Flagged as an outlier and excluded from subsequent analysis

**Figure 7.** The mean tebuthiuron contents for the ten spatial replicates analysed. The numbers over each sample location provide the mean tebuthiuron content in mg/g.

#### *3.4. Analysis of FTIR Spectra*

Figure 8 shows the FTIR spectra of one of the Regain samples, overlaid with the FTIR spectra of a sample of pure tebuthiuron powder. The tebuthiuron spectrum appeared similar to literature records on SpectraBase (https://spectrabase.com/spectrum/Gwi6 TOkmgNy; accessed on 22 June 2022), with a complex pattern of peaks. The major peaks and their aetiological bonds are summarised in Table 5. Not all of the peaks between 1500–900 cm−<sup>1</sup> were able to be confidently assigned identities, due to the complexity in this region. The Regain400 sample showed additional peaks at 1026, 1002, 910 and <600 cm−1, due to absorptions from the other matrix constituents. In contrast, the pure tebuthiuron powder samples showed minimal absorbance in this region (1026–910 cm−1).

**Figure 8.** FTIR spectra of a typical powdered 400 mg/g Regain sample and a sample of pure tebuthiuron powder (99.6% purity).


**Table 5.** The major peaks observed in the FTIR spectrum of pure tebuthiuron powder and their assigned bonds.

Partial least squares regression performed on the FTIR spectra indicated that this technique could be used to predict the moisture content of the samples with reasonable accuracy (R2 cv = 0.85; RMSECV = 1.12%; Table 6). The RPD (ratio of performance to deviation; equal to the standard deviation of the entire calibration set divided by the RMSECV) was 2.55, indicating good predictive ability for this analyte [28].

**Table 6.** Optimum pre-processing methods for the prediction of moisture and tebuthiuron content in powdered Regain samples using FTIR spectroscopy. The best-performing model for each analyte is highlighted in bold.


However, FTIR spectroscopy was unable to predict the tebuthiuron content of the Regain400 samples with a high level of accuracy (R2 cv = 0.43; RMSECV = 8.0 mg/g; RPD = 1.32). The best-performing results for both moisture and tebuthiuron content were found using standard normal variate (SNV) pre-processing of the FTIR spectra (Figures 9 and 10).

**Figure 9.** The results of the best-performing PLSR model for the prediction of moisture content using the FTIR spectra of the Regain samples (using SNV pre-processing). The blue points show the model calibration, while the red points show the model cross-validation.

**Figure 10.** The results of the best-performing PLSR model for the prediction of tebuthiuron content using the FTIR spectra of the powdered Regain400 samples (using SNV pre-processing). The blue points show the model calibration, while the red points show the model cross-validation.

#### *3.5. Analysis of Benchtop NIR Spectra—Granules*

The NIR spectra collected from the whole Regain granules is shown in Figure 11. Four of the oldest samples (shown in blue) had much lower absorbances across the NIR spectrum; possibly due to a different matrix composition for the non-active ingredients.

These samples also contained the lowest moisture contents. However, the NIR spectra of the remaining samples were more consistent.

**Figure 11.** The benchtop NIR spectra of the Regain granules, coloured by moisture content.

The main peaks observed in a pure (99.6% assay) sample of tebuthiuron powder (Figure 12) were at 1190, 1380, 1506, 1733 (shoulder), 2042 and 2269 nm. These were attributed to CH3 (second overtone), CH3 (second overtone), amide bond (second overtone), S-H (first overtone), C-S-C stretch and amide bond (amide I and III region), respectively [29].

**Figure 12.** The NIR spectra of a sample of pure tebuthiuron powder (99.6% assay), measured using the benchtop Antaris instrument.

The Regain400 samples showed additional peaks at 1396 (weak), 1415 (strong), and 2210 nm (strong) (cf. Figure 11). These are most likely attributable to CH3 (second overtone), H2O (second overtone) and CHO or CH3 combination bands, arising from non-active ingredients of the formulation (e.g., from organic matter in clay components or from polymer-like ingredients) [30].

The best-performing PLSR models for the prediction of moisture and tebuthiuron content using the benchtop Antaris NIR instrument are shown in Table 7. The highest accuracy for the prediction of moisture content was found using no pre-processing (Figure 13); however, the best model for tebuthiuron content used 2d11 pre-processing (Figure 14).

The model accuracy was quite high for moisture content (R<sup>2</sup> cv of 0.93 and RMSECV of 0.85%); however, the prediction results for tebuthiuron content were somewhat poorer (R<sup>2</sup> cv of 0.46 and RMSECV of 7.9 mg/g). However, the RMSECV of the tebuthiuron calibration was only slightly lower than the standard deviation of the analytical (HPLC) reference method (7.0 mg/g). Consequently, this indicates that the accuracy of the NIR method was approaching the maximum accuracy expected using this reference method.

**Table 7.** Optimum pre-processing methods for the prediction of moisture and tebuthiuron content in Regain granules using the benchtop Antaris NIR instrument. The best-performing model for each analyte is highlighted in bold.


**Figure 13.** The results of the best-performing PLSR model for the prediction of moisture content using the benchtop NIR spectra collected from the Regain granules (using no pre-processing). The blue points show the model calibration, while the red points show the model cross-validation.

**Figure 14.** The results of the best-performing PLSR model for the prediction of tebuthiuron content using the benchtop NIR spectra collected from the Regain400 samples (using 2d11 pre-processing). The blue points show the model calibration, while the red points show the model cross-validation.

The model loadings for the moisture content PLSR model showed positive contributions from the peaks at 1936, 1450 and 1190 nm (Figure 15). These regions correspond to the H2O 1st overtone, amide second overtone and C-H second overtone, respectively. The greatest contribution was observed from the H2O 1st overtone, confirming that the model was principally measuring moisture in the samples. The smaller contribution from the amide region should be considered in future investigations, as it appears that the tebuthiuron content may be moderately influencing the model.

**Figure 15.** Loadings plot for the prediction of moisture content in Regain granules using the benchtop Antaris NIR instrument.

The loading plot for the tebuthiuron content model (Figure 16) showed major contributions in the 2200–2400 nm region, corresponding to the CH3 combination band and amide I and III regions. Contributions were also observed around 1420 nm (H2O second overtone) and 1700 nm (S-H first overtone). Again, this confirmed that the model was principally looking at tebuthiuron in the sample, with a possible minor influence of moisture content.

**Figure 16.** Loadings plot for the prediction of tebuthiuron content in Regain granules, using the benchtop Antaris NIR instrument.

#### *3.6. Analysis of Benchtop NIR Spectra—Powder*

Generally, the accuracy of results obtained by NIR spectroscopy is improved by having a more homogenous sample matrix. Large particle sizes can lead to light scattering effects, which in turn biases the resultant NIR spectra obtained [31,32]. Using finely powdered matrices reduces this scattering effect and may consequently provide improved insight into the true sample composition. Consequently, the granule samples were ground to a fine powder and the NIR spectra collected from the powdered samples using the benchtop Antaris instrument. Figure 17 shows the NIR spectra of the powdered Regain samples, which were quite similar to the spectra collected from the granules (cf. Figure 11).

**Figure 17.** The benchtop NIR spectra of the powdered Regain samples, coloured by moisture content.

The results for the prediction of moisture and tebuthiuron content from the powdered samples are shown in Table 8. Interestingly, the moisture prediction was less accurate when using the powdered samples compared to the whole granules, with an RMSECV of 1.02% for the best-performing model, compared to an RMSECV of 0.85% for the granules. This is an important observation from a rapid quality control viewpoint, as it indicates that NIR spectra obtained from the intact/whole Regain granules will perform just as well—if not better—than spectra obtained from the powdered samples, thus removing the need for the time-consuming grinding process.

In contrast, the prediction accuracy for tebuthiuron content was slightly increased compared to NIR spectra collected from the whole granules (RMSECV of 7.7 mg/g, compared to 7.9 mg/g). However, this slight increase in accuracy would not be significant enough to justify the additional sample preparation time in most practical applications.


**Table 8.** Optimum pre-processing methods for the prediction of moisture and tebuthiuron content in powdered Regain samples using the benchtop Antaris NIR instrument. The best-performing model for each analyte is highlighted in bold.

#### *3.7. Analysis of Handheld NIR Spectra—Granules*

Given the promising results observed for the benchtop NIR spectra collected from the granules, the MicroNIR spectra were only collected from the granular Regain samples, not the powdered Regain samples. Additionally, the analysis of powder samples using this handheld instrument would require the instrument to be cleaned after every sample to prevent cross-contamination, adding to the time taken to collect the NIR spectra from each sample.

The NIR spectra of the Regain granules collected with the MicroNIR instrument are shown in Figure 18. The wavelength range of this instrument (908–1676 nm) is narrower than that of the benchtop Antaris instrument (1000–2500 nm); however, it still contains important information in the regions of 1185 nm (CH3 second overtone), 1360 nm (shoulder; CH3 second overtone), 1408 nm (H2O second overtone) and 1509 nm (amide bon d second overtone). Consequently, it was expected that this wavelength range could still be used for the prediction of moisture and tebuthiuron content.

**Figure 18.** The handheld MicroNIR spectra of the Regain granules, coloured by moisture content.

The performance statistics for the models developed for the prediction of moisture and tebuthiuron content using the MicroNIR spectra are shown in Table 9. Notably, the best models for moisture and tebuthiuron content were both slightly better (lower mean error) than those previously found using the benchtop NIR instrument. This supports the proposition of using handheld NIR instrumentation for real-time quality assessment of Regain products.

**Table 9.** Optimum pre-processing methods for the prediction of moisture and tebuthiuron content in Regain granules using the handheld MicroNIR instrument. The best-performing model for each analyte is highlighted in bold.


As seen from the calibration graph in Figure 19, the MicroNIR instrument was able to predict the moisture content of the Regain samples with reasonably high accuracy across most of the range tested. Towards the higher range of moisture contents (>12%), the accuracy flattened off, with the model under-predicting the moisture content of all of these samples.

**Figure 19.** MicroNIR prediction of moisture content. The blue points show the model calibration, while the red points show the model cross-validation.

The loading plot of the PLSR model for moisture content (Figure 20) shows the greatest contribution at 1410 nm, corresponding to the second overtone of H2O. A minor positive contribution was also observed at 1156 nm (CH3 second overtone) as well as

negative contributions around 1497 nm (likely corresponding to the second overtone of the amide region).

**Figure 20.** Loadings plot for the prediction of moisture content using the MicroNIR instrument.

The calibration graph for tebuthiuron content showed increased variability compared to the moisture content prediction (Figure 21), with a few samples which could potentially be outliers. The R<sup>2</sup> for cross-validation was not particularly high (0.53), but the RMSECV was quite good (7.5 mg/g), particularly compared to the mean laboratory error of 7.0 mg/g for tebuthiuron analysis using HPLC. Inclusion of a larger number of samples, particularly covering a wider calibration range, could potentially improve the prediction accuracy.

**Figure 21.** MicroNIR prediction of tebuthiuron content. The blue points show the model calibration, while the red points show the model cross-validation.

The loadings plot for tebuthiuron prediction (Figure 22) showed the largest contributions at 1379 and 1453 nm. These likely correspond to the CH3 second overtone and amide band overtone. Both of these bonds are found in the structure of tebuthiuron (Figure 1).

**Figure 22.** Loadings plot for the prediction of tebuthiuron content using the MicroNIR instrument.

*3.8. Comparison of Different Instruments and Sample Matrices*

Table 10 provides a succinct summary of the best-performing models from all the different methods of infrared spectroscopy trialled, as well as the different matrix types (powder vs. granules).

**Table 10.** Optimum pre-processing methods for the prediction of moisture and tebuthiuron content in powdered and granular Regain samples, using three different IR instruments. The best-performing model for each analyte is highlighted in bold.


The best-performing models were found using the handheld MicroNIR instrument, applied to the whole Regain granule samples. This was true for both moisture and tebuthiuron content. Consequently, this demonstrates that the handheld MicroNIR instrument should be highly suitable for the rapid, in-situ quality assessment of Regain granules throughout the manufacturing process.

#### *3.9. Independent Test Set—Handheld NIR*

The final stage of this work was to test the performance of the handheld NIR models using an independent set of samples (i.e., ones not used in the model calibrations). NIR spectra were collected from the granules using the MicroNIR instrument and predictions made using the optimum model for each analyte.

All of the PLS scores of the test set lay within the bounds of the scores of the calibration set (Figure 23), indicating that the spectra of the test set samples were similar enough to the spectra of the calibration set to allow accurate predictions to be made.

**Figure 23.** Scores plot of the calibration and test set spectra.

The prediction results are shown in Table 11, alongside the reference measurements. Most of the moisture predictions were relatively close, with a mean prediction error of +0.36% *w*/*w*. The RMSEP (root mean square error of prediction) was 0.93% *w*/*w*, with a R2 pred of 0.52.


**Table 11.** Prediction and reference results for the independent test set.
