A high-performance detection model ISA-YOLO for eggplant pests and diseases

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Pathology

Background:

• Eggplant (Solanum melongena) is a vital cash crop facing significant challenges in field pest and disease detection.
• Existing methods struggle with occluded foliage, blurred lesion boundaries, and scale variations, particularly for fruit rot due to surface properties and gradual infection spread.
• Accurate detection is crucial for timely intervention and crop yield preservation.

Purpose of the Study:

• To introduce a novel principle for eggplant disease detection by jointly modeling contextual cues, lesion boundaries, and scale variations.
• To develop and evaluate an improved detection model, ISA-YOLO, based on this principle.
• To enhance the accuracy and efficiency of automated pest and disease identification in eggplant cultivation.

Main Methods:

• Proposed a context-boundary-scale coupling principle for integrated analysis of disease symptoms.
• Developed ISA-YOLO, an enhanced YOLOv13 model incorporating coordinated context modeling, boundary-aware aggregation, and progressive multi-scale fusion.
• Conducted experiments on two public datasets to validate the model's performance.

Main Results:

• ISA-YOLO achieved high mean Average Precision (mAP) of 78.1% and 77.7% at approximately 30 FPS.
• The model demonstrated superior accuracy and speed trade-offs compared to mainstream detectors.
• Post-optimization (pruning and quantization), inference speed reached ~75 FPS with maintained accuracy.

Conclusions:

• The context-boundary-scale coupling principle offers an effective framework for accurate eggplant disease detection.
• ISA-YOLO presents a deployable solution for smart agriculture, enabling efficient pest and disease identification.
• The findings pave the way for advanced automated crop monitoring systems.