SARCDNet-an enhanced deep learning network for change detection from bi-temporal SAR images

Area of Science:

• Remote Sensing
• Geoscience
• Artificial Intelligence

Background:

• Change detection analysis is vital for monitoring Earth's surface using microwave remote sensing.
• Bi-temporal Synthetic Aperture Radar (SAR) imaging offers all-weather, day-and-night capabilities, outperforming optical methods challenged by cloud cover and daylight.
• Advancements in SAR data from satellites like Sentinel-1 have propelled change detection research.

Purpose of the Study:

• To introduce SARCDNet (SAR Change Detection Network), an enhanced deep learning model for change detection in bi-temporal SAR images.
• To improve the accuracy and efficiency of change detection, particularly in environments prone to speckle noise.

Main Methods:

• Developed SARCDNet, a deep learning network incorporating an adaptive fusion block.
• The adaptive fusion block integrates adaptive global filtering in the frequency domain and a channel attention mechanism.
• The model leverages both spatial and frequency domain information to extract and enhance relevant features.

Main Results:

• SARCDNet demonstrated improved performance on public datasets (Yellow River, Farmland) known for high speckle noise.
• The model achieved significant performance gains in flood change detection on the Chao Lake dataset, including a 4.26% increase in F1 score.
• SARCDNet effectively mitigates speckle noise, enhancing prediction accuracy and computational efficiency.

Conclusions:

• SARCDNet represents a significant advancement in deep learning for SAR image change detection.
• The model's adaptive fusion block enhances feature relevance and robustness against noise.
• SARCDNet shows strong potential for applications in environmental monitoring, disaster response, and urban planning.