Reinforcement learning (RL) has demonstrated remarkable effectiveness in addressing task offloading decisions within highly complex network environments. Particularly multi-agent reinforcement learning (MARL), new methodologies have emerged to tackle real-world issues involving both cooperation and competition. We focus on the challenge of task offloading for vehicular users operating in remote areas. To mitigate the scarcity of computational resources resulted by the impracticality of deploying traditional infrastructure like Base Stations (BSs) and roadside units (RSUs) in remote area, we first introduce the use of Unmanned Aerial Vehicles (UAVs) and Low Earth Orbit (LEO) satellites to provide edge service for vehicular users. Furthermore, we put forward an optimization algorithm for allocating computational resource, UAV trajectory optimization, and task offloading based on Multi-Agent Twin Delayed Deep Deterministic Policy Gradient (MATD3) by making real-time strategies to achieve minimal task completing delay and energy consumption. Finally, we conduct simulation experiments to prove proposed algorithm. The experimental results proved our designed algorithm can reduce task completion delay by 44.4% and task processing energy consumption by 46.8% compared to the MADDPG algorithm.