Answers are grounded in this paper and cite their sources.
Satellite staring video provides high-resolution, long-duration observations of local regions, but the resulting data volume and strict onboard computation budgets make efficient compression particularly challenging. Existing neural video compression (NVC) methods are mostly developed for generic natural videos and do not consider the distinct spatiotemporal statistics of satellite staring imagery. Their optical-flow-based motion branches and large contextual and entropy networks are therefore overdesigned for this scenario, leading to high FLOPs and memory usage that are difficult to accommodate in satellite environments and excessive relative to the underlying spatiotemporal redundancy. In this work, we first perform a modulewise parameter and complexity analysis of representative NVC architectures, revealing that motion, context, and entropy branches dominate the computation cost and exhibit strong feature redundancy. Guided by these observations and the spatiotemporal properties of staring video, we propose LiteSVC-global context prior (GCP), a lightweight, flow-free NVC framework tailored for satellite staring video. LiteSVC-GCP replaces dense computation modules with the proposed parameter-efficient LiteBlocks and adopts a simple context-based codec, substantially reducing model size and computation. In addition, a GCP module aggregates long-term scene information and fuses it with short-term reference features through a dual-mask mechanism, enabling the codec to better capture temporally varying regions while leveraging slowly changing scene content. Experiments on Jilin-1 and Carbonite-2 satellite videos demonstrate that LiteSVC-GCP outperforms the traditional H.265/HEVC codec on most sequences while using only 5.1 M parameters and roughly 5– 6× fewer FLOPs than strong NVC baselines to achieve comparable or better rate–distortion performance.
No takes yet. Share an insight, caveat, or question.
Qi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: