AUTHOR=Zang Hecang , Wang Congsheng , Zhao Qing , Zhang Jie , Wang Junmei , Zheng Guoqing , Li Guoqiang 

TITLE=Constructing segmentation method for wheat powdery mildew using deep learning

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 16 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2025.1524283

DOI=10.3389/fpls.2025.1524283

ISSN=1664-462X

ABSTRACT=Powdery mildew is an important factor affecting wheat yield and global food security as well as a leading factor restricting the sustainable development of agriculture. Timely and accurate segmentation of wheat powdery mildew image is an important practical significance for disease-resistant breeding and precise control. In this study, RSE-Swin Unet was proposed based on the Swin-Unet architecture to address the complex morphology of wheat powdery mildew lesions, blurred boundaries between lesions and non-lesions, and low segmentation accuracy. The method combines ResNet and SENet to solve the abovementioned problem. Firstly, the attention mechanism module SENet is introduced into Swin-Unet, which can effectively capture global and local features in images and extract more important information about powdery mildew. Secondly, the output of the SENet module add to the corresponding feature tensor of the decoder for subsequent decoder operations. Finally, in the deep bottleneck of Swin-Unet network, ResNet network layers are used to increase the expressive power of feature. The test results showed that in the experiment with the self-built wheat powdery mildew dataset, the proposed RSE-Swin Unet method achieved MIoU, mPA, and accuracy values of 84.01%, 89.96%, and 94.20%, respectively, which were 2.77%, 3.64%, and 2.89% higher than the original Swin-Unet method. In the wheat stripe rust dataset, the proposed RSE-Swin Unet method achieved MIOU, MPA, and accuracy values of 84.91%, 90.50%, and 96.88%, respectively, which were 4.64%, 5.38%, and 2.84% higher than those of the original Swin-Unet method. Compared with other mainstream deep learning methods U-Net, PSPNet, DeepLabV3+, and Swin-Unet, the proposed RSE Swin-Unet method can detect wheat powdery mildew and stripe rust image in a challenging situation and has good computer vision processing and performance evaluation effects. The proposed method can accurately detect the image of wheat powdery mildew and has good segmentation performance, which provides important support for the identification of resistance in wheat breeding materials.