Using bundles as the grading unit for tobacco leaves is a key strategy to improve efficiency of tobacco procurement.However,due to occlusion and curling characteristics within the bundles,existing mainstream classification methods struggle to accurately extract their fine-grained features for grading.To address this,we propose a two-stage grading model for bundled tobacco leaves based on strongly supervised data augmentation.Our approach achieves precise grading in a progressive manner.First,a dual-attention residual network is designed to adaptively fuse multi-dimensional features for coarse-grained information extraction,and a soft channel attention module is proposed to generate attention maps highlighting key regions of the bundled leaves,thereby enabling coarse grading.Then,guided by these attention maps,we perform strongly supervised data augmentation on the global images to crop local images with distinctive features.This step facilitates a more refined grading by encouraging the network to focus on discriminative fine-grained features.The proposed method is compared with current mainstream general and fine-grained classification approaches on a bundled tobacco leaf dataset.Experimental results show that the proposed method achieves a grading accuracy and macro-F1 score of 98.54%,significantly outperforming the compared methods,and better meets the practical needs of industrial bundled tobacco leaf grading.