Due largely to the presence of the pathogenic fungus Magnaporthe oryzae, rice blast disease (RBD) poses a serious danger to Karnataka, India’s main rice crop. Rapid and effective diagnosis and control of RBD are required to protect food production and ensure the viability of agriculture in this area. The identification and classification of rice blast disease in rice fields may be accomplished by using convolutional neural networks (CNNs). This method can hasten diagnosis and allow for prompt actions to reduce crop losses. In the current study, a novel technique for anticipating and controlling RBD in Karnataka is introduced. The proposed model evaluates the regional distribution of RBD using information gathered over a two-year period across 120 sample locations spanning various rice ecosystems in Karnataka, India. This required rigorous data processing, which included managing samples and normalizing data. The creation of the Hybrid Pelican Komodo Optimization (HPKO), a cutting-edge metaheuristic feature selection technique created expressly for tackling the problems presented by rice blast disease, is a significant leap in this study. The Pelican Optimization Algorithm (POA) in addition Komodo Mlipir Algorithm (KMA), two extensively used metaheuristic methods, work together to create HPKO, a revolutionary strategy that adds a number of improvements. These improvements include managing the local problem space by changing the agent’s current position according to the dimensions of that problem area, stochastic selection of four possible movements within a single phase and changing the random target of the initial phase to a predetermined one. The paper also used the VTHR (Vision Transformer with Hybrid High-Resolution Network) technique for rice image categorization and feature selection. This method combines an attention mechanism with the Vision Transformer (ViT) and Deep High-Resolution Network (HRNet). The VTHR network makes use of the Vision Transformer’s self-attention mechanism to recognize global relationships in pictures, improving the precision of feature classification. Notably, the attention algorithm gives more weight to crucial rice properties, increasing the accuracy of picture capture. With an amazing accuracy rate of 99%, the model successfully categorizes multi-scale data while maintaining minute details and increasing semantic information. This remarkable outcome demonstrates the VTHR classifier and HPKO’s complementary efficacy. The use of cutting-edge technology and ground-breaking algorithms provides a feasible plan for controlling rice blast disease and guaranteeing food sustainability in Karnataka, India.