Elevating Paddy Crop Health Monitoring: A Vision Transformer Approach with Comparative Analysis

Rice serves as a fundamental dietary staple for a substantial portion of the global populace, constituting a primary sustenance source for a majority of individuals worldwide. Management practice using smart agriculture techniques to tackle factors affecting rice yield is a necessary step in the present scenario to negate any possibility of world hunger. To maximize the yield of rice, crop health monitoring is crucial. Computer vision plays a key role in non-invasive crop health monitoring by analyzing drone-based agricultural crop image data. Vision transformer (ViT) is the current popular vision-based deep learning technique to solve image-related classification problems. In this study, we employ two base variant ViT models b16 and b32 and one ResNet-based hybrid ViT model to classify healthy and unhealthy rice crops, utilizing input images of size of 224 × 224 and batch size of 64. The models demonstrate accuracies ranging from 97 to 100%. Significantly, the hybrid ViT model emerges as the most proficient. Notably, the ViT b32 model exhibits comparatively lesser performance in this context. The findings from the models depict a promising future of vision transformer in agricultural image classification tasks, especially in plant growth detection.