Deep-broad learning network model for precision identification and diagnosis of grape leaf diseases

Area of Science:

• Agricultural Technology
• Computer Science
• Plant Pathology

Background:

• Grape leaf disease identification is crucial for crop yield and quality.
• Existing methods struggle with rapid, precise diagnosis, especially for subtle or small disease spots.
• The need for efficient and accurate automated disease detection systems in agriculture is growing.

Purpose of the Study:

• To develop an advanced model for rapid, precise, and efficient identification and diagnosis of grape leaf diseases.
• To enhance the accuracy and efficiency of disease detection by integrating deep learning and Broad Learning techniques.
• To improve the recognition of small and irregular disease features on grape leaves.

Main Methods:

• Proposed the Deep-Broad Learning Network Model (ABLSS), integrating Broad Learning (BLS) with deep learning.
• Optimized the model using the Adam algorithm and incorporated the LTM mechanism for enhanced learning.
• Integrated a SENet attention mechanism and U-Net segmentation with dilated spatial pyramid pooling and feature pyramid networks.

Main Results:

• The ABLSS model achieved higher recognition accuracy for grape leaf diseases compared to BLS and standard deep learning networks (7.69% and 4.48% improvement, respectively).
• The segmentation component achieved a Mean Intersection over Union (MIOU) of 86.61% and a Mean Pixel Accuracy (MPA) of 90.23%, outperforming the original U-Net model.
• The ABLSS model demonstrated a significant speed improvement of 72.12% (0.375 seconds faster recognition) over deep learning networks.

Conclusions:

• The ABLSS model effectively combines the high accuracy of deep learning with the processing speed of Broad Learning.
• The proposed method significantly enhances the recognition of complex and small disease features, overcoming limitations of existing models.
• This research provides a valuable tool for developing smart orchard technologies and optimizing agricultural learning network models.