<p>Future 6G networks will integrate non-terrestrial communication with mobile edge computing to enable wide-area edge intelligence, providing ubiquitous communication and computation services for everyone and everything. In this paper, we introduce the intelligent cloud-edge-device architecture into low Earth orbit (LEO) satellite-supported remote Internet of Things (IoT) networks to facilitate latency-sensitive and compute-intensive IoT applications. However, the dynamic spatio-temporal characteristics inherent in LEO satellite networks, such as the high volatility of task traffic and the high mobility of space nodes, pose severe challenges. For efficient task execution, we investigate the joint optimization of cross-region task offloading and cross-domain resource allocation, and innovatively propose a spatio-temporal attention-based proximal policy optimization (STA-PPO) algorithm. Specifically, the temporal attention-based actor network analyzes the timing dependencies of task arrival and makes time-variant decisions. Simultaneously, the spatial attention-based critic network captures the spatial variations of satellite motion and evaluates the topology-related value. We verify the effectiveness of the cloud-edge-device collaboration through extensive simulations. Numerical results demonstrate that the STA-PPO algorithm outperforms benchmarks, showing the lowest system delay and the highest link throughput.</p>