Commonality Feature Representation Learning for Unsupervised Multimodal Change Detection.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Multimodal change detection (MCD) faces challenges due to the direct incomparability of multimodal bitemporal images (MBIs).
• Existing methods struggle to effectively align and compare features across different modalities for accurate change identification.

Purpose of the Study:

• To propose a novel commonality feature representation learning (CFRL) framework for unsupervised MCD.
• To develop a CFRL-based framework that enables direct comparison of features from MBIs for robust change detection.

Main Methods:

• A Siamese-based encoder and two decoders are utilized for mapping MBIs into a shared feature space.
• Modality alignment is achieved by reconstructing pseudo-MBIs through decoder swapping.
• Latent commonality features are extracted by minimizing feature distance, enabling comparable representations.

Main Results:

• The CFRL framework successfully extracts comparable commonality features from MBIs.
• Two change magnitude images (CMIs) are generated simultaneously, facilitating binary change map creation.
• Extensive experiments on six datasets demonstrate superior performance over state-of-the-art approaches.

Conclusions:

• The proposed CFRL-based unsupervised MCD framework offers a robust solution for identifying changes in multimodal bitemporal images.
• The CFRL method effectively addresses the challenge of feature comparability across different modalities.
• The framework achieves state-of-the-art performance, highlighting its potential for practical MCD applications.