Application of ECV-UNet Model in Semantic Segmentation of Wheat Stripe Rust

Accurate semantic segmentation of wheat stripe rust is crucial for achieving mechanized disease control and precise pesticide application. Addressing challenges such as uneven lighting, complex backgrounds, and diverse lesion morphologies in field images, this study proposes the ECV-UNet model, which integrates attention mechanisms and Transformer. By incorporating ECA-Skip connections, C-Max Pooling, and ViT Blocks, the model enhances channel-spatial feature modeling and global semantic capture capabilities, while a dedicated dataset is constructed. Experiments demonstrate that the model's segmentation accuracy (Dice) and mean Intersection over Union (mIoU) improve by 4.77% and 7.79%, respectively, compared to the original UNet, effectively enhancing lesion detection precision and providing technical support for precise disease management.