Deep Ordinal Classification in Forest Areas Using Light Detection and Ranging Point Clouds.

Area of Science:

• Geospatial Science
• Computer Science
• Forestry

Background:

• Deep Learning and aerial Light Detection And Ranging (LiDAR) advance 3D point cloud analysis for environmental monitoring.
• Accurate classification of forest structures is crucial for ecological studies and management.

Purpose of the Study:

• To develop an ordinal classification model for LiDAR point clouds in forest areas.
• To classify point clouds into four distinct classes: ground, low, medium, and high vegetation.
• To enhance the PointNet network architecture using a novel soft labeling technique.

Main Methods:

• Application of a novel generalized exponential function (CE-GE) for soft labeling.
• Utilizing the PointNet deep learning architecture.
• Statistical validation using Kolmogorov-Smirnov and Student's t-test.

Main Results:

• The CE-GE method demonstrated superior performance across all evaluation metrics compared to existing methodologies.
• Ordinal classification using smoothed labels resulted in more consistent results than nominal classification.
• Significant improvement in point-by-point classification accuracy, particularly in differentiating between low and medium vegetation.

Conclusions:

• The proposed CE-GE soft labeling technique effectively refines the classification of forest LiDAR point clouds.
• The methodology enhances the accuracy of deep learning models like PointNet for complex environmental monitoring.
• This approach offers a more robust solution for distinguishing vegetation strata in 3D point cloud data.