A multi-class hybrid variational autoencoder and vision transformer model for enhanced plant disease identification

Abstract

Agriculture is considered as the propeller of economic growth as it accounts for 6.4 % of global gross domestic product (GDP) and in low-income countries, it can account for more than 25 % of GDP. Plants supply more than 80 % of the food consumed by humans and are the main source of nutrition for animals. Plant diseases pose a major risk to global food security as they account for losses of between 10 to 30 % of the global harvest every year. Deep learning techniques like convolutional neural networks successfully identify image-based diseases but struggle with capturing long-range contextual information. This makes them less robust in noisy or high-resolution images. Their high computational and memory demands also limit scalability for large datasets. To overcome these issues, we propose a hybrid model with the potential to combine Variational Autoencoders and Vision Transformers for enhanced accuracy and robustness of plant disease classification. Variational Autoencoder reduces image dimensionality while preserving essential features, and Vision Transformer captures global relationships to enhance accuracy and scalability, especially in multi-class disease classification. The experiment used images of corn, potato, and tomato plant leaves from the publicly available PlantVillage dataset. On-the-fly data augmentation was applied to further increase the robustness of the model. The proposed model achieved a classification accuracy of 93.2 %. This technique provides a reliable and efficient solution for identifying multiple plant diseases across various crops. It enhances agricultural productivity and supports food security efforts.