Location-guided lesions representation learning via image generation for assessing plant leaf diseases severity

Area of Science:

• Plant Pathology
• Computer Vision
• Precision Agriculture

Background:

• Accurate plant leaf disease severity assessment is vital for effective pesticide application and crop yield optimization.
• Existing methods struggle with background interference, misidentifying non-lesion areas and reducing accuracy.
• Precision agriculture demands robust disease detection systems that can overcome complex visual challenges.

Purpose of the Study:

• To propose a location-guided lesion representation learning (LLRL) method for accurate plant leaf disease severity assessment.
• To overcome the challenge of background interference in automated disease detection.
• To enhance the precision of pesticide application through improved disease severity analysis.

Main Methods:

• Developed a three-part approach: image generation network (IG-Net), location-guided lesion representation learning network (LGR-Net), and hierarchical lesion fusion assessment network (HLFA-Net).
• Utilized a diffusion model within IG-Net to generate paired healthy and diseased leaf images.
• Employed LGR-Net to focus on lesion areas by contrasting paired images, creating a dual-branch feature encoder (DBF-Enc).
• Integrated HLFA-Net to fuse and optimize features from DBF-Enc for precise severity classification.

Main Results:

• The proposed LLRL method demonstrated superior performance compared to existing classification models.
• Achieved at least a 1% improvement in accuracy for plant leaf disease severity assessment.
• Experimental validation was conducted on a dataset of 12,098 images covering apple, potato, and tomato plant diseases.

Conclusions:

• The LLRL method effectively addresses background interference, leading to more accurate plant leaf disease severity assessment.
• The approach shows significant potential for enhancing precision agriculture practices.
• The developed method offers a robust solution for automated plant disease detection and management.