Dual-Coupled CNN-GCN-Based Classification for Hyperspectral and LiDAR Data.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Deep learning, particularly Convolutional Neural Networks (CNNs) and Graph Convolutional Networks (GCNs), has advanced remote sensing image classification.
• CNNs excel at short-range spatial-spectral feature extraction from hyperspectral (HS) images.
• GCNs are adept at modeling middle- and long-range spatial relationships and structural features.

Purpose of the Study:

• To propose an effective classification method for combined HS and Light Detection and Ranging (LiDAR) data.
• To leverage the complementary spatial-spectral information from HS data and elevation information from LiDAR data.
• To enhance the accuracy of remote sensing image classification in complex environments.

Main Methods:

• A dual-coupled CNN-GCN structure is proposed, comprising a coupled CNN and a coupled GCN.
• The coupled CNN uses a weight-sharing mechanism for fusing and simplifying dual CNN models to extract spatial features from HS and LiDAR data.
• The coupled GCN concatenates HS and LiDAR data, constructs a uniform graph, and fuses structural information via shared graph structures and weights.

Main Results:

• The proposed dual-coupled CNN-GCN method demonstrates superior performance compared to state-of-the-art methods like two-branch CNN and context CNN.
• Extensive experiments on two real-world HS and LiDAR datasets validate the method's effectiveness.
• The model achieved an outstanding overall accuracy of 99.11% on the Trento dataset, setting a new performance record.

Conclusions:

• The dual-coupled CNN-GCN approach effectively integrates HS and LiDAR data for high-accuracy remote sensing image classification.
• The synergistic use of CNNs and GCNs within a coupled framework enhances feature extraction from multi-source data.
• The proposed method offers a significant advancement in pixel-level classification for complex remote sensing scenes.