Investigation of Fourier transform infrared spectroscopy potential in cotton fiber micronaire measurement and distribution

Reference fibers, HVI MIC reference and ATR FTIR spectral measurement

A total of 104 subsamples from a pool of well-defined 104 reference materials was used. ⁵ They were well homogenized and clean fiber samples representing the two principal cultivated species (Upland and Pima) at various grown locations (US domestic and foreign countries), and a wide range of fiber maturity and MIC attributes. Average MIC values were obtained from 40 measurements on each sample bale by a Uster HVI 900A system following a standard protocol. ⁵

An FTS 3000MX FTIR spectrometer (Varian Instruments, Randolph, MA, USA) equipped with a ceramic source, KBr beam splitter, deuterated triglycine sulfate detector, and an ATR attachment was used to scan the fiber spectra. An ATR sampling device utilized a DuraSamplIR single-pass diamond-coated internal reflection accessory (Smiths Detection, Danbury, CT, USA), and a consistent contact pressure was applied by way of a stainless-steel rod and an electronic load display. The background was recorded without the sample presence before scanning the samples. The spectra were collected over the range of 600–4000 cm⁻¹ at 8 cm⁻¹ and 64 co-added scans in the absorbance unit. Five spectra were collected on each sample by subsampling, and the mean spectrum was obtained for the analysis. During the spectral collection of individual subsamples, a small bundle of fibers (approx. 0.5 mg) was put into the optical window (2 mm in diameter) of an ATR device.

Fibers in seed cotton, HVI MIC reference and ATR FTIR spectral measurement

There is a total of 38 seed cotton fiber samples that were divided into two sets on the basis of their spectral sampling – ginning sequence. The first seed cotton fiber set includes 11 commercial Upland seed cotton samples from nine cultivars, ²³ grown in three US states (Mississippi, Missouri, and New Mexico) and harvested in 2019. For each sample, 50 locules with seed presence (approx. 1.5 g each) were randomly selected across a sample box containing more than 4.5 kg seed cottons, in which fibers were mingled together within neighboring locules, as well as between locule fibers and nonlint materials (seed, trash, seed-coat, etc.). The ATR FTIR spectrum was collected directly on each locule fiber with cotton seed presence over the range of 4000–600 cm⁻¹ at 4 cm⁻¹ resolution, and 16 co-added scans in the absorbance unit, with extreme attention to avoid the spectral interference of any nonlint materials in a locule. Hence, there are no contributions of any nonlint materials on FTIR spectra. Meanwhile, seed cottons (approx. 1 kg each) were ginned by using a laboratory roller gin. The ginned fibers were conditioned at 21 ± 1°C temperature and 65 ± 2% RH for 48 h prior to routine HVI testing. The average HVI MIC value was determined by an HVI 1000 (Uster Technologies Inc., Knoxville, TN, USA) with five replications per sample at the Southern Regional Research Center (SRRC, ARS, USDA) routinely.

The second seed cotton fiber set consists of 27 commercial Upland and Pima seed cotton samples that differ completely from those in the first seed cotton fiber set. These 27 samples were from five crop years (2016, 2018, 2020, 2021, and 2022), grown in two US states (Mississippi and New Mexico) and from 25 cotton cultivars. After ginning of seed cottons (approx. 250 g) by using a laboratory tabletop 10-saw gin, the fibers were conditioned at standard climate conditions for 48 h before subsequent measurements. The same HVI and procedure as in the first seed cotton fiber set was used. The FTIR spectra of five subsamples randomly taken across each ginned fiber sample were collected in the same way as in the first seed cotton fiber set described above, with great caution to keep any nonlint materials from spectral sampling.

HVI MIC model development from PLS multivariate regression analysis and algorithmic M_IR approach of ATR FTIR spectra

All ATR FTIR spectra were imported into the GRAMS/AI program (version 9.1, Thermo Fisher Scientific, Waltham, MA, USA) and thw mean spectrum was calculated for each sample. The spectral set was exported into Microsoft Excel for Office 365 for spectral normalization, by dividing the intensity of individual bands in the 1800–600 cm⁻¹ region with the average intensity in the 1800–600 cm⁻¹ region. Next, the normalized spectra were loaded into the GRAMS IQ application in Grams/AI for PLS regression model development. On the order of the smallest to largest in the MIC value of the 104 reference materials, 70 spectra (or samples) were selected for calibration equation development, and the remaining 34 spectra (every third sample) were used for model validation. To optimize the accuracy of the validation models, the spectra were subjected to different combinations of spectral pretreatments (e.g. mean centering (MC), multiplicative scatter correction (MSC), standard normal variate, and the first and second derivatives). A full (one-sample-out rotation) cross-prediction method was used, and the number of optimal factors chosen for the regression equation generally corresponded to the minimum of the predicted residual error sum of squares. The saved regression equations were subsequently applied to: (a) the calibration and validation samples that were clean and well-blended cotton fiber samples known as the 104 reference materials; and (b) the test samples that were not blended following a laboratory ginning of seed cottons. Model efficiencies were assessed in the calibration, validation, and independent test set on the basis of the coefficient of determination ( R c 2 , R v 2 , R T 2 ), root mean square error of calibration (RMSEC), validation (RMSEV) or test (RMSET), bias (or differences) between spectral predicted and referenced MIC values, as well as the number (or percentage) of samples being predicted accurately. A sample was considered to be predicted accurately if the difference between FTIR predicted and the referenced MIC value was within the 95% agreement range of bias ±1.96 RMSEV. ²⁴ The higher the R V 2 and R T 2 , the lower the RMSEV and RMSET, less bias, and greater percentage of correctly predicted samples, the better and more robust the FTIR method to predict fiber MIC. Both the P value at a confidence level of 95% and the F-test were performed using the respective Regression and F-test function under data analysis in Microsoft Excel for Office 365.

Separately, the spectral set was retained in Microsoft Excel for Office 365 to assess the M_IR index by using the previously proposed algorithm analysis. ²¹ Then the M_IR value was related to the MIC value of the 104 reference materials, and a general conversion between M_IR and MIC was established. MIC values predicted from the algorithmic M_IR approach were utilized to assess MIC modeling on identical sample sets as PLS models for a comparison between two analytical strategies.

Cotton fiber MIC component and ATR FTIR spectral response

Figure 1(a) shows the typical and normalized ATR FTIR spectral average of the 104 cotton materials in the spectral region between 1800 and 600 cm⁻¹. For the purpose of only exhibiting the infrared spectral response to the fiber MIC component, these spectra were obtained by averaging the spectra of neighboring MIC values in the respective range of less than 3.4, 3.4–3.7, 3.7–4.2, 4.2–5.0, and over 5.0 MIC units. In general, the spectra of cotton fibers with low MIC have common infrared bands with those of fibers having high MIC, but there is some degree of intensity changes induced by the rising MIC in this spectral region, as highlighted in difference spectra in Figure 1(b). It clearly indicates a nearly equal number of relatively large negative and positive intensity peaks in the 1200–750 cm⁻¹ region, mainly assignable to the C-O stretching modes and crystal forms in fiber cellulose. With fiber MIC increasing, intensities of the bands between 1200 and 990 cm⁻¹ decrease, whereas those in the 990–750 cm⁻¹ region increase (Figure 1(b)). Spectral intensity variations suggest the capability of the FTIR technique to estimate the fiber MIC component by the analysis of characteristic bands in this infrared region. Previous studies^{25 –27} have assigned a broad band centered at 1620 cm⁻¹ to the OH bending mode of adsorbed water and protein amide I mode, a number of bands in the 1500–1200 cm⁻¹ region to both CH₂ deformations and C-O-H bending vibrations, at least five intense bands in the 1200–900 cm⁻¹ region to the stretching modes of C-O and C-C vibrations, a strong band at 895 cm⁻¹ to the β-glycosidic linkage in cellulose, and weak bands from 800 to 700 cm⁻¹ to crystal forms of native cellulose in cotton fibers. In addition, there were intense absorptions due to the O-H and C-H stretching vibrations at 3340, 3280, 2910, 2895, and 2845 cm⁻¹ (inset in Figure 1(a)).

OPEN IN VIEWER

Figure 1.

(a) Representative of normalized attenuated total reflection (ATR) Fourier transform infrared (FTIR) spectra of cotton fibers with various micronaire (MIC) readings, by averaging the spectra of neighboring MIC values in the respective range of less than 3.4, 3.4–3.7, 3.7–4.2, 4.2–5.0, and over 5.0 MIC unit and (b) Difference in ATR FTIR spectra of an average spectrum of samples with a MIC range of less than 3.4, 3.4–3.7, 3.7–4.2, 4.2–5.0, or over 5.0 MIC minus that of a mean spectrum calculated from all spectra in (a).

MIC model development and assessment

Seventy spectra (or samples) of the 104 reference materials were selected for MIC calibration equation development (range 2.60–5.65; average 4.22; standard deviation 0.58), and the remaining 34 spectra (every third sample) were used for model validation (range 2.90–5.12; average 4.22; standard deviation 0.51). Variations of MIC values among the 104 reference materials could represent the variability of fiber MIC in commercial cotton bales and cotton breeding programs, because these materials were used in fiber MIC, maturity, and fineness model development from the NIR region considerably.^{14 –18}

PLS models were developed from different combinations of spectral pretreatments and spectral regions. The statistics of optimal results in calibration and validation sets from different spectral regions are listed in Table 1 for comparison. In addition to the entire 3600 to 615 cm⁻¹ region, the spectra were analyzed subjectively in two narrow regions: 1800–615 and 1200–800 cm⁻¹. The reason for choosing a narrow spectral region was to compare the model performance from different spectral absorptions that are indicative of unique vibration modes in cotton fiber celluloses, and also to facilitate the development of portable and handheld infrared sensors. The optimal calibration models were obtained from the combination of MC + MSC and Savitzky–Golay second derivative (two degrees and 13 points) spectral preprocessing in each spectral region. The model with a narrow spectral region (1200 to 800 cm⁻¹) becomes a little worse in RMSEC, RMSEV, and bias in both the calibration set and validation set. In other words, two models with an extension either to a high wavenumber side (>1200 cm⁻¹) or to a low wavenumber side (<800 cm⁻¹) may not obviously improve the R C 2 , RMSEC, R V 2 , RMSEV or bias, probably suggesting the minimal effect of adding spectral regions beyond the 1200–800 cm⁻¹ region into models on fiber MIC determination. Another parameter to examine the robustness and accuracy of the spectral model is the use of the 95% agreement range, defined as bias ± 1.96 × RMSEV, in which the bias is the difference between the averages of the predicted and referenced MIC values while RMSEV is the root mean square error of validation. Within 70 calibration samples and 34 validation samples, there are four (5.7%) and one (2.9%) samples that have absolute prediction errors (or differences) greater than the 95% agreement range, irrespective of spectral regions for the modeling. That is, three PLS models result in over 90% of MIC predictions that are within the 95% agreement range. Overall, the MIC property could be predicted by any PLS model given by Table 1. A scatter plot of FTIR predicted against measured MIC in the validation set from the 1200 to 800 cm⁻¹ model is given in Figure 2(a) (P value <0.001). The F-test showed a nonsignificant difference statistically in the variances of predicted and measured MIC (P value >0.05). The corresponding Bland–Altman plot in Figure 2(b) shows the visualization of individual samples in comparison with the 95% limits of agreement between the predicted and referenced MIC.

OPEN IN VIEWER

Table 1.

Statistics of optimal MIC models in calibration, validation, and test sets from PLS model and algorithmic M_IR analysis of ATR FTIR spectra ^a

Spectral region (cm−1)	Calibration set				Validation set				Test set
	R C 2	RMSEC b	bias c	no. (%) d	R V 2	RMSEV b	bias c	no. (%) d	R T 2	RMSET b	bias c	no. (%) d
PLS, 3600-615	0.87	0.208	0.000	66 (94.3)	0.83	0.227	−0.054	33 (97.1)	0.76	0.309	0.049	31 (81.6)
PLS,1800-615	0.87	0.209	0.000	66 (94.3)	0.83	0.226	−0.054	33 (97.1)	0.76	0.312	0.055	31 (81.6)
PLS, 1200-800	0.86	0.215	0.000	66 (94.3)	0.82	0.232	−0.060	33 (97.1)	0.75	0.310	0.015	35 (86.8)
Algorithmic MIR	0.84	0.234	0.032	65 (92.8)	0.72	0.280	−0.066	33 (97.1)	0.70	0.340	−0.046	37 (97.4)

MIC: micronaire; M_IR: infrared maturity.

a

All partial least squares (PLS) spectral processing with mean centering (MC), multiplicative scatter correction (MSC), and Savitzky–Golay second derivative. Two optimal factors were used for each model.

b

Root mean square error of calibration (RMSEC), validation (RMSEV) and test (RMSET).

c

Bias = attenuated total reflection (ATR) Fourier transform infrared (FTIR) predicted – high volume instrument (HVI) measured.

d

Number and percentage (%) of samples with the difference between predicted and referenced value within the 95% agreement range of bias ± 1.96 × RMSEC in calibration set or bias ± 1.96 × RMSEV in validation set and independent test set.

OPEN IN VIEWER

Figure 2.

(a) Comparative correlation and (b) Bland–Altman plot of Fourier transform infrared (FTIR) predicted and referenced micronaire (MIC) in the validation set (n = 34) from the partial least squares (PLS) model in the 1200–800 cm⁻¹ region (●) and from the algorithmic infrared maturity (M_IR) approach (●). The 95% agreement range of bias ± 1.96 × root mean square validation (RMSEV) was from the PLS model in the 1200–800 cm⁻¹ region, differing from the algorithmic M_IR approach slightly.

Instead of utilizing a specific spectral region for PLS analysis, another strategy is to select the unique wavebands that are correlated highly with the target property. As an example of this concept, Liu et al. ²¹ reported the identification of two characteristic infrared bands (at 956 and 1032 cm⁻¹) in the 1200 to 900 cm⁻¹ region, and the development of simple ratio algorithms for assessing the cotton fiber maturity (M_IR). The M_IR and MIC values on the 104 reference materials in Figure 3 display a linear correlation with R² = 0.80 (P < 0.001). From the relationship contained in Figure 3, MIC values of the 104 reference materials were estimated from the individual M_IR value and then analyzed in an identical way to PLS models. Caution should be taken when interpreting the relationship in Figure 3 for other cultivar samples, because cotton MIC is a function of fiber maturity and fineness rather than one of the two. ²⁸ Contrasted to PLS models, statistics from the algorithmic M_IR approach given by Table 1 suggest a poor R C 2 , RMSEC, R V 2 , RMSEV or bias in both the calibration set and validation set, but a nearly unchanged percentage of samples within the 95% agreement range. This is expected, because the algorithmic M_IR approach utilizes fewer datapoints than PLS models; for example, 20 datapoints in the algorithmic M_IR approach versus 104 datapoints in the 1200–800 cm⁻¹ model. The algorithmic M_IR approach predicted MIC in the validation set are inserted in Figure 2(a) (R² = 0.72, P value <0.001) and Figure 2(b) for comparison. The difference in the variances of M_IR predicted and measured MIC values was not statistically significant (P value >0.05).

OPEN IN VIEWER

Figure 3.

Comparison of high volume instrument (HVI) micronaire (MIC) against Fourier transform infrared (FTIR) infrared maturity (M_IR) on the 104 reference materials.

Furthermore, both three PLS models and the algorithmic M_IR approach were applied to a test set composed of 38 independent commercial cottons that represents different cultivars, crop years, and grown locations, and the results are presented in Table 1. Compared with the calibration or validation set from three PLS models, the test set reveals a little inferiority in R T 2 , RMSET, bias, and the percentage of samples within the 95% agreement range. Notably, three PLS models show a negative bias (–0.060 to –0.054) in the validation set but a positive bias (0.015 to 0.055) in the test set, causing the reduced percentages of samples within the 95% agreement range from approx. 97% (97.1%) in the validation set to approx. 86% (81.6–86.8%) in the independent test set. This is not surprising, probably due to such factors as: (a) the 104 reference materials were clean and well prepared/blended, whereas fiber samples from seed cottons were not blended; and (b) all HVI and spectral data of the 104 reference materials were collected much earlier than those of the seed cotton samples. In practice, new samples were added to the original calibration sample set for robust NIR model development.^17,18 If the test set samples were divided into calibration and validation samples, and then compiled into the respective sets from the 104 reference materials, all statistics from the re-calibrated model were improved (data not shown).

While from the algorithmic M_IR approach, the validation set and test set share a similarity in R_C² versus R_V², bias, and the percentage of samples within the 95% agreement range, except for a discrepancy in RMSEV versus RMSET. For comparison with those in the validation set, all samples in the test set are included in the scatter plot (Figure 4(a), R² = 0.70, P value <0.001) and the Bland–Altman plot (Figure 4(b)). The difference in the variances of M_IR predicted and measured MIC values was insignificant (P value >0.05). However, there are more samples in the test set than in the validation set having a large MIC difference (e.g. >0.40 MIC unit) between M_IR predicted and referenced MIC values, mostly because of the sample blending issue. Further examination in the test set reveals that fiber sampling methods in seed cotton (i.e. locule fiber sampling in cotton bolls vs. fiber subsampling in ginned samples) does not impact the tendency of MIC differences between M_IR predicted and referenced MIC values. Independent of PLS models or the algorithmic M_IR approach, there was one sample that was constantly determined to be outside of the 95% agreement range. Nevertheless, the algorithmic M_IR approach avoids the need to re-calibrate the model with new samples, and is consistent over a lengthy spectral data acquisition, and also ATR FTIR technique is advantageous in analyzing a small amount of fiber samples at a single cotton boll/locule level that is not sufficient for routine HVI MIC test occasionally during cotton breeders’ variety trials. Therefore, the result reveals a potential of the FTIR instrument with the algorithmic M_IR approach for rapid and accurate measurements of fiber MIC at a laboratory environment, and also for finished cotton yarns and fabrics that are impossible to know their MIC values from the HVI test.

OPEN IN VIEWER

Figure 4.

(a) Comparative correlation and (b) Bland–Altman plot of Fourier transform infrared (FTIR) predicted and referenced micronaire (MIC) from the algorithmic infrared maturity (M_IR) approach in the validation set (n = 34, ●) and test set (n = 38, ●). The 95% agreement range of bias ± 1.96 × root mean square validation (RMSEV) was from the algorithmic M_IR approach in the validation set.

It is of interest to look into the MIC distribution in a maturity variable fiber sample. For instance, MIC values of fiber samples at the locule level in seed cottons were estimated from their corresponding M_IR values, and the MIC frequency of 50 locules on two seed cotton cultivars (DP1646 and NG3522) that have the close HVI MIC (4.63 for DP1646 vs. 4.66 for NG3522), and FTIR predicted MIC (4.58 for DP1646 vs. 4.50 for NG3522) are compared in Figure 5. From low to high MIC, MIC distribution of the DP1646 fiber samples is similar to that of the NG3522 samples (P < 0.05) – that is, there is a nonsignificant difference statistically in MIC distribution between two samples. However, Figure 5 shows a relatively great difference in the MIC range of 5.50 or greater. Conversely, two cultivars (ST4848 and DG3385) that have different HVI MIC (5.08 for ST4848 vs. 3.42 for DG3385) and FTIR predicted MIC (4.96 for ST4848 vs. 3.59 for DG3385) indicate the dissimilarity in the pattern of MIC distribution (P > 0.05), with apparent differences in the MIC ranges of 3.75 or less and 4.75 or greater (Figure 6). In contrast to the ST4848 cultivar with a MIC frequency of more than 20% occurring at two adjacent MIC ranges (4.75 and 5.25), the DG3385 cultivar shows a MIC frequency of over 20% at the low MIC range (≤2.25). Regardless of the MIC frequency in Figures 5 and 6, four cultivars exhibit a broad range of MIC variation within a maturity variable seed cotton sample. In general, a high MIC cotton cultivar screened from the average of 50 locules should have relatively high MIC fiber frequency, and vice versa. Indeed, the ST4848 cotton (HVI MIC = 5.08) has a greater MIC (≥5.0) fiber percentage (approx. 56%), followed by that of approx. 40% for DP1646 and NG3522 (HVI MIC = 4.63–4.66) and of approx. 14% for DG3385 (HVI MIC = 3.42).

OPEN IN VIEWER

Figure 5.

Comparison of micronaire (MIC) frequency (%) between DP1646 (■, MIC = 4.63) and NG3522 (■, MIC = 4.66) cultivar from the algorithmic infrared maturity (M_IR) approach on 50 locule fiber samples, after assigning the MIC in the range of 2.0–6.5 into nine bins (bin width 0.5) and counting the frequency number of each bin.

OPEN IN VIEWER

Figure 6.

Comparison of micronaire (MIC) frequency (%) between ST4848 (■, MIC = 5.08) and DG3385 (■, MIC = 3.42) cultivar from the algorithmic infrared maturity (M_IR) approach on 50 locule fiber samples, after assigning the MIC in the range of 2.0–6.5 into nine bins (bin width 0.5) and counting the frequency number of each bin.

To verify the MIC frequency in Figures 5 and 6 from locule fibers in seed cottons, ATR FTIR spectra for each of GM-39 (HVI MIC = 2.6) and DM-9 (HVI MIC = 4.0) samples were collected on 50 subsamples and the result is illustrated in Figure 7. The MIC distribution of the GM-39 samples differs from that of the DM-9 samples (P > 0.05), indicating a significant difference in the MIC distribution between the two samples. Unlike the GM-39 sample with a MIC frequency of over 85% at low MIC range (<3.0), the DM-9 sample has the MIC frequency of more than 20% at two neighboring MIC ranges (4.25 and 4.75). Hence, MIC frequency in Figures 5 and 6 could reflect a general fiber MIC occurrence in seed cottons, because of analyzing 50 subsampling locules in a maturity variable seed cotton sample.

OPEN IN VIEWER

Figure 7.

Comparison of micronaire (MIC) frequency (%) between DM-9 (■, MIC = 4.0) and GM-39 (■, MIC = 2.6) from the algorithmic infrared maturity (M_IR) approach on 50 fiber subsamples, after assigning the MIC in the range of 2.0–6.5 into nine bins (bin width 0.5) and counting the frequency number of each bin. Both samples are used as AMS MIC calibration cottons and are not normal cottons. They were selected for their specific MIC values and their high uniformity around those MIC values.