Research on urban tree classification method based on YOLO-CNGD.

Area of Science:

• Remote Sensing
• Urban Ecology
• Computer Vision

Background:

• Accurate urban tree species classification is crucial for effective urban green space management and ecological assessments.
• Detecting small and overlapping tree crowns in high-resolution remote sensing data presents significant challenges.

Purpose of the Study:

• To develop a novel framework, YOLO-CNGD, for enhanced urban tree species classification.
• To address limitations in detecting small and overlapping tree crowns in remote sensing imagery.

Main Methods:

• The proposed YOLO-CNGD framework integrates the Convolutional Block Attention Module (CBAM) for improved feature representation.
• It utilizes Normalized Wasserstein Distance (NWD) loss for robust small-object localization and Deformable Convolution v3 (DCNv3) for adaptability to irregular shapes.
• Standard convolutions are replaced with GhostConv for a lightweight and efficient model design.

Main Results:

• YOLO-CNGD achieved a precision of 94.8%, a recall of 91.1%, and an mAP@0.5 of 93.7% on a self-built urban tree dataset.
• The model demonstrated a balance between high accuracy and computational efficiency.
• Experimental results indicate significant improvements in small and overlapping object detection.

Conclusions:

• YOLO-CNGD offers a promising solution for automated urban tree inventory and management.
• The framework's performance highlights its potential for large-scale ecological assessments using remote sensing data.
• The integration of attention mechanisms, specialized loss functions, and efficient convolutions enhances object detection capabilities.