This paper proposes a novel approach for extracting deep features and classifying diseased plant leaves. The agriculture industry is negatively impacted by plant diseases causing crop and economic loss. Accurate and timely diagnosis is crucial for managing and controlling plant diseases, as traditional methods can be costly and time-consuming. Deep learning-based tools effectively detect plant diseases depending on the qualitative of extracted features. In this regard, a hybrid model for plant disease classification based on a Transfer Learning-based model followed by a vision transformer (TLMViT) is proposed. TLMViT has four stages: 1) data acquisition, where the PlantVillage and wheat datasets are used to train and evaluate the proposed model, 2) image augmentation to increase the number of training samples and overcome the overfitting issue, 3) leaf feature extraction by two consecutive phases: initial features extraction by using pre-trained based model and deep features extraction by using ViT model, and 4) classification by using MLP classifier. TLMViT is experimented with using five pre-trained-based models followed by ViT individually. TLMViT performs accurately in plant disease classification, obtaining 98.81% and 99.86% validation accuracy for VGG19 followed by the ViT model on PlantVillage and wheat datasets respectively. Moreover, TLMViT is compared with pre-trained-based architecture. The comparison result illustrates that TLMViT achieved an enhancement of 1.11% and 1.099% in validation accuracy, 2.576% and 2.92% in validation loss compared with the transfer learning-based model for PlantVillage and wheat datasets respectively. Thereby proposed model proves the efficiency of using ViT for extracting deep features from the leaf.INDEX TERMS Plant disease, image processing, deep learning, transfer learning, vision transformer.

No abstract

In this paper, different types of plant diseases in the PlantVillage dataset are getting focused for classification. In the realm of machine vision, plant disease identification is one of the most crucial tasks in the agricultural sector. It is a technique that employs equipment to capture images to detect and classify different types of diseases in plants. However, nakedeye monitoring of plants is impractical due to long processing times and a lack of specialists on farms in remote locations. Hence, combining image processing techniques with machine learning provides a solution to the problem of agricultural production while also ensuring food security. The plant features are extracted using a modified gray-level co-occurrence matrix (GLCM) technique and based on various statistical features. Both of these approaches were applied on original images with background and segmented images without background. Wavelet transform is also used with segmented images to decompose the image into sub-bands. All the features obtained are combined and SMOTE technique is used to balance the dataset prior to classification. For the purpose of classification, six machine learning models were compared, namely Light Gradient Boosting Machine (LGBM), Random Forest (RF), Decision Trees (DT), Logistic Regression (LG), AdaBoost, and Support Vector Machine (SVM). Further, different combinations of features were experimented and the experimental results prove that employing LGBM and SVM models resulted in attaining higher accuracy values i.e. 94.39% and 93.15%, respectively.