Segmentation and Multi-Scale Convolutional Neural Network-Based Classification of Airborne Laser Scanner Data.

Area of Science:

• Geospatial science
• Computer vision
• Remote sensing

Background:

• Point cloud classification is crucial for airborne laser scanning (ALS) data processing.
• Complex scenes and irregular point distributions pose significant challenges to traditional methods.
• Existing point-based classification methods can be computationally intensive and lack accuracy.

Purpose of the Study:

• To develop a novel classification method for ALS point clouds that enhances accuracy and reduces computational burden.
• To address the challenges of complex scenes and irregular point distributions in point cloud data.
• To improve the efficiency and effectiveness of point cloud classification.

Main Methods:

• A three-step region-growing segmentation method was employed to minimize under- and over-segmentation.
• A feature image generation technique transformed 3D neighborhood features into 2D images.
• A multi-scale convolutional neural network (CNN) was utilized for training and testing with the generated feature images.

Main Results:

• The proposed method achieved 84.9% overall accuracy on the ISPRS WG II/4 benchmark dataset.
• An average F1 score of 69.2% was obtained, demonstrating competitive performance.
• The framework showed satisfactory performance compared to existing approaches analyzed.

Conclusions:

• The segmentation and multi-scale CNN-based classification method offers a robust solution for ALS point cloud processing.
• This approach effectively handles complex scenes and irregular point distributions.
• The method provides a significant improvement in classification accuracy and computational efficiency.