GANInSAR: Deep Generative Modeling for Large-Scale InSAR Signal Simulation

Abstract

Interferometric synthetic aperture radar (InSAR) technology is widely used to create digital elevation models and measure dynamics on the Earth’s surface, including monitoring ground displacements. The lack of or limited-collected ground-truth data, however, often poses a bottleneck in validating the research outcome, particularly at high precision and resolution levels. To mitigate the gap, we introduce a new deep generative model (DGM) for the simulation of linear deformation rate maps. We demonstrate that our adversarial DGM architecture with carefully designed preprocessing and postprocessing modules performs well for InSAR deformation signal synthesis, even when limited data are available. We also introduce a dimensionality reduction method, based on the distance between the real-world and generated image feature vectors, to address the lack of quantitative evaluation for data simulation. Furthermore, we introduce a hybrid evaluation metric integrating quantitative and qualitative measures, which is more intuitive than the existing methods and makes it easier for domain experts to participate in the evaluation. We compare the results of our model with established methods. The comparison result illustrates the superior performance of our proposed method.

Keywords

• Deep generative modeling
• generative adversarial network (GAN)
• interferometric synthetic aperture radar (InSAR)
• radar signal processing
• remote sensing images
• simulation