A Deep Convolutional Coupling Network for Change Detection Based on Heterogeneous Optical and Radar Images.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Change detection traditionally relies on homogeneous images.
• Growing interest in heterogeneous image change detection due to complementary sensor data (optical and radar).
• Existing methods often require supervised learning or struggle with data from different sensors.

Purpose of the Study:

• To develop an unsupervised deep learning network for change detection using heterogeneous optical and radar images.
• To create a method that aligns feature representations from different sensor types.
• To enable effective change detection without labeled data.

Main Methods:

• Proposed a symmetric deep convolutional coupling network.
• Input images (optical and radar) are processed through convolutional and coupling layers.
• Features are transformed into a common space for difference calculation and thresholding.

Main Results:

• The unsupervised network learns by optimizing a coupling function.
• Achieved promising performance on both homogeneous and heterogeneous image datasets.
• Demonstrated superior results compared to several existing change detection methods.

Conclusions:

• The proposed unsupervised network is effective for change detection with heterogeneous images.
• The method offers a novel approach to leveraging complementary sensor data.
• This unsupervised technique advances the field of remote sensing change detection.