A Novel Deep-Learning-Based Approach for Estimating High-Resolution InSAR Parameters

Nowadays, SAR Interferometry (InSAR) represents one of the most powerful tools to perform complex tasks, such as retrieving Earth’s surface topography and monitoring its deformations. However, denoising strategies are crucial for the generation of reliable and high-quality InSAR products. In this framework, we propose a complete and generalized theoretical and statistical model to physically describe and simulate a noise-free interferogram, together with its corresponding noisy version, starting from the knowledge of the InSAR acquisition geometry and of the underlying topography. A deep-learning-based model is then trained in a fully-supervised manner to estimate high-resolution InSAR parameters, i.e. coherence and interferometric phase, leveraging the proposed statistical model. In particular, an accurate assessment of the network’s estimation performance is conducted by comparing it to state-of-the-art denoising filters, such as Boxcar and Phi-Net and showing the added value of the proposed theoretical modifications to the state-of-the-art literature.