Non-destructive assessment of cannabis quality during drying process using hyperspectral imaging and machine learning

2.2. Drying conditions

Cannabis flowers were dried in six open plastic trays each under two drying and relative humidity conditions: hot-air drying (59 ± 3.6°C and 10 ± 3.7%) and cool-air drying (19 ± 1.2°C and 44 ± 8.2%). The cool-air drying conditions were similar to the traditional air-drying conditions (Challa et al., 2021). The air temperature and relative humidity were measured and recorded at 20-min intervals using a temperature and humidity data logger (RC-51H; Elitech Technology, Inc., Milpitas, CA, USA). The changes over time under these drying conditions are shown in Supplementary Figure S1. The flowers in each tray were weighed before drying and at 2, 4, 7, and 9 days post-drying. The relative weight change and relative water content (RWC) were calculated as follows:

where w_tis the weight at time t, w₀ is the fresh weight at harvest, and DM is the ratio of dry matter (DM = 0.21894) measured from flowers of the same cultivar. Flower samples were collected twice from five trays per treatment to obtain hyperspectral images and cannabinoid data for model development. For tracking data, weight and hyperspectral imaging data were collected from a single tray under each of the two drying conditions.

2.3. UHPLC analysis for cannabinoids

Flower samples were collected prior to drying and at 2, 4, 7, and 9 days after drying. Subsequently, the samples were promptly immersed in liquid nitrogen and were freeze-dried at –80°C. The lyophilized samples were ground finely, and the powder (1 g) was extracted using methanol (16 mL) under sonication at 50°C for 20 min. The extracts were centrifuged, filtered through a 0.22 µm membrane filter (Whatman, Maidstone, UK), and concentrated using a nitrogen gas evaporator. These concentrated extracts were then re-dissolved in DMSO to achieve a concentration of 10 mg/mL and stored at –80°C before analysis. The samples were diluted to 0.1 mg/mL with methanol before injection into an ultra-high-performance liquid chromatography (UHPLC) system. To quantify the target compounds, four standards (CBDA, CBD, THCA, and THC) were purchased from Cerilliant (Cerilliant Corp., Round Rock, TX, USA) and dissolved in acetonitrile at a concentration of 1 mg/mL. UHPLC analysis was performed using a Shimadzu Nexera X3 UHPLC system (Shimadzu Corp., Kyoto, Japan), equipped with two pumps (LC-40B), a column oven (CTO-40C), an autosampler (SIL-40C), and a photodiode array (SPD-M40). Separations were achieved on a YMC-Triart C18 column (100 × 2.0 mm, 1.9 µm; YMC Co., Ltd., Kyoto, Japan), with a mobile phase composed of 0.2% formic acid in both water (A) and acetonitrile (B). The gradient elution was set as follows: 75% B for 0–4 min, linear increase to 90% B in 4–10 min, decreased to 75% B within the next 0.5 min and re-equilibrated to initial gradient 75% B until 13 min. The column temperature was 30°C. The sample injection volume was 5 µL.

2.4. Hyperspectral image collection and processing

The hyperspectral imaging system consisted of a hyperspectral imaging camera (MicroHSI 410 SHARK, Corning, NY, USA) and eight 20 W halogen lamps placed within a movable stage in a dark chamber (Figure 1). The camera captured line-scan images with 150 spectral bands in the 400–1,000 nm range at a rate of 100 mm/s. The hyperspectral images were obtained at a spatial resolution of 682 × 1,540 pixels. Each round of scanning involved five samples for the experimental data and one tray for the tracking data. Hyperspectral data were examined within Python 3.9 environment (Python Software Foundation, Wilmington, DE, USA) using the Spectral Python (SPy) library.

Figure 1

Hyperspectral imaging system (A) and description of hyperspectral data processing (B) in this study. The regions of interest (ROI) were selected through a threshold technique.

A threshold technique was used to eliminate the background from the hyperspectral cube data. A normalized band difference (NBD) was calculated using reflectance values at 764.74 and 684.69 nm, where NBD = (R_764.74 – R_684.69)/(R_764.74 + R_684.69) and R denotes the reflectance values at the wavelength in a single pixel. A threshold was applied to the images to enhance the contrast between the plants and background, and pixels with NBD > 0.3 were selected as regions of interest (ROI) (Figure 1B). Classification data were collected by extracting multiple ROI from 90 hyperspectral images. Only regions with 20 × 20 pixels covering more than 70% of the ROI were selected without overlapping were selected, resulting in 27–73 data points for each sample. An average spectrum was extracted from each data point, and a total of 4,707 spectral data were used to develop the classification model.

2.5. Model development for quality classification

Four quality characteristics, namely dryness, CBDA conversion, THCA conversion, and CBD : THC, were categorized into two or three classes based on RWC and cannabinoid content (Table 1). To determine the classes of each characteristic, we conducted a sensitivity analysis for each range using the raw spectrum and a logistic regression model (Supplementary Figure S2).

To predict and classify the quality traits of harvested cannabis flowers, we compared combinations of different spectral preprocessing methods and machine learning algorithms. Five spectral preprocessing methods were used: raw spectrum, SG filter with a third-order polynomial fit with five data points: 1st Der, 2nd Der, and MSC. The pre-processed average spectra are shown in Supplementary Figure S3. Five machine learning algorithm classifiers were used: LR, SVM, KNN, RF, and GNB. We applied the one-vs.-rest method to binary classification algorithms, such as LR and SVM, for multiclass classification. The model implementation was programmed in Python 3.9 based on SciPy and the scikit-learn package.

3.2. Spectrum and color changes of cannabis flowers during drying process

The average spectra of the hyperspectral images revealed the spectral reflectance of cannabis flowers, and the variations were more closely associated with the drying method than to the drying duration (Figure 3). The drying-induced changes in reflectance were categorized into four ranges, and representative spectral images are shown in Figure 4. The reflectance at wavelengths below 552.61 nm was lower after hot-air drying compared to other drying conditions. In contrast, higher values were observed in the range of 556.61–612.65 nm after cool-air drying in comparison to alternative conditions. For spectra within the wavelength range of 616.65–708.7 nm, the highest values were observed after cool-air drying, followed by hot-air drying, with the lowest levels prior to drying. Notably, reflectance in the range of 720.71–884.81 nm experienced a rapid decrease under both drying conditions, with the most pronounced decline observed after only two days of hot-air drying. Conversely, cool-air drying resulted in gradual decrease over a longer drying period.

Figure 3

Spectral changes of cannabis flowers during the drying period (0, 2, 4, 7, and 9 days) according to drying method (A, hot-air drying; B, cool-air drying). Solid lines and shaded areas indicate mean ± SD (n = 10).

3.3. Changes in cannabinoids of cannabis flowers during drying process

Cannabinoids, such as CBDA, CBD, THCA, and THC, underwent significant changes during the drying process (Figure 5). The results of the two-way ANOVA revealed that the drying method, drying period, and their interaction considerably influenced cannabinoid content (p< 0.01), except for total CBD. Total CBD concentration was significantly affected by the duration of drying, whereas the drying method (p = 0.065) and their interaction (p = 0.098) were not significant. Under hot-air conditions, CBDA gradually declined between days 4–9 of drying, resulting in a 65.4% decrease at day 7 compared to the initial value. However, the CBD content rapidly increased after hot-air drying, surging by a 20.1-fold after 7 days of drying compared to the initial value. The total CBD content peaked at 48 h post hot-air drying, with a 59.1% increase from the initial value. After 7 days of hot-air drying, THCA, THC, and total THC concentrations reached their highest levels, increasing by 1.7-, 41.1-, and 2.7-fold, respectively, compared to their pre-drying levels. Meanwhile, under cool-air conditions, CBDA and THCA reached their maximum values after 4 days of exposure, showing increases of 62.3% and 81.2%, respectively, compared with their levels before drying. However, no significant differences were observed in CBD and TCH levels during the cool-air drying period. After 4 days of exposure to cool-air drying, the total CBD and total THC levels increased by 64.8% and 87.2%, respectively, compared to their initial levels before drying.

3.4. Cannabis quality assessment models based on hyperspectral imaging

As a result of the 10-fold cross-validation (CV), the spectral preprocessing method and machine learning model were determined for each quality characteristic (Figure 6). The LR model had the highest overall accuracy when coupled with the MSC, SG filter, or raw reflectance, followed by the RF model with 2nd Der. Regarding the classification of dryness, THCA conversion, and CBD : THC, the LR model with MSC preprocessing achieved the highest 10-fold CV accuracies of 0.9979, 0.9450, and 0.9570, respectively. To classify the CBDA conversion, the LR model with the SG filter method was selected owing to its CV accuracy of 0.9862.

The selected models were evaluated using five metrics, and they demonstrated high accuracy in predicting each quality characteristic (Table 2). The prediction accuracy closely aligned with the CV accuracy of each model. All confusion matrices and ROC curves used to calculate these metrics are detailed in the (Supplementary Figures S4, S5). Only the THCA conversion model exhibited precision, recall, and F1-score values lower than the accuracy values, indicating an imbalance among the classes. Among the THCA conversion classes, the high THC% class contained a relatively small amount of data (10% of all data), which consequently led to lower precision, recall, and F1-score values (Table 1, Supplementary Figure S4).

4.1. Changes in cannabinoid composition in cannabis plants

Major cannabinoids, including CBDA, CBD, THCA, and THC, share a biosynthetic pathway that connects to the precursor molecule, CBGA (Tahir et al., 2021; Govindarajan et al., 2023). This biosynthesis predominantly occurs within the trichomes of cannabis plants (Livingston et al., 2020; Tanney et al., 2021), which develop in various parts of female cannabis flowers. While trichome development may vary among different genotypes, it typically begins gradually after the onset of flowering, with a significant increase observed at approximately 3–4 weeks as female flowers take form. As the trichomes continue to mature, they synthesize and accumulate cannabinoids. However, the senescence phase begins at approximately 8 weeks after flower anthesis, and resin secretion gradually decreases (Punja et al., 2023).

Cannabinoid trichomes can be categorized into four types: non-secretory, sessile capitate, pre-stocked capitate, and stocked-capitate trichomes (Hammond and Mahlberg, 1973). The heads of the stocked-capitate trichomes are protected by a cuticle layer. Within the lower part of these trichome heads, 12–16 disc cells can be found where cannabinoid synthesis occurs (Hammond and Mahlberg, 1973; Small and Naraine, 2016; Livingston et al., 2020). In contrast, the upper part of the resin accumulated various secondary metabolites, including cannabinoids, terpenes, organic acids, and polysaccharides (Jin et al., 2020; Livingston et al., 2021; Tanney et al., 2021). Cannabidiolic acid synthase (CBDAS) and Δ9-tetrahydrocannabinolic acid synthase (THCAS), responsible for the synthesis of THCA and CBDA, respectively, from CBGA, are equipped with secretory signal peptides that guide them to the resin. CBDAS and THCAS, once secreted into the extracellular space, catalyze the conversion of CBGA to Δ9-THCA and CBDA (Taura et al., 2007).

Remarkably, trichomes maintain their physical integrity even after the drying process owing to the protective cuticle layer covering their heads. This preservation of trichome heads suggests that no spatial limitations hinder the catalytic activity of THCAS and CBDAS during the drying period (Taura et al., 2007; Meija et al., 2022). This study provides limited evidence to support the preserved functional capacity of cannabinoid synthesis (Figure 5). Further research is needed to understand the relationship between trichome preservation and precursor turnover during drying.

Cannabinoid acids, such as CBDA and THCA, are readily decarboxylated and stabilized by heat during the drying process (Tahir et al., 2021; Govindarajan et al., 2023). Excessive heat can lead to the loss of the synthesized cannabinoids. Although the drying temperatures applied in this experiment were not high enough to cause significant cannabinoid loss (Wang et al., 2016; Das et al., 2022), prolonged drying can cause such loss (Addo et al., 2021; Chen et al., 2021). Therefore, determining the drying endpoint supports a smooth and stable transition to the next step.

4.2. Cannabis quality in industrial processes

The industrial decarboxylation process is crucial for extracting the active components, CBD and THC, through heating at a relatively high temperature, approximately 100°C, for a short reaction duration (Wang et al., 2016). In this study, CBDA conversion gradually increased with longer drying times, reaching 83.2% after 9 days of hot-air drying (Figure 5). CBD chemotype plants with a CBD : THC ratio of approximately 25:1 are commonly used in medical-grade products (Chandra et al., 2017). The ‘Cherry Blonde’ cultivar used in this study is a CBD chemotype cannabis with low THC levels, with a CBD : THC ratio of 36:1 at harvest. Under hot-air conditions, the ratio increased to 55:1 at 2 days after drying and subsequently decreased to 22–15:1 at 4–9 days after drying (Figure 5). The THCA conversion followed a similar pattern, peaking at 62.5% 2 days after hot-air drying and subsequently decreasing to 41.5–33.5% during days 4–9. In contrast, under cool air conditions, the conversion rates of CBDA and THCA and the CBD : THC ratio remained below 6% and in the range 29–36:1, respectively. These findings suggest that cool air can effectively maintain the conversion rates and desired ratios. Although additional decarboxylation process is required for medical-grade cannabis production, preservation is a suitable postharvest strategy, particularly for extended transportation and storage, ensuring a longer shelf life (AL Ubeed et al., 2022).

Each class of the four qualities in the present study was determined as the range that could be best classified through sensitivity analysis (Supplementary Figure S2). A criterion for classifying the industrial quality of cannabis is required to accelerate the development of quality control technologies.

4.3. Hyperspectral imaging analysis with spectral pre-processing and machine learning

We established a hyperspectral imaging-based model for evaluating cannabis quality during the drying process, including dryness, CBDA conversion, THCA conversion, and CBD : THC ratio. Extracting valuable information from hyperspectral data is challenging because of high dimensionality, redundancy, and noise (Saha and Manickavasagan, 2021). To make hyperspectral imaging applications more adaptable for real-time use, efficient algorithms and data processing techniques are necessary.

The most common practices of spectral preprocessing used in this study were derived from chemometric techniques, including the SG filter, 1st and 2nd derivatives, and MSC (Supplementary Figure S3). The SG filter is one of the most well-known smoothing methods for denoising, such as instrumental noise or extreme band rejection. It is also used to interpolate spectral data that can be used for other transformations, such as derivatives. In particular, the hyperspectral imaging of plants requires additional correction techniques because of the variability arising from these complex geometries (Mishra et al., 2017). Derivative techniques effectively reduce the additive effects as a constant offset and linear baseline shift. These techniques not only emphasize spectral features but also increase noise levels in the data (Vidal and Amigo, 2012). MSC is widely used to compensate for additive or multiplicative effects, i.e., both light scattering and baseline shift corrections. In this study, preprocessing methods, except for the SG filter, significantly affected the accuracy of the models (Figure 6, Supplementary Figure S3). The similarity in model accuracy between the SG filter and the raw spectrum may be attributed to the stable conditions in the dark chamber, which indicate minimal noise from the measurement environment (Figure 1).

Among the various machine-learning algorithms, we used common supervised classification methods, including LR, SVM, KNN, RF, and GNB. In this study, LR models exhibited the highest accuracy in predicting the four qualities when coupled with the MSC or SG filters (Figure 6). LR is a probability-based algorithm based on a logistic (sigmoid) function that calculates the probability of a binary outcome (Saha and Manickavasagan, 2021). The probability (P) with multiple variables has the following general form:

where X ₁ to X _n are distinct independent variables; β ₀ to β _n are the regression coefficients. When the number of samples or predictors is limited, e.g., in the field of clinical prediction models (Christodoulou et al., 2019; Nusinovici et al., 2020), LR is considered more suitable than other machine learning models. However, the predictor variables in the hyperspectral image-based model were wavelength bands, and the number was not small (150 in this study). Because of the assumption of a linear relationship between the features and class labels, LR may not effectively capture complex nonlinear relationships. For instance, LR models have limitations in inferring relationships between gene expressions in large-scale profiling (Chen et al., 2016). In this study, the accuracy of the LR model was low only when combined with the 2nd Der pre-processing, likely due to the increased feature complexity (Figure 6). In contrast, the accuracy of the SVM, KNN, RF, and GNB models increased when combined with 1st or 2nd derivatives compared with other preprocessing methods (Figure 6). These models are appropriate for handling non-linear relationships. SVM is an algorithm that determines the hyperplane that maximizes the distance between different classes in the data. It was designed to resolve overfitting issues when dealing with high-dimensional data (Noble, 2006). KNN is an instance-based algorithm that assigns data to the major class among its k nearest neighbors, where k = 5 in this study (Rehman et al., 2019). RF is a bagging algorithm that constructs multiple decision trees and combines their predictions (Breiman, 2001). In previous studies using hyperspectral data, SVM and RF models were more accurate than the KNN model in predicting nitrogen accumulation in legume plants (Flynn et al., 2023). GNB uses Bayes’ theorem with the assumption of feature independence and the Gaussian distribution of each class and classifies data based on probability (Frank et al., 2000). These results might be attributed to the linear relationship between spectral features and cannabis quality, or because the preprocessing makes the data more linear or removes nonlinear variations that better match the assumption of LR.