Individual Tree Crown Segmentation and Crown Width Extraction From a Heightmap Derived From Aerial Laser Scanning...

Area of Science:

• Forestry
• Remote Sensing
• Computer Vision

Background:

• Accurate individual tree crown (ITC) information is crucial for forest management.
• Existing methods using aerial imagery face challenges with illumination and complex canopy structures.
• Deep learning shows promise for analyzing remote sensing data.

Purpose of the Study:

• To develop and validate a novel deep learning method for ITC segmentation using LiDAR-derived heightmaps.
• To overcome limitations of traditional methods relying on aerial photographs.
• To enhance the accuracy and efficiency of forest resource assessment.

Main Methods:

• Utilized the YOLO-v4 deep learning network with LiDAR-generated heightmaps for ITC segmentation.
• Incorporated a computer graphics algorithm to refine segmentation of overlapping crowns.
• Employed generative adversarial networks (WGAN, CycleGAN, SinGAN) to create synthetic training data.
• Validated the approach on diverse forest plots: tree nursery, forest landscape, and mixed plantation.

Main Results:

• Achieved an overall recall of 83.6% and precision of 81.4% for ITC detection across all plot types.
• Obtained a coefficient of determination (R²) ≥ 79.93% for tree crown width estimation.
• Demonstrated accuracy unaffected by key parameter values, outperforming the watershed method by 3.9%.

Conclusions:

• The proposed deep learning method effectively enhances tree crown segmentation using LiDAR heightmaps.
• This approach provides a robust alternative to visual-based methods, avoiding complexities of aerial imagery.
• The findings support improved forest assessment and smart management through advanced remote sensing techniques.