Hybrid Explainable DeiT-Based Framework for Plant Disease Classification and Severity Estimation

The detection of plant diseases is essential to the preservation of agricultural productivity and food security, yet the existing technologies are likely to have low interpretability and low generalization in practice. This work suggests a hybrid deep learning model in the form of Data-efficient Image Transformers (DeiT) to detect and classify plant diseases and estimate their severity. The framework takes advantage of DeiT-Base, DeiT-Small, and DeiT-Tiny models to embrace global contextual dependencies of plant leaf images. The proposed hybrid Explainable Artificial Intelligence (XAI) module aims to enhance interpretability by combining Gradient-weighted Class Activation Mapping (Grad-CAM) as a local feature attribution model and Attention Rollout as a global dependency visualization model. In addition, a leaf segmentation method, a HSV-based method, is employed, which isolates disease-relevant regions and minimizes noise to increase the classification accuracy and level of explanation. The damage ratio analysis is combined with attention maps generated by XAI to build a severity estimation module. Experiments on the New Plant Diseases Dataset (Augmented) with large-scale experiments demonstrate that the proposed DeiT-Base model can achieve a maximum accuracy of 99.13, a better result compared to a variety of CNNs, such as ResNet50, DenseNet121, MobileNetV3, EfficientNet, InceptionV3. Also, hybrid XAI framework has better interpretability performance, such as focus score, noise, signal-to-noise ratio (SNR), and entropy, than single explanation procedures. The system proposed is not only capable of improving the accuracy of classification but also has better transparency and meaningful severity estimation which makes it appropriate to the real-world application of precision agriculture.