LGMNet: A Hybrid Multi-Scale Architecture with Local-Global Integration for UAV-Based Wheat Fusarium Head Blight Detection

Fusarium Head Blight (FHB) poses a significant threat to global wheat security. Traditional expertise-dependent methods struggle to address wheat FHB detection in real-world production due to issues like accuracy, efficiency, and cost. Recent advancements in deep learning and Unmanned Aerial Vehicle (UAV)-based remote sensing have facilitated large-scale and accurate wheat disease detection. In light of these advances, we propose LGMNet, a semantic segmentation model designed for FHB detection using UAVs. LGMNet effectively combines Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs) to address challenges such as small affected regions, unclear boundaries, and complex background interference in UAV-scale images. Key innovations include the Enhanced Multi-scale Convolutional Attention (E-MSCA) module for multi-scale feature extraction and fusion, supporting precise feature representation, and the Global Capture and Local Preservation (GCLP) module that employs a novel self-attention mechanism for global context capturing to reduce interference from complex backgrounds, while combining a convolutional branch to maintain local finegrained features. Furthermore, the use of a semi-automatic annotation method significantly reduces the manual annotation workload for UAV images, constructing a wheat FHB dataset with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1, 8 0 7}$</tex> pixel-level annotated images. Notably, with only <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{5. 9 2}$</tex> million parameters, LGMNet achieves a mean Accuracy (mAcc) of 85.41 % and a mean Intersection over Union (mIoU) of 73.21 %, surpassing ten comparative models. This method provides an effective solution for large-scale wheat FHB detection in practical production environments.