Synthesis of a multispectral image dataset for ML-based space surveillance
Files
(Published version)
Date
2025
Authors
Kildare, J.
Knight, J.
Evans, M.
Law, Y.W.
Editors
Advisors
Journal Title
Journal ISSN
Volume Title
Type:
Conference item
Citation
76th International Astronautical Congress (IAC), 2025, pp.1-1
Statement of Responsibility
Conference Name
76th International Astronautical Congress (IAC) (29 Sep 2025 - 3 Oct 2025 : Sydney, Australia)
Abstract
Space surveillance involves detecting and tracking high-speed vehicles, such as re-entry vehicles. Due to the excessive speeds of atmospheric re-entry, objects entering the atmosphere from space are heated significantly and emit strongly in both the short-wavelength infrared (SWIR) and mid-wavelength infrared (MWIR) regimes. However, MWIR satellite imagery is significantly limited in the public domain, thus limiting capability for training machine learning (ML) algorithms to detect atmospheric entry in this band. Despite the limited MWIR imagery, there is a wealth of data from the Landsat program that encompasses both SWIR and long-wavelength infrared (LWIR) bands. The availability of this data provides some scope for interpolation between these measurements to generate alternative bands. This work proposes a diffusion model that incorporates existing spectral data to generate SWIR and MWIR band data from the surrounding spectral measurements. A hyperspectral data source — Earth Observation 1’s Hyperion instrument — is used to develop the training dataset, where the conditional input bands are selected to match those available from the Landsat instruments, and target bands are selected outside the Landsat band ranges. The methodology of an existing hyperspectral image generator, Hyper LDM, is adapted for small-scale inputs and rapid image synthesis, without the requirement of known spectral endmembers for radiance prediction. The developed model is capable of high-quality generation of images in the specified bands, typically with less noise artefacts than the ground truth bands. The decoder structure is further analysed using a class activation mapping method to improve interpretability of the model outputs. In particular, the conditional input is shown to provide the small-scale details of a scene during reconstruction, whereas the quantised latent code provides major spatial features.
School/Discipline
Dissertation Note
Provenance
Description
Access Status
Rights
Copyright 2025 International Astronautical Federation (IAF)