FCA_Unet: An Enhanced UNet with Feature Coordination Attention for Wheat Leaf Disease Severity Assessment

Abstract

Accurate monitoring of wheat leaf disease severity, critical for crop protection and precision agriculture, has attracted extensive attention in the context of smart agriculture. Existing methods face notable limitations: manual assessment is inefficient and subjective; classification-based approaches lack generalisability; segmentation models struggle with poor small-lesion detection, low boundary accuracy, and high parameter counts. To tackle these issues, we propose FCA_Unet, an enhanced UNet framework. The model consists of two key components: (1) the integration of the designed A_CBAM attention mechanism module and multi-scale feature fusion (MFFE) module. The attention module, based on deformable convolution, enhances focus on lesion-specific attributes while suppressing irrelevant background information. In contrast, the MFFE enables cross-layer fusion, integrating high-level semantic features for disease classification and low-level contextual features for boundary optimisation. (2) An optimised backbone network: The Inception multi-branch structure, coordinated attention mechanism (INCA), and MFFE are integrated into ResNet50 to reduce parameter redundancy while preserving feature extraction efficiency, thus meeting the lightweight deployment requirements of agricultural field equipment. Experimental results demonstrate that FCA_Unet achieves 89.85% mIoU on wheat stripe rust and powdery mildew, representing an 11.8% improvement over UNet, with a parameter count of 29.95M (68.2% of UNet's). Combined with disease severity indices, the model proves its superiority in both segmentation accuracy and severity assessment, providing robust support for automated wheat disease monitoring in practical agricultural scenarios.