Language Models Help Tune the Search for Better Satellite Imaging Schedules

Agile Earth observation satellites can change their pointing direction to photograph different ground targets. That flexibility creates competing choices. Taking one image can consume the time or energy needed for another, potentially more valuable observation. A useful schedule must therefore coordinate which requests to accept, which satellite should handle them and when each image should begin.

In their preprint, He Wang and colleagues propose LLMACO, a method combining a large language model with ant colony optimization. Their objective is to maximize “observation benefit”: the sum of each scheduled task’s priority multiplied by the quality associated with its selected observation opportunity. This is a mathematical scheduling score, not simply the number of images collected or a demonstrated measure of scientific value.

The language model’s role is deliberately limited. Rather than generating an entire observation sequence, it creates a heuristic—a scoring rule that helps the search choose among feasible opportunities. The researchers use DeepSeek-Flash to produce an initial rule from information about the tasks, observation windows and satellite resources. Candidate features include benefit, urgency, maneuvering requirements, energy, load and flexibility.

The rule combines weighted features and selected interactions between them. Its structured output is checked for valid feature names, coefficient ranges and numerical values before use. Scores are also bounded. These checks constrain what the language model can contribute, but they are not a general guarantee of operational safety.

Ant colony optimization then constructs schedules using both the scoring rule and artificial “pheromone” information that is updated during the search. Here, pheromones are numerical signals that influence later choices, not physical substances. A separate insertion procedure checks whether an observation fits between already scheduled tasks without moving them. It accounts for visibility windows, release times, deadlines, turning time and energy. Each imaging task can be performed at most once.

The method also lets the language model revise its advice during the search. At checkpoints set at 20% and 45% of the search, it receives feedback including the best benefit found, stagnation, unassigned high-priority tasks, energy use and pheromone statistics. It proposes a bounded correction to the original scoring rule.

A proposed correction is not adopted automatically. The current and proposed rules undergo four paired schedule constructions using the same pheromone states and random seeds. Acceptance requires both a positive average benefit difference and more improved pairs than degraded pairs. An accepted correction replaces the previous correction; it does not overwrite the initial rule. This is a local screening mechanism, not proof that every accepted change will improve all subsequent schedules.

The experiments cover six scenarios with 50, 100 or 200 imaging tasks. Each task scale has a moderate-competition and a high-competition case. All scenarios use three satellites over a 24-hour planning horizon, with 60-second observations and randomly assigned priorities from 1 to 10. Each algorithm runs independently 20 times per scenario, for up to 100 iterations.

The five comparison methods are conventional ant colony optimization, artificial bee colony, genetic algorithm, particle swarm optimization and whale optimization. Particle swarm optimization has the highest mean benefit among those baselines in every scenario. The authors report that LLMACO exceeds it by 4.32%–9.63%, with larger gains under high competition at each task scale. In the most competitive 200-task case, the mean score rises from 716.20 to 785.20. LLMACO also has the smallest reported standard deviation across runs in all six cases.

These results support a narrower conclusion than general superiority in satellite scheduling. The experiments use fixed task sets and observation windows, not changing mission requests or conditions. The reported comparisons do not include the learning-assisted scheduling methods discussed elsewhere in the paper. Nor does the reported evaluation separate the contribution of language-model initialization from that of later adaptation. Means and standard deviations describe performance and variability, but the paper reports no statistical significance tests.

Runtime evidence is also qualified. The authors describe LLMACO as slightly slower than conventional ant colony optimization in two illustrated scenarios and faster in the largest illustrated case; their figure distinguishes search time from wall time including language-model API latency. That does not establish a universal speed advantage. The central finding is that, in these simulations, language-model-generated guidance helped a constraint-checking search find higher-scoring schedules. Dynamic task arrivals, changing observation conditions and transferring rules between instances remain future work.