Reconfigurable Earth Observation Satellite Scheduling

• REOSSP is a formal optimization framework that integrates dynamic reconfiguration stages to enhance imaging, communication, and resource management for satellite constellations.
• It employs MILP, CP, and hybrid methodologies to sequentially schedule orbital maneuvers, sensor-mode switching, and energy allocation under tight resource constraints.
• The framework boosts disaster response and environmental monitoring by enabling adaptive, high-yield satellite operations, as validated through empirical performance benchmarks.

The Reconfigurable Earth Observation Satellite Scheduling Problem (REOSSP) is a formal optimization framework for selecting and sequencing imaging, communication, and resource-management actions for maneuverable Earth observation satellite constellations. In contrast to classical formulations that treat satellite orbits as fixed throughout the planning horizon, REOSSP explicitly models opportunities for on-orbit constellation reconfiguration—enabling dynamic orbit maneuvers, sensor-mode switching, power management, and task prioritization in response to time-varying observation objectives and environmental constraints. REOSSP has emerged as the central scheduling problem in applications such as disaster observation, rapid-response environmental monitoring, and autonomous, adaptive satellite operations leveraging onboard AI and multi-satellite coordination.

1. Formal Definition and Mathematical Formulation

REOSSP extends the standard Earth Observation Satellite Scheduling Problem (EOSSP) by incorporating satellite reconfiguration stages. The planning horizon $T$ is discretized into $S$ blocks (“stages”), between which each satellite $k\in \mathcal{K}$ can transition between discrete orbital/attitude “slots” $j\in \mathcal{J}^{s,k}$. Let $\mathcal{P}$ and $\mathcal{P}'$ denote priority and auxiliary targets, and $\mathcal{G}$ the set of ground stations. For each timeslot within a stage, binary variables indicate the choice of orbital slot ($x^{sk}_{ij}$), imaging decision on priority/auxiliary targets ($y^{sk}_{tp}$, $\alpha^{sk}_{tp}$), downlink ($S$0), and charging ($S$1). Continuous variables track on-board data ($S$2) and battery levels ($S$3).

The canonical REOSSP formulation is a mixed-integer linear program (MILP):

$S$4

subject to (a) flow constraints enforcing physically consistent reconfiguration sequences within $S$5 propellant and battery budgets, (b) visibility and task exclusivity, (c) onboard data storage dynamics between observations and downlink, and (d) battery recharging/consumption from imaging, communication, maneuvers, and housekeeping (Pearl et al., 9 Feb 2026, Pearl et al., 14 Jul 2025).

The model further supports mission-specific enhancements—multiple imaging directions and flexible observation durations via interval variables in constraint programming (Caleiras et al., 17 Jan 2026), dynamic task arrival/cancellation, and joint on-board processing/edge computing (Soret et al., 7 Apr 2026).

2. Reconfiguration Models and Constraints

REOSSP encapsulates diverse forms of physical and operational reconfigurability:

• Orbital Maneuvering: Binary variables $S$6 capture transitions between orbital slots, with constraints enforcing continuity and $S$7/propellant thresholds $S$8.
• Attitude Pointing and Sensor Modes: Sequence-dependent reorientation costs and sensor-mode transitions are modeled via binary exclusivity constraints and, for super-agile platforms, interval variables and disjunctive/cumulative constraints using start time, duration, and alternative imaging senses (Caleiras et al., 17 Jan 2026).
• Resource Dynamics: Data storage and power budgets are updated at each timestep based on imaging, downlink, housekeeping, charging opportunities, and maneuver costs.

Such models are extensible to “edge intelligence” architectures in which the set of feasible actions at each time step expands to include selection of local versus distributed processing, dynamic voltage-frequency scaling, and allocation of semantic-processing workloads across a constellation (Soret et al., 7 Apr 2026).

3. Solution Methodologies

The computational complexity of REOSSP is dominated by its integrality and large action space, growing with the product of satellites, time steps, and reconfiguration options. Solution strategies include:

• Direct MILP Approach: Exacts solves (via Gurobi, CPLEX, YALMIP, or SCIP) are feasible for moderate-scale instances (e.g., $S$9 satellites, $k\in \mathcal{K}$0 blocks, $k\in \mathcal{K}$1 slots) with proven optimality within 1 hour per planning block (Pearl et al., 9 Feb 2026).
• Rolling Horizon Procedure (RHP): The horizon is decomposed into overlapping control/lookahead windows, sequentially optimizing over sub-blocks to maintain tractable variable counts at the expense of potentially $k\in \mathcal{K}$2 objective degradation (Pearl et al., 14 Jul 2025).
• Constraint Programming (CP): The use of interval, alternative, and sequencing global constraints captures flexible durations, multiple imaging senses, and sequence-dependent transitions. IBM ILOG CP Optimizer achieves proven-optimal solutions for instances with up to 200 targets and 180 VTWs in seconds to minutes, with optimality gaps below 10% in harder cases (Caleiras et al., 17 Jan 2026).
• Column Generation: Each orbit’s feasible schedule is formulated as a resource-constrained shortest path; efficient label-setting algorithms generate new columns for a restricted master LP, enabling near-optimality for large instances (gap $k\in \mathcal{K}$3 for $k\in \mathcal{K}$4 in $k\in \mathcal{K}$5 minutes) (Han et al., 2018).
• Maximum Independent Set (MIS) Approaches: The schedule feasibility graph (vertices: collects, edges: conflict) is built, and a (weighted) MIS is computed to maximize non-conflicting observations; this scales to $k\in \mathcal{K}$6 collections, outperforming MILP in both profit and solve time (e.g., $k\in \mathcal{K}$7 scheduled in $k\in \mathcal{K}$8 time for 24 satellites, 10,000 requests) (Eddy et al., 2020).
• Reinforcement Learning + OR Hybrids: Two-stage frameworks use RL (MDP or deep Q-learning) to restrict the prior assignment search space, followed by MIP or DP for sequencing/timing, yielding near-optimal scheduling for problems up to 400 tasks in under 25 seconds (He et al., 2021).

4. Application Domains and Integration in Operational Pipelines

REOSSP serves as the scheduling backbone for autonomous, adaptive spaceborne sensing. In the WildFIRE-DS wildfire detection pipeline, for example, REOSSP (“Phase 4”) ingests updated priority and auxiliary target lists generated from passive imaging, CNN-based detection, and Bayesian target confirmation (Phases 1–3) (Pearl et al., 9 Feb 2026). The scheduling output specifies slew maneuvers, revisit sequences, downlink schedules, and charging plans, thereby dynamically steering the constellation through multi-day event response cycles. The resultant increases in data yield and detection rates—e.g., $k\in \mathcal{K}$9 data collected and $j\in \mathcal{J}^{s,k}$0 true positives compared to non-reconfigurable baselines—demonstrate REOSSP’s operational impact.

Similarly, in edge-intelligent EO pipelines, the two-stage REOSSP (observation plus real-time distributed processing allocation) enables the system to balance energy, latency, and bandwidth tradeoffs under heterogeneous in-orbit and ground computing resources (Soret et al., 7 Apr 2026).

5. Computational Results and Empirical Performance

Empirical benchmarking shows that REOSSP offers substantial performance gains:

Instance	Downlink GB	True Positives	ΔV Used (km/s)	Solve Time (min)	% Opt. Gap
REOSSP (CNN)	25.3	123	1.71/2.00	< 60 (per block)	0–5%
EOSSP (fixed)	7.9	35	–	< 20	–
RHP (fast)	2.4–3.2	–	2.85	6.7	10–20%
Column Gen (AEOS)	–	–	–	< 10	<4%
MIS Graph	–	–	–	< 15	–
CP (SAEOS-ISP)	–	–	–	< 1 (easy inst.)	0–10%

Performance scaling with problem size varies by method: direct MILP and CP solvers are effective for $j\in \mathcal{J}^{s,k}$1, rolling-horizon and MIS-based solvers scale to $j\in \mathcal{J}^{s,k}$2 and dozens of satellites. RL-OR hybrids maintain near-optimality up to hundreds of tasks with minimal computational expense (Pearl et al., 9 Feb 2026, Pearl et al., 14 Jul 2025, Eddy et al., 2020, He et al., 2021, Caleiras et al., 17 Jan 2026, Han et al., 2018).

6. Model Extensions and Research Directions

Research threads within REOSSP include:

• Integrated Attitude and Orbit Control: Current models typically alternate between orbit and attitude reconfiguration; joint AEOSSP–REOSSP formulations are identified as future work (Pearl et al., 14 Jul 2025).
• Stochastic and Robust Scheduling: Real disasters generate uncertain, time-varying targets; stochastic programming and online/incremental scheduling techniques are proposed for responsive adaptation.
• Edge-Processing Coupling: Dynamic allocation of on-board semantic extraction and cross-orbit data movement introduces new convex/nonconvex optimization subproblems (Soret et al., 7 Apr 2026).
• Scalability and Real-Time Operation: Rolling-horizon, graph-theoretic, and RL-OR hybrids offer practical, operationally relevant strategies for large constellations with dynamic task streams.
• Multi-Objective Optimization: Weighted objectives in profit, latency, energy, and coverage are readily modeled as max-weight MIS or multi-criteria programming.

7. Summary of Core Contributions and Impact

REOSSP rigorously formalizes the scheduling of maneuverable Earth observation satellites in an era of high-agility, high-autonomy constellations. By explicitly modeling orbital and attitudinal reconfiguration, onboard resource dynamics, and adaptability for complex event-driven scenarios, REOSSP enables dramatic improvements in data yield, timeliness, and operational flexibility. Its MILP, CP, graph-theoretic, and AI-augmented solution methods have established new empirical and algorithmic baselines for satellite constellation tasking, with demonstrated applicability in disaster response, resource monitoring, and next-generation autonomous observation architectures (Pearl et al., 9 Feb 2026, Pearl et al., 14 Jul 2025, Soret et al., 7 Apr 2026, Caleiras et al., 17 Jan 2026, Han et al., 2018, Eddy et al., 2020, He et al., 2021).