Hybrid Vision Transformer And Optimized Neural Network Approach For Accurate Sugarcane Leaf Disease Classification

Abstract

Early and precise identification of sugarcane leaf diseases is vital in timely treatment and enhanced crop productivity. Manual inspection processes are usually prone to errors and time consuming, making it necessary to have intelligent automated systems. Current image-based disease detection methods have difficulties in dealing with real world variations in image quality, light and noise in the background. Moreover, the models which are used traditionally do not have the power of the feature representation they need to distinguish between similar classes of diseases. The proposed study presents a strong methodology that combines high-quality preprocessing, feature extraction based on the wavelet, and Vision Transformer (ViT) models that are optimized with the help of metaheuristic algorithms. Preprocessing involves resizing, filtering noise and enhancing contrast. Discrete Wavelet Transform (DWT) and Stationary Wavelet Transform (SWT) are used to extract spatial-frequency features. The mean, median, entropy and variance are statistical measures that are calculated using the transformed images. They are then classified with the help of three variants of ViT, which are ViT-RSA (Reptile Search Algorithm), ViT-BES (Bald Eagle Search), and ViT-HBA (Honey Badger Algorithm). ViT-HBA statistical output parameters are then input to machine learning classifiers such as Bayesian Optimized Twin Neural Network (BO-TNN) and Bayesian Optimized Gaussian Process Regression (BO-GPR). As experimental results indicate, ViT-HBA has the highest classification accuracy of 98.90 as compared to the other variants of ViT. This reflects good generalization and strength in detection of diseases. The suggested hybrid architecture, consisting of ViT-HBA and BO-TNN, is much more effective in terms of accuracy and validity of sugarcane leaf disease classification. This end-to-end pipeline offers a highly useful resource in terms of real-time disease monitoring and assistance in the smart agricultural practice.