Prediction-Enhanced Cloud-Aware Scheduling for Micro Agile Maritime Observation Satellite

A micro agile Earth observation satellite is increasingly tasked with time-sensitive maritime monitoring using electro-optical (EO) payloads, where feasible imaging opportunities are jointly constrained by target motion, cloud transmittance, solar illumination conditions, and limited attitude maneuver capability. This paper presents a prediction-enhanced, cloud-aware mission-planning framework that explicitly couples environmental forecasting with bi-objective scheduling. Vessel motion is forecast from Automatic Identification System streams using a long short-term memory (LSTM) predictor, while cloud transmittance is nowcast from Himawari-8 observations using a U-Net and convolutional LSTM (ConvLSTM) model. The resulting multistep predictions are integrated into a unified observation-profit model and a dynamics-consistent maneuver-energy model, yielding a bi-objective formulation that maximizes transmittance-weighted imaging profit and minimizes attitude maneuver energy under nonlinear rotational dynamics. To adapt the evolutionary search to nonstationary conditions, a reinforcement-learning (RL)-assisted Nondominated Sorting Genetic Algorithm II (NSGA-II) is developed, in which a lightweight Q-learning mechanism selects genetic operators based on dominance improvement signals. Experiments with real Automatic Identification System trajectories and Himawari-8 cloud data show improved hypervolume, convergence stability, and robustness over representative multi-objective baselines (NSGA-II and Multi-Objective Evolutionary Algorithm Based on Decomposition) and an RL baseline, proximal policy optimization. The results indicate that incorporating multistep environmental prediction into adaptive evolutionary planning can substantially improve decision quality for EO maritime observation under rapidly changing cloud conditions.