Mitigation of atmosphere noise in satellite interferometry through machine learning

DongChan Ahn, & Zhen Liu

Published August 8, 2019, SCEC Contribution #9385, 2019 SCEC Annual Meeting Poster #216

Interferometric synthetic aperture radar (InSAR) is a radar imaging technique widely used to track surface deformation. InSAR can measure surface deformation with centimeter to millimeter accuracy, but is susceptible to noise. One major noise source is radar phase delay caused by variations in atmospheric conditions over the region of interest. Convolutional neural networks have shown great performance in spatial pattern recognition tasks. Our objective is to use convolutional filters to aid in identifying spatially-correlated noise in InSAR interferograms. Using a deep learning-based approach, our method employs deep convolutional autoencoders to estimate and correct atmospheric noise. As a proof of concept, we trained and validated autoencoders on synthetic test cases with realistic noise. Preliminary results indicate this novel approach can denoise interferograms effectively. Ongoing work involves benchmarking the robustness of the autoencoder through more synthetic test cases and applying it to real-world InSAR observations.

Key Words
InSAR, Noise, Machine Learning, Deep Learning

Ahn, D., & Liu, Z. (2019, 08). Mitigation of atmosphere noise in satellite interferometry through machine learning. Poster Presentation at 2019 SCEC Annual Meeting.

Related Projects & Working Groups
Tectonic Geodesy