A Method for Generating Large-Scale Spaceborne Long-Wave Infrared Point-Target Images Using Landsat 8 TIRS Data

Deep learning has demonstrated remarkable potential in spaceborne infrared small target detection; however, its application is severely bottlenecked by the scarcity of high-quality, annotated on-orbit data. To bridge this “Sim-to-Real” domain gap, this paper proposes a radiometrically consistent image generation framework capable of synthesizing large-scale, high-fidelity Long-Wave Infrared (LWIR) datasets. Unlike traditional methods that rely on simplistic image injection or computationally expensive ray-tracing, our approach adopts a hybrid strategy: it leverages the global archive of calibrated Landsat-8 data to ensure authentic background diversity and couples it with physically modeled targets derived from Computational Fluid Dynamics (CFD). A key innovation is the Radiance-Domain Fusion mechanism, which integrates targets and backgrounds strictly in physical units considering atmospheric transmission and sensor degradation effects (PSF, motion blur, NEDT), thereby guaranteeing energy conservation for sub-pixel targets. Quantitative validation against real on-orbit imagery from the SDG satellite demonstrates that the proposed framework achieves superior fidelity, with intensity profiles and statistical metrics (HP-PSNR) closely matching real observations. This work provides a scalable solution to the data scarcity problem, offering a valuable foundation for training and benchmarking robust deep learning models in spaceborne surveillance tasks.