An Improved Ningxia Desert Herbaceous Plant Classification Algorithm Based on YOLOv8.

Area of Science:

• Ecology
• Computer Vision
• Artificial Intelligence

Background:

• Desert grasslands present unique challenges for plant species detection due to diverse habitats and uneven distribution.
• Existing plant recognition models are computationally expensive and imprecise for these environments.

Purpose of the Study:

• To develop a lightweight and fast plant species detection system for complex desert grassland environments.
• To improve the precision and efficiency of plant recognition in challenging ecological settings.

Main Methods:

• Introduced a dynamic convolutional KernelWarehouse method for kernel dimensionality reduction and increased kernel count.
• Incorporated triplet attention into the feature extraction network to capture channel and spatial relationships.
• Developed a dynamic detection head to address target detection head and attention non-uniformity.

Main Results:

• The YOLOv8s-KDT model demonstrated a 50.8% reduction in FLOPs (floating-point operations per second), a 4.5% increase in accuracy, and a 5.6% increase in mAP (mean Average Precision).
• The system achieved rapid and effective identification of desert grassland plants.

Conclusions:

• The YOLOv8s-KDT system is suitable for deployment in mobile applications and ecological observation platforms for desert grassland research.
• Facilitates large-scale vegetation distribution investigation and long-term ecological information tracking in regions like Ningxia.