Remote Sensing Task-oriented Task Partitioning and Data Distribution in Satellite Edge Computing Networks with Multipath Offloading

The advancement of Earth observation-based applications (e.g., geographical monitoring and disaster warning) has triggered a surging demand for timely processing of remote sensing data (e.g., high-resolution images). Traditionally, the sensing data are directly transmitted from remote sensing satellites to ground stations for processing, often resulting in high latency caused by long waiting times for satellite overpasses. In addition, given the limited bandwidth of satellite-to-ground links, the time for transmitting the raw sensing data from remote satellites to the ground station would be excessively high as the data size keeps growing, which also affects the timeliness of data processing. A promising approach to address this issue is satellite edge computing networks, which leverage multiple computing satellites to execute the computational tasks for processing remote sensing data. In particular, by partitioning each remote sensing task into multiple subtasks and assigning them to different computing nodes (computing satellites and ground stations) via multipath offloading, task execution and task offloading can be performed in parallel, resulting in reduced computing and communication latencies, respectively. Meanwhile, given the expected large number of various types of satellites and the dynamic topology of satellite networks, how to properly partition each task and distribute a proper amount of task data to each transmission path remains a challenge. In this paper, we investigate the problem of task partitioning and data distribution in satellite edge computing networks. Aiming to minimize the overall end-to-end latency, such a problem is formulated as a sequential decision-making problem. Considering the features of satellite edge computing networks, we transform the formulated problem into a partially observable Markov decision process and propose a dual multiagent deep deterministic policy gradient approach to solve it. Simulation results demonstrate that our proposed scheme reduces the average end-to-end latency by at least 30% compared to benchmark schemes.