Statistical Difference Representation-Based Transformer for Heterogeneous Change Detection.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Heterogeneous change detection compares images from different sensors over time.
• Deep learning and domain adaptation methods are current mainstream approaches.
• Lack of credible labels hinders practical application of existing methods.

Purpose of the Study:

• To develop a weakly supervised framework for heterogeneous change detection.
• To overcome the limitation of insufficient labeled data.
• To improve the accuracy and robustness of change detection in heterogeneous imagery.

Main Methods:

• Proposed a structure similarity-guided sample generating (S3G2) strategy for reliable pseudo-label generation.
• Introduced a Statistical Difference representation Transformer (SDFormer) to mitigate modality differences.
• Employed differential structure similarity for prior information acquisition.

Main Results:

• The proposed S3G2 strategy iteratively generates reliable pseudo-labels.
• SDFormer effectively reduces the influence of modality differences in bitemporal heterogeneous imagery.
• Experimental results demonstrate competitive performance against state-of-the-art methods.

Conclusions:

• The developed weakly supervised framework addresses the challenge of limited labeled data in heterogeneous change detection.
• The proposed methods show significant potential for real-world applications requiring change detection from diverse image sources.
• Further investigation into parameter influences confirmed the robustness of the approach.