Detection of Rice Blast Disease Using DCNN and Vision Transformer Model with PSO-Lemon-OVT-NetX Architecture Based Effective Attention Fusion

Abstract

Rice blast disease is one among the major threats which directly affects the production of rice in global manner. In
order to maximize the yield several methods are developed in recent times but still it needs improvement in terms of
detection accuracy. For that purpose, in this article a hybrid model is developed with the combination of the models like Deep Convolutional Neural Networks (DCNN), Vision Transformer (ViT), and an optimized feature selection strategy using a PSO-Lemon-OVT-NetX architecture. Initially, image data are acquired from the tilleringstage samples using single push-broom hyperspectral camera. Followed by that data leakage is prevented using effective preprocessing which performed certain activities such as normalization, resizing and Principal Component Analysis (PCA). Then the DCNN model is employed to perform initial feature extraction and classification process to avoid the vanishing gradient problem. To enhance feature representation, an effective attention fusion is introduced which is the combination of multiscale detection heads and attention modules. The extracted features are further refined using the proposed PSO-Lemon-OVT-NetX architecture, which combines Particle Swarm Optimization (PSO) with an improved optimization strategy. Subsequently, a Vision Transformer (ViT) model is utilized for final classification and as well the Modified Grey Wolf Optimization algorithm is integrated to fine-tune the model parameters. Experimentation is performed in Python and the parameters which are calculated for the performance analysis are loss, accuracy, F1-Score, precision and recall. Overall, this proposed model provides an effective solution to increase the crop yield by attaining maximum accuracy in the early detection of rice disease.