Theoretical framework and key technologies of spatio-temporal intelligence for satellite remote sensing constellations

As networked Earth observation systems, satellite remote sensing constellations serve as the core support for global aerospace information acquisition. However, the traditional open-loop, feed-forward remote sensing paradigm dominated by the “planned observation to scene interpretation” pipeline struggles to cope with the complete sampling of non-stationary events, spatio-temporal reasoning of complex physical processes, and autonomous platform coordination. This restricts constellation systems from leaping from “global coverage” to “global cognition.” Targeting these bottlenecks, this paper systematically constructs a unified theoretical framework for the spatio-temporal intelligence of satellite remote sensing constellations. We formalize the constellation as a discrete-time dynamic multi-agent collaborative network and examine its three core technologies: dynamic event perception, physics-driven cognitive modeling, and autonomous cross-domain scheduling. Crucially, by combining terrestrial pre-training with on-orbit model manifold splitting and quantized Directed Acyclic Graph (DAG) offloading, this architecture achieves sub-second local operator inference on edge nodes while compressing the system-level end-to-end event response loop from traditional hours/days to minutes, thereby drastically decreasing inter-satellite communication overhead. On this basis, we further explore the deployment potential of this intelligent architecture in high-impact scenarios, such as the early warning of major ecological disasters, the tracking of high-threat maneuvering targets, and complex urban emergency responses. We also discuss its limitations encountered during deployment: the vulnerability of cross-modal perception to non-stationary spatial noise, the multi-year compute gap between space-grade radiation-hardened hardware (e.g. FPGAs/ASICs) and terrestrial GPUs, and the dynamic instability of Inter-Satellite Links (ISLs). Ultimately, this lays a theoretical foundation for expanding the cognitive boundaries of the physical world’s evolutionary mechanisms and building high-precision, highly credible aerospace decision-support systems, while providing a clear development roadmap for the evolution of next-generation autonomous intelligent aerospace systems.