Parameter-Driven Simulation of Synthetic InSAR Data Using Deep Learning

Synthetic simulation of Interferometric Synthetic Aperture Radar (InSAR) data plays a critical role in algorithm development, mission design, and machine learning pretraining. However, most existing approaches rely on handcrafted simulation pipelines or rigid data generation setups that lack adaptability to varying sensor configurations. In this work, we propose a novel deep learning-based framework for generating realistic, complex-valued InSAR data that is explicitly conditioned on acquisition parameters, such as wavelength, baseline, incidence angle, and revisit time. The proposed multitask model jointly synthesizes scene reflectivity, coherence magnitude, and interferometric phase while enabling the simulation of temporal decorrelation and noise characteristics via both analytical and data-driven modules. We further introduce a reparameterizable noise model for speckle and phase noise, whose statistics are tuned using Particle Swarm Optimization to match real sensor characteristics. Extensive experiments on the high-resolution GeoNRW TDX dataset demonstrate the realism and flexibility of our simulation framework, which can generalize to unseen acquisition scenarios and support SAR mission design.