Diffusion Models for Interferometric Satellite Aperture Radar

However, their performance relative to non-natural images, like radar-based satellite data, remains largely unknown. Generating large amounts of synthetic (and especially labelled) satellite data is crucial to implement deep-learning approaches for the processing and analysis of (interferometric) satellite aperture radar data. Here, we leverage PDMs to generate several radarbased satellite image datasets. We show that PDMs succeed in generating images with complex and realistic structures, but that sampling time remains an issue. Indeed, accelerated sampling strategies, which work well on simple image datasets like MNIST, fail on our radar datasets. Probabilistic Diffusion Models (PDMs) are a recent family of deep generative models which have demonstrated state-of-the-art performance in image translation [e.g., SWB21] and generation [e.g., DN21, MFNK In addition, PDMs have the major advantage of being very versatile. They are less prompt to various failures often encountered with other generative approaches, such as mode collapse during the training of GANs or posterior collapse for VAEs [LTGN19]. Consequently, this considerably reduces the engineering work required to train generative models, and paves the way for fully automated data analysis pipelines in remote sensing.