ED-Swin Transformer: A Cassava Disease Classification Model Integrated with UAV Images.

Area of Science:

• Agricultural Science
• Computer Vision
• Remote Sensing

Background:

• Cassava diseases threaten tropical agriculture and food security.
• Manual monitoring is inefficient and lacks spatial coverage.
• Drone imagery offers high resolution but faces challenges with complex backgrounds and irregular disease morphology.

Purpose of the Study:

• To develop an intelligent classification method for cassava diseases using drone imagery.
• To address limitations of existing methods in handling complex backgrounds and irregular disease shapes.
• To enhance the accuracy and efficiency of cassava disease identification.

Main Methods:

• Proposed an ED-Swin Transformer model integrating EMAGE and DASPP modules.
• EMAGE module uses multi-scale grouped attention for feature extraction, mitigating background noise.
• DASPP module employs deformable atrous convolution to adapt to irregular disease boundaries.

Main Results:

• The ED-Swin Transformer model achieved high performance across five evaluation metrics.
• Demonstrated significant improvements in accuracy compared to existing methods.
• Achieved scores of 94.32%, 94.56%, 98.56%, 89.22%, and 96.52%.

Conclusions:

• The ED-Swin Transformer model shows superior performance for cassava disease classification using drone imagery.
• The proposed method effectively handles complex backgrounds and irregular disease morphology.
• This technology can significantly improve agricultural monitoring and disease management.