MaCon: A Generic Self-Supervised Framework for Unsupervised Multimodal Change Detection.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Multimodal change detection (MCD) is crucial for Earth observation and emergency response.
• Unsupervised MCD methods are needed due to limited labeled data.
• Existing methods struggle with precise feature extraction from diverse data modalities.

Purpose of the Study:

• To develop an unsupervised framework for effective multimodal change detection.
• To synergistically distill common and discrepancy representations from multimodal data.
• To improve the accuracy and robustness of change detection algorithms.

Main Methods:

• Proposed the MaCon framework unifying mask reconstruction (MR) and contrastive learning (CL).
• Implemented an optimal sampling strategy within the CL architecture for enhanced discrepancy representation.
• Introduced a silent attention mechanism to improve output contrast and training stability.

Main Results:

• MaCon effectively distills intrinsic common representations between different data modalities.
• Achieved state-of-the-art performance on both multimodal and monomodal change detection datasets.
• Demonstrated the framework's robustness and effectiveness in complex scenarios.

Conclusions:

• The MaCon framework provides a powerful unsupervised solution for multimodal change detection.
• It shows potential as a unified framework for various change detection applications.
• The approach enhances feature extraction and discrimination for improved detection accuracy.