TropoDeep: a deep learning-based model for InSAR tropospheric correction on large-scale interferograms using GNSS and WRF outputs

Abstract This study introduces TropoDeep, an advanced model based on Super-Resolution Generative Adversarial Networks (SRGAN), developed to downscale the outputs of the Weather Research and Forecasting (WRF) model in order to improve displacement measurements from Interferometric Synthetic Aperture Radar (InSAR) in California and Nevada. TropoDeep improves differential Slant Tropospheric Delay (dSTD) resolution by leveraging High-Resolution Sentinel-1 InSAR interferograms (IFGs) and Low-Resolution (LR) WRF dSTD data, while reducing the temporal mismatch between atmospheric model outputs and SAR acquisition times. It employs SRGAN’s generator and discriminator framework, supported by Visual Geometry Group-19 (VGG19) model to ensure high perceptual quality. Performance evaluation shows that TropoDeep significantly enhances InSAR data quality, achieving a root-mean-square error (RMSE) improvement of up to approximately 40%, with an average improvement of 21% compared to the Global Atmospheric Correction Online Service (GACOS). Validation of time-series displacement fields demonstrated that InSAR results corrected with TropoDeep align more closely with Global Navigation Satellite Systems (GNSS) measurements compared to those corrected with GACOS. RMSE improvements for InSAR time-series data corrected with TropoDeep ranged from approximately 10% to 29% at nearly 84% of the GNSS stations. In addition, applying the proposed model to the subsidence signal in California shows that TropoDeep can reduce the intruder tropospheric effect in subsidence time series by up to 66% compared to GACOS, illustrating TropoDeep’s enhanced capability in refining tropospheric corrections.