Classification of Some Iranian Vicia Species Using SEM Image Analysis Coupled with Conventional Texture Analysis and Deep Learning

Micromorphological characteristics of seed sculpturing might be effective in circumscribing the infra-specific taxa in the genus Vicia. The present study was conducted to determine whether microstructural and seed coat texture data obtained from SEM images can serve as sufficient tools for delimiting Vicia genus. Other than visual inspections, a variety of texture-based methods, including the four conventional approaches of GLCM, LBP, LBGLCM, and SFTA, and the four pre-trained convolutional neural networks, namely, ResNet50, VGG16, VGG19, and Xception models were employed to extract features and to classify the species of Vicia genus using SEM images. In a subsequent step, the four unsupervised k-means, Mean-shift, agglomerative, and Gaussian mixture classification methods were used to group the identified Vicia spices based on the underlying features thus extracted. Moreover, the three supervised classifiers of Multilayer Perceptron Network (MLP), Support Vector Machine (SVM), and k-Nearest Neighbor (kNN) were compared in terms of capability in discriminating the different visually-identified classes. SEM results showed that three classes might be identified based on the micromorphological character-species connections and that the differences among the species in the Vicia genus and the validity of Vicia sativa could be confirmed. Regarding the performance of the classifiers, SFTA textural descriptor outperformed the GLCM, LBP, and LBGLCM algorithms, but yielded a decreased accuracy compared with deep learning models. The combined Xception model and a MLP classifier was successful to discriminate the species in the Vicia genus with the best classification performances of 99 and 96% in training and testing, respectively.

Classification of Some Iranian Vicia Species Using SEM Image Analysis Coupled with Conventional Texture Analysis and Deep Learning Keywords:

Convolutional neural networks , Micromorphology , Plant taxonomy , Seed sculpturing , Scanning Electron Microscope (SEM).