A modular vision transformer and machine learning framework for scalable multi-crop plant disease classification

Precision agriculture relies on accurate plant disease classification, but current deep learning methods struggle with scalability, generalization, and computational demands across various crops. This research introduces a two-step hybrid framework for multi-crop plant disease classification that merges deep feature extraction with machine learning-based classification.

The initial phase assesses four pretrained CNN models (VGG16, ResNet50, InceptionV3, and MobileNetV2) as feature extractors on a dataset of 21 disease classes for tomato, chilli, and potato plants. The subsequent phase employs a Vision Transformer (ViT) to extract global feature representations from a 25-class dataset encompassing seven crop species, followed by classification using machine learning classifiers.

Results indicate that MobileNetV2-SVM yields the best performance among CNN-based setups, achieving 89.44% accuracy, while the ViT-SVM framework attains 98% accuracy for multi-crop disease classification. These findings underscore the efficacy of transformer-derived features for generalized plant disease recognition and underscore the promise of hybrid deep feature learning frameworks for precision agriculture.