Integrated of Hyperspectral Imaging and Machine Learning Algorithms for Nondestructive Detection of Therapeutic...

The approaches used to determine the medicinal properties of the plants are often destructive, labor-intensive, time-consuming, and expensive, making it impossible to analyze their quality analysis online. Performance of hyperspectral imaging (HSI) integrated with intelligent techniques to overcome these problems was investigated in this research. For this purpose, three classification methods-support vector machine, random forest (RF), and extreme gradient boosting-were studied for the classification of plants in three classes of medicinal, edible, and ornamental for the organs of leaf, stem, flower, and root. The medicinal effects of the plant organs were determined by measuring different biochemical properties of the organs. The spectral reflectance of the samples was used to train and test the classification methods in which output targets were the plant types. The results showed that amounts of the biochemical factors except oil content of the medicinal plants were higher than the other types of plants. Further, the biochemical factors of flowers and leaves were higher than the other organs indicating that the most therapeutic effect of the plants is through the flowers and leaves. Using HSI, a similar spectral trend was appeared in each organ, whereas it was different among the organs. Using the RF as the best method (precision and accuracy were higher than 0.95), the lowest misclassification rates were related to the stem and leaf datasets, indicating that these two organs were most suitable to classify the plants aromatically. The most misclassifications of the organs were occurred between medicinal and edible plants related to the spectra having higher correlations with flavonoid, phenol, and antioxidant compounds. Overall, the misclassification rates were negligible, and thus, the methods developed in this study can be used online in postharvest processes of the medicinal plants.