Non-Destructive Near-Infrared Technology for Efficient Cannabinoid Analysis in Cannabis Inflorescences

2.3. PLSR Modeling for the Prediction of Total CBD, Total THC, and Total CBG Concentrations

The partial least-squares regression (PLSR) model is based on training (70%), data scaling, and cross-validation, ensuring a consistent and reproducible approach for assessing predictive accuracy. The predictive performances of the PLSR models for total CBD, total THC, and total CBG concentrations were compared between raw data, SNV, and SG smoothing (Table 2).

In Table 2, the performance of different pre-processing techniques in predicting total CBD concentration is detailed. The application of standard normal variate (SNV) pre-processing exhibited superior performance, as evidenced by the lowest root mean square error of prediction (RMSEP) of 2.228 and the highest coefficient of determination for prediction (R²P) of 0.792. Additionally, the ratio of performance to deviation (RPD) for total CBD was highest (2.195) with SNV, indicating its robustness in accurately predicting CBD concentration. In contrast, the raw data resulted in the highest RMSEP (2.379) and the lowest R²P (0.763) and RPD (2.055) compared to both pre-processed datasets. Figure 2 visually represents the comparison of these results, showing that the predicted values with SNV (Figure 2b) are more closely aligned with the observed values, indicated by their proximity to the diagonal line, compared to raw data (Figure 2a) and SG smoothing (Figure 2c).

The PLSR model demonstrated optimal performance when coupled with SNV pre-processing for predicting total THC concentration (Table 2). This combination yielded the highest coefficient of determination for prediction (R²P) of 0.847 and ratio of performance to deviation (RPD) of 2.555, along with the lowest root mean square error of prediction (RMSEP) of 1.498. In contrast, both raw data and Savitzky–Golay (SG) smoothing techniques fell short in accuracy. Specifically, raw data resulted in an R²P of 0.812, an RPD of 2.306, and an RMSEP of 1.651, indicating less precise predictions compared to SNV pre-processing. These findings underscore the superiority of SNV pre-processing in enhancing the accuracy of total THC concentration predictions (Figure S1).

In predicting the concentration of total CBG, the results exhibited unique characteristics (Figure S1). Raw data produced the highest prediction correlation coefficient (R²P) of 0.806, accompanied by a ratio of performance to deviation (RPD) of 2.267 and a root mean square error of prediction (RMSEP) of 0.623. Notably, SNV pre-processing, which typically performed favorably for other cannabinoids in this study, yielded lower predictions with an R²P of 0.763, an RPD of 2.057, and an RMSEP of 0.687 (Table 2). Interestingly, SG smoothing demonstrated competitive performance comparable to the highest raw data prediction, achieving an R²P of 0.804, an RPD of 2.256, and an RMSEP of 0.627.

NIR spectroscopy has already been applied to the quantitative analysis of CBD [21], THC [13,21], and CBG [14] in fine powder [22], semi fine powder [8], and whole inflorescences [13,23]. Heterogeneity in the whole cannabis inflorescence presents a challenging analytic target for NIR spectroscopy [13] due to its sensitivity to both chemical and physical properties [24]. Therefore, a whole inflorescence with 10 spectra per sample was used in this study. In addition, a large sample size is pivotal in the training of NIR prediction to underline the spectral pattern more effectively [25], strengthen robustness to variations within the dataset [26], and mitigate the risk of overfitting [27]. In this investigation, we utilized a substantial sample size (n = 890), emphasizing the effectiveness of highlighting spectral patterns and dataset variations and reducing the risk of overfitting.

In the study conducted by [10], a predictive model was developed for the whole inflorescences, covering five cannabinoids within the wavelength range of 950 to 1650 nm. The PLSR model employed in their research demonstrated notable predictive accuracy, yielding the highest R²P of 0.89 for CBD concentration (SD = 1.84) and a contrasting R²P of 0.11 (SD = 0.20) for THC across a total of 194 samples [10]. For CBG concentration, which involved 187 whole inflorescence samples, the R²P was 0.43 (SD = 0.04). The lower variations observed in the CBG and THC datasets were identified as potential contributors to the lower prediction accuracy in their study [10]. In our investigation, where the total THC concentration exhibited substantial variation, with a standard deviation of 3.88 (Table 1), pre-processing with SNV resulted in the highest R²P of 0.847 (Table 2). Conversely, despite lower variations in total CBG concentration (SD = 1.34) compared to both total THC and total CBD concentrations (SD = 5.12), the raw data still yielded a noteworthy R²P of 0.806. Importantly, variation in cannabinoid concentrations alone does not translate into higher prediction accuracy. As previously mentioned [10], demonstrated that CBG concentration, characterized by extremely low variation (SD = 0.04), exhibited a higher R²P compared to THC (R²P = 0.11), which had relatively higher variation (SD = 0.20). Our study aligns with these findings, where total CBD concentration, marked by higher variation (Table 1), exhibited lower prediction accuracy in comparison to total THC and total CBG concentrations (Table 2).

In the comparison conducted by [13], low-cost NIRSG1 and mid-cost MicroNIR devices were evaluated on a total of 26 samples for THC concentration prediction using PLSR models, considering variations in whole inflorescences and different grinding levels. In their study, MicroNIR outperformed NIRSG1, achieving the highest prediction accuracy for whole inflorescences with an R²P of 0.93 and an RPD of 4.54, while NIRSG1 showed lower accuracy with R²P = 0.73 and RPD = 1.95 [13]. In our current study, we obtained the highest prediction accuracy for total THC concentration using the low-cost NIRSG1 device (R²P = 0.847 and RPD = 2.555). This was achieved through the application of the pre-processing SNV technique (Table 2). Notably, our research, conducted with a considerably larger sample size (n = 890), reinforces the robustness of our findings while utilizing the cost-effective NIRSG1 device.

The findings in this study underscore the critical influence of pre-processing techniques on the predictive performance of PLSR models for cannabinoid concentration. SNV pre-processing consistently produced superior prediction of both total CBD and total THC concentrations, emphasizing its effectiveness in reducing noise and enhancing the interpretability of spectral data. The raw data, while showcasing competitive prediction for CBG concentration, lagged when applied to total CBD and total THC concentration models. The unexpected outcome observed for the prediction of total CBG concentration requires further investigation. While SNV pre-processing was advantageous for other cannabinoids, it exhibited decreased performance for total CBG concentration. This suggests that the underlying spectral characteristics of total CBG may differ substantially from total CBD and total THC, necessitating alternative pre-processing strategies.

The reference values were obtained using high-performance liquid chromatography (HPLC), a widely trusted method. However, the findings from NIR spectroscopy, as indicated above, continue to hold relevance within the cannabis industry. Typically, only a single inflorescence (apical bud) per plant is utilized for cannabinoid analysis to represent the whole plant. Nevertheless, research studies have revealed significant variations in CBD concentrations among different inflorescences within the same plant [7,28]. For instance, a recent study [7], demonstrated significant differences in CBD concentration between top, mid, and low buds. To conduct a comprehensive and representative examination of the cannabis plant, multiple samples per plant are necessary, which can be highly expensive when using HPLC. Incorporating NIR spectroscopy alongside HPLC offers a robust, cost-effective, and representative approach that aligns with the standards of the cannabis industry.

The regulations that impose stringent limits on THC levels in cannabis genotypes, as seen in Europe with a threshold of 0.3%, present significant hurdles for the cultivation of non-psychoactive dominant varieties, such as CBD-dominant varieties. Considering these restrictions, breeding programs encounter substantial challenges when attempting to develop novel cultivars [29]. The risk of pollen contamination from a single THC-dominant plant can jeopardize the entire breeding initiative. NIR spectroscopy offers a potential solution by enabling comprehensive cannabinoid profiling at early growth stages [11]. This allows for a swift assessment to avoid potential contamination issues and aids in the early selection of relevant genotype profiles.

In conclusion, the results obtained in this study highlight the importance of tailoring pre-processing techniques to the specific cannabinoid of interest with an NIRSG1 device. The choice of pre-processing technique can significantly impact model accuracy and generalization; careful consideration should be given to the unique spectral characteristics of each compound. Further research is required with extended wavelength ranges to uncover the underlying factors contributing to the varying performance of pre-processing methods across different cannabinoids. Overall, this study demonstrates the versatility of PLSR modeling and the critical role of pre-processing in optimizing predictions for cannabinoid concentration with NIRSG1, offering valuable insights for researchers in the field of spectroscopy and analytical chemistry.

3.5. Partial Least-Squares Regression (PLSR)

Utilizing partial least-squares regression (PLSR), the relationship between the predictor variables (wavelength spectra) and the target response variable (total CBD, total THC, and total CBG) was modeled. Ten-fold cross-validation was employed to determine the optimal number of latent variables. The dataset was randomly divided into ten subsets, and the PLSR model was trained and validated iteratively. Equation (3) represents the PLSR model equation used in our analysis:

where T represents the K × A matrix of orthogonal latent variables, X is a I × K matrix of wavelength , and W is a K × A matrix of weights. The matrix T has the property that errors in X = TP + E and Y = TC + F are small [30,31], where Y is the vector of the response variable, P and C are some weighting matrices, and E and F are error matrices. Small means that it maximizes the coefficient of determination (R²) and minimizes the root mean square error of prediction (RMSEP), ensuring they effectively capture the underlying patterns in the I ×K that are most relevant for predicting the response variable vector Y. In this study, A = 20 was chosen as the optimal number of latent variables. Predicting the content of the target response variable (ˆ) for a new dataset, (pred) is achieved using Equation (4):

where ˆ represents the predicted content of the target response variable, pred represents the set of latent variables for these new observations, and ˆ is the estimated matrix of regression coefficients. To ensure the reproducibility of our results, a random seed was set before using any modeling techniques. Additionally, we performed a random split of the dataset, selecting 70% for the training set and reserving 30% for the test set. This random data split allows one to train the PLSR model on one portion of the data and validate its performance on an independent dataset, thereby assessing its predictive accuracy. The prediction ability of the model was evaluated through several metrics, including the root mean square error of cross-validation (RMSECV), the root mean square error of prediction (RMSEP), correlation coefficients (R²), and the relative predictive deviation (RPD) for each model [32]. These metrics collectively assess the accuracy and reliability of our predictive models.

In addition to PLSR modeling, we used the expolinear function for visualization purposes. This function is defined by the Equation (5):

where  represents the output variable (cannabinoid) and  denotes the input variable (wavelength spectra). Parameters ,,and are used for curve fitting. The ‘expolinear’ function was primarily employed to enhance the visual representation of our data.