Hardware-efficient satellite change detection via hybrid Mamba decoder with differentiable lookup table classifier and progressive knowledge distillation

The rapid expansion of Low Earth Orbit satellite networks unlocks new frontiers in real-time remote sensing, but deploying resource-efficient change detection systems on constrained on-board hardware remains a significant challenge. This work proposes a compact hybrid Mamba student model with a differentiable lookup table classifier that achieves exceptional change detection accuracy with minimal computational overhead, enabled by innovative architectural designs and progressive knowledge distillation. The architecture features a lightweight encoder paired with a hybrid Mamba decoder that, despite a nearly identical parameter count to the teacher decoder, delivers 54.7% fewer floating-point operations and 33% lower peak memory, enabling linear-complexity long-range dependency modeling ideally suited for high-resolution satellite imagery on embedded platforms. The differentiable lookup table thresholding classifier enables ultra-low-footprint inference, with Field-programmable gate array implementations demonstrating 6.5×to 35.9×latency reductions and complete elimination of multiplier resources. The progressive knowledge distillation framework incorporates adaptive logit alignment, multi-scale feature matching, and phase-adaptive training to effectively transfer robust change detection capabilities from a high-capacity teacher to the hardware-optimized lightweight student. Experiments on three standard benchmarks yield F1-scores of 91.18%, 93.42%, and 96.78%, respectively, using only 1.78 million parameters and 3.59 billion floating-point operations (GFLOPs), confirming its suitability for on-orbit deployment. The code is available at https://github.com/etii75/MambaLUT-CD .