SPN-Geo: A Stream-Oriented Physics-Neural Framework for On-Board Geolocation of Fengyun Meteorological Satellites

Accurate and low-latency geolocation is a prerequisite for real-time onboard processing of Fengyun meteorological satellites. However, onboard geolocation of wide-swath observations must simultaneously preserve physical modeling fidelity, represent nonlinear coordinate variations, and satisfy deterministic latency constraints under limited computational resources. Existing acceleration-oriented methods mainly improve efficiency through model simplification, hardware parallelization, or feature databases, but they often fail to jointly address physical consistency, nonlinear coordinate-field representation, and continuous stream-level execution. To address these limitations, this paper proposes SPN-Geo, a stream-oriented physics-neural geolocation framework for onboard meteorological satellite real-time processing. In SPN-Geo, rigorous geometric modeling provides physically consistent sparse geolocation references, an implicit neural representation reconstructs dense and continuous geographic coordinate fields, and a CPU–GPU heterogeneous execution scheme supports continuous frame-stream processing. Experiments using FY-3F MERSI-III and MWHS-II observations demonstrate sub-pixel geolocation accuracy for both optical and microwave payloads, with planar RMSEs below 0.80 and 0.55 pixels, respectively. Compared with conventional interpolation methods, the INR-based reconstruction achieves lower cross-track and planar errors, showing improved capability in capturing nonlinear geolocation distortions across the swath. The complete SPN-Geo pipeline maintains stable per-frame latency of approximately 0.1 s on an embedded heterogeneous platform, satisfying the real-time requirement of continuous onboard observations. These results indicate that SPN-Geo provides an effective solution for accurate, low-latency, and multi-sensor onboard geolocation of wide-swath meteorological satellite observations.