FHB-Net: a severity level evaluation model for wheat Fusarium head blight based on image-level annotated aerial RGB...

Area of Science:

• Agricultural Science
• Plant Pathology
• Computer Vision

Background:

• Fusarium head blight (FHB) significantly threatens global wheat production, causing up to 50% yield loss.
• Resistant wheat varieties are crucial for disease management, necessitating accurate FHB severity evaluation.
• Current methods struggle with fine-grained feature learning, leading to unreliable predictions for early-stage infections.

Purpose of the Study:

• To develop an end-to-end deep learning model (FHBNet) for accurate FHB severity assessment using RGB aerial images.
• To improve upon existing methods that lack fine-grained feature learning capabilities for distinguishing wheat infection levels.
• To facilitate high-throughput and non-destructive phenotype analysis for accelerating disease resistance breeding.

Main Methods:

• Proposed FHBNet model utilizing multi-scale criss-cross attention (MSCCA) for global contextual relationships and spatial context modeling.
• Incorporated bi-level routing attention (BRA) module to accurately locate small FHB lesions by suppressing irrelevant information.
• Trained and evaluated the model on a dataset of 6035 RGB aerial images with light, moderate, and severe FHB labels.

Main Results:

• FHBNet achieved 79.49% accuracy on the test set, outperforming several mainstream neural networks (MobileViT, MobileNet, EfficientNet, RepLkNet, ViT, ConvNeXt).
• Visualization heatmaps confirmed FHBNet's ability to precisely locate FHB lesions under varying severity and illumination conditions.
• Demonstrated the feasibility of rapid, non-destructive FHB severity evaluation using only image-level annotated aerial RGB images.

Conclusions:

• FHBNet offers a robust and accurate solution for evaluating Fusarium head blight severity in wheat.
• The model's performance validates the potential of deep learning with aerial imagery for agricultural disease monitoring.
• This approach can significantly accelerate wheat breeding programs by providing efficient and precise phenotype analysis.