Code-Aligned Autoencoders for Unsupervised Change Detection in Multimodal Remote Sensing Images.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Convolutional autoencoders are used for multimodal change detection in bitemporal satellite images.
• Aligning code spaces and reducing change pixel influence are key challenges.
• Supervised change area information is often unavailable for training.

Purpose of the Study:

• To develop an unsupervised method for change detection using relational pixel information.
• To enforce alignment of code spaces by reducing the impact of change pixels.
• To improve the accuracy of change detection in bitemporal satellite imagery.

Main Methods:

• Extracting relational pixel information using domain-specific affinity matrices.
• Deriving an unsupervised change prior from comparable pixel pair affinities.
• Enforcing code space alignment by correlating pixels with similar affinity relations across domains.
• Utilizing cycle consistency in the image translation process.

Main Results:

• The proposed method effectively aligns code spaces and reduces the impact of change pixels.
• Unsupervised change prior derivation proved effective.
• Demonstrated improved performance compared to state-of-the-art algorithms on four real datasets.

Conclusions:

• The methodology offers an effective unsupervised approach for multimodal change detection.
• Relational pixel information and enforced code space alignment are crucial for accurate change detection.
• The approach shows significant potential for applications in bitemporal satellite image analysis.