EConv-ViT: A strongly generalized apple leaf disease classification model based on the fusion of ConvNeXt and Transformer

Abstract

The accurate recognition of apple leaf diseases is crucial for ensuring crop health and agricultural productivity. However, deep learning models often suffer from poor generalization across diverse environments due to variations in lighting, background complexity, and leaf appearance. To address these challenges, we proposed EConv-ViT, a novel robust generalization model integrating ConvNeXt and Vision Transformer (ViT), enhanced with Efficient Channel Attention (ECA) for superior feature extraction and DropKey to improve generalization and applied the mode on image dataset both captured in laboratory and natural environments for healthy apple leaves, alternaria blotch, grey spot, rust, and mosaic disease. The propsed EConv-ViT model was tested on an independent dataset and achieved accuracy of 99.2% on laboratory-captured image dataset and 79.3% on images captured in natural environments. The classification accuracy for EConv-ViT model exhibited 18.6%, 36.1% and 37.8% improvements compared with ViT, ConvNeXt, and ResNet50 models on a dataset captured in natural environments. EConv-ViT can effectively capture both local and global features and demonstrate its potential for the application on related automated disease monitoring systems.