Joint Offloading, Trajectory and Deployment Optimization for Multi-UAV Cooperative Regional Search in SAGINs: A Hybrid DRL-GA Framework

Abstract

Multi-UAV systems in Space-Air-Ground Integrated Networks (SAGINs) offer solutions for diverse applications, but realizing their full potential in search and rescue (SAR) is challenged by complex terrains, limited infrastructure, and dynamic interferences. These demanding environments reveal shortcomings in jointly optimizing task offloading, flight trajectories, and UAV deployment, and limits of idealized simulations. This problem is formulated as a multi-objective optimization to maximize UAV search coverage and minimize task energy cost under resource constraints. To solve this, we propose a two-stage Hybrid Convolutional Deep Reinforcement Learning (HCDRL) and a Genetic Algorithm (GA) framework. In HCDRL, we employ a novel feature-fusing multi-modal state encoding. By individualizing per-UAV perception and using Convolutional Neural Networks (CNNs) for visual features and Graph Convolutional Networks (GCNs) for network topology and offloading features, this encoding linearizes the multi-agent action state space’s exponential growth, enhancing training efficiency and robustness. The GA component then utilizes the learned HCDRL policy as a fitness evaluator to optimize global UAV deployment. In addition, incorporating uncertainty-aware terrain modeling and NOAA-derived realistic wind-field data substantially improves simulation realism. Extensive simulations provide evidence of robustness and scalability across the evaluated scenarios. Notably, under strong wind conditions, the proposed GA-HCSAC framework improves mission lifetime by up to 38% and search coverage by 33% compared to standard baselines, while the GA-optimized deployment alone contributes to a nearly 18% coverage lift. Finally, offloading heatmaps and UAV visit-frequency maps provide interpretable evidence of spatially structured coordination and directional adaptability under dynamic wind fields.