Mamba-YOLO-ML: A State-Space Model-Based Approach for Mulberry Leaf Disease Detection

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Mulberry cultivation faces significant yield and quality losses due to pest and disease infestations.
• Traditional detection methods are inefficient and unsustainable, necessitating advanced solutions.
• Existing computer vision models struggle with small targets, occlusions, and feature details in natural environments.

Purpose of the Study:

• To develop an optimized deep learning model, Mamba-YOLO-ML, for accurate and efficient detection of mulberry pests and diseases.
• To address limitations of current models, including low recognition rates, computational inefficiency, and poor adaptability.
• To enhance the identification of structural features like leaf veins for improved disease diagnosis.

Main Methods:

• Proposed Mamba-YOLO-ML model incorporating Phase-Modular Design (PMSS) with dual blocks and Mamba Block for enhanced feature representation.
• Utilized Haar wavelet downsampling to preserve critical texture details and Normalized Wasserstein Distance loss for small-target detection.
• Employed GradCAM for visualization analysis to assess model focus on disease characteristic regions.

Main Results:

• Mamba-YOLO-ML achieved state-of-the-art detection accuracy with 78.2% mAP50 and 59.9% mAP50:95.
• The model demonstrated superior performance compared to YOLO variants and Transformer-based models.
• Visualization confirmed earlier and more effective focus on disease-specific regions by the enhanced model.

Conclusions:

• Mamba-YOLO-ML offers a significant advancement in vision-based pest and disease detection for mulberry.
• The model's lightweight architecture enables real-time deployment on embedded devices for precision agriculture.
• This provides a sustainable solution for efficient crop management and improved mulberry cultivation.