Ingo Waldmann, Spaceflux; Marco Rocchetto, Spaceflux; Marcel Debczynski, Spaceflux
Keywords: diffusion models, deep learning, generative models, inpainting, background modelling, data fusion
Abstract:
The field of space situational awareness (SSA) has seen significant growth in recent years, driven by the increasing number of objects in orbit around the Earth and the need to track and monitor their movements. Accurate detection and tracking of these objects is critical for the safety and security of space operations. However, the detection of faint objects is challenging due to the presence of noise in the images, which can make it difficult to distinguish between real signals and background noise.
In this paper we will discuss the use of generative deep learning models to learn the local detector characteristics, astronomical scene and weather patterns to produce adaptive background removal and optimally suppress noise in SSA images. Our approach is based on recent advances in deep learning and in particular the use of generative models to de-noise existing images. Architectures such as Noise2Self (Batson et al. 2019), Generative Inpainting (e.g. Yu et al. 2018) and denoising diffusion networks (Ho et al 2020) find broad applicability in the fields of image processing and image restoration.
The generative network is learned on the focal plane images derived from our Spaceflux optical sensor arrays and fine-tuned to the individual detector characteristics. A significant advantage of denoising diffusion generative models is their ability to accurately model non-linear relations within images. This characteristic can be exploited to represent astronomical backgrounds, time variable flat fielding (e.g. dawn/dusk gradients) and the effects of seeing and partial cloud coverage in one uniform model. We demonstrate the ability of Denoising Diffusion Neural Network (DDNN) to mimic real observations at a photorealistic level. These simulations can then be used to, for example, de-trend SSA observations with high-background levels.
Improvements in the quality and reliability of SSA observations, ultimately contribute to better space situational awareness and space traffic management. Furthermore, optimal noise suppression in SSA images allows for the detection of smaller and fainter targets, a more robust detection during sub-optimal weather conditions and more accurate absolute photometry. Improvements in photometric calibration lead to a better ability to characterise the RSOs morphologies and time-dependent behaviour through the study of their reflected light. Optimal photometry is also particularly important in low signal-to-noise conditions, such as the detection of small RSOs in high-altitude orbits or eccentric cis-lunar orbits.
Overall, the use of DDNN for optimal noise suppression in SSA images has the potential to significantly enhance the capabilities of space situational awareness and improve our understanding of the space environment. The proposed method can be further developed and integrated into existing SSA systems, leading to more accurate and reliable detection and tracking of space objects, and ultimately improving the safety and security of space operations.
Date of Conference: September 19-22, 2023
Track: Machine Learning for SDA Applications