Denoising Method for SHARAD Radar Signal Based on Diffusion Probabilistic Model

SHARAD on the MRO emits high-frequency signals that can penetrate the Martian surface and detect layered structures in the subsurface, but the radargram data is often contaminated by background noise. Due to the complex noise generation mechanism, it is difficult to denoise it by traditional filters or numerical methods which analyze the properties of the noise. Besides, common deep learning methods can not solve this problem either because SHARAD lacks groundtruth. In this paper, we propose a deep learning denoising algorithm based on the idea of Denoising Diffusion Probabilistic Model (DDPM). Our model implicitly learn the distribution of the radargram noise by using a Markov chain process, then estimate the noise statistical properties and denoise the radargram in a process called sampling. We have conducted numerical experiments on both synthetic data and real SHARAD radargram. Both the signal-to-noise ratio on synthetic data and the information entropy on the SHARAD radar-gram data are obviously improved, which proves that our model can eliminate the noise in the background of the radargram effectively. in the future, this model is expected to be applied to denoising tasks of other kind of GPR (Ground pene trating radar) data.