Multiclass Fruit Disease and Maturity Classification Using Hybrid Deep Neural and Transformer Models With Explainable AI Techniques

The principal aim of this research work is to develop an efficient and explainable deep learning (DL) framework for the classification of fruit diseases. Timely and accurate detection of diseases in plants is essential in order to reduce the loss of crops and high-quality yields. Existing traditional methods of inspecting the crops manually are laborious and prone to errors. Most existing DL solutions are limited in the sense that they are trained on a single fruit type, depend on heavy models that are hard to use in real farm conditions, and do not explain their predictions. The proposed FruitViTNet is used to overcome these limitations by handling five fruits and sixteen disease classes in one model and by using a lightweight architecture that can be deployed on mobile and edge devices. Hence, this study aims for high classification accuracy with interpretable means obtained from explainable AI techniques so that they can be reliably used for applications in precision agriculture. The experiments were performed using public, labeled datasets of fruit diseases, namely, Apple (4 classes), Orange (3 classes), Pomegranate (2 classes), Banana (4 classes), and Papaya (3 classes), obtained from the online data sources. The dataset was divided into training (65%), validation (15%), and testing (20%) so as to ensure a balanced evaluation of the models. In this study, two pretrained convolutional neural network (CNN) models, two pretrained Transformer models, and the proposed hybrid deep transformer model were presented for plant disease classification. The evaluation framework was generated in such a way that it contained model training, 6-fold cross-validation in order to develop a robust model, and final testing. To improve explainability and trust in the results from the classification model, some Explainable AI techniques, such as Gradient-weighted Class Activation Mapping (Grad-CAM) and Local Interpretable Model-agnostic Explanations (LIME), were employed so that visual explanations of the model predictions could be made. The proposed hybrid deep transformer model showed better performance over existing CNN and Transformer background models, achieving a testing accuracy of 99.31%. The results from cross-validation also indicate that the model is stable and robust across various folds in the dataset. In conclusion, the results show that the proposed hybrid model leads to significant enhancement in the classification of fruit diseases, possibly due to the combination of local feature extraction from CNNs with global attention mechanism from Transformer models. The integration of explainable AI techniques leads to transparency in the model predictions, thereby enabling its use in real-time agriculture for applications in precision farming.