CGLCS-Net: Addressing Multi-Temporal and Multi-Angle Challenges in Remote Sensing Change Detection.

Area of Science:

• Geosciences
• Computer Science
• Artificial Intelligence

Background:

• Deep learning models like CNN and Transformer advance remote sensing image change detection.
• Existing models struggle with multi-sensor data, varied viewing angles, and long time spans, limiting irregular change area analysis.

Purpose of the Study:

• To introduce the Context-Aware Global-Local Subspace Attention Change Detection Network (CGLCS-Net) for improved remote sensing change detection.
• To enhance the modeling of dynamic interactions and feature representations in complex change regions.

Main Methods:

• Developed the Global-Local Context-Aware Selector (GLCAS) for dynamic receptive field selection using joint pooling attention and depthwise separable convolution.
• Introduced the Subspace-based Self-Attention Fusion (SSAF) module for dynamic dual-temporal feature interaction via decomposition and self-attention.
• Enhanced global context and local feature extraction for multi-scale and irregular change regions.

Main Results:

• CGLCS-Net achieved significant IoU improvements over ChangeFormer on LEVIR-CD, SYSU-CD, and S2Looking datasets (0.95%, 9.23%, 13.16% respectively).
• Reduced model parameters by 70.05%, floating-point operations by 7.5%, and inference time by 11.5%.
• Demonstrated superior performance in handling sensor variations and long-term spectral inconsistencies.

Conclusions:

• CGLCS-Net effectively addresses limitations in current deep learning models for remote sensing change detection.
• The proposed GLCAS and SSAF modules enhance feature representation and dynamic interaction capabilities.
• The network's efficiency and accuracy make it suitable for continuous land use and land cover change monitoring.