Point-of-generation sorting and near-infrared spectroscopy govern recycling feedstock purity

While cutting-table segregation solves purity for a disciplined CMT plant's own offcuts, a regional recycling hub aggregating bales from smaller garment workshops, tailoring clusters, or post-consumer take-back programmes cannot trust paper declarations alone [S24] [S26] [S27]. To verify incoming bales rapidly and non-destructively, industrial sorting centres rely on near-infrared (NIR) reflectance spectroscopy [S26] [S27] [S0404]. The near-infrared spectrum spans wavelengths from roughly 750 to 2,500 nanometres, with industrial indium gallium arsenide (InGaAs) photodiode array sensors typically operating between 900 and 1,700 nanometres [S26] [S0404]. When a halogen light source illuminates a textile sample, photons in this window excite overtone and combination vibrational modes of covalent molecular bonds, specifically carbon-hydrogen ($\text{C-H}$), oxygen-hydrogen ($\text{O-H}$), and nitrogen-hydrogen ($\text{N-H}$) functional groups [S26] [S27].

Because every textile polymer possesses a distinct chemical backbone, each fibre family absorbs near-infrared radiation at characteristic resonant wavelengths and reflects the remaining light into the spectrometer's optical grating [S26] [S27] [S0404]:

Until recently, proprietary lock-in of spectral libraries held by equipment vendors hindered smaller regional recyclers from deploying automated spectroscopic sorting [S26] [S27]. In January 2025, the United States National Institute of Standards and Technology (NIST) addressed this barrier by releasing an open-access reference database of near-infrared spectral fingerprints for textile recycling [S27]. Expanded through July 2026 as the Near Infrared Spectra of Origin-defined and Real-world Textiles (NIR-SORT) repository (versions 1.0 and 2.0), the dataset recorded 516 downloads by 395 unique users worldwide by February 2026 [S0404].

NIST's NIR-SORT 2.0 library provides a calibration standard for regional sorting facilities [S0404]. The repository contains reflectance spectra for 90 known-provenance fabrics, 12 pre-consumer manufacturing waste fabrics, and 11 post-consumer garments, supplemented in version 2.0 by 12 custom-dyed fabrics, 61 custom-blended samples, 8 undyed fabrics spanning three additional fibre families, and 12 additional pre-consumer fabrics [S0404]. Crucially for African industrial deployment, NIST recorded spectra across both benchtop Fourier-transform NIR (FT-NIR) instruments and compact handheld NIR devices, accompanied by ground-truth physical measurements of fabric thickness, areal density, air permeability, yarn linear density, twist, thread count and colour coordinates [S27] [S0404]. By downloading NIR-SORT 2.0, an engineering team at a Cameroon regional textile hub can pre-train and validate chemometric models across benchtop and portable spectrometers before scanning local bales [S26] [S0404].

In industrial practice, NIR spectroscopy is deployed across two hardware configurations whose capital cost and throughput differ by two orders of magnitude [S20] [S26] [S0402] [S0404]:

If every textile sample were an undyed sheet of 100 percent pure polymer, a two-wavelength ratio threshold could separate cotton from polyester with near-perfect reliability [S26] [S27]. Real-world textiles violate these ideal conditions in three ways [S26] [S0404]. First, fabrics vary in surface geometry, thickness, weave tightness and moisture content, causing baseline offsets and light-scattering variations across the reflectance spectrum, while dark carbon-black pigments attenuate signal-to-noise ratios [S26] [S0404]. Second, in intimate polyester-cotton blends (such as 65/35 PES/CO shirting) and plated knits, the reflectance spectrum is a non-linear superposition of cellulosic $\text{O-H}$ and ester aromatic $\text{C-H}$ bands whose apparent ratio shifts with yarn surface migration [S26] [S35]. Third, elastane filaments in stretch cotton fabrics are typically present at only 1 to 5 percent by mass and are physically buried as the core of a core-spun cotton yarn [S26] [S35]. Because near-infrared photons penetrate only a few hundred micrometres into the fabric surface, the cotton sheath masks the polyurethane core, leaving faint spectral perturbations in the 1,180 to 1,220 nanometre and 1,500 to 1,700 nanometre bands that linear threshold rules miss [S26].

To overcome spectral masking and non-linear blend interference, modern optical sorting pipelines combine spectral preprocessing with supervised machine learning [S26]. In an April 2026 research investigation at the University of Groningen, Temmo Rijken evaluated machine-learning classification pipelines across 181 certified textile samples scanned in the 900 to 1,700 nanometre NIR wavelength range [S26]. Rather than feeding raw reflectance values into a classifier, the pipeline extracted a multi-domain feature set combining smoothed spectral derivatives (removing baseline offsets from texture and colour), targeted band-ratio indices, and discrete wavelet transform coefficients that isolate localized high-frequency absorption shoulders [S26].

Across multi-class composition identification spanning mono-materials and commercial blends, a Weighted K-Nearest Neighbours (Weighted KNN) classifier achieved the highest overall test accuracy at 91.7 percent, demonstrating that common mono-materials and multi-fibre blends can be distinguished reliably in the 900 to 1,700 nanometre window ({{xr:wp11-04-nir-ml-accuracy}}) [S26]. For binary detection of hidden elastane, a one-dimensional Convolutional Neural Network (1D-CNN) operating on the preprocessed spectral curve achieved an overall accuracy of 92.6 percent, comprising 95.2 percent specificity on elastane-absent samples and 83.3 percent sensitivity on elastane-present samples [S26].

A practical advantage of deploying machine-learning NIR classification within GFT's digital manufacturing blueprint is that the required compute infrastructure already exists on the factory floor [R28] [R29]. In GFT's Digital Strategy (Section 3, Tier-2 Quality Inspection, and Section 5, Reference Architecture), the plant deploys automated computer-vision camera bars above fabric inspection and spreading tables to detect weaving and knitting defects [R28] [R29]. Academic benchmarks show that deep-learning fabric inspection models achieve 92.3 to 97.2 percent defect detection accuracy at speeds up to 60 metres per minute (versus 60 to 75 percent accuracy at 10 to 15 metres per minute for human inspectors), while commercial systems such as Smartex on circular knitting machines cut defective fabric production by 85 percent [R28] [R29].

Executing a Weighted KNN or compact 1D-CNN model on a 256-channel 1D near-infrared spectrum requires fewer than 50,000 floating-point operations and completes in under 2 milliseconds on an industrial edge gateway [S26]. As detailed in our companion white paper on edge-to-dashboard reference architecture (wp18), the NIR sensor connects via USB or Ethernet to the same local edge node that runs the computer-vision inspection pipeline and logs results directly into the shop-floor PostgreSQL/TimescaleDB database, operating offline without cloud latency.

Upstream sorting does not end with polymer identification; physical preprocessing must also match the mechanical state of the textile waste to the opening equipment [S20] [S24] [S0405]. Cutting-table clips generated before the sewing line consist purely of flat, single-ply fabric pieces free of metal fasteners, buttons, zippers, interlinings or foreign sewing threads [S0402] [S0405]. Provided they are segregated at the cutting table, these clips bypass mechanical de-trimming and proceed directly to baling and guillotine size reduction into 20 to 50 millimetre squares prior to mechanical tearing [S20] [S0405].

By contrast, sewing-line quality rejects (the 1 to 3 percent of garments that fail Acceptable Quality Limit inspection and cannot be reworked), overlock edge trimmings, unsold deadstock, and post-consumer garments are composite assemblies [S20] [S26] [S0401]. An assembled cotton shirt or trouser contains brass or steel zippers, copper denim rivets, polyester buttons, woven polyester care labels, fusible resin collar interlinings, elastane waistbands, and polyester core-spun seam threads [S20] [S26] [S35]. Mixing whole garment rejects into a bale of clean cutting clips contaminates the batch with hard trims and synthetic seam threads, downgrading clean clips to the processing cost of post-consumer rags [S24] [S0402].

Feeding un-trimmed garments containing metal zippers, rivets or hard buttons into a high-speed mechanical tearing line creates severe mechanical and fire hazards [S20] [S23] [S0405]. Tearing cylinders rotate at high peripheral speeds with sub-millimetre clearances between hardened steel pins or saw-tooth wire [S20]. When a metal zipper slider or rivet strikes the rotating cylinder, it snaps steel pins, gouges carding wire, and throws incandescent friction sparks into an atmosphere saturated with dry, aerated cotton lint [S20]. Dust-spark ignition inside tearing and pneumatic ducts is a primary fire hazard in mechanical recycling plants, requiring magnetic metal traps, heavy-part gravimetric drops, and infrared spark-detection suppression systems on lines processing whole garments [S20].

Textile Exchange's Guide to Recycled Materials V2.0 (GRS-202-V2.0-2025.07.01 and RCS-202-V2.0-2025.07.01, effective July 2025) codifies the regulatory boundary between physical preprocessing and certified recycling [S0405]. Under the standard, opening, cleaning, sorting, screening, contamination removal, baling and size conversion are classified as material concentration steps prior to mechanical shredding [S0405]. The guide also establishes strict audit thresholds based on fibre physics: pre-consumer garment cutting waste typically yields reclaimed cotton staple lengths between 6 and 21 millimetres after tearing (and 5 to 18 millimetres for post-consumer waste) [S0405]. Consequently, Textile Exchange mandates that only products containing mechanically recycled cotton with a yarn count coarser than Ne 34 (<Ne 34) and a fibre length shorter than 22 millimetres (<22 mm) may be certified to the Recycled Claim Standard (RCS) or Global Recycled Standard (GRS) without mandatory additional due diligence by the certification body [S0405].

To protect downstream tearing cylinders and satisfy GRS v4 and European Extended Producer Responsibility (EPR) standards, factory preprocessing must physically separate three restricted fractions before baling [S20] [S23] [S28] [S0405]: