Hybrid local–global vision transformer framework for robust plant disease detection in precision agriculture

Correctly identifying plant diseases has significant consequences for sustainable agricultural productivity as well as for national and global food security. Even though traditional convolutional neural networks have achieved great success detecting diseases using images, the receptive field is inherently limited which in turn limits the ability simply to model long-range dependencies. However, vision transformers can capture global relationships but often fail to maintain fine-grained local texture information that may be critical for distinguishing subtle symptoms caused by diseases variably. To mitigate these issues, we introduce ViT-LGFNet, a Hybrid Local-Global Fusion Vision Transformer that uses local convolutional feature extraction along with global self-attention encoding. The proposed architecture is built with the following proposed modules: a Local CNN Stem module for learning fine textures followed by a Transformer Encoder module for modeling contextual features. We propose a dual Cross-Attention Fusion Module that performs bi-direction direction exchanges between local and global tokens to increase the discriminative ability for the hybrid representation. The PlantVillage dataset results demonstrate ViT-LGFNet mean validation accuracy of 95.15% ± 0.74 across multiple runs, with a best-case performance of 96.28%. Grad-CAM and Attention Rollout maps were placed for visualization purposes, demonstrating that the model accurately attends to the disease-affected part of an image. The proposed model offers a reliable and explainable benchmark for plant disease recognition and shows potential for application in smart farming systems and AI-based crop health monitoring applications.