Graph Attention Feature Fusion Network for ALS Point Cloud Classification.

Area of Science:

• Geospatial data processing
• Computer vision
• Machine learning for remote sensing

Background:

• Airborne laser scanning (ALS) point cloud classification is crucial but challenging due to scene complexity and irregular data distribution.
• Existing deep learning methods often suffer from complex preprocessing, high sampling costs, and limited receptive fields.

Purpose of the Study:

• To propose a Graph Attention Feature Fusion Network (GAFFNet) for improved ALS point cloud classification.
• To enhance the capture of wider contextual information and increase the receptive field for each point.

Main Methods:

• Developed a neighborhood feature fusion unit and an extended neighborhood feature fusion block utilizing a graph attention mechanism.
• Designed an encoder-decoder neural network architecture to extract multi-level semantic features from point clouds.
• Evaluated the method on a public ISPRS ALS point cloud dataset.

Main Results:

• GAFFNet effectively classifies nine types of ground objects in ALS point clouds.
• The proposed method demonstrates superior performance across various evaluation metrics compared to existing approaches.
• The graph attention mechanism successfully expands the receptive field, capturing wider contextual information.

Conclusions:

• GAFFNet offers a robust and efficient solution for airborne laser scanning point cloud classification.
• The method achieves high accuracy by effectively leveraging contextual information through graph attention and an encoder-decoder architecture.