Field investigations were conducted during the phenological vegetation phase of saffron, spanning February to April, across five saffron fields in Torbat-Heydarieh (Khorasan-Razavi Province, northeast of Iran), situated at an average altitude of 1450 meters above sea level (Table 2). These fields employed furrow irrigation. The primary objective was to identify saffron plants damaged by mites. Prior to the investigation, manual inspections confirmed the presence of mite pests in the examined fields during the previous growing season.

This research employed a variety of imaging techniques. Initially, hyperspectral images of saffron leaves in both healthy and damaged states were captured within a laboratory setting. A desktop HYSPIM alpha visible-near infrared (Vis-NIR) hyperspectral camera (Hyspim, Sweden) was utilized for this purpose. Each pixel of the 2D image contained spectral data within the wavelength range of 400-950 nm at a resolution of 2 nm. The camera was positioned in a controlled dark room, and illumination was provided by four full-spectrum 50W halogen lamps (Osram, Germany). Subsequently, on-site imaging was conducted during ground and aerial surveys. A Survey2 NIR camera (MAPIR, USA), equipped with a Sony Exmor IMX206 16MP (Bayer RGB) sensor and an FOV of 82° 23mm f/2.8, was used to capture images in 850 nm (NIR) and 660 nm (visible region) bands. A Canon Powershot RGB camera, featuring a 1/2.55” CMOS sensor with 12MP effective pixels and a field of view (FOV) of 94° and f/2.2, was also employed.

A Phantom 3 SE (DJI, China) drone was utilized for aerial imaging. This drone is equipped with a 4K camera capable of recording video at 30 frames per second, a 1/2.3-inch CMOS sensor with 12 million effective pixels, a maximum flight time of 25 minutes, and a maximum transmission range of 5 kilometers. Its specially designed lens with an FOV of 94° is well-suited for aerial imaging. The NIR camera was subsequently mounted on a drone for aerial imaging.

To assess the feasibility of using spectral data to differentiate between healthy and infested saffron plants and to select the most appropriate spectral bands, twenty plants were sampled from each field. These plants were divided into two groups: healthy and infested. Hyperspectral imaging was conducted within the 400-900 nm spectral range. To identify the optimal bands for classification and validate camera performance, spectral signatures of healthy and infested saffron leaves were compared.

The field survey phase comprised two stages. During the first year, ground surveys were conducted to identify and harvest mite-infested saffron plants. Corms were separated from the surrounding soil and analyzed. Infested areas were marked with GPS coordinates for subsequent study of the infestation's effects on plants in the second year. Plants were divided into two groups: healthy and infected. Images were captured under consistent lighting conditions in both visible light and NIR spectra. Generally, the NIR images were preferred over RGB images due to their greater sensitivity to plant photosynthetic properties (Aslahishahri, et al., 2021). Image quality and resolution were evaluated at three flight altitudes: 5, 10, and 15m. Given the requirement for high-resolution images, a flight altitude of five meters was selected.

In the second stage, infested plants were labeled, and aerial images were acquired at the study site at a 5-meter altitude with 12-megapixel resolution. Images were captured using both RGB and NIR cameras in two consecutive years, March 2021 and 2022. At a height of 5 m, 12-megapixel images covered an approximate area of 7.7 × 6 m, resulting in a spatial resolution of 593 pixels/m² (6 pixels/cm).

The first-stage images of saffron plants (Figure 1) and the aerial images (Figure 2) were captured on sunny days between 10 a.m. and 12 p.m. During the first stage, 20 images were acquired, while the second stage involved capturing 20 aerial images from each of the five fields in both RGB and NIR bands. A total of 100 images (from the first stage) were used to train the classification algorithm. To create datasets of healthy and infested plants, plants were removed from the soil, and damage was confirmed. The first-stage images were employed to process and train the machine learning classification model, while the second set was used to evaluate the model’s performance. Raw images underwent color analysis (extraction of color channels, histogram stretching, edge detection) before being processed. Image processing comprised four steps: pre-processing, segmentation, determination of vegetation indices, and image classification.

Image processing was conducted using MATLAB software (version 2017). Two distinct codes were developed, one for analyzing RGB images and the other for NIR images. These codes followed a similar procedure, each consisting of five components:

Image acquisition.

Prior to creating the dataset, NIR images required calibration, which was performed using “Mapir Camera Control” software (MAPIR, USA). The flowchart illustrating the dataset creation process is depicted in Figure 3. To separate infested and healthy saffron leaves, segmentation was performed on the images to create masks. This process was carried out individually for each image using MATLAB code or through manual cutting and separation of infested leaf sections. In some cases, where saffron plants exhibited minimal corm damage caused by mites, healthy and infested leaves were juxtaposed. For these instances, masking and separation were performed manually in Photoshop software, guided by information from field inspections and physically placed markers on the leaves.

Two datasets were generated for each image: the first containing a raw dataset of infested leaf pixels extracted from the main bands, and the second comprising a raw dataset of the main bands relating to healthy leaves. Both groups were utilized to calculate spectral indices (Table 3) for training the classifier.

In order to identify the required indices for a sound description of the infested areas and to maximise segmentation during the classification process, some indices were selected based on the available spectra extracted from the cameras (red, green, blue and 850nm NIR) according to the literature (Table 3). Since the infested and healthy plants were identical, the shape-based indices were discarded and the classification process was performed based on the spectral colour-based indices. As shown in Table 3, seven indices, including the Red, NIR, NDVI, RVI, difference vegetation index (DVI), difference red-nir ratio (DRN), and infrared percentage vegetation index (IPVI) were used to classify the infrared images. Besides, the reflectance in the main spectra of red, blue, green, and visible-band difference vegetation index (VDVI), visible atmospheric resistant index (VARI), triangular greenness index (TGI), normalized difference greenness index (NDGI), normalized green blue difference index (NGBDI), modified green red vegetation index (MGRVI), red green blue vegetation index (RGBVI), vegetative index (VEG), and excess of green index (EXG) were used to classify the RGB images.

The SVM, a supervised learning algorithm based on statistical learning theory, was adopted to detect and classify the infested plants from the healthy ones (Vani et al., 2017). Assuming class separability based on spectral index differences, hyperplanes with minimal margins were developed to distinguish between the classes. A dataset comprising randomly selected pixels and their corresponding indices from the total NIR images dataset for each group was created. These datasets included vegetation indices of 4000 pixels associated with healthy plants and 4000 pixels associated with infested ones. Ground survey data validated the accuracy of these groups. A similar procedure was followed for RGB images, resulting in two separate SVM structures: one for RGB images and one for NIR images. Eighty percent of the data was randomly selected for training, while the remaining 20% was used for validation (Baradaran Motie et al., 2021). The radial basis function (RBF) kernel was utilized in the SVM classifier. It is important to note that SVM models assign a class to each datum based on a decision-making function (Equation 1). Equation 2 presents the RBF kernel (Vani et al., 2017).

No significant differences were found between clusters in the composition, hygiene-sanitary parameters, and antioxidant capacity of the milk studied, as the animals in all farms were similarly managed based on the use of natural pastures. However, significant changes were observed in the milk composition by calendar month because of lactation effects and differences in feeding regimens. The percentage of lactose and the milk component yields (g d−1) patterns throughout the lactation months were similar, while the relationship between the fat and protein percentages was inverse to milk yield. During the end of summer and autumn months, the highest number of bacteria and somatic cells in the milk were observed, but the bacteriological count levels were below the legal limit set by EC regulations. The TAC was significantly higher in winter and spring milks than in the other milk samples.

The negative correlation between the antioxidant capacity and the somatic cells shows the important role of antioxidants in maintaining optimal udder health. One of the compounds involved in this antioxidant mechanism could be vitamin A due to the positive correlation between the antioxidant capacity and retinol determined in a previous study.

Finally, the information generated in this and previous studies on the quality of goat milk from the autochthonous Payoya breed will contribute to establishing the records of the traceability system to guarantee that the animal products obtained are of the native breed. All of this will help the consumer to easily identify these products and increase their demand, which will result in the conservation and promotion of the genetic heritage of these breeds and the foods derived from them.

In these equations, α is a constant coefficient, y is the group vector (labels), S_i is the support vector, Z is the input vector, K (Z, S_i) is the kernel of the SVM model, D(Z) is the decision function and σ is the variance.

To identify infested saffron plants, the generated SVM-based codes first processed the images, calculated vegetation indices, and applied the classification model. Furthermore, these developed SVM classifiers enabled the calculation of infestation proportions and the identification of corresponding areas within the field.

To validate the SVM classifier models, two new datasets were created using images (NIR and RGB) from different fields acquired during the second year of imaging and not utilized in model development. These datasets comprised 30 aerial images, three RGB and three NIR, representing six images from each field.

The confusion matrix method was employed to evaluate classifier performance. To address the class imbalance resulting from a higher prevalence of healthy plants, an under-sampling technique was implemented. This involved randomly sampling from the healthy class to maintain approximately equal class sizes. The developed model classified the images, generating a binary matrix indicating the positions of infected pixels. This matrix was termed “Model Prediction”. Subsequently, a second matrix, referred to as “Actual,” was manually created based on field survey data, containing the positions of pixels corresponding to infected plants at the same imaging points. By comparing the Actual and Model Prediction matrices, classification accuracy parameters were calculated. These parameters included ‘true positive’ (TP), ‘true negative’ (TN), ‘false positive’ (FP), and ‘false negative’ (FN), as presented in Table 6. Based on these data, classification quality criteria were calculated.

Classifier algorithm accuracy is represented as the ratio of correctly classified pixels to the total number of pixels. This is also known as Sensitivity (SNS) or True Positive Rate (TPR). The F1 score, or F Measure, reflects the balance between precision and TPR. In classification models, the ideal scenario is when both FP and FN approach zero. The F1 score ranges from zero to one, with higher values indicating better performance.

To develop a classification structure using SVM, model results were continually compared with actual data to optimize the model and minimize deviations from the original data. The SVM model aimed to reduce this deviation iteratively by identifying discrepancies from actual values and determining optimal hyperparameters for separating infested and healthy plants. Model optimization was conducted using the “fitsvm” function with 5-fold cross-validation. This process identified the optimal values for the hyperparameters gamma and C in the RBF kernel function, which correspond to the selection of values for box constraints and kernel scale. A larger C value results in a narrower boundary between the two datasets (infested and healthy groups). While higher classification quality can be achieved with larger C values, it increases the risk of misclassification during the testing phase due to potential errors in classifying borderline data.

The kernel scale parameter, inversely related to gamma, indicates the influence of each observation on the model’s output during training. Smaller kernel scale values are preferable as they suggest greater model stability. Additionally, the receiver operating characteristic (ROC) curve was employed as another evaluation metric to assess model performance. The ROC curve represents the area under the curve (AUC), and a value close to 1 indicates superior model performance.

To initiate the analytical process, preliminary correlation coefficients between all indices (features) and descriptive statistics were calculated. A dataset was prepared, encompassing images with selective masks for healthy and infested plants, including all five fields across two years of imaging. Table 4 presents the mean, standard deviation, and t-test results for the calculated indices (Table 3) of healthy and infested plants in NIR images. The results revealed that damaged plants exhibited higher average reflectance in the red band compared to healthy plants, while conversely, NIR band reflectance values were lower. The NDVI index, indicative of plant greenness, was on average 23.7% lower in damaged plants than in healthy plants (Devadas et al., 2009). Significant differences (p < 0.01) were observed between the values of the two groups (healthy and infested) for all indices. However, the standard deviation analysis indicated a wide dispersion of data, resulting in overlapping values between the two groups, as visualized in Figure 5. This suggests that no single index could effectively differentiate between healthy and infested saffron plants.

Figure 5 summarizes the pearson correlation coefficient (PCC) between indices and presents scatter plots depicting index changes relative to each other within the two groups (healthy and infested). These plots revealed distinct trends between the indices of the two groups, particularly for NIR-NDVI, NIR-RVI, NIR-DVI, NIR-DRN, and NIR-IPVI, suggesting potential for differentiating and classifying the groups. To distinguish between healthy and infested saffron plants, identifying distinct trends in the cross-changes of indices between the two groups is crucial. For instance, the NIR-NDVI plot (Figure 5) demonstrates a lack of correlation between the NIR index and NDVI of the infested group (correlation coefficient -0.012), whereas a correlation coefficient of -0.66 is observed for the healthy group. This highlights the importance of including both indices in the classification model. Conversely, in the cross-correlation plot of certain indices, such as RVI-NDVI, a high correlation coefficient between the values of the two groups (0.994 for healthy and 0.999 for infested) is evident. However, the pattern of change remains similar, with significant overlap between the data (scatter plot - Figure 5). Similar observations can be made for DRN-NDVI, DRN-RVI, IPVI-RVI, IPVI-NDVI, and IPVI-DRN.

Likewise, for RGB images, the descriptive statistics as well as the results of the t-test for healthy and infested plants are presented in Table 5. It reveals that, while there was no statistically significant difference between the mean values of infested and healthy plants in MGRVI, EXG, and VEG indices, the mean value of EXG in healthy plants was approximately 25% higher than in infested plants. However, the standard deviation indicates some overlap between the two groups, suggesting that this index may not be sufficient for individual group differentiation.

The PCC matrix and the scatter plot of the features which selected in the feature selection process relative to healthy and infested groups are presented in Figure 6. The feature selection process identified Red, Green, Blue, VDVI, VARI, NGBDI, MGRVI, and RGVBI as essential features for the classification model, a finding supported by t-test results (p < 0.01) (Table 5). The average values of the primary RGB bands ((red+green+blue)/3) were higher in infested plants compared to healthy plants, indicating a brighter color in infected leaves.

The training and testing steps for evaluating the best objective function after 30 iterations of NIR images revealed optimal C (Box-Constraint) and kernel scale values of 386.4 and 34.41, respectively. The model’s accuracy in the testing phase reached 82.3%. Similarly, the SVM classifier with the radial basis function was used to classify RGB images. The training and testing steps for evaluating the best objective function identified optimal C (Box-Constraint) and kernel scale values of 205.6 and 74.23, respectively. The resulting classifier model achieved a testing stage accuracy of 91.4%.

The performance of the classifier models is presented in two confusion matrices, with their parameters summarized in Table 6 for both RGB and NIR models. The confusion matrices shown in Figures 7 and 8 demonstrate that over 75.6% and 80.3% of the data were correctly categorized for NIR and RGB images, respectively.

The RGB-SVM model exhibited a marginally higher AUC compared to the NIR-SVM model, as evidenced by the ROC curves (Figure 9). Therefore, based on the obtained results and the ease of using an RGB camera, the RGB-SVM detection system is recommended.

The developed SVM models could discriminate mite infested plants with an acceptable level of performance and accuracy. Low FN (false negative) values (Table 6) suggest that there is little chance of missing the infested plants. However, the results for FP (false positive) show that, because of the small variation in their spectral reflectance, healthy plants could be identified as infested ones. Additionally, the F1 and sensitivity values are close to 0.7, indicating that the SVM classifier is performing as expected. To sum up the findings, the analysis shows that the classification model generates results that are both highly accurate and broadly applicable. The MATLAB code, running on a Core i7 CPU with 4 gigabytes of RAM, processed each aerial image in less than 15 seconds.