Building Extraction Based on an Optimized Stacked Sparse Autoencoder of Structure and Training Samples Using LIDAR...

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Artificial Intelligence

Background:

• Building extraction is crucial for urban planning and construction.
• Existing methods face challenges like resolution limits, data correction issues, and terrain influences.
• Multi-sensor data fusion, including LiDAR and optical imagery, is key to improving extraction accuracy.

Purpose of the Study:

• To propose an improved building extraction method using a stacked sparse autoencoder (SSAE).
• To optimize SSAE network structure and training sample selection for enhanced performance.
• To accurately extract buildings from fused Digital Surface Model (DSM) and optical image data.

Main Methods:

• Utilized a stacked sparse autoencoder (SSAE) neural network for in-depth feature learning.
• Combined Light Detection and Ranging (LiDAR) derived Digital Surface Models (DSM) with optical imagery as input.
• Developed optimized strategies for SSAE network architecture and training sample selection.

Main Results:

• The optimized SSAE method demonstrated high accuracy in building extraction.
• The approach showed good robustness in handling complex urban environments.
• Fusion of DSM and optical data with SSAE effectively overcame limitations of traditional methods.

Conclusions:

• The proposed SSAE-based method offers a robust and accurate solution for building extraction.
• Optimized network structure and training samples significantly improve SSAE performance.
• This technique advances automated building extraction for urban construction and planning.