Background and Motivation

A. EO Satellites

EO satellites collect raw sensor data for geospatial analytics. These satellites capture images along their ground track, generating geospatial images that cover hundreds of square kilometers and contain hundreds of millions of pixels. The level of detail present in these images is described by the ground sample distance (GSD) [6], which is determined by orbit altitude, sensor size, and camera characteristics [6]. Besides, the high velocity of satellites, up to 7.9 km/s, results in brief periods of visible contact with ground stations. These periods typically last only a few minutes, sometimes less than 8 minutes, and may occur infrequently.

During a single orbit revolution, a single satellite captures more images than it is capable of downlinking [10]. This is due to the downlink capacity of current satellite sensors, which is insufficient to support their data rates. Currently, the majority of Earth-observation satellites are organized in a bentpipe architecture [6], where raw observations are transmitted to ground stations and then processed by machine learning algorithms. However, satellites only achieve downlinking rates of no more than hundreds of Mbps using Ka-band [11], resulting in a limited daily downlink data volume. For instance, given that a contact session lasts for 6 minutes, the system can downlink a maximum data quantity of 4.39 GB at a downlinking speed of 100 Mbps [12]. The downlink bottleneck prevents these daily global observations from being transmitted to the ground. What’s more, not all raw observations contain high-value data. Bent-pipe satellites may waste the precious downlink bandwidth as they indiscriminately transmit raw observations independent of the value of data. Statistics show that 67% of the observations are obscured by clouds, thus becoming low-value to users [13].

对地观测（EO）卫星为地理空间分析采集原始传感器数据。这些卫星沿其地面轨迹捕获图像，生成覆盖数百平方公里、包含数亿像素的地理空间影像。图像中的细节程度由地面采样距离（Ground Sample Distance, GSD） [6] 来描述，它取决于轨道高度、传感器尺寸和相机特性 [6]。此外，卫星高达7.9公里/秒的高速导致其与地面站的可见接触时间非常短暂。这些接触窗口通常仅持续几分钟，有时甚至不足8分钟，且可能不频繁发生。

在单次轨道周期内，一颗卫星捕获的图像量远超其下行传输能力 [10]。这是因为当前卫星传感器的下行链路容量不足以支持其数据生成速率。目前，大多数对地观测卫星采用 “弯管”（bent-pipe）架构 [6]，即: 原始观测数据被传输到地面站，然后由机器学习算法进行处理 。然而，使用Ka波段，卫星的下行速率最多只能达到数百Mbps [11]，导致每日可下行的数据总量非常有限。例如，在一个持续6分钟的接触窗口内，以100 Mbps的下行速度， 系统最多能传输4.39 GB的数据 [12]。下行链路瓶颈阻碍了每日全球观测数据的回传。

更重要的是， 并非所有原始观测数据都包含高价值信息。“弯管”式卫星可能会浪费宝贵的下行带宽，因为它们无差别地传输原始观测数据，而不考虑其价值 。统计数据显示，67%的观测数据被云层遮挡，因此对用户而言价值很低 [13]。

B. Orbital Edge Computing

OEC [6] has been proposed to address the downlink bottleneck, in which satellites process raw data in space. OEC aims to address the limitations of “bent-pipe” architectures [6] by distributing processing across a constellation. Nowadays, each satellite can be equipped with hundreds of high-datarate cameras, sensors, and commercial, off-the-shelf (COTS) hardware [14]. The satellite is lightweight, small-sized, and expensive, with each satellite weighing a few kilograms, measuring a few centimeters, and costing millions of USD.

However, satellites still face limitations in providing highlycapable onboard processors. Currently, available space-grade processors are often decades-old, “flight heritage”. The satellite systems may operate for decades in the space environment, which means that COTS hardware may be low-risk and highly reliable at the expense of performance. Hence, recent trends in space systems have started to consider COTS-embedded systems [15], which enable the use of in-orbit processing. Applying these terrestrial techniques directly to space is appealing, but computational capacity is subject to unique operating constraints. For instance, unlike on Earth, all energy expended in space must be harvested from solar panels, which is backed by a rechargeable energy buffer [16]. In line with typical 3U Cubesat systems [6], the size of the tested satellite limits the area of the solar panel to 57.2 × 20.6 cm and thus limits the power to a range of 34-118 W.

The limited and inelastic computational resources also pose a major space systems challenge. Though the satellites are exposed to sunlight for about 60% of each orbit period (e.g., approximately 90 minutes) [17]. Only 30% of the solar energy is initially converted into battery power, and less than 50% of the battery capacity is utilized for daily satellite operation over its lifetime [13]. We collected real-world data from Baoyun satellite and observed that computing in operation accounts for approximately 50%, compared to other subsystems such as basic and electrical operations, as depicted in Fig. 2. Therefore, during runtime, operating system (OS) of the satellite allocates an energy budget to the computing modules by adjustmenting in input performance and energy parameters.

为解决下行链路瓶颈，轨道边缘计算（Orbital Edge Computing, OEC） [6] 被提出，其核心思想是让卫星在太空中直接处理原始数据。OEC旨在通过将处理任务分布于整个星座，来解决“弯管”架构的局限性 [6]。如今，每颗卫星都可以配备数百个高数据速率的相机、传感器以及商用现成（Commercial, Off-The-Shelf, COTS）硬件 [14]。这些卫星具有轻量化、小尺寸和高成本的特点，每颗卫星重几公斤，尺寸仅为几十厘米，但成本高达数百万美元。

然而，在提供高性能星上处理器方面，卫星仍面临限制。目前可用的航天级处理器通常是具备“飞行遗产”（flight heritage） 的、技术上已沿用数十年的产品。卫星系统可能需要在太空环境中运行数十年，这意味着COTS硬件的选择会优先考虑低风险和高可靠性，而牺牲性能。因此，空间系统的近期趋势已开始考虑采用COTS嵌入式系统 [15]，这使得在轨处理成为可能。尽管将这些地面技术直接应用于太空很有吸引力，但计算能力受到独特的运行约束。例如，与地球不同，太空中消耗的所有能量都必须通过太阳能电池板收集，并由一个可充电能量缓冲器（即电池）支持 [16]。参照典型的3U立方星系统 [6]，我们测试的卫星尺寸将其太阳能电池板的面积限制在57.2 × 20.6平方厘米，从而将功率限制在34-118瓦的范围内。

有限且缺乏弹性的计算资源也构成了空间系统的一大挑战。尽管在每个轨道周期（例如约90分钟）中，卫星约有60%的时间暴露在阳光下 [17]，但最初只有30%的太阳能被转换为电池电能，并且在其整个生命周期中，只有不到50%的电池容量被用于日常卫星操作 [13]。我们从“宝云”卫星收集的真实数据显示，如图2所示，与其他子系统（如基础和电气操作）相比，运行中的计算任务约占总能耗的50%。因此，在运行时，卫星的操作系统（OS）会通过调整输入性能和能量参数，为计算模块分配一个能量预算。

C. Challenges for EO Computing

EO computing is a crucial application of OEC. However, two challenges arise at the orbital edge: a downlink bottleneck that hinders the transmission of all raw data and a computational bottleneck that restricts the processing of all data in orbit. OEC has been proposed to address computational needs by distributing tasks across a constellation. Although effective in reducing per-satellite compute workload to meet full ground track coverage, this approach is in nature designed for vertically-integrated constellations for a single purpose, which requires a large pipeline population and incurs high monetary costs. Existing OEC work can hardly reduce the per-satellite workload without increasing the constellation population [6]; this shortcoming is a key motivation for our work.

Processing raw observations in space presents significant challenges due to limited computational resources. Satellites have an energy limit that prevents them from processing all images, which creates a computational bottleneck that limits OEC’s ability to address the downlink bottleneck. Unfortunately, improving the computational capacity of COTS hardware is difficult because of physical constraints. And there are no feasible options for adjusting the computational capacity of the hardware already in space. Furthermore, inorbit computing alone is insufficient because each satellite has natural constraints such as volume, mass, and energy that prevent it from processing all images.

Compared to the space environment, a more collaborative approach between satellite and ground station is feasible by utilizing the relatively more favorable computing capacity and higher energy availability of ground stations. As a result, we consider the two bottlenecks of in-orbit processing jointly to adapt the computing hardware of the target satellite in space.

对地观测计算是OEC的一项关键应用。然而，在轨道边缘出现了两大挑战：阻碍所有原始数据传输的下行链路瓶颈，以及限制所有数据在轨处理的计算瓶颈。OEC已被提议通过在星座内分配任务来满足计算需求。尽管这种方法能有效降低单星计算负载以实现完整的地面轨迹覆盖，但它本质上是为单一目的的垂直一体化星座设计的，这需要庞大的在轨卫星数量并产生高昂的经济成本。现有的OEC工作在不增加星座规模的情况下，几乎无法降低单星工作负载 [6]；这一缺陷是我们工作的关键动机。

由于计算资源有限，在太空中处理原始观测数据面临巨大挑战。卫星的能量限制使其无法处理所有图像，这造成了计算瓶颈，从而限制了OEC解决下行链路瓶颈的能力。不幸的是，由于物理约束，提升COTS硬件的计算能力十分困难，并且对于已在太空中的硬件，没有可行的方法来调整其计算能力。此外，仅靠在轨计算是不足够的，因为每颗卫星都有其固有的体积、质量和能量限制，这些都使其无法处理所有图像。

与太空环境相比，地面站拥有相对优越的计算能力和更充裕的能源。因此，一种卫星与地面站之间更紧密的协同方法是可行的。为此，我们联合考虑在轨处理的这两个瓶颈，以期更好地适应目标卫星在太空中的计算硬件。