Spatio-Temporal Refinement for InSAR Phase Unwrapping Using Untrained Neural Networks

Synthetic aperture radar interferometry (InSAR) enables high-precision ground deformation detection by measuring phase differences between SAR images. However, the challenging phase unwrapping is required to resolve inherent ambiguities. Disturbances like spatio-temporal decorrelation, atmospheric artifacts and multipath interference introduce noise, leading to unwrapping errors that affect ground motion estimation accuracy. In order to correct InSAR phase unwrapping errors made by traditional phase unwrapping methods, we propose utilizing an untrained deep neural network combined with early stopping during the learning process to more effectively capture the spatio-temporal priors of InSAR time series. Our model is untrained, meaning that it is learned only on the noisy wrapped InSAR data to be processed, without requiring any additional training data. Experimental results show that our method achieves competitive results compared to the European Ground Motion Service (EGMS) product and the reference methods.