A Novel LiDAR Data Classification Algorithm Combined CapsNet with ResNet.

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Machine Learning

Background:

• LiDAR data is crucial for land classification, providing detailed ground target information like height and shape.
• Convolutional Neural Networks (CNNs) excel at feature extraction but struggle with spatial relationships.
• Capsule Networks (CapsNets) effectively capture spatial variations and feature directionality.

Purpose of the Study:

• To develop an advanced deep learning model for improved LiDAR-based land classification.
• To address the limitations of CNNs in adequately resolving spatial feature relationships in LiDAR data.

Main Methods:

• A novel deep network, ResCapNet, was designed by integrating Capsule Networks (CapsNet) with Residual Networks (ResNet).
• CapsNet's vector representation captures feature direction and relative positions for detailed information extraction.
• ResNet preserves information integrity, mitigating degradation issues common in traditional CNNs.

Main Results:

• Comparative experiments were conducted using two distinct LiDAR datasets against several classic machine learning algorithms.
• The proposed ResCapNet model demonstrated significant improvements in LiDAR classification performance.
• The integration of CapsNet and ResNet effectively enhanced the extraction of detailed spatial features from LiDAR data.

Conclusions:

• ResCapNet offers a superior approach for land classification using LiDAR data compared to existing methods.
• The model's ability to capture complex spatial relationships is key to its enhanced performance.
• This research contributes a more accurate and robust deep learning framework for geospatial analysis.