The t-test showed that the following wavenumbers are significantly different, at 5% significance, in the 32 scans analysis: 3 997.77 cm⁻¹, 3 993.92–3 986.20 cm⁻¹, 1 861.00–1 832.07 cm⁻¹, 1 828.21–1 797.36 cm⁻¹, 1 789.64–1 785.79 cm⁻¹, and 798.39–781.04 cm⁻¹. As for the 100 scans resolution of 4 cm⁻¹, the wavenumbers 3 997.77–3 991.99 cm⁻¹, 3 988.13–3 984.27 cm⁻¹, 1 857.14–1 849.43 cm⁻¹, 1 839.78–1 832.07 cm⁻¹, 1 822.43–1 803.14 cm⁻¹, 1 797.36 cm⁻¹, 1 787.71 cm⁻¹, and 1 785.79 cm⁻¹, 794.54–784.89 cm⁻¹ differ between the oven-dried and air-dried samples, whereas the wavenumbers 2 159.91 cm⁻¹, 2 153.16 cm⁻¹ and 2 152.20 cm⁻¹, 2 137.74 and 2 136.77 cm⁻¹, 805.14 cm⁻¹ and 804.18 cm⁻¹, 792.61–789.72 cm⁻¹, 763.68–753.07 cm⁻¹ make the 100 scans resolution of 2 cm⁻¹ analysis statistically different, concerning the drying method.

Regarding the LDA, the 32 scans analysis as well as the 100 scans and resolution of 4 cm⁻¹ analysis showed 94.44% accuracy, whereas the 100 scans and resolution of 2 cm⁻¹ was 100% accurate (Table 7).

The sum of the first two linear discriminants (LD1 and LD2) accounts for 89.57% of data variability in the 32 scans and resolution of 4 cm⁻¹, 87.77% for the 100 scans and resolution of 4 cm⁻¹ and 87.63% for the 100 scans and resolution of 2 cm⁻¹ analyses.

When the t-test was performed using the samples from Colorado in the spectral range between 4 000–500 cm⁻¹ and only considering the first reading, i.e., the first replicate of each sample, the oven-dried and air-dried samples differ from each other between 790.68 cm⁻¹ and 784.89 cm⁻¹ when analyzed with 32 scans, whereas in the analysis with 100 scans and resolution of 4 cm⁻¹ the samples differ in the spectral ranges between 790.68–786.82 cm⁻¹, 806.11 and 804.13 cm⁻¹, and 831.18–827.32 cm⁻¹, and in the analysis with 100 scans and resolution of 2 cm⁻¹ the samples differ in the spectral ranges between 799.36 cm⁻¹ and 792.61 cm⁻¹.

Therefore, it is noticeable that the wavenumbers responsible for differing the samples regarding the drying method lie between a narrower range when the aforementioned parameters change, concentrating in the 780–830 cm⁻¹ spectral range, which is part of the fingerprint region [23] of the infrared spectrum. Moreover, when the analysis performed with 32 scans, resolution of 4 cm⁻¹, spectral range between 4 000–500 cm⁻¹ without the baseline correction, the oven-dried and the air-dried samples did not show statistical differences.

The LDA showed that both the 32 scans and resolution of 4 cm⁻¹ and the 100 scans and resolution of 2 cm⁻¹ analyses achieved 100% accuracy, whereas the 100 scans and resolution of 4 cm⁻¹ had 94.44% accuracy (Table 8). The sum of the LD1 and LD2 resulted in 79.69%, 78.60% and 89.19% of data variability, respectively.

When the t-test was performed in the samples from Colorado in the spectral range between 4 000–500 cm⁻¹ and only considering the first reading, i.e., the first replicate of each sample, and without the baseline correction, i.e., the oven-dried and air-dried samples did not differ from each other when analyzed with 32 scans, whereas in the analysis with 100 scans and resolution of 4 cm⁻¹ the samples differ in the spectral ranges 1 783.86–1 785.79 cm⁻¹, 1 793.50 cm⁻¹, 1 803.14–1 818.57 cm⁻¹, 1 822.43–1 849.43 cm⁻¹, 1 855.21–1 870.64 cm⁻¹, 1 880.28–1 882.21 cm⁻¹, 1 899.57–1 903.42 cm⁻¹, 1 909.21–1 911.14 cm⁻¹, 2 375.91–2 377.83 cm⁻¹, 2 399.05 cm⁻¹, 2 485.83–2 607.33 cm⁻¹, and in the analysis with 100 scans and resolution of 2 cm⁻¹ the samples differ in the spectral ranges 2 011.42–2 012.38 cm⁻¹, 2 015.28 and 2 016.24 cm⁻¹, 2 021.06 cm⁻¹, 2 029.74 cm⁻¹, 2 157.98 cm⁻¹. When the parameters change, it caused the wavenumbers responsible for differing the samples regarding the drying method to lie in the 1 780–2 600 cm⁻¹ spectral range, which is not part of the fingerprint region [23]. Additionally, as seen from the analysis performed with 32 scans, resolution of 4 cm⁻¹, spectral range between 4 000 and 500 cm⁻¹, changing the parameters strongly affected the t-test results.

Regarding the LDA, only group C had all their samples correctly classified in all approaches tested. Group E had just one sample misclassified in the approach with 100 scans and resolution of 4 cm⁻¹ (Table 9). The accuracy for both the 32 scans analysis and the 100 scans and resolution of 2 cm⁻¹ was of 88.89%, whereas for the 100 scans and resolution of 4 cm⁻¹ it was slightly lower, 83.33%.

The LDA performed with the ATR-FTIR “regular” methodologies consisting of 32 scans and resolution of 4 cm⁻¹, 100 scans and resolution of 2 cm⁻¹, and 100 scans and resolution of 4 cm⁻¹ presented accuracies of 94.44%, 100% and 94.44%, respectively. These results are similar to those of the Brazilian samples seized in the 2017 police operation, but not to those seized in 2014. As mentioned before, the particle size of all the samples from Brazil and from the United States was visually identical; therefore, this factor can be excluded as a possible explanation to the results from the “regular” methodology. Similar plant age and temperature treatment seem to have an impact on the chemical and/or physical aspects of the samples, and consequently on the data. However, when the samples were assessed by the “modified” approach, the accuracy results were considerably lower for the 32 scans and the 100 scans resolution of 4 cm⁻¹ approaches, being 88.88% for the first and 77.77% for the latter, but remained the same for the 100 scans resolution of 2 cm⁻¹. These results indicate that doing replicates [38] and baseline correction improve accuracy and should be employed whenever possible.