MOMFNet: A Deep Learning Approach for InSAR Phase Filtering Based on Multi-Objective Multi-Kernel Feature Extraction.

Area of Science:

• Geosciences
• Remote Sensing
• Artificial Intelligence

Background:

• Interferometric Synthetic Aperture Radar (InSAR) is crucial for Earth observation, measuring ground deformation.
• Phase noise in InSAR interferograms critically degrades data quality and accuracy.
• Existing denoising methods struggle with complex noise patterns and preserving subtle deformation signals.

Purpose of the Study:

• To introduce MOMFNet, a deep learning model for advanced InSAR phase filtering.
• To improve the quality of InSAR interferograms by effectively suppressing phase noise.
• To enhance the accuracy of ground deformation measurements derived from InSAR data.

Main Methods:

• Developed MOMFNet, a deep learning network employing multi-objective multi-kernel feature extraction.
• Utilized a multi-objective loss function considering spatial and statistical properties of denoised results.
• Incorporated weighted residual blocks for adaptive feature importance and a novel interferogram simulation strategy for training.

Main Results:

• MOMFNet demonstrated superior noise suppression and phase recovery compared to traditional and other deep learning methods.
• The model excelled in challenging scenarios with large gradients and random noise.
• Empirical validation using Sentinel-1 data from the Yanzhou coal mine confirmed significant improvements in interferogram quality.

Conclusions:

• MOMFNet effectively removes noise while preserving critical phase details in InSAR data.
• The proposed deep learning approach offers a significant advancement in InSAR phase denoising.
• This research highlights the potential of deep learning for enhancing remote sensing applications.