Diffusion Probabilistic Models for Compressive SAR Imaging

Odysseas Pappas, Perla Mayo, Andrew Austin, Alin Achim · 2025

Read Diffusion Probabilistic Models for Compressive SAR Imaging by Odysseas Pappas, Perla Mayo, Andrew Austin, Alin Achim (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-17053.txt.

Introduction

Odysseas Pappas and co-authors test in 2025 diffusion models for forming radar images from heavily reduced data. Synthetic aperture radar imaging from compressed samples is usually solved by slow iterative methods. Their approach guides a denoising diffusion model with a rough first reconstruction made from the sparse data by standard means.

They ran the method on real radar data. The diffusion reconstructions were compared against standard methods that used the full data set. Image quality gains suggested the generative approach can offset missing samples. Faster, lighter imaging matters for platforms that cannot downlink everything. The study is an early proof of concept.

Catalog record

CREATOROdysseas Pappas, Perla Mayo, Andrew Austin, Alin Achim
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE18 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-17053

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Diffusion Probabilistic Models for Compressive SAR Imaging Odysseas Pappas, Perla Mayo, Andrew Austin, Alin Achim Abstract Compressed sensing Synthetic Aperture Radar (SAR) image formation, formulated as an inverse problem and solved with traditional iterative optimization methods can be very computationally expensive. We investigate the use of denoising diffusion probabilistic models for compressive SAR image reconstruction, where the diffusion model is guided by a poor initial reconstruction from sub-sampled data obtained via standard imaging methods. We present results on real SAR data and compare our compressively sampled diffusion model reconstruction with standard image reconstruction methods utilizing the full data set, demonstrating the potential performance gains in imaging quality. a]School of Computer Science, University of Bristol, Bristol, UK b]School of Electrical, Electronic and Mechanical Engineering, University of Bristol, Bristol, UK keywords Synthetic Aperture Radar, Compressive Sensing, Denoising Diffusion Probabilistic Models † † authorinfo: Corresponding author: Odysseas Pappas, [email protected] 1 INTRODUCTION Synthetic Aperture Radar (SAR) is one of the most commonly employed remote sensing modalities of the modern era, being utilized in a great variety of monitoring tasks due to its ability to produce highly detailed images of the Earth regardless of the day-night cycle and weather conditions. SAR is a coherent imaging radar system (typically airborne/spaceborne) that transmits frequency-modulated signals to a scene on the ground and records the reflected radar echoes while in flight, achieving impressive spatial resolution due to its long synthetic aperture. As with all coherent radar systems, the collected reflected radar echoes, often referred to as phase history data, must be processed to form an image suitable for visual interpretation. Several classical methods exist for the formation of SAR images from phase history data [ 1 ] , each offering distinct advantages and disadvantages in terms of speed, computational cost and image quality. The sheer volume of the generated phase history data, especially for modern high-resolution SAR systems, imposes severe costs in terms of data storage and transfer as well as processing time. There is therefore significant interest in the formation of SAR images from subsampled phase history data, with compressed sensing (CS) methods for SAR imaging gathering a lot of research interest. CS [ 2 , 3 ] is a mathematical framework that allows for the reconstruction of signals from a parsimonious set of measurements (weighted linear combinations of samples) far smaller than what the Nyquist theorem would define as the minimum total of samples required. Applications of CS in SAR [ 4 ] allow for the reconstruction of a SAR image from a sub-sampled set of phase history data, introducing obvious advantages in terms of data acquisition speed and storage requirements while also allowing for the treatment of erroneously incomplete data. Algorithms for CS SAR image formation often employ an inverse problem formulation combined with sparse regularization methods [ 5 , 6 , 7 ] . The design of a suitable signal model and measurement matrix for various SAR imaging modes can however be a non-trivial task, and this limits the practicability of such methods. Additionally, due to the iterative nature of the solvers employed in such inverse problem formulation, they can be extremely computationally intensive, especially for high-resolution imagery. In this paper we examine an alternative approach taking inspiration from CS SAR but utilizing generative image models such as Denoising Diffusion Probabilistic Models (DDPMs) instead of classical CS reconstruction theory. DDPMs rely on the modelling of a denoising process to generate images from noise samples, with the capability of conditioning the generated image on a class label or on an input signal. This conditioning input signal can be a low resolution copy or a poor quality reconstruction of the desired image, with the DDPM effectively achieving super-resolution or image enhancement respectively. We explore the use of a CS measurement matrix via which a sub-sampled phase history can be obtained, with this sub-sampled phase history in turn formed into an image using classical imaging methods. While the resulting image is expectedly of very poor quality, we demonstrate how it can serve as conditioning input to a DDPM which in turn can produce an image of comparable quality to what would have been obtained from classical methods operating on the fully sampled phase history. The paper is organized as follows: Section 2 provides some additional information on CS and discusses the design of an appropriate measurement matrix. Section 3 discusses DDPMs and their use in our proposed method. Section 4 presents experimental results of our CS SAR DDPM imaging method while concluding remarks are provided in Section 5. 2 MEASUREMENT MATRICES FOR CS SAR IMAGING CS [ 2 , 3 ] is a novel sampling framework that allows for the reconstruction of sparse signals from a subsampled set of measurements. CS theory posits that the signal of interest x is sparsely representable in some domain, i.e. there exists a K -sparse signal s that satisfies x=\Psi s,x\in\Re^{N} (1) for some sparsifying transform \Psi . CS then provides us with a framework for the reconstruction of x from a set of measurements y obtained via y=Cx=C\Psi s,y\in\Re^{M} (2) where C is the CS measurement matrix and M<<N . Typical CS reconstruction methods rely on an inverse problem formulation whereby a minimization problem \hat{s}=\arg\min||s||_{1} (3) is solved such that C\Psi s=y . Figure 1: Subsampling patterns for a SAR phase history matrix containing M x N samples. (a) Regular subsampling (at 50%) along the azimuth. (b) Regular subsampling across azimuth with random subsampling across range. (c) Random subsampling across azimuth. (d) Aperture gap subsampling. Central to any CS method is the design of a suitable measurement matrix C , as it must satisfy two properties, namely the restricted isometry property (RIP) and incoherence (with respect to \Psi ). Random C matrices can satisfy these properties, however their use can be impractical in many applications and therefore the design of application-specific CS measurement matrices holds significant research interest. SAR-specific CS measurement matrices have been a topic of discussion in the literature [ 7 , 6 , 8 ] . A pattern easily realizable in practice is that of regular undersampling along the azimuth direction, as a simple modification to the A/D converter at the SAR groundstation can allow for sampling at lower rates along the azimuth, in effect discarding entire aperture positions (phase history matrix columns) at regular intervals. There is some debate as to whether regular subsampling across the azimuth is sufficient for CS, as it may not satisfy the restricted isometry property [ 6 ] . Introducing random dropping of samples along the range direction (on the aperture positions that are not outright discarded) can help alleviate such concerns, though it necessitates much more complex modifications to the SAR hardware [ 7 ] . Random sampling in azimuth alone is also feasible, as is sampling with continuous gaps in the total aperture [ 8 ] .Figure 1 demonstrates these subsampling patterns. For the experiments discussed in this paper, we have opted to follow the regular azimuth subsampling pattern shown in Fig. 1 (a) as it is easily realizable in practice. Note RIP adherence is not a major concern since we do not follow conventional CS reconstruction methods but rather rely on DDPM for reconstruction. Conventional imaging algorithms can

Read the complete text