Vision Transformer-Based Multi-Head Self-Attention for Early Recognition and Classification of Paddy Leaf Diseases in Rice Fields

Abstract

Rice has become an essential food source for a large portion of the world's population, greatly enhancing global food security. One of the fundamental staple crops, paddy, is especially susceptible to diseases primarily caused by bacteria and viruses. The source of the rice blast, Magnaporthe oryzae, poses a severe danger to the world's rice supply, mainly in South India. Both yield and quality are at risk due to the continuous threat of different diseases. However, a few diseases can drastically lower crop yields and quality, making agricultural productivity extremely vulnerable. Therefore, it is crucial to detect diseases at an early stage to effectively manage these risks. Scalable and effective solutions are required because conventional approaches are laborious, expensive, and frequently inaccessible to smallholder farmers. Data-driven strategies like machine learning (ML) and deep learning (DL), can assist in addressing these issues and increasing agricultural sustainability and crop yield. This study presents a new Vision Transformer-based hyperparameter optimization approach for the classification and detection of paddy leaf diseases in rice crops field (VTMHSA-RCPRF). The VTMHSA-RCPRF model comprises data preprocessing, ViT multi-head self-attention-based feature extraction, MLP-based Focal Loss for classification and detection, and Population-Based Training (PBT) as hyperparameter tuning. A wide range of experiments have been carried out to exhibit the promising performance of the VTMHSA-RCPRF method. The simulation outcomes highlighted that the VTH-RCPFRF approach reaches better performance over its recent approaches in terms of distinct measures.