Christopher Grant, Hyperion Vector Systems; Geoffrey Lake, Icebox Engineering
https://doi.org/10.64861/MAPH7677
Keywords: infrared detection, training dataset, WISE, machine learning, space domain awareness, thermal emission, labeled data, dim-object detection, multi-domain, NEO Surveyor, coast-phase tracking
Abstract:
The space domain awareness community develops detection algorithms almost exclusively on visible-band electro-optical data, synthetic or ground-based. This constrains detection to solar-illuminated objects and limits characterization to reflected-light photometry. The gap is practical: objects in Earth shadow are invisible to visible sensors. IR detection through thermal self-emission operates independently of illumination geometry (the physics isn’t new), but the SSA community lacks ML-ready training data at these wavelengths. NASA’s infrared science archives contain over two decades of space object observations that astronomy pipelines reject as contamination. Nobody has converted these into labeled training data.
Fitzmaurice et al. (2021) developed the WISEstreakDET algorithm to detect and correlate GEO satellite streaks across all four WISE bands without a priori position knowledge. Lee et al. (2017) extracted over 860 labeled observations of 245 box-wing and 18 cylindrical GEO satellites with multi-band IR photometry. These efforts produced detection catalogs and photometric classifications, not ML training datasets (no image patches, no bounding boxes, no negative samples). We build on this foundation.
This paper presents a framework and initial results for constructing the first ML-formatted training datasets for space object detection from WISE/NEOWISE single-exposure observations. The labeling pipeline cross-references observation metadata (timestamp, pointing, FOV, exposure duration) with TLE-propagated catalog positions to tag RSO signatures in real IR imagery from 3.4 to 22 µm.
We use single-exposure images rather than coadded Atlas products. The WISE processing pipeline treats satellite streaks as contamination and rejects them during coaddition. Single-exposure frames (8.8 seconds for W1/W2, 11 seconds for W3/W4) preserve the original streak and point-source signatures that coaddition removes. The pipeline propagates catalog TLE elements to each exposure’s epoch, projects predicted RSO positions through the detector’s astrometric solution, and correlates predictions with detected sources or streak features.
The WISE/NEOWISE archive is large. Across cryogenic (2010) and post-cryogenic NEOWISE-Reactivation (2013-2024) phases, the mission accumulated over 21 sky coverages in an archive exceeding 8 petabytes, hosted at IRSA and mirrored on AWS S3. W1 (3.4 µm) and W2 (4.6 µm) operated throughout; W3 (12 µm) and W4 (22 µm) collected data during the cryogenic phase only. At these wavelengths, RSOs produce detectable thermal emission regardless of solar illumination. Visible sensors cannot do this.
We validate on a targeted sample of single-exposure frames selected along ecliptic latitudes where geosynchronous objects concentrate. Results include detection statistics, false positive characterization from asteroid contamination in the ecliptic plane, and a preliminary completeness assessment relative to the public catalog. We also extract noise floor measurements, background spatial structure, and PSF behavior from the WISE infrared detector arrays. To our knowledge, this characterization data has not previously been published in the SSA context. Synthetic generators like SatSim do not currently model it.
The resulting dataset fills a specific gap: real labeled infrared imagery formatted for ML training pipelines. Synthetic generators produce high-fidelity simulated imagery, but the synthetic-to-real transfer gap for IR phenomenology is unvalidated by operational data. Dark current shot noise and thermal background dominate IR noise budgets rather than the read noise that characterizes visible sensors. Real IR training data is the empirical ground truth needed to calibrate domain adaptation and confirm that synthetic noise models match physical reality.
Beyond WISE, NASA archives from SOFIA/FORCAST (5-40 µm, 2010-2022) and Spitzer/IRAC (3.6-8.0 µm, 2003-2020) contain complementary observations that the labeling pipeline can process with modest adaptation. We haven’t applied the pipeline to these archives yet, but the potential is clear: SOFIA’s 2.5-meter aperture has angular resolution roughly 6× better than WISE, while Spitzer has superior sensitivity at 3.6-8.0 µm. Systematic exploitation of these archives as ML training sources for SDA is future work.
We have initiated collaboration discussions with EO Solutions to extend prior SatSim-based synthetic imagery work. Pairing real WISE data with SatSim’s synthetic generation closes the validation loop: train on synthetic, validate on real, feed noise discrepancies back into the simulator. A companion short course submission describes a hands-on workshop where participants apply detection algorithms against paired visible and IR validation sets derived from this work.
The paper concludes with a roadmap for scaling to the full NEOWISE-Reactivation archive, estimated processing resource requirements, and recommendations for mixed-domain ML training pipelines that handle the distinct noise regimes of visible and infrared imagery. The IR noise characterization and dim-object detection performance bounds established here apply directly to coast-phase midcourse tracking, where thermally dim objects against space background present the same detection physics. With Golden Dome and space-based interceptor architectures now in open program development, we intend to evaluate this applicability as a near-term follow-on topic. The framework also produces validated IR noise models and detection algorithms directly applicable to NEO Surveyor (NASA’s $1.2B infrared space telescope approaching launch), where WISE wavelength overlap at 3.4-4.6 µm means algorithms trained on this archival data are ready for initial NEO Surveyor observations on day one. RSO contamination cataloging in these archives also benefits astronomy data quality across current and future IR missions. The framework, noise characterization products, and labeled datasets serve SSA persistent custody, midcourse IR tracking, planetary defense, and NASA archival data quality. Four domains, same detection physics, zero shared training infrastructure.
Date of Conference: September 15-18, 2026
Track: AI for SDA Applications