 This paper proposes a novel attention-based convolutional neural network SU-Net for winter wheat identification in multiband GF-2 images. The SU-Net network consists of three parts, a shallow feature extractor, a deep feature extractor, and a jump phase. The shallow feature extractor uses the batch normalization layer to accelerate network convergence and improve accuracy. The deep feature extractor uses the upsampling method to obtain deeper features. Finally, the jump phase uses shuffled attention to optimize the features and then combines them with the deep features obtained by upsampling. The SU-Net network outperforms the UNET network in terms of mean intersection over union, overall classification accuracy, recall, F1 score, and CAPA coefficient. This article was authored by Kujo, Jinyan Zhang, Lulu, and others.