A tree species classification model based on improved YOLOv7 for shelterbelts

Area of Science:

• Forestry Science
• Remote Sensing Technology
• Artificial Intelligence in Ecology

Background:

• Shelterbelt management requires accurate tree species classification, which is challenging with current satellite and drone-based methods due to complex backgrounds and similar tree crown sizes.
• Existing methods struggle to differentiate individual tree species in mixed-growth protective forests, impacting management strategies.
• The You Only Look Once (YOLO) algorithm shows promise in forestry applications, including tree species identification.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for precise tree species classification in shelterbelts using drone RGB imagery.
• To enhance the YOLOv7 architecture for improved feature extraction and fusion in complex forest environments.
• To provide a robust tool for scientific theoretical basis for shelterbelt management, particularly in Northwest China.

Main Methods:

• A specialized dataset of shelterbelt tree species was created, incorporating data augmentation to address limited training data.
• The K-means++ algorithm was utilized for optimal anchor box clustering.
• The YOLOv7 backbone network was modified with Coordinate Convolution (CoordConv) in the ELAN module and the Convolutional Block Attention Module (CBAM) integrated into the PANet for enhanced feature representation.

Main Results:

• The proposed YOLOv7-Kmeans++_CoordConv_CBAM (YOLOv7-KCC) model achieved a mean average precision@0.5 of 98.91%.
• YOLOv7-KCC significantly outperformed established models including Faster RCNN (VGG16, Resnet50), SSD, YOLOv4, and the base YOLOv7 model.
• The model demonstrated a 5.6% increase in F1 metrics compared to YOLOv7, with computational costs of 105.07 GFlops and 143.7MB parameters.

Conclusions:

• The YOLOv7-KCC model offers a highly effective solution for classifying tree species within shelterbelts from drone imagery.
• This advancement provides a crucial scientific foundation for optimizing shelterbelt management practices, especially in arid and semi-arid regions.
• The integration of CoordConv and CBAM modules enhances the model's capability to handle complex visual data in forestry applications.