Hybrid CNN-ViT Framework for Enhanced Plant Illness Identification: Accuracy and Interpretability Improvements

Abstract

The timely detection of plant health status is essential to obtaining several advantages, such as increasing crop yield, reducing the toxic pollutants used in crops, enhancing healthy crops, and enhancing economic returns. A computer-aided plant status identification system enables plant health identification using plant leaves because plant leaves are the most prominent in the plant. Deep learning (DL) and convolutional neural networks (CNN) are distinguished in the area of plant health identification. Still, CNN cannot handle images of variable size and fails to extract features efficiently if the image has complex backgrounds. To avoid these problems, this research proposes a hybrid model for plant (groundnut) illness status identification using the CNN and vision transformer (HCVT). CNN is efficient in identifying local features, and ViTs are efficient in identifying global features from leaf images; leveraging these advantages enhances the model's efficacy. The HCVT model used the two datasets, groundnut and rice datasets, for experimentation. The HCVT model obtained an accuracy of 96.40% on the groundnut leaf image dataset, with 99.86% accuracy on the rice dataset. The ablation analysis performed the necessity of each component in the HCVT model. The LIME (local interpretable model-agnostic explanations) technique is utilized to comprehend the HCVT method functionality and its results showed that the HCVT model accomplished better than the present cutting-edge models in identifying plant illness and exhibiting its generalization potential. The HCVT model is affordable and widely accessible for analysing plant leaf images. Utilizing the HCVT model offers a robust and easily accessible approach for diagnosing plant diseases by analysing leaf images.