Related Work¶

Satellite networking has witnessed substantial growth, with a predominant focus in research on inter-satellite networking [31], [32]. However, the challenges posed by the scarce downlink bandwidth and the associated bottleneck are also crucial. OEC [6] focuses on the downlink bottleneck and shifts it to the inelastic computation capacities. Another recent work, Kodan [10] has proposed filtering low-value data and prioritizing high-value data for downlinking to mitigate the downlink bottleneck. However, Kodan considers the constraints of scarce satellite-ground bandwidth, a limitation that we aim to address in this work. This work not only tackles the downlink bottleneck through bandwidth-aware downlinking throttling but also addresses the computational bottlenecks.

卫星网络技术已取得长足发展，研究重点主要集中在星间网络 [31], [32]。然而，由稀缺的下行链路带宽及其引发的瓶颈所带来的挑战也同样至关重要。轨道边缘计算（OEC） [6] 专注于解决下行链路瓶颈，但它将瓶颈转移到了缺乏弹性的计算能力上。另一项近期工作 Kodan [10] 提出通过过滤低价值数据并优先下行传输高价值数据来缓解下行链路瓶颈。然而， Kodan 仅考虑了稀缺的星地带宽这一约束，而本工作旨在解决这一局限。本工作不仅通过带宽感知的下行传输节流来处理 下行链路瓶颈 ，同时也解决了 计算瓶颈问题。

The computational bottleneck represents a major challenge for satellite systems. Some works, such as [33] and [34], have explored the viability of utilizing DNN models for in-orbit processing. However, these approaches do not directly tackle the specific challenges addressed by TargetFuse, such as operating with real-world energy budgets. Moreover, several works focus on optimizing DNN models for accuracy or speed in terrestrial applications [35]–[41]. Various terrestrial and embedded systems that operate on harvested energy [42], [43] can transmit data at any time within energy constraints. However, this continuous data transmission capability is impractical for satellites as they can only transmit data when they are in proximity to ground stations. Moreover, the limited bandwidth available for satellite communication is significantly smaller than that of ground-based connections.

计算瓶颈是卫星系统面临的一大主要挑战。一些工作，如 [33] 和 [34]，已探索了利用深度神经网络（DNN）模型进行在轨处理的可行性。然而，这些方法并未直接应对 TargetFuse 所解决的特定挑战，例如在真实的能量预算下运行。此外，一些工作专注于在地面应用中优化DNN模型的精度或速度 [35]–[41]。各种依靠收集能量运行的地面和嵌入式系统 [42], [43] 可以在能量约束内随时传输数据。然而，这种连续数据传输的能力对于卫星而言是不切实际的，因为它们只有在接近地面站时才能传输数据。此外，卫星通信可用的有限带宽也远小于地面连接的带宽。

Vision tasks in EO satellites have been extensively studied and proven valuable for scientific investigations [44]–[46]. These applications span various domains, including computer systems, satellite systems, satellite networks, and machine learning systems. Developing a comprehensive scheduling system that considers image size, processing speed, energy constraints, and orbital mechanics poses a challenging research problem in computer systems. We are actively working to resolve the computer systems design challenges associated with satellite computing under real-world satellite constraints. Our system leverages a tradeoff between accuracy and execution time, and effectively addresses downlink bottleneck by bandwidth-aware downlinking throttling.

对地观测（EO）卫星中的视觉任务已被广泛研究，并被证明对科学调查极具价值 [44]–[46]。这些应用横跨计算机系统、卫星系统、卫星网络和机器学习系统等多个领域。开发一个综合考虑图像尺寸、处理速度、能量约束和轨道力学的全面调度系统，是计算机系统领域一个富有挑战性的研究问题。我们正积极致力于解决在真实卫星约束下，与卫星计算相关的计算机系统设计挑战。我们的系统利用了精度与执行时间之间的权衡，并通过带宽感知的下行传输节流有效解决了下行链路瓶颈。