Two-stage offline knowledge distillation for onboard registration of multispectral satellite images

Abstract

Multi-band optical sensors onboard modern Earth observation satellites capture complementary spectral responses across varying spatial and spectral resolutions. To effectively fuse this information for downstream applications, accurate band co-registration is critical. However, for real-time processing, such registration must be performed onboard, where sensor distortions, platform-induced motion, and spectral disparities introduce significant challenges. Traditional feature matching algorithms struggle to cope with these variations or are often too computationally intensive for the constrained hardware typically found on small satellites. As a result, real-time onboard multimodal fusion has remained largely impractical in operational settings. With the emergence of next-generation satellites equipped with AI-enabled onboard processing, such as Australia’s Kanyini mission, there is now an opportunity to overcome these limitations. In this work, we introduce a deep learning-based, lightweight band registration framework specifically designed for real-time onboard deployment. Our approach features a band-independent teacher network that jointly leverages adversarial learning and supervised regression to estimate affine registration parameters across spectral bands. To meet hardware constraints, we employ a two-stage knowledge distillation strategy that produces a compact yet accurate student model. Experimental results demonstrate that our method delivers robust and efficient registration performance, enabling real-time spectral alignment and significantly enhancing the potential for onboard multimodal data fusion in Earth observation missions.