Using "bundles" as the grading unit for tobacco leaves is a key strategy to improve the efficiency of tobacco pro-curement. Due to the occlusion and curling characteristics among bundled tobacco leaves, existing mainstream clas-sification methods struggle to accurately extract their fine-grained grading features. To address this, this paper pro-poses a two-stage bundled tobacco grading model based on strongly supervised data augmentation, achieving pre-cise grading of bundled tobacco leaves in a progressive manner. Firstly, a dual-attention residual network is de-signed to adaptively fuse multi-dimensional features for extracting coarse-grained information, and a soft channel attention module is proposed to generate attention maps reflecting key parts of the bundled tobacco leaves, thereby achieving coarse grading. Then, to encourage the network to focus on discriminative fine-grained features, strongly supervised data augmentation is performed on the global images guided by the coarse grading attention maps, ob-taining local images with distinctive features, thus achieving more refined grading results. The proposed method is compared with current mainstream general classification methods and fine-grained classification methods on a bun-dled tobacco leaf dataset. Experimental results show that the grading accuracy and macro-F1 score of the proposed method both reach 98.54%, significantly outperforming the comparison methods, and can better meet the practical needs of industrial bundled tobacco grading.