Robust multiclass classification of crop leaf diseases using hybrid deep learning and Grad-CAM interpretability

Area of Science:

• Agricultural Science
• Computer Science
• Machine Learning

Background:

• Plant diseases significantly impact crop yield and food security.
• Accurate and early disease detection is crucial for effective crop management.
• Deep learning offers promising solutions for automated plant disease identification.

Purpose of the Study:

• To develop and evaluate an effective deep learning framework for detecting and classifying diseases in banana, cherry, and tomato leaves.
• To compare the performance of a novel hybrid model against state-of-the-art pre-trained models.
• To validate the proposed model's accuracy and practical applicability in agriculture.

Main Methods:

• Utilized a publicly available dataset of healthy and diseased plant leaves (banana, cherry, tomato).
• Pre-processed data for deep learning architectures and split into training, validation, and testing sets.
• Implemented and compared baseline models (EfficientNetV2, ConvNeXt, Swin Transformer, ViT) with a novel Hybrid ConvNet-ViT model.
• Employed 5-fold cross-validation to enhance classifier performance and prevent overfitting.

Main Results:

• The proposed Hybrid ConvNet-ViT model achieved a testing accuracy of 99.29%, surpassing all evaluated pre-trained models.
• Demonstrated the efficacy of combining Convolutional Neural Network (ConvNet) local feature extraction with Vision Transformer (ViT) global context capabilities.
• The hybrid approach proved superior in classifying plant leaf diseases compared to individual state-of-the-art models.

Conclusions:

• The Hybrid ConvNet-ViT model is a highly effective and accurate solution for plant leaf disease detection and classification.
• The model's outstanding performance positions it as a valuable tool for practical agricultural applications.
• Integrating ConvNet and transformer frameworks enhances image-based disease detection capabilities.