Emergency-Oriented Multi-Satellite SAR Incremental Replanning Using Multi-Relational Graph Convolution and Constrained Proximal Policy Optimization

Floods, earthquakes, and landslides may require a multi-satellite synthetic aperture radar (SAR) system to insert new observations after a regular plan has been prepared but before upload. The revised plan must respect visibility, imaging chronology, look-direction switching, orbit resources, strip formation, and plan stability. We propose a multi-relational graph convolutional network with constrained proximal policy optimization (GCN-PPO). Five opportunity relations encode conflict, feasible sequence, cross-satellite competition, demand alternatives, and strip synergy. A three-level policy selects the action category, logical completion unit, and observation opportunity, while feasibility-normalized constraints regulate emergency-service shortfall and plan perturbation. In a 300 s equal-budget experiment, all 750 returned schedules were feasible and free of hard-constraint violations, and 96.4% passed the preregistered timing contract. Under natural termination, GCN-PPO responded in 102.4 s on average, 9.35% of the ALNS time, and met the normalized emergency target in 98.1% of evaluations. Its per-edit perturbation was 31.4% lower than MLP-PPO and 35.9% lower than ALNS. Timely executable schedules support rapid SAR data acquisition for downstream disaster-object detection and emergency monitoring.