemical
fingerprints. We combine the AI model with Raman spectroscopy for qualitative and
1.3. Vibrational Spectroscopy and AI for Textile Sorting
quantitative fiber analysis to classify fibers into purer categories and increase the flow of
Vibrational spectroscopy [6], including near-infrared (NIR) [7] and Raman tech-
closed-loop recycling, as shown in Figure 6.
niques [8–10], is widely used to analyze molecular compositions. NIR spectroscopy relies
on vibrational bond absorption (from 780 nm to 2526 nm), while Raman spectroscopy
detects vibrational states through inelastic scattering. We utilize informative Raman spectra
as a vibrational spectrum for analyzing textiles’ compositions, which serve as chemical
 Vibrational spectroscopy [6], including near-infrared (NIR) [7] and Raman tech-
niques [8–10], is widely used to analyze molecular compositions. NIR spectroscopy relies
on vibrational bond absorption (from 780 nm to 2526 nm), while Raman spectroscopy de-
Sensors 2025, 25, 57 6 of 27
tects vibrational states through inelastic scattering. We utilize informative Raman spectra
as a vibrational spectrum for analyzing textiles’ compositions, which serve as chemical
fingerprints. Wefingerprints.
combine theWeAIcombine
model thewithAIRaman
model withspectroscopy for qualitative
Raman spectroscopy and
for qualitative and
quantitative fiberquantitative
analysis tofiber
classify fibers
analysis into purer
to classify categories
fibers into purerand increase
categories the
and flow of
increase the flow of
closed-loop recycling, as shown
closed-loop in Figure
recycling, 6. in Figure 6.
as shown

Figure 6. The high-efficiency automatic textile-sorting technology combines Raman spectroscopy and
artificial intelligence technologies of data mining, machine learning, and deep learning.

1.4. Importance of Sorting for rPET Recycling

Polyester, a petroleum-based synthetic fiber, remains in high demand. The recyclability
potential index (RPI) [4] ranks PES first, for its environmental and economic advantages,
among ten commonly used fibers, as reported by Muthu et al. [11]. This ranking highlights
its recyclability potential for ecological sustainability.
Efficient textile sorting is essential for successful closed-loop recycling, particularly
for polyester textiles, as it ensures higher purity in material streams. Accurate sorting en-
ables pure polyester (PES) and cotton (CO) fibers to enter the recycling process, producing
high-quality recycled polyethylene terephthalate (rPET). Improperly sorted or low-purity
materials result in contamination, limiting high-value recycling and often leading to down-
cycling or disposal.
Sorting textiles into six distinct classes improves the efficiency of the recycling process
by separating materials with different properties, enabling the use of the most suitable
recycling method for each category.

1.5. Sorting and Recycling Processes for Different Textile Classes

The classification of textiles into six categories is essential for optimizing the recycling
process in Figure 7. By sorting textiles based on fiber composition and material content,
each category can be treated using the most effective recycling method. This ensures higher
purity, better material recovery, and efficient closed-loop recycling. The six categories are
based on the materials’ compositions, which determine the suitable recycling technique
and the quality of the recycled output [12–16].
 process in Figure 7. By sorting textiles based on fiber composition and material content,
each category can be treated using the most effective recycling method. This ensures
higher purity, better material recovery, and efficient closed-loop recycling. The six cate-
Sensors 2025, 25, 57 gories are based on the materials’ compositions, which determine the suitable recycling
7 of 27
technique and the quality of the recycled output [12–16].

Figure 7.
Figure Waste textile
7. Waste textile collection,
collection, sorting,
sorting, and
and recycling
recycling via
via mechanical
mechanical and
and chemical
chemical processes
processes for
for
pure-composition textiles and blended textiles, as well as the outputs after recycling.
pure-composition textiles and blended textiles, as well as the outputs after recycling.

1.5.1. 100% Cotton (CO)

1.5.1. 100% Cotton (CO)
• Reason for Sorting: Cotton is a natural fiber that can be efficiently recycled when not
• Reason for Sorting: Cotton is a natural fiber that can be efficiently recycled when not
contaminated by synthetic materials. Sorting cotton ensures higher purity during
contaminated by synthetic materials. Sorting cotton ensures higher purity during re-
recycling, leading to better-quality recycled cotton yarns or fibers;
cycling, leading to better-quality recycled cotton yarns or fibers;
• Recycling Method: Mechanical Recycling;
• Recycling Method: Mechanical Recycling;
• Process: Cotton is shredded, cleaned, carded, and spun into new yarns for reuse;
• Process: Cotton is shredded, cleaned, carded, and spun into new yarns for reuse;
• Recycling Outcome: High-quality recycled cotton yarns for new textile production.
• Recycling Outcome: High-quality recycled cotton yarns for new textile production.
1.5.2. 100% Polyester (PES)
• Reason for Sorting: Pure polyester can be recycled more efficiently compared to blends.
The absence of cotton or other materials ensures that the recycling process can focus
on re-polymerizing the polyester into rPET, maintaining its properties for reuse in
high-quality textile products;
• Recycling Method: Mechanical Recycling;
• Process: Polyester textiles are shredded, cleaned, carded, and spun into yarns for reuse;
• Recycling Outcome: High-quality recycled polyester (rPET).

1.5.3. Polyester–Cotton Blends (PES/CO ≥ 70%)

• Reason for Sorting: Polyester–cotton blends with a higher percentage of polyester
(≥70%) are common in the market because of their cost effectiveness and durability.
Sorting these textiles allows for the efficient separation of polyester and cotton for
recycling, enabling a higher yield of recycled polyester for rPET production;
• Recycling Method: Chemical Recycling;
• Separation Methods:
• Polyester Dissolution: Chemicals like DMSO dissolve polyester, leaving cotton intact.
The polyester is recovered, purified, and re-polymerized into rPET;
• Cellulose Degradation: Cotton is broken down using hydrolysis, leaving behind pure
polyester for rPET production;
• Recycling Outcome: High-quality rPET and cellulose for reuse.

1.5.4. Polyester–Cotton Blends (PES/CO < 70%)

• Reason for Sorting: Blends with less than 70% polyester are more challenging to recycle
because of the increased presence of cotton. Sorting them separately ensures that a
more appropriate recycling method is applied to maximize material recovery. This
Sensors 2025, 25, 57 8 of 27

group often requires additional chemical or combined methods to effectively separate

the fibers;
• Recycling Method: Mechanical or Combined Recycling;
• Process: Textiles are shredded and treated with chemical processes to separate
polyester and cotton fibers;
• Recycling Outcome: Lower-quality recycled materials, potentially requiring virgin
fibers to enhance product performance.

1.5.5. Polyester with Elastane (PES/EA)

• Reason for Sorting: Elastane is typically added in small quantities (<5%) to enhance
the elastic properties of textiles, improving stretchability and comfort. Sorting these
textiles ensures that the elastane content is taken into account when applying recycling
methods, thereby preventing the contamination of recycled polyester and maintaining
the quality of the final product;
• Recycling Method: Solvent-Based Recycling or Mechanical Shredding;
• Process:
# Solvent-Based Techniques: Selective solvents dissolve polyester while isolating
elastane for further treatment;
# Mechanical Shredding: Elastane-contaminated polyester can be shredded into
fibers, but this reduces the quality of recycled materials;
• Recycling Outcome: Recycled polyester with minimal elastane contamination, ensur-
ing better-quality rPET.

1.5.6. Others: Blends with Polyamide (PA) or Other Fibers (Three or More Materials)
• Reason for Sorting: Textiles containing polyamide (PA), nylon, or three or more
materials (such as polyester mixed with elastane, cotton, and polyamide) are highly
complex and difficult to recycle. Sorting these textiles allows for the identification
of the most appropriate recycling method for each fiber type, whether it involves
downcycling or valorization to value-added products;
• Recycling Method: Downcycling or Valorization;
• Process: Textiles that cannot be efficiently separated are often downcycled to lower-
quality products, like wood–plastic composites and insulation materials, or used for
energy recovery;
• Recycling Outcome: Limited reuse and potential disposal if not sorted correctly.

1.6. Valorization and the Circular Economy

By incorporating automatic sorting technology, higher purity in recycling streams can
be ensured, allowing for materials to be reused more efficiently and in higher-value appli-
cations. Higher purity leads to better-quality recycled outputs, such as recycled polyester
(rPET), which can be reintroduced to the textile production cycle. Waste textile valorization
to value-added products, like wood–plastic composites, insulation, or industrial materials,
is also possible as open-loop recycling.

1.7. Enhancing the Circular Economy and Sustainability

• Utilizing automatic sorting technology—by integrating Raman spectroscopy with
AI techniques, such as data mining, machine learning, and deep learning—we can
enhance the circular economy and sustainability, as shown in Figure 8, as follows:
# Enhance closed-loop recycling efficiency with high-purity PET textiles;
# Divert low-purity textiles to open-loop recycling, ensuring alternative uses
instead of disposal;
 Sensors 2025, 25, 57 9 of 27

# Support waste textile valorization, turning waste into value-added products

that contribute to resource sustainability;
# This approach transforms the textile life cycle from an un-recycling linear
Sensors 2025, 25, x FOR PEER REVIEW 9 of 28
economy to a circular economy, minimizing waste, reducing environmental
pollution, and ensuring sustainable textile management.

Figure 8. Proposed waste textile life cycle assessment with automatic sorting to achieve closed-loop
Figure 8. Proposed waste textile life cycle assessment with automatic sorting to achieve closed-loop
recycling and open-loop recycling with extended life cycle and a comparison with un-recycling.
recycling and open-loop recycling with extended life cycle and a comparison with un-recycling.
2. Literature Review
2. Literature Review
2.1. Raman Spectroscopy in Fiber and Textile Analysis
2.1. Raman Spectroscopy in Fiber and Textile Analysis
Raman scattering spectroscopy was introduced by D.A. Long [8] as a robust analytical
Raman scattering
methodology spectroscopy
for molecular wasanalysis.
vibration introduced by D.A. Long
Symmetric [8] as vibrations
molecular a robust analyti-
may be
cal
inactive in infrared spectroscopy, but they can still be observed in Raman scattering. Inmay
methodology for molecular vibration analysis. Symmetric molecular vibrations 1998,
beL.A.
inactive
Lyoninetinfrared spectroscopy,
al. [9] focused on Raman but spectroscopy
they can still be observed
with in Raman
a microscope forscattering. In
the biological
1998, L.A. Lyon
application of et al. [9] focused onRaman
surface-enhanced Ramanscattering
spectroscopy withIna 2000,
(SERS). microscope for the bio-
S.P. Mulvaney [10]
logical application of surface-enhanced Raman scattering (SERS).
reviewed Raman spectroscopy from 1997 to 1999. Since then, Raman spectroscopy In 2000, S.P. Mulvaney has
[10] reviewed Raman spectroscopy from 1997 to 1999. Since then, Raman
kept growing as an analytical chemical tool in various fields. In 2008, A. Kudelski [17] spectroscopy
has kept growing
reviewed as an analytical
the applications of Ramanchemical tool in various
spectroscopy fields.
from 2004 In 2008,
to 2008, A. Kudelski
showing [17]
that Raman
reviewed the applications
measurements can create ofmolecular
Raman spectroscopy
vibrationalfrom 2004that
spectra to 2008, showing
act like that Raman
a “fingerprint” for
measurements can create molecular vibrational spectra that act like
identifying compounds. Since then, more and more studies have used Raman spectroscopy a “fingerprint” for
identifying
to analyzecompounds.
materials like Since then,
fibers, more and
cellulose, andmore studies
textiles. havestudies
Notable used Raman
include spectros-
those by
copy
J.H. Wiley et al. [18] in 1987, H.G.M. Edwards et al. [19] in 1997, K. Kavkler et al. those
to analyze materials like fibers, cellulose, and textiles. Notable studies include [20] in
by2011,
J.H. and
Wiley
C. et al. [18]
Carey in [21]
et al. 1987,inH.G.M. Edwards
2015. Raman et al. [19]
scattering in 1997,
provides K. Kavkler
distinct spectralet al. [20]
patterns
inof2011, and C. Carey et al. [21] in 2015. Raman scattering
textiles, helping to differentiate between pure and blended fabrics. provides distinct spectral pat-
terns of textiles, helping to differentiate between pure and blended fabrics.
2.2. Machine-Learning and Deep-Learning Models for Raman Spectroscopy
2.2. Machine-Learning and Deep-Learning
Artificial intelligence (AI), includingModels for Raman Spectroscopy
machine-learning (ML) and deep-learning (DL),
hasArtificial
been used as a powerful
intelligence (AI),tool in many
including fields, such as image
machine-learning processing
(ML) and and chemical
deep-learning (DL),
analytical
has been used fields
as a[22]. In Raman
powerful tool inspectroscopy applications
many fields, such as image[23], AI models
processing such as k-
and chemical
nearest neighbors
analytical fields [22].(KNNs) [24],spectroscopy
In Raman decision treesapplications
(DTs) [25], random
[23], AIforests
models(RFs)
such[26], support
as k-near-
est neighbors (KNNs) [24], decision trees (DTs) [25], random forests (RFs) [26], support
vector machines (SVMs) [27–29], Bayesians [30], artificial neural networks (ANNs) [31],
convolutional neural networks (CNNs) [32], recurrent neural networks (RNNs) [33], and
generative adversarial networks (GANs) [34] have all made significant contributions.
 Sensors 2025, 25, 57 10 of 27

sors 2025, 25, x FOR PEER REVIEW 10 of

vector machines (SVMs) [27–29], Bayesians [30], artificial neural networks (ANNs) [31],
convolutional neural networks (CNNs) [32], recurrent neural networks (RNNs) [33], and
generative adversarial
[35], detecting bacterial networks
[36], and(GANs)
aiding[34] have all diagnosis
medical made significant
[37].contributions.
For environmental co
In the past decade, Raman spectroscopy combined with deep learning to extract
cerns, in 2022, E.R.K. Neo Neo [38] used Raman spectroscopy with KNNs and SVMs
spectral features for classification has proven to be effective for identifying component [35],
sort plastic wastes. In 2020, M.K. Maruthamuthu [39] applied CNN to detect microb
detecting bacterial [36], and aiding medical diagnosis [37]. For environmental concerns, in
contamination
2022, E.R.K. in
Neowater.
Neo [38]Inused
2022, P.Y.spectroscopy
Raman Kow [40] withproposed
KNNs andusing CNN
SVMs forplastic
to sort real-time a
quality estimation
wastes. based
In 2020, M.K. on images. [39] applied CNN to detect microbial contamination
Maruthamuthu
in water. many
Today, In 2022,studies
P.Y. Kowon [40]using
proposed usingspectroscopy
Raman CNN for real-time air-quality
with estimation
AI models are been pu
based on images.
lished in journals focused on sustainability.
Today, many studies on using Raman spectroscopy with AI models are been published
in journals focused on sustainability.
3. Methodology
3. Methodology
Due to the limited availability of waste-textile samples and dye-related fluorescen
Due to the limited availability of waste-textile samples and dye-related fluorescence
in the Raman data, we carefully manage Raman data through a series of steps. These ste
in the Raman data, we carefully manage Raman data through a series of steps. These steps
include datadata
include collection,
collection,data
data preprocessing, data
preprocessing, data mining,
mining, machine
machine learning
learning (ML) or de
(ML) or deep
learning (DL)(DL)
learning training,
training,and
andmodel training
model training andand testing,,
testing„ as shown
as shown in9.Figure 9.
in Figure

FigureFigure
9. The
9. experimental
The experimentalflow includes
flow includes data
data collection,
collection, preprocessing,
preprocessing, mining,and
mining, modeling, modeling,
testing. and te
ing. 3.1. Textile Labeling with FTIR Spectroscopy

The textile samples are collected from three textile factories in Taiwan. In addition to
3.1. Textile Labeling with FTIR Spectroscopy
using the composition labels on the washing tags for the ground-truth labels of the samples,
The
we usetextile
FTIRsamples are collected
(Thermo brand) fromfor
spectroscopy three textile factories
compositional in Taiwan.
verification to ensureInthe
addition
textile compositions’ correctness. This FTIR instrument is located in the
using the composition labels on the washing tags for the ground-truth labels of the sa laboratory of
the Material and Chemical Research Laboratories (MCLs) at the Industrial Technology
ples, we use FTIR (Thermo brand) spectroscopy for compositional verification to ensu
Research Institute (ITRI) [41] and is calibrated annually.
the textile compositions’ correctness. This FTIR instrument is located in the laboratory
3.2. Ramanand
the Material Online Hardware Research Laboratories (MCLs) at the Industrial Technolo
Chemical
Research We collect (ITRI)
Institute online Raman textile
[41] and data from hardware
is calibrated annually.at the Industrial Technology
Research Institute (ITRI) (US patent: US202/013730A1) in the actual fieldwork scenario, as
shown in Figure 10 [42].
3.2. Raman Online Hardware
We collect online Raman textile data from hardware at the Industrial Technolo
Research Institute (ITRI) (US patent: US202/013730A1) in the actual fieldwork scenario,
shown in Figure 10 [42].
 We collect online Raman textile data from hardware at the Industrial Techn
Research Institute (ITRI) (US patent: US202/013730A1) in the actual fieldwork scenar
Sensors 2025, 25, 57 shown in Figure 10 [42]. 11 of 27

Sensors 2025, 25, x FOR PEER REVIEW 11 of 28

Figure 10. The

Figure10. The hardware
hardware of the Raman
of the Raman detector
detectorand
andsoftware
softwareofofthe
theautomatic
automaticsorting
sortingML/DL
ML/DL model
model
(redrawn with permission from [42]).
(redrawn with permission from [42]).

•• Conveyer
Conveyer speed: 40 cm/s;
speed: 40 cm/s;
•• Camera
Camera integration time: 11 s;
integration time: s;
•• Excitation
Excitation laser
laser wavelength:
wavelength: 10641064nm;
nm;
•• Raman spectral range: − 1775~3597 cm −1 ;
Raman spectral range: −1775~3597 cm −1;

•• Sampling
Sampling Z-height
Z-height scan for signal
scan for signal optimization;
optimization;
•• Detection speed: 1 s per piece.
Detection speed: 1 second per piece.

3.3.
3.3. Data
Data Collection and Label
Collection and Label Distribution
Distribution
AAtotal
totalofof225
225textiles
textilesare scanned
are scanned using
using anan
online Raman
online Raman textile sorter,
textile withwith
sorter, scanscan
taken
at around ten different positions on each sample to collect Raman spectra.
taken at around ten different positions on each sample to collect Raman spectra. These These samples
include
samplessix types six
include textiles:
typespolyester (PES), cotton
textiles: polyester (CO),
(PES), a polyester-cotton
cotton blend with
(CO), a polyester-cotton over
blend
70% polyester (PES/CO with PES ≥ 70%), a polyester-cotton blend with
with over 70% polyester (PES/CO with PES ≥ 70%), a polyester-cotton blend with less than less than 70%
polyester (PES/CO
70% polyester (PES/CO withwith
PESPES< 70%), a polyester-elastane
< 70%), a polyester-elastaneblend
blend(PES/EL),
(PES/EL),andandnylon
nylonor
other mixed
or other mixedblends
blends(Others).
(Others).The Theoverall
overalldistribution thesesamples
distribution of these samplesisisshown,
shown,along
along
with
withindividual
individual class distribution chartisisshown
distribution chart shownininFigure
Figure11a,b.
11a,b.TheThedata
datafor
for the
the classes
classes
“PES/CO
“PES/CO with PES << 70%” 70%”and and“Others”
“Others”are areimbalanced
imbalancedandandneed
needtotobebe addressed
addressed before
before
training the model.
training the model.

Figure 11. Cont.

Sensors 2025,
Sensors 25, 25,
2025, 57 x FOR PEER REVIEW 12 of1228
of 27

(b)

Figure 11. These are the Raman data for six classes. (a) The six classes are the distribution of the
collected data. (b) The distribution in each class of blend combination.
 man spectrum preprocessing flow is required to remove fluorescence normalization and
filter out noise in spectra, as shown in Figure 12a.
To separate out the Raman signal, we perform a baseline correction with the im-
Sensors 2025, 25, 57 13 of 27
proved asymmetrically reweighted penalized-least-squares (IarPLS) method, proposed
by J. Ye et al. [44] in 2020, for Raman spectra, as shown in Figure 12b. After that, we per-
3.4. Data Preprocessing
formed the normalization for Fluorescence
after the baseline Background
removal. Reduction
In addition, toand Outlierthe
smooth Removal
data while
The dyeing of textiles is a process to transfer dyes from a
keeping the signal’s shape and width, we applied the Savitzky–Golay filter [45], proposed solution to fiber surfaces. In
that case, the laser will interact not only with textiles showing non-elastic Raman scattering
in 1964 by Savitzky and Golay in Analytical Chemistry, using a selected sliding window
but also with dyes showing a solid fluorescence background in spectra [43]. A Raman
with linear least squares [46]. The filtered Raman spectrum is the final data preprocessing
spectrum preprocessing flow is required to remove fluorescence normalization and filter
result. The preprocessed
out noise Raman
in spectra,spectra
as shown forinaFigure
total 12a.
of six classes are shown in Figure 12c.

Figure 12. Data

Figure 12. Data preprocessing preprocessing
of Raman of Raman
spectra spectra with background
with background correction,correction, S-G filtering,
S-G filtering, and in-and
intensity normalization. (a) The preprocess flow. (b) IaLS background correction before and after.
tensity normalization. (a) The preprocess flow. (b) IaLS background correction before and after. (c)
(c) The preprocessed Raman spectra of six classes.
The preprocessed Raman spectra of six classes.
To separate out the Raman signal, we perform a baseline correction with the improved
asymmetrically
3.5. Data Mining with reweighted
PCA and Outlier penalized-least-squares (IarPLS) method, proposed by J. Ye
Removal
et al. [44] in 2020, for Raman spectra, as shown in Figure 12b. After that, we performed the
The processed Raman spectra
normalization after thespan from
baseline 560 cm
removal.
−1 to 3430 cm−1 and have 342 dimen-
In addition, to smooth the data while keeping the
sions. To visualizesignal’s
the data,
shape we
andapplied principal
width, we applied component analysis
the Savitzky–Golay filter(PCA) to reduce
[45], proposed the by
in 1964
dimensionality fromSavitzky
342 and Golay
to 15, in Analytical
while retaining Chemistry,
over 95% using
of athe
selected sliding
original window with
information. linear
We,
least squares [46]. The filtered Raman spectrum is the final data preprocessing result. The
preprocessed Raman spectra for a total of six classes are shown in Figure 12c.

3.5. Data Mining with PCA and Outlier Removal

The processed Raman spectra span from 560 cm−1 to 3430 cm−1 and have 342 dimen-
sions. To visualize the data, we applied principal component analysis (PCA) to reduce
FOR PEER REVIEW
Sensors 2025, 25, 57 14 of14 28
of 27

the dimensionality from 342 to 15, while retaining over 95% of the original information.
then used the 15-dimensional
We, then used thedata to reconstruct
15-dimensional thereconstruct
data to Raman spectrum, and shown
the Raman spectrum, and in Fig-in
shown
ure 13a, the critical peak
Figure 13a,information was
the critical peak effectively
information waspreserved.
effectively preserved.

Figure 13. Data mining with

Figure 13. PCA
Data and
mining violin
with plots.
PCA and (a)plots.
violin Raman spectra
(a) Raman before
spectra PCA
before PCAand afterPCA
and after PCAwith
PC =violin
with PC = 15. (b) The 15. (b)plots
The violin plots for
for PC1, PC2,PC1,
andPC2,
PC3andfor
PC3six
forclasses.
six classes.

With the reduced dimensionality of 15-D, the violin plots for the top three PCs in
With the reduced dimensionality
the six classes of 15-D, the is
show that between-class violin plotsafter
separable for PCA,
the top three PCs
as shown in the
in Figure 13b.
six classes show Therefore,
that between-class
to achieve aishigher
separable afterofPCA,
efficiency as shown
the online system,in we
Figure 13b. There-
introduce AI to ML
fore, to achieve aand DL forefficiency
higher feature learning.
of theWe combine
online PCA with
system, ML models,AI
we introduce including
to MLKNN,and DL SVM,
for feature learning. We combine PCA with ML models, including KNN, SVM, and RF,the
and RF, while using deep-learning models of ANN and CNN with no reduction in
dimensionality of the 342-D Raman spectrum. AI technology is introduced to learn the
while using deep-learning models of ANN and CNN with no reduction in the dimension-
features of textile combinations with high-efficiency and high-accuracy classification tasks.
ality of the 342-D RamanIn dataspectrum.
mining, we AI technology
found is introduced
that the dope-dyed garment’sto learn
Ramanthe features
signal of
was buried
textile combinations
underwith high-efficiency
the dye’s and high-accuracy
fluorescence signal, as shown in Figure classification tasks. garment’s
14a, and the piece-dyed
spectrum
In data mining, we isfound
in the Figure 14b. dope-dyed
that the The spectrum garment’s
of the dope-dyed
Ramangarment exhibits
signal wasfluorescence
buried
signals about five times stronger than those of the piece-dyed garment. To keep the dataset
under the dye’s fluorescence signal, as shown in Figure 14a, and the piece-dyed garment’s
clean and prevent the dope-dyed garment’s spectrum from biasing the training process, we
spectrum is in the Figure 14b. The spectrum of the dope-dyed garment exhibits fluores-
remove them from our dataset.
cence signals about fiveDatatimes stronger
imbalance thaninduce
will highly thoseinherent
of the piece-dyed garment.
bias while training, To keep
and training willthe
tend
dataset clean and to prevent the adope-dyed
the class with large amountgarment’s spectrum
of data. In the from biasing
textile classification task, the training
having a biased
process, we removedataset
themin waste
fromcollection is unavoidable. In our collection, two classes are imbalanced:
our dataset.
PES/CO with PES ≥ 70% and Others. The small sample of PES/CO with PES ≥ 70% is
Data imbalance will highly induce inherent bias while training, and training will tend
because of its limited volume in the market [13]. First, we leave the test data out and then
to the class withperform
a largeup-sampling
amount of for data. In the textile classification task, having a biased
these two classes, as shown in the histogram in Figure 14c. The
dataset in waste test
collection is unavoidable.
data are used for performance In checks
our collection, two classes
in the test stage, while theare imbalanced:
augmented training
PES/CO with PES set≥is70%
used and Others.
for training and The small in
validation sample of PES/CO
the training stage. with PES ≥ 70% is be-
cause of its limited volume in the market [13]. First, we leave the test data out and then
perform up-sampling for these two classes, as shown in the histogram in Figure 14c. The
test data are used for performance checks in the test stage, while the augmented training
set is used for training and validation in the training stage.
R PEER REVIEW 15 of 28
Sensors 2025, 25, 57 15 of 27

Figure
Figure 14. Data mining for14. Data removal
outlier mining forand
outlier
dataremoval and data augmentation.
augmentation. (a) Raman
(a) Raman spectra of spectra of the
the doped-
doped-dyed garment. (b) Raman spectra of the piece-dyed garment. (c) Training/testing split and
dyed garment. (b) Raman spectra of the piece-dyed garment. (c) Training/testing split and up-sam-
up-sampling for PES/CO with PES ≥ 70% and “Others” bar and circular plots.
pling for PES/CO with PES ≥ 70% and “Others” bar and circular plots.
After data are preprocessed, the training data are used for training models of machine
learning and deep learning. However, with the limited data size, we need to consider the
After data are preprocessed, the training data are used for training models of ma-
model fitness of our limited training dataset and obtain sound generation while testing
chine learning andindeep learning.
both the However, with
under-parameterized the limited
machine-learning dataand
model size,
the we need to con-
over-parameterized
sider the model fitness of our model.
deep-learning limited training dataset and obtain sound generation while
Therefore, we will discuss
testing in both the under-parameterized the training strategy in
machine-learning Sections
model and3.6.1–3.6.3
the over-parame-
terized deep-learning model.
3.6. Model-Training Strategy
Therefore, we3.6.1.
willModel
discuss the training strategy in Sections 3.6.1.–3.6.3.
Fitness
If the model fits the data well, it usually exhibits a lower training error rate in the
3.6. Model-Trainingtraining
Strategy
process. However, when applied to testing unknown data, there may be a variance
in the performance. Hence, a validation set from the training set for the fitted-model check
3.6.1. Model Fitness
is expected to be split. This is known as the “bias–variance” tradeoff property in machine
If the model learning,
fits the as shown
data in Figure
well, 15b. Belkin
it usually et al. [47]
exhibits proposed
a lower a double-descent
training risk curve
error rate in thein
2019, as shown in Figure 15a. It shows evidence that increasing the function class capacity
training process. However, when applied to testing unknown data, there may be a vari-
improves the classifier’s performance with no concern for the inductive bias because the
ance in the performance.
function Hence, a validation
can explicitly set
or implicitly fromit,the
handle training
as shown set for
in Figure 15c.the fitted-model
check is expected to be split. This is known as the “bias–variance” tradeoff property in
machine learning, as shown in Figure 15b. Belkin et al. [47] proposed a double-descent
risk curve in 2019, as shown in Figure 15a. It shows evidence that increasing the function
class capacity improves the classifier’s performance with no concern for the inductive bias
because the function can explicitly or implicitly handle it, as shown in Figure 15c.
Sensors 2025, 25, x FOR PEER REVIEW 16 of 28

OR PEER REVIEW 16 of 28
Sensors 2025, 25, 57 16 of 27

Figure 15. Model fitness for the dataset. (a) Double-descent risk curve in the classical ML- and DL-
error-curve regime (redrawn based on [47]). (b) The classical regime of the ML model’s error curve
(c) Over-parameterized DL model’s error curve.
Figure
Figure 15. Model fitness Model
for15.the fitness
dataset. (a)for the dataset. (a) Double-descent
Double-descent risk curve in risk curve in the
the classical ML-classical
and DL-ML- and
DL-error-curve regime (redrawn based on [47]). (b) The classical regime of the ML model’s error
3.6.2.
error-curve regime Machine-Learning
(redrawn based on [47]).Optimization Strategy
(b) The classical regime of the ML model’s error curve.
curve. (c) Over-parameterized DL model’s error curve.
(c) Over-parameterized
In DL model’s
machine error curve.
learning, a bias–variance tradeoff needs to be handled. With the limita-
3.6.2. Machine-Learning Optimization Strategy
tion of the training-data size, to fit the under-parameters of the machine-learning model
3.6.2. Machine-Learning In machine learning,Strategya bias–variance tradeoff needs to be handled. With the limitation
we reduceOptimization
the input parameter dimensionality from 342D to 15D with PCA, as shown in
of the training-data size, to fit the under-parameters of the machine-learning model, we
In machineFigure 16a.the
learning,
reduce This provides
a bias–variance
input parameter extra benefits
tradeoff forfrom
needs
dimensionality noise
to be reduction
handled.
342D and
Withfast
to 15D with theinference
PCA, shown efficiency
limita-
as in
In classical
Figure 16a. This provides extra benefits for noise reduction and fast inference efficiency. In over-
tion of the training-data machine
size, to fit learning,
the cross-validation
under-parameters of is
thecommonly conducted
machine-learning to deal
model, with
fitting
we reduce the input and underfitting.
classical
parameter In our
machinedimensionality
learning, case, we perform
cross-validation
from to fivefold
is commonly
342D cross-validation
15Dconducted
with PCA,to deal forinmodeling, as
with overfitting
as shown
and in
underfitting. In ourto case,
Figure 16a. Thisshown
provides Figure
extra16b,benefits forwe
obtain perform
the
noise “sweet fivefold
reduction cross-validation
point”and(without for modeling,
overfitting
fast inference andas shown
efficiency. underfitting)
in Figure
The training 16b,
flow to obtain the
plotted in Figure“sweet point” (without
16c is usedconducted overfitting
to train a good and underfitting).
generalization The
model by
In classical machine learning, cross-validation is commonly to deal
training flow plotted in Figure 16c is used to train a good generalization model by fitting
with over-
fitting training
fitting and underfitting. ourdata
Indata towe generate thefivefold
model and test the validation set with aashigh degree
training tocase,
generate perform
the model and test thecross-validation
validation set withfor modeling,
a high degree accuracy
shown in Figure accuracy
16b, to (in the
obtain case
the of underfitting)
“sweet point” and
(withoutslight variance
overfitting (in
andthe case of overfitting). This
underfitting).
(in the case of underfitting) and slight variance (in the case of overfitting). This is a recursive
is a process
The training flow recursive
plotted for process
Figure for
insearching for searching
16c an used tofor
is optimal an optimal
a good under-parameterized
under-parameterized
train machine model
generalization that machine
model includes
by mode
thatKNN,
includes
SVM,KNN,
and RF. SVM, and RF.
fitting training data to generate the model and test the validation set with a high degree
accuracy (in the case of underfitting) and slight variance (in the case of overfitting). This
is a recursive process for searching for an optimal under-parameterized machine model
that includes KNN, SVM, and RF.

Figure 16. Cont.

Sensors 2025, 25, x FOR PEER REVIEW 17 of 28
Sensors 2025, 25, 57 17 of 27

Figure
Figure 16.16. Machine-learning model-training
Machine-learning model-training strategies
strategiesof of
(a)(a)
PCA-based dimensionality
PCA-based reduction
dimensionality reduction
andand
(b)(b) fivefold cross-validation. (c) Training, validation, and test flow.
fivefold cross-validation. (c) Training, validation, and test flow.
3.6.3. Deep-Learning Optimization Strategy
3.6.3. Deep-Learning Optimization Strategy
Recently, deep learning has been proven to be powerful in pattern recognition [48], includ-
ingRecently, deep learning
image classification has been proven
[49], recommendation to be
systems [50],powerful
and objectin pattern[51]
detection recognition
(Zhao Z.Q. [48],
et al., 2019).
including imageWithclassification
the double-descent
[49], risk curve, the training
recommendation loss of the
systems over-parameterized
[50], and object detection
[51] (Zhao Z.Q. et al., 2019). With the double-descent risk curve, theThere
deep-learning model will go further after the implementation of a threshold. are several
training loss of the
commonly used strategies for deep-learning model training based on
over-parameterized deep-learning model will go further after the implementation of athe studies and findings
in recent years, not only for 3D image data [42] but also for 1D spectral data [43].
threshold. There are several commonly used strategies for deep-learning model training
In the deep-learning model optimization, we use the validation set for early stop-
based on the studies and findings in recent years, not only for 3D image data [42] but also
ping [52] once the loss is going to the minimum. The hyperparameter of the learning
forrate
1D [53]
spectral data [43].
is tuned to make the model converge to the local minimum position. Another
In the deep-learning
hyperparameter, model
batch size optimization,
[54], is used to searchwe foruse
the the validation
optimal set for earlywith
model convergence stopping
[52]anonce the loss
acceptable is going to the
computational cost.minimum. The hyperparameter of the learning rate [53]
is tuned to make the model converge to the local minimum position. Another hyperpa-
3.7. Model-Training
rameter, and -Testing
batch size [54], is usedAccuracy
to search for the optimal model convergence with an ac-
3.7.1. Machine-Learning Training and Testing Accuracies
ceptable computational cost.
In this section, we compare the performance of three widely used machine learning
models—K-Nearest
3.7. Model-Training and Neighbors
-Testing(KNN),
AccuracySupport Vector Classifier (SVC), and Random Forest
(RF)—based on their training and testing accuracies. Each model was optimized using
3.7.1. Machine-Learning Training and Testing Accuracies
cross-validation and evaluated on a test set to determine how well it generalizes to unseen
In this
data. Belowsection, we compare
is a detailed theofperformance
evaluation of three widely
each model’s performance, used their
including machine learning
strengths,
models—K-Nearest Neighbors
weaknesses, and specific (KNN),
challenges Supportduring
encountered Vectorthe
Classifier
evaluation(SVC), and Random
process.
Forest
• (RF)—based on their(KNNs)
K-Nearest Neighbors training and testing accuracies. Each model was optimized
[24]
using cross-validation and evaluated on a test set to determine how well it generalizes to
KNN is a well-known machine learning model that classifies an unknown sample based
unseen data. Below is a detailed evaluation of each model’s performance, including their
on its nearest neighbors. After optimizing the model using cross-validation, the value of n (the
strengths,
number ofweaknesses, andsetspecific
neighbors) was challenges
to 4. This resulted in encountered
a high trainingduring
accuracythe evaluation
of 97%. However, pro-
cess.
• K-Nearest Neighbors (KNNs) [24]
KNN is a well-known machine learning model that classifies an unknown sample
based on its nearest neighbors. After optimizing the model using cross-validation, the
 The SVM is a machine learning model that works by maximizing the margin be
classes using support vectors [55]. We used the radial-basis-function (RBF) kernel
one-vs-one approach, where each class is compared against others. After tuning t
Sensors 2025, 25, 57 18 of 27
perparameters (C = 50,000, and gamma is 0.01) using fivefold cross-validation, we
model evaluation. The SVM shows strong performance but with some trade-offs in
of stability and generalization.
when evaluated on test data, KNN achieved a testing accuracy of 94%. A notable challenge
with KNN is that• it tends
Random Forest (RF)
to misclassify [26] textiles as either 100% PES or PES/CO when
PES/EA
the polyester contentRandom
is over 90%,Forest (RF) is
leading toaconfusion
machine between
learning these
modelcategories.
that uses multiple decision t
an ensemble. Each decision trees [25] splits the data based on key features at each n
• Support Vector Machine (SVM) [27,28]
reduce entropy. To avoid over-fitting, we optimize the n_estimators parameter usin
The SVM is a machine
fold learning model
cross-validation. that works
The optimized by maximizingare
hyperparameters themax_depth
margin between
= 25, and n_esti
classes using support vectors [55]. We used the radial-basis-function
= 150. Random Forest stands out as the best-performing model, achieving (RBF) kernel withthe highe
a one-vs-one approach,
ing accuracy (90.0%) and the lowest variance (0.43%) across 5-foldthe
where each class is compared against others. After tuning cross-valid
hyperparameters (C = 50,000, and gamma is 0.01) using
demonstrating strong generalization and stability. fivefold cross-validation, we do the
model evaluation. The SVM shows strong performance but with some trade-offs in terms
• Machine learning performance comparison:
of stability and generalization.
The confusion matrix in Figure 17a shows how well KNN, RF, and SVC perfo
• Random Forest (RF) [26]
classifying the test data. It highlights the strengths and weaknesses of each mode
box plot
Random Forest in Figure
(RF) 17b illustrates
is a machine learningthe variation
model in accuracy
that uses multiple and variance
decision for KNN, R
trees
in an ensemble. SVC across
Each 5-foldtrees
decision cross-validation,
[25] splits theshowing
data basedhowon each
keymodel performs
features at eachunder di
node to reducedata splits.To avoid over-fitting, we optimize the n_estimators parameter
entropy.
SVC achievesThe
using fivefold cross-validation. theoptimized
highest validation accuracy are
hyperparameters (99.0%) but has= higher
max_depth 25, andvariance
as shown
n_estimators = 150. Random in Table
Forest1, suggesting
stands out itasisthelessbest-performing
stable across different
model,data splits.
achieving
KNN has(90.0%)
the highest testing accuracy a high training
and theaccuracy (97.0%) but
lowest variance struggles
(0.43%) acrosswith overfitting,
5-fold cross- as ind
by its lowerstrong
validation, demonstrating testinggeneralization
accuracy (89.5%) andand higher variance (5.7%) in Table 1.
stability.
• Machine learning Random Forest (RF)
performance delivers the best overall performance with a high testing
comparison:
racy (90.0%) and very low variance (0.43%), as shown in Table 1, making it the mo
The confusion
sistentmatrix in Figure
and reliable 17a shows how well KNN, RF, and SVC perform in
model.
classifying the test data. It highlights
Among these three ML the strengths
models,andRandom
weaknesses of each
Forest (RF)model.
stands The
out box
as the best
plot in Figure 17b illustrates the variation in accuracy and variance for KNN,
among the three. It balances high accuracy and low variance, making it the most RF, and SVC
across 5-fold cross-validation,
and reliable choice showing how each model
for generalization inperforms
this task.under different data splits.

Sensors 2025, 25, x FOR PEER REVIEW 1

Figure 17.
Figure 17. Performance Performance of machine-learning
of machine-learning model
model training and training
testing. and confusion
(a) The testing. (a)matrix
The confusion
showing
showing the results the machine
for three results for three machine
learning models:learning
KNN, SVC,models:
and KNN,
RF. (b)SVC, and RF. (b) The cross-val
The cross-validation
box plot, trainingbox
andplot, training
testing and testing
accuracies of theaccuracies
KNN, SVC, ofand
the RF
KNN, SVC, and RF models.
models.

Table 1. Comparison of KNN, SVC, and RF Models: Training, Validation, and Testing Accu

KNN SVC R
Validation accuracy 88.40% 99.00% 100
Five fold cross-validation
Sensors 2025, 25, 57 19 of 27

SVC achieves the highest validation accuracy (99.0%) but has higher variance (4.9%)
as shown in Table 1, suggesting it is less stable across different data splits.

Table 1. Comparison of KNN, SVC, and RF Models: Training, Validation, and Testing Accuracies.

KNN SVC RF

Five fold Validation accuracy 88.40% 99.00% 100.00%

cross-validation Validation STD 5.70% 4.90% 0.43%
Training accuracy 97.00% 90.40% 91.40%
Testing accuracy 89.50% 88.00% 90.00%

KNN has a high training accuracy (97.0%) but struggles with overfitting, as indicated
by its lower testing accuracy (89.5%) and higher variance (5.7%) in Table 1.
Random Forest (RF) delivers the best overall performance with a high testing accuracy
(90.0%) and very low variance (0.43%), as shown in Table 1, making it the most consistent
and reliable model.
Among these three ML models, Random Forest (RF) stands out as the best model
among the three. It balances high accuracy and low variance, making it the most stable and
reliable choice for generalization in this task.

3.7.2. Deep-Learning Training and Testing Accuracies

The ANN and CNN models were trained using optimized hyperparameters to effi-
ciently classify Raman spectra, achieving fast convergence without overfitting, as shown in
the training/validation curves.
• Artificial Neural Network (ANN)
We used an artificial neural network (ANN) model, as neural networks have been
proven to perform well for Raman classification tasks [56]. The architecture of the ANN
consists of four fully connected layers, with hidden layers of 128, 64, and 32 neurons. Each
hidden layer uses the Mish activation function [57], which introduces nonlinearity to the
model. The input spectrum has a dimension of 342, and the output layer has six categories,
with a total of 54,438 trainable parameters.
• Learning rate: 0.001;
• Number of learning epochs: 200;
• Batch size: 128;
• Optimizer: Adam [58];
• Convolutional neural network (CNN).
A convolutional neural network (CNN) [32] model was chosen for feature extraction
from the Raman spectra. CNNs use a shared kernel that convolves with a sliding window
through the data and utilizes backpropagation to optimize the kernels. These learned
kernels serve as effective feature extraction tools. CNNs are widely used in image classifi-
cation, particularly with large datasets, such as CIFAR [59] and ImageNet [60], and have
shown strong performance. In our task, we used three layers of 1D CNNs with a kernel
size of 5 × 5 and filter numbers of 32, 64, and 128. Each convolution output uses the Swish
activation function [61]. The model has a total of 68,198 trainable parameters.
• Learning rate: 0.001;
• Number of learning epochs: 200;
• Batch size: 128;
• Optimizer: Adam [58].
25, 25, x FOR PEER REVIEW 20 of 28
Sensors 2025, 25, x FOR PEER REVIEW 20 of 28
Sensors 2025, 25, 57 20 of 27

The training and validation curves for both models, shown in Figure 18, demonstrate
The training and validation curves for both models, shown in Figure 18, demonstrate
that both models converge quickly
The training within 200
and validation epochs
curves without
for both overfitting,
models, shown inbased
Figureon
18,the
demonstrate
that both models converge quickly within 200 epochs without overfitting, based on the
hyperparametersthatprovided.
both models converge quickly within 200 epochs without overfitting, based on the
hyperparameters
hyperparametersprovided.
provided.

Figure 18. Deep-learning model’s loss and accuracy curves. (a) ANN’s training/validation loss and
Figure 18.
18.Deep-learning
Deep-learning model’s loss
loss and
andaccuracy
accuracycurves.
curves.(a)(a) ANN’s training/validation
ANN’s lossloss
andand
accuracy withinFigure
200 epochs. (b) CNN’s training/validation loss and accuracy withintraining/validation
200 epochs.
accuracywithin
accuracy within200
200epochs.
epochs. (b) CNN’s
CNN’s training/validation
training/validationloss
lossand
andaccuracy
accuracywithin 200
within epochs.
200 epochs.
The confusionThematrix in Figure
confusion matrix 19 shows that the that
ANN performs betterbetter
and and
has has higher
The confusion matrixininFigure 19
Figure shows
19 shows the
thatANN
the performs
ANN performs better and has
higher accuracy on the on
accuracy testthe
data
testcompared to the to
data compared CNNthe model. The CNN
CNN model. The model misclas-
CNN model misclassifies
higher accuracy on the test data compared to the CNN model. The CNN model misclas-
sifies some PES/EL
someas pure polyester,
PES/EL likely because
as pure polyester, the blended
likely because samples
the blended containcontain
samples less than
less than 10%
sifies some PES/EL as pure polyester, likely because the blended samples contain less than
10% elastane, making
elastane,them
makingharder
themfor the model
harder for theto distinguish
model accurately.
to distinguish accurately.
10% elastane, making them harder for the model to distinguish accurately.

Figure 19. Deep-learning-model confusion matrices

Figure 19. Deep-learning-model of thematrices
confusion ANN and CNN
of the with
ANN andthe
CNNtestwith
dataset.
the test dataset.
Figure 19. Deep-learning-model confusion matrices of the ANN and CNN with the test dataset.
x FOR PEER REVIEW 21 of 28
Sensors 2025, 25, 57 21 of 27

3.8. Misclassification
3.8.Check for the ANN
Misclassification Model
Check for the ANN Model
The confusion matrix for the matrix
The confusion ANN for
model, illustrated
the ANN in Figurein20a,
model, illustrated reveals
Figure threethree types
20a, reveals
types of misclassifications:
of misclassifications:

Figure 20. (a) Confusion

Figurematrix
20. (a) of the ANN
Confusion model
matrix with
of the the model
ANN test dataset. (b)test
with the Polyester wasPolyester
dataset. (b) wronglywas wrongly
classified as PES/COclassified
with PES as≥PES/CO
70%. (c)with
PES/CO ≥ 70%.
PES with PES(c)≥ PES/CO
70% waswith PES ≥
wrongly 70% wasas
classified wrongly
Polyes-classified as
Polyester. (d) PES/CO with PES ≥ 70% was wrongly classified as PES/CO with PES < 70%.
ter. (d) PES/CO with PES ≥ 70% was wrongly classified as PES/CO with PES < 70%.
Five samples were misclassified, where 100% polyester (PES) was incorrectly identified
Five samplesas were misclassified,
PES/CO with PES ≥where 100%
70%. This polyester (PES)
misclassification was incorrectly
occurred because of identi-
the absence of the
fied as PES/CO withC-OPES
bond’s≥ 70%.
peakThis misclassification
at 2903 − 1
cm in the spectrum,occurred because
which of the absence
is characteristic of Raman
of cotton’s
the C-O bond’s peak at 2903 cm⁻ 1 in the spectrum, which is characteristic of cotton’s Ra-
signal (Figure 20b) [62].
man signal (Figure 20b) [62].
Two samples were misclassified from PES/CO with PES ≥ 70% to 100% PES, as
Two samplesshown in Figure 20c. from
were misclassified This likely
PES/CO resulted
with from
PES ≥the limited
70% sample
to 100% PES,size for PES/CO with
as shown
PES ≥
in Figure 20c. This likely resulted from the limited sample size for PES/CO with of
70%, biasing the model toward PES. This issue reflects the scarcity such
PES ≥ textiles in
the market.
70%, biasing the model toward PES. This issue reflects the scarcity of such textiles in the
Similarly, three samples were misclassified from PES/CO with PES ≥ 70% to PES/CO
market.
with PES < 70% (Figure 20d). The C-H vibrational bond at 2903 cm−1 in spectrum for
Similarly, three samples were misclassified from PES/CO with PES ≥ 70% to PES/CO
PES/CO with PES < 70% exhibits a high peak intensity [62], depicted as a green line.
with PES < 70% (Figure 20d). The C-H vibrational bond at 2903 cm⁻1 in spectrum for
The intensity offset between the green and red lines likely caused this misclassification.
PES/CO with PESImproving
< 70% exhibits a high peakratio
the signal-to-noise intensity [62], depicted
by adjusting as atime
the exposure green line.
could Theto eliminate
help
intensity offset between the green
this background andHowever,
offset. red linesthis likely caused this ismisclassification.
misclassification Im- process,
tolerable in the recycling
proving the signal-to-noise
as both cases ratio by adjusting
involve the exposure
chemical recycling time could
and require help to
only minor eliminate to process
adjustments
this background offset. However,
parameters, ensuringthisminimal
misclassification is tolerable
loss in the rPET in the
output ratio [12].recycling pro-
Additionally, two samples of 100% polyethylene
cess, as both cases involve chemical recycling and require only minor adjustments (PE) were misclassified
to pro- as PE/EA.
However,minimal
cess parameters, ensuring advancements
loss ininthe
recycling technology,
rPET output proposed in 2024, allow for effective
ratio [12].
Additionally, two samples of 100% polyethylene (PE) were misclassified as PE/EA.
However, advancements in recycling technology, proposed in 2024, allow for effective
separation of PE from PE/EA blends, yielding rPET with chemical properties comparable
to those of pure polyester. Thus, this misclassification is acceptable within the PE/EA re-
Sensors 2025, 25, 57 22 of 27

separation of PE from PE/EA blends, yielding rPET with chemical properties comparable
to those of pure polyester. Thus, this misclassification is acceptable within the PE/EA
recycling process and maintains reasonable output yield [13].

4. Results and Discussion

The accuracies of the three machine-learning (ML) models and two deep-learning
(DL) models were evaluated through cross-validation, training, and testing. Among the
ML models, The RF exhibited the lowest variance during training. However, all three ML
models achieved an approximately 90% testing accuracy. Misclassifications were the most
frequent in classes with the lowest data proportions, such as PES/CO with PES ≥ 70% and
PES/EA blended textiles. The RF delivered the best performance because of its ensemble
nature, which is well suited for classification tasks.
In comparison, The ANN achieved the highest testing accuracy of 96.9%, while the
CNN achieved 93.5%, as summarized in Table 2.

Table 2. Classification accuracies of three machine-learning models and two deep-learning models.

KNN SVC RF ANN CNN

Training accuracy 97.00% 90.40% 91.40% 96.20% 98.00%
Validation accuracy 88.40% 99.00% 100.00% 93.50% 97.00%
Testing accuracy 89.50% 88.00% 90.00% 96.90% 93.50%

The integration of vibrational spectroscopy and AI demonstrates the feasibility of

addressing the challenging task of textile material classification, including the classification
of blended compositions. The Raman spectra’s high resolution provides detailed vibrational
information that AI models can leverage to extract features, map them to the feature domain,
and optimize the classification through learned weights.
The results of both classical ML and modern DL models align with the double-descent
risk curve phenomenon. In ML models, such as KNN, SVM, and RF, the accuracy exceeds
90%, reflecting a balance between bias and variance. In DL models, such as ANN and
CNN, the losses diminish beyond the interpolation threshold, achieving accuracies above
95%. Among the models, the ANN emerged as the most accurate for the proposed Raman-
spectroscopy-based Textile Sorter, achieving a 96.9% testing accuracy. The ANN captures
global attention through the connections between neurons, whereas the CNN employs
local attention via sliding convolutions for feature extraction from Raman spectra.
Table 3 shows the number of correctly classified samples and the classification accuracy
by fabric type using the Artificial Neural Network (ANN) model.

Table 3. Number of correctly classified samples and classification accuracy by fabric type using
ANN model.

Number of Samples Accuracy of the ANN

Fabric Type Number of Samples
Correctly Classified Model (%)
Cotton (CO) 90 90 100.0%
Polyester (PES) 88 81 92.0%
PES/CO with PES ≥ 70% 14 9 64.3%
PES/CO with PES < 70% 41 40 97.5%
PES/EA 37 37 100.0%
Others 24 24 100.0%

The low accuracy for PES/CO with PES ≥ 70% is primarily because of data imbalance
in this class. Addressing this limitation through expanded training data collection could
Sensors 2025, 25, 57 23 of 27

enhance the model’s performance. Additionally, PES/CO textiles with PES ≥ 70% are less
prevalent in the market, consistent with findings in [13], which reported that Hennes &
Mauritz AB (H&M)’s samples included limited PES/CO blends, such as CO/PES = 80/20,
60/40, and 40/60. Despite this, the misclassification is manageable in the recycling process,
as both cases involve chemical recycling with minimal adjustments to process parameters,
ensuring negligible rPET output loss [12].
Overall, the remaining five fabric classes are classified with a high degree of accuracy,
confirming the robustness of the sorting process. This ensures that the subsequent recycling
phase achieves a closed-loop flow with high efficiency and purity.

5. Conclusions
Currently, without sorting, textile waste faces significant challenges. The majority
ends up in landfills or is incinerated, with around 85% of textiles not being meaningfully
recycled. Textiles in landfills are non-biodegradable, leaching harmful chemicals into soil
and groundwater, and taking up valuable space. Incineration releases toxic pollutants into
the air and contributes to climate change, while valuable materials are lost. Both landfill
disposal and incineration represent un-recycling processes, as they do not recover or reuse
these materials, contributing to environmental harm and resource depletion.
With effective sorting, both closed-loop recycling (fiber to fiber) and open-loop recy-
cling become achievable. Sorting allows for high-quality fiber separation, enabling the
recovery and reuse of materials in new textile products (closed loop), or repurposing waste
into value-added products, like insulation or construction materials (open loop). This
reduces environmental impacts, conserves resources, and supports a more sustainable,
circular textile economy.

5.1. Raman Spectroscopy with AI for Waste Textile Sorting

We demonstrate a three-fold benefit of integrating AI with Raman spectroscopy to
achieve a circular economy:
• High Degree of Accuracy in Fiber Type Classification: Achieving over 95% accuracy
in distinguishing between fiber types;
• High Degree of Accuracy in Blend Compositional Analysis: Achieving over 95%
accuracy in analyzing blended-fiber compositions;
• High Throughput: Enabling automatic sorting at a speed of one piece per second,
replacing manual sorting processes.
Artificial intelligence (AI) implemented through the ANN model in Raman spec-
troscopy excels in achieving a classification accuracy exceeding 96%. Both machine-learning
(ML) and deep-learning (DL) models effectively learn the patterns of textiles’ Raman spec-
tra. Notably, DL models outperformed ML models, with an 8.9% increase in accuracy.
Among DL models, the global attention mechanism of the ANN further improved the accu-
racy by 3.4% compared to that of the local attention mechanism of the CNN. This integrated
system enables real-time, high-efficiency, automatic textile sorting, with a Raman signal
integration time of one second and a conveyor detection speed of one piece per second.

5.2. Data Preprocessing for Enhanced AI Modeling

Data preprocessing is crucial for optimizing Raman spectroscopy’s performance in AI
modeling. The key steps include the following:
• Dimensionality Reduction: Enhancing machine-learning models by reducing complexity;
• Balanced Dataset Preparation: Ensuring clean and relatively balanced datasets for
effective model training;
Sensors 2025, 25, 57 24 of 27

• Fluorescence Signal Reduction: Removing dyes’ fluorescence signals to retain only

vibrational spectral information essential for AI modeling.
A current limitation of Raman spectroscopy in separating waste textiles is the signal
interference caused by dope-dyed fabrics with high concentrations of dark colors [63,64].
However, growing global attention to water pollution caused by dyeing processes [43,65,66]
suggests potential solutions. Pretreatment processes for dye removal could mitigate this
limitation, facilitating clearer Raman signals for accurate classification.

5.3. Achieving Qualitative and Quantitative Sorting with High Degrees of Accuracy and Efficiency
Raman spectroscopy combined with the ANN model enables both qualitative and
quantitative sorting with high degrees of accuracy and efficiency. Unlike NIR spectroscopy,
which lacks precise quantification, Raman spectroscopy can accurately measure fiber
composition and contamination levels. This system supports:
• Closed-Loop Recycling: Sorting purer textiles for recycling into high-quality recycled
PES fibers;
• Open-Loop Recycling: Valorizing waste textiles to value-added products, such as
wood–plastic composites [14];
• By improving recycling rates and extending textile lifespans, this technology helps to
meet the demand for recycled polyester and supports a circular economy.

5.4. Future Directions

To further enhance the Raman-spectroscopy-based autoclassification system, future
work could explore:
• Transformer Models for Feature Extraction: Implementing transformers [67] in Raman
spectroscopy [68] to visualize critical features and understand wavenumber shifts in
the spectra of waste textiles;
• Expanding Textiles’ Raman Spectral Datasets: Reducing data imbalance, particularly
for textile classes like PES/CO with PES ≥ 70%, through dataset expansion;
• Integrated Background Correction: Utilizing autoencoders, for self-adaptive learn-
ing to correct for background interference, seamlessly connected to ANN or CNN
networks for one-stage classification.
In conclusion, the integration of Raman spectroscopy with ANNs delivers a powerful
solution for waste-textile sorting, combining high degrees of accuracy, efficiency, and
adaptability. This technology paves the way for sustainable recycling practices and a robust
circular economy for textiles.

Author Contributions: Conceptualization, P.-F.T. and S.-M.Y.; methodology, P.-F.T.; software, P.-
F.T.; validation, P.-F.T. and S.-M.Y.; formal analysis, P.-F.T.; writing—original draft preparation,
P.-F.T.; writing—review and editing, S.-M.Y.; visualization, P.-F.T.; supervision, S.-M.Y.; project
administration, P.-F.T. All authors have read and agreed to the published version of the manuscript.

Funding: This research received no external funding.

Institutional Review Board Statement: Not applicable.

Informed Consent Statement: Not applicable.

Data Availability Statement: The data are available upon request.

Conflicts of Interest: The authors declare no conflict of interest.

Sensors 2025, 25, 57 25 of 27

Abbreviations
Abbreviation Definition
3R reduction, reuse, and recycling
AI artificial intelligence
ANN artificial neural network
CE circular economy
CLR closed-loop recycling
CNN convolutional neural network
CO cotton
DL deep learning
EA elastance (spandex)
ESG environmentally sustainable governance
FTIR Fourier-transform infrared
IRS infrared spectroscopy
ITRI Industrial Technology Research Institute
kNN k-nearest neighbor
LCA life cycle assessment
ML machine learning
NIRS near-infrared spectroscopy
OLR open-loop recycling
PC principal component
PCA principal component analysis
PES polyester
PES/CO polyester and cotton blend
PES/EA polyester and elastance blend
PA polyamide
RF random forest
rPET recycled polyethylene terephthalate
S-G filter. Savitzky–Golay filter
SVM support vector machine

References
1. Stegmann, P.; Daioglou, V.; Londo, M.; Vuuren, D.P.V.; Junginger, M. Plastic futures and their CO2 emissions. Nature 2022, 612,
272–276. [CrossRef]
2. Geyer, R.; Jambeck, J.R.; Law, K.L. Production, use, and fate of all plastics ever made. Sci. Adv. 2017, 3, e1700782. [CrossRef]
[PubMed]
3. MDBC. A New Life for Fabrics and Plastics; Malaysia Dutch Business Council: Kuala Lumpur, Malaysia, 2021. Available online: https:
//www.mdbc.com.my/wp-content/uploads/2021/12/M4M-Environmental-Business-Kloth.pdf (accessed on 19 December 2024).
4. Payne, A. Open-and closed-loop recycling of textile and apparel products. In Handbook of Life Cycle Assessment (LCA) of Textiles
and Clothing; Woodhead Publishing Series in Textiles: Delhi, India, 2015; pp. 103–123. [CrossRef]
5. Rudisch, K.; Jüngling, S.; Mendoza, R.C.; Woggon, U.K.; Budde, I.; Malzacher, M.; Pufahl, K. Paving the road to a circular textile
economy with AI. In Informatik 2021; Gesellschaft für Informatik: Bonn, Germany, 2021; pp. 313–320. [CrossRef]
6. Sathyanarayana, D.N. Vibrational Spectroscopy: Theory and Applications; New Age International Publisher: Delhi, India, 2015.
7. Pasquini, C. Near infrared spectroscopy: A mature analytical technique with new perspectives—A review. Anal. Chim. Acta 2018,
1026, 8–36. [CrossRef]
8. Long, D.A. The Characterization of Chemical Purity Organic Compounds; Staveley, L.A.K., Ed.; Elviser: Amsterdam, The Netherlands,
1977; pp. 149–162.
9. Lyon, L.A.; Keating, C.D.; Fox, A.P.; Baker, B.E.; He, L.; Nicewarner, S.R.; Mulvaney, S.P.; Natan, M.J. Raman spectroscopy. Anal.
Chem. 1998, 70, 341–362. [CrossRef] [PubMed]
10. Mulvaney, S.P.; Keating, C.D. Raman spectroscopy. Anal. Chem. 2000, 72, 145–158. [CrossRef]
11. Muthu, S.S.; Li, Y.; Hu, J.Y.; Mok, P.Y. Recyclability Potential Index (RPI): The concept and quantification of RPI for textile fibres.
Ecol. Indic. 2012, 18, 58–62. [CrossRef]
Sensors 2025, 25, 57 26 of 27

12. Ling, C.; Shi, S.; Hou, W.; Yan, Z. Separation of waste polyester/cotton blended fabrics by phosphotungstic acid and preparation
of terephthalic acid. Polym. Degrad. Stab. 2019, 161, 157–165. [CrossRef]
13. Wang, H.; Kaur, G.; Pensupa, N.; Uisan, K.; Du, C.; Yang, X.; Lin, C.S.K. Textile waste valorization using submerged filamentous
fungal fermentation. Process Saf. Environ. Prot. 2018, 118, 143–151. [CrossRef]
14. Subramanian, K.; Sarkar, M.K.; Wang, H.; Qin, Z.H.; Chopra, S.S.; Jin, M.; Kumar, V.; Chen, C.; Tsang, C.W.; Lin, C.S.K. An
overview of cotton and polyester, and their blended waste textile valorisation to value-added products: A circular economy
approach–research trends, opportunities and challenges. Crit. Rev. Environ. Sci. Technol. 2022, 52, 3921–3942. [CrossRef]
15. Hou, W.; Ling, C.; Shi, S.; Yan, Z.; Zhang, M.; Zhang, B.; Dai, J. Separation and characterization of waste cotton/polyester blend
fabric with hydrothermal method. Fibers Polym. 2018, 19, 742–750. [CrossRef]
16. Islam, S.; Ahmed, S.; Arifuzzaman, M.; Islam, A.S.; Akter, S. Relationship in between strength and polyester content percentage
of cotton polyester blended woven fabrics. Int. J. Cloth. Sci. 2019, 6, 1–6.
17. Kudelski, A. Analytical applications of Raman spectroscopy. Talanta 2008, 76, 1–8. [CrossRef] [PubMed]
18. Wiley, J.H.; Atalla, R.H. Band assignments in the Raman spectra of celluloses. Carbohydr. Res. 1987, 160, 113–129. [CrossRef]
19. Edwards, H.; Farwell, D.; Webster, D. FT Raman microscopy of untreated natural plant fibres. Spectrochim. Acta Part A Mol.
Biomol. Spectrosc. 1997, 53, 2383–2392. [CrossRef]
20. Kavkler, K.; Demšar, A. Examination of cellulose textile fibres in historical objects by micro-Raman spectroscopy. Spectrochim.
Acta Part A Mol. Biomol. Spectrosc. 2011, 78, 740–746. [CrossRef]
21. Carey, C.; Boucher, T.; Mahadevan, S.; Bartholomew, P.; Dyar, M. Machine learning tools formineral recognition and classification
from Raman spectroscopy. J. Raman Spectrosc. 2015, 46, 894–903. [CrossRef]
22. Guo, S.; Popp, J.; Bocklitz, T. Chemometric analysis in Raman spectroscopy from experimental design to machine learning-based
modeling. Nat. Protoc. 2021, 16, 5426–5459. [CrossRef]
23. Qi, Y. Recent progresses in machine learning assisted Raman spectroscopy. Adv. Opt. Mater. 2023, 11, 2203104. [CrossRef]
24. Peterson, L.E. K-nearest neighbor. Scholarpedia 2009, 4, 1883. [CrossRef]
25. Myles, A.J.; Feudale, R.N.; Liu, Y.; Woody, N.A.; Brown, S.D. An introduction to decision tree modeling. J. Chemom. 2004, 18,
275–285. [CrossRef]
26. Biau, G.; Scornet, E. A random forest guided tour. Test 2016, 25, 197–227. [CrossRef]
27. Burges, C.J. A tutorial on support vector machines for pattern recognition. Data Min. Knowl. Discov. 1998, 2, 121–167. [CrossRef]
28. Gunn, S.R. Support vector machines for classification and regression. ISIS Tech. Rep. 1998, 14, 5–16.
29. Furey, T.S.; Cristianini, N.; Duffy, N.; Bednarski, D.W.; Schummer, M.; Haussler, D. Support vector machine classification and
validation of cancer tissue samples using microarray expression data. Bioinformatics 2000, 16, 906–914. [CrossRef]
30. Finkelstein, M.O.; Fairley, W.B. A Bayesian approach to identification evidence. Harv. Law Rev. 1970, 83, 489–517. [CrossRef]
31. Yegnanarayana, B. Artificial Neural Networks; PHI Learning Pvt. Ltd.: Delhi, India, 2009.
32. Albawi, S.; Mohammed, T.A.; Al-Zawi, S. Understanding of a convolutional neural network. In Proceedings of the 2017
International Conference on Engineering and Technology (ICET), Antalya, Turkey, 21–23 August 2017; Volume 10.
33. Schuster, M.; Paliwal, K.K. Bidirectional recurrent neural networks. IEEE Trans. Signal Process. 1997, 45, 2673–2681. [CrossRef]
34. Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; Bengio, Y. Generative adversarial
nets. Adv. Neural Inf. Process. System 2014, 27, 1–9. Available online: https://proceedings.neurips.cc/paper_files/paper/2014
/file/5ca3e9b122f61f8f06494c97b1afccf3-Paper.pdf (accessed on 19 December 2024).
35. Fan, X.; Ming, W.; Zeng, H.; Zhang, Z.; Lu, H. Deep learning-based component identification for the Raman spectra of mixtures.
Analyst 2019, 144, 1789–1798. [CrossRef]
36. Ho, C.S. Rapid identification of pathogenic bacteria using Raman spectroscopy and deep learning. Nat. Commun. 2019, 10, 4927.
[CrossRef]
37. Ralbovsky, N.M.; Lednev, I.K. Towards development of a novel universal medical diagnostic method: Raman spectroscopy and
machine learning. Chem. Soc. Rev. 2020, 49, 7428–7453. [CrossRef] [PubMed]
38. Neo, E.R.K.; Yeo, Z.; Low, J.S.C.; Goodship, V.; Debattista, K. A review on chemometric techniques with infrared, Raman and
laser-induced breakdown spectroscopy for sorting plastic waste in the recycling industry. Resour. Conserv. Recycl. 2022, 180,
106217. [CrossRef]
39. Maruthamuthu, M.K.; Raffiee, A.H.; Oliveira, D.M.D.; Ardekani, A.M.; Verma, M.S. Raman spectra-based deep learning: A tool
to identify microbial contamination. MicrobiologyOpen 2020, 9, e1122. [CrossRef] [PubMed]
40. Kow, P.Y.; Hsia, I.W.; Chang, L.C.; Chang, F.J. Real-time image-based air quality estimation by deep learning neural networks. J.
Environ. Manag. 2022, 307, 114560. [CrossRef]
41. Industrial Technology Research Institute. Available online: https://www.itri.org.tw/english/index.aspx (accessed on 19
December 2024).
42. Online Raman Textile Sorter in ITRI. Available online: https://www.itri.org.tw/ListStyle.aspx?DisplayStyle=01_content&SiteID=
1&MmmID=1036233376244004650&MGID=1163436641542114306 (accessed on 19 December 2024).
Sensors 2025, 25, 57 27 of 27

43. Castro, M.; Pereira, F.; Aller, A.; Littlejohn, D. Raman spectrometry as a screening tool for solvent-extracted azo dyes from
polyester-based textile fibres. Polym. Test. 2020, 91, 106765. [CrossRef]
44. Ye, J.; Tian, Z.; Wei, H.; Li, Y. Baseline correction method based on improved asymmetrically reweighted penalized least squares
for the Raman spectrum. Appl. Opt. 2020, 59, 10933–10943. [CrossRef] [PubMed]
45. Savitzky, A.; Golay, M.J. Smoothing and differentiation of data by simplified least squares procedures. Anal. Chem. 1964, 36,
1627–1639. [CrossRef]
46. Schafer, R.W. What is a Savitzky-Golay filter? IEEE Signal Process. Mag. 2011, 28, 111–117. [CrossRef]
47. Belkin, M.; Hsu, D.; Ma, S.; Mandal, S. Reconciling modern machine-learning practice and the classical bias-variance trade-off.
Proc. Natl. Acad. Sci. USA 2019, 116, 15849–15854. [CrossRef] [PubMed]
48. Mathew, A.; Amudha, P.; Sivakumari, S. Deep learning techniques: An overview. In Advanced Machine Learning Technologies and
Applications: Proceedings of AMLTA 2020; Springer: Singapore, 2021; pp. 599–608.
49. Li, S.; Song, W.; Fang, L.; Chen, Y.; Ghamisi, P.; Benediktsson, J.A. Deep learning for hyperspectral image classification: An
overview. IEEE Trans. Geosci. Remote Sens. 2019, 57, 6690–6709. [CrossRef]
50. Cheng, H.T.; Koc, L.; Harmsen, J.; Shaked, T.; Chandra, T.; Aradhye, H.; Anderson, G.; Corrado, G.; Chai, W.; Ispir, M.; et al. Wide
& deep learning for recommender systems. In Proceedings of the 1st Workshop on Deep Learning for Recommender Systems,
New York, NY, USA, 15 September 2016; pp. 7–10.
51. Zhao, Z.Q.; Zheng, P.; Xu, S.T.; Wu, X. Object detection with deep learning: A review. IEEE Trans. Neural Netw. Learn. Syst. 2019,
30, 3212–3232. [CrossRef]
52. Prechelt, L. Automatic early stopping using cross validation: Quantifying the criteria. Neural Netw. 1998, 11, 761–767. [CrossRef]
[PubMed]
53. Gotmare, A.; Keskar, N.S.; Xiong, C.; Socher, R. A closer look at deep learning heuristics: Learning rate restarts, warmup and
distillation. arXiv 2018, arXiv:1810.13243. [CrossRef]
54. Smith, S.L.; Kindermans, P.J.; Ying, C.; Le, Q.V. Don’t decay the learning rate, increase the batch size. arXiv 2017, arXiv:1711.00489.
[CrossRef]
55. Scholkopf, B.; Sung, K.K.; Burges, C.J.; Girosi, F.; Niyogi, P.; Poggio, T.; Vapnik, V. Comparing support vector machines with
Gaussian kernels to radial basis function classifiers. IEEE Trans. Signal Process. 1997, 45, 2758–2765. [CrossRef]
56. Luo, R.; Popp, J.; Bocklitz, T. Deep learning for Raman spectroscopy: A review. Analytica 2022, 3, 287–301. [CrossRef]
57. Misra, D. Mish: A self regularized non-monotonic activation function. arXiv 2019, arXiv:1908.08681. [CrossRef]
58. Kingma, D.P.; Ba, J. Adam: A method for stochastic optimization. arXiv 2014, arXiv:1412.6980. [CrossRef]
59. Krizhevsky, A.; Hinton, G. Convolutional deep belief networks on cifar 10. Unpubl. Manuscr. 2010, 40, 1–9.
60. Deng, J.; Dong, W.; Socher, R.; Li, L.J.; Li, K.; Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, Miami, FL, USA, 20–25 June 2009; pp. 248–255. [CrossRef]
61. Ramachandran, P.; Zoph, B.; Le, Q.V. Swish: A self-gated activation function. arXiv 2017, arXiv:1710.05941. [CrossRef]
62. Puchowicz, D.; Cieslak, M. Raman spectroscopy in the analysis of textile structures. In Recent Developments in Atomic Force
Microscopy and Raman Spectroscopy for Materials Characterization; IntechOpen: London, UK, 2022; pp. 1–21. [CrossRef]
63. Grady, A.; Dennis, A.C.; Denvir, D.; Mcgarvey, J.J.; Bell, S.E. Quantitative Raman spectroscopy of highly fluorescent samples
using pseudosecond derivatives and multivariate analysis. Anal. Chem. 2001, 73, 2058–2065. [CrossRef] [PubMed]
64. Chequer, F.D.; de Oliveira, G.A.R.; Ferraz, E.A.; Cardoso, J.C.; Zanoni, M.B.; de Oliveira, D.P. Textile dyes: Dyeing process and
environmental impact. In Eco-Friendly Textile Dyeing and Finishing; IntechOpen Limited: London, UK, 2013; Volume 6, pp. 151–176.
Available online: https://www.intechopen.com/chapters/41411 (accessed on 19 December 2024).
65. Varadarajan, G.; Venkatachalam, P. Sustainable textile dyeing processes. Environ. Chem. Lett. 2016, 14, 113–122. [CrossRef]
66. Yaseen, D.; Scholz, M. Textile dye wastewater characteristics and constituents of synthetic effluents: A critical review. Int. J.
Environ. Sci. Technol. 2019, 16, 1193–1226. [CrossRef]
67. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A.N.; Kaiser, Ł.; Polosukhin, I. Attention is all you need. In
Proceedings of the Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA, 4–9 December 2017; Volume 30.
Available online: https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf
(accessed on 19 December 2024).
68. Chang, M.; He, C.; Du, Y.; Qiu, Y.; Wang, L.; Chen, H. RaT: Raman Transformer for highly accurate melanoma detection with
critical features visualization. Spectrochim. Acta Part A Mol. Biomol. Spectrosc. 2024, 305, 123475. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.