Constructing segmentation method for wheat powdery mildew using deep learning.

Area of Science:

• Agricultural Science
• Computer Vision
• Deep Learning

Background:

• Powdery mildew significantly impacts wheat yield and global food security, necessitating accurate disease detection for sustainable agriculture.
• Effective segmentation of wheat powdery mildew images is crucial for disease-resistant breeding and precise agricultural control strategies.

Purpose of the Study:

• To develop an advanced deep learning model for accurate segmentation of wheat powdery mildew and stripe rust images.
• To address challenges in wheat disease image analysis, including complex lesion morphology and blurred boundaries, using an improved Swin-Unet architecture.

Main Methods:

• Proposed RSE-Swin Unet, integrating ResNet and SENet modules into the Swin-Unet architecture.
• Utilized SENet (Squeeze-and-Excitation Network) for enhanced feature extraction and attention.
• Incorporated ResNet (Residual Network) layers in the bottleneck to improve feature representation.

Main Results:

• RSE-Swin Unet achieved superior segmentation performance on a self-built wheat powdery mildew dataset, outperforming original Swin-Unet by up to 3.64% in mPA.
• On a wheat stripe rust dataset, RSE-Swin Unet also demonstrated significant improvements, with MIoU, mPA, and accuracy higher than Swin-Unet by up to 5.38%.
• The proposed method showed competitive and improved results compared to other mainstream deep learning models like U-Net, PSPNet, and DeepLabV3+.

Conclusions:

• The RSE-Swin Unet model provides accurate and robust segmentation for wheat powdery mildew and stripe rust images, even in challenging conditions.
• This method offers significant support for identifying resistance in wheat breeding materials and advancing precise disease management in agriculture.
• The enhanced computer vision capabilities contribute to improved performance evaluation and disease detection in crop monitoring.