Locality Preservation for Unsupervised Multimodal Change Detection in Remote Sensing Imagery.

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Multimodal change detection (MCD) is challenging due to differing imaging mechanisms.
• Direct comparison of multimodal images for change detection is often infeasible.
• Existing methods frequently rely on intermediate difference images, adding complexity.

Purpose of the Study:

• To develop a novel, robust framework for multimodal change detection.
• To establish modality-invariant links between image superpixels for change analysis.
• To formulate MCD within a mathematical framework that directly outputs change maps.

Main Methods:

• Exploration of the topological structure of multimodal images.
• Construction of modality-invariant links between pairwise superpixel class relationships and change labels.
• Formulation of the locality-preserving energy model (LPEM) to maintain local consistency constraints.
• Direct generation of change maps (CM) without intermediate difference images (DI).

Main Results:

• The proposed locality-preserving energy model (LPEM) demonstrates robustness across various MCD scenarios.
• LPEM effectively utilizes modality-invariant links based on topological structure.
• Experimental results on diverse real datasets confirm the method's effectiveness.

Conclusions:

• The developed LPEM offers a robust and direct approach to multimodal change detection.
• The modality-invariant links and mathematical framework provide a universal solution.
• The direct generation of change maps simplifies the MCD process and enhances efficiency.