Residual Noise Learning for Atmospheric Correction of InSAR Unwrapped Maps

Atmospheric artifacts can obscure tectonic deformation in interferometric synthetic aperture radar (InSAR) observations, while conventional correction methods often depend on external atmospheric data. This study proposes a supervised residual-learning framework that directly predicts noise components in unwrapped InSAR maps rather than reconstructing deformation signals. Physically informed synthetic datasets were generated by combining Okada and Mogi deformation models with topography-correlated tropospheric delays, spatially correlated turbulent noise, and long-wavelength ramps. A network-depth sensitivity analysis identified a 20-layer denoising convolutional neural network as the optimal balance between accuracy and model complexity. Tests on independent synthetic datasets showed that the model reliably distinguished deformation from noise when the signal-to-noise ratio exceeded approximately 10−1, whereas performance degraded under extremely noise-dominated conditions. The framework was further evaluated using ALOS/PALSAR and ALOS-2/PALSAR-2 observations of post-eruptive deformation at Mt. Ontake, Japan, and coseismic deformation associated with the 2009 L’Aquila earthquake, Italy. Compared with uncorrected, GACOS-corrected, and linear-corrected results, the CNN correction reduced topography-correlated and long-wavelength artifacts, improved temporal consistency, and generally achieved closer agreement with GNSS observations. These results suggest that residual noise learning provides an efficient approach for automatic atmospheric correction of unwrapped InSAR observations, with the potential for transferability across different deformation-source settings.