Multimodal Change Detection in Remote Sensing Images Using an Unsupervised Pixel Pairwise Based Markov Random Field...

Area of Science:

• Remote Sensing
• Computer Vision
• Statistical Modeling

Background:

• Multimodal change detection in remote sensing is challenging due to varying image modalities.
• Existing methods often struggle with heterogeneity across different imaging types.

Purpose of the Study:

• To develop a robust Bayesian statistical approach for unsupervised multimodal change detection.
• To introduce a novel Markovian model that is invariant to imaging modality.

Main Methods:

• Formulation of the multimodal change detection problem within an unsupervised Markovian framework.
• Development of a pixel pairwise modeling for a robust observation field.
• Utilizing an iterative estimation technique for Markovian mixture model parameters.
• Employing a stochastic optimization process for Maximum a posteriori (MAP) solution computation.

Main Results:

• The proposed Markovian model demonstrates robustness across heterogeneous satellite image pairs.
• The pixel pairwise modeling provides a quasi-invariant visual cue, effective across multiple imaging modalities.
• Experimental results confirm the approach's effectiveness with mixed imaging modalities.

Conclusions:

• The Bayesian statistical approach offers a robust solution for multimodal change detection.
• The novel Markovian model effectively handles variations in imaging modalities for improved change detection.
• The method provides a reliable tool for analyzing bitemporal heterogeneous satellite imagery.