Deep Learning–Based Mitigation of Atmospheric Noise in InSAR Unwrapped Phase Maps

Interferometric Synthetic Aperture Radar (InSAR) provides millimeter-scale measurements of line-of-sight (LOS) surface displacement, enabling detailed investigation of tectonic and volcanic processes [1]. However, interferograms are frequently contaminated by atmospheric delays and other noise sources whose amplitudes can be comparable to the deformation signal, especially over regions with complex topography [2]. Effective noise mitigation is therefore essential for extracting reliable geophysical information.

We developed a supervised deep-learning framework based on a modified Denoising Convolutional Neural Network (DnCNN) [3], with residual learning [4], designed to learn and remove atmospheric noise embedded in unwrapped interferograms automatically. The model was trained to estimate noise components directly and subtract them from the original interferograms, avoiding explicit physical modeling of atmospheric effects.

To evaluate performance, we applied the model to two ALOS-2 PALSAR-2 datasets: an ascending track (path/frame 126-710, 8 images) and a descending track (20-2890, 15 images) spanning 2014–2017. After baseline filtering (720 days, 150 m), 18 and 59 interferograms were generated. Linear correction [5], Generic Atmospheric Correction Online Service for InSAR (GACOS) [6], and the deep-learning (DL) method were applied, followed by conversion to 8-bit (uint8) format to standardize contrast for comparison. For ascending interferograms, the DL method produced the lowest mean standard deviation (SD = 11.59), outperforming GACOS (15.49), linear correction (16.78), and uncorrected results (15.89). Similar improvements were observed for descending interferograms (DL: 14.37; GACOS: 14.46; linear: 16.53; uncorrected: 16.45).

These results demonstrate that the proposed deep-learning approach effectively mitigates atmospheric noise in InSAR unwrapped phase maps and can outperform conventional correction methods.

References:

[1] Bürgmann, Roland, Paul A. Rosen, and Eric J. Fielding. "Synthetic aperture radar interferometry to measure Earth’s surface topography and its deformation." Annual review of earth and planetary sciences, 2000.

[2] Chaussard E, Wdowinski S, Cabral-Cano E, et al. Land subsidence in central Mexico detected by ALOS InSAR time-series. Remote sensing of environment, 2014.

[3] Zhang K, Zuo W, Chen Y, et al. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising. IEEE transactions on image processing, 2017.

[4] He, Kaiming, et al. "Deep residual learning for image recognition." Proceedings of the IEEE conference on computer vision and pattern recognition, 2016.

[5] Takada Y, Sagiya T, Nishimura T. Interseismic crustal deformation in and around the Atotsugawa fault system, central Japan, detected by InSAR and GNSS. Earth, planets and space, 2018.

[6] Yu C, Li Z, Penna N T. Interferometric synthetic aperture radar atmospheric correction using a GPS-based iterative tropospheric decomposition model. Remote Sensing of Environment, 2018.