Background and Motivation

We provide context for Kodan and characterize challenges for Earthobservation satellites.

Earth-observation satellites: LEO, Earth-observation satellites collect sensor data — e.g., multispectral images — for geospatial analytics. These satellites often deploy to polar orbits (i.e., orbits crossing near the poles of Earth). As the satellite travels through its orbit, it accesses nearly all latitudes; as the planet rotates, the satellite accesses all longitudes. LEO altitudes are hundreds of kilometers, and LEO periods are about 90 min.

Historically, Earth-observation satellites are large and monolithic. Recently, inexpensive nanosatellites have proliferated. The monolithic Worldview [13], Earth-Observing 1 (EO-1) [33], and Landsat [27] satellites cost hundreds of millions of US dollars each (e.g., $855,000,000 [20]). Now, many missions use cubesats [32], chipsats [42], and pocketqubes [9, 36] to increase hardware refresh cadence and avoid the high costs of monolithic satellites. Lower costs enable Earth-observing constellations consisting of hundreds of devices [5, 26].

Orbital mechanics determine both access to and the quality of satellite sensor data. For images, a satellite captures a frame along its ground track. A frame is a large geographic region; geospatial applications often split frames into many smaller tiles for analysis. Satellite image quality is characterized by ground sample distance (GSD) — geographic distance between adjacent pixels — which may range from km/px to cm/px [13, 27] and is determined by altitude and camera characteristics. Figure 1 illustrates these concepts.

对地观测卫星： 位于近地轨道 (LEO) 的对地观测卫星为地理空间分析任务采集传感器数据，例如多光谱图像。这些卫星通常部署在极地轨道（即靠近地球两极飞行的轨道）。当卫星沿其轨道运行时，它几乎能访问所有纬度；随着地球的自转，卫星则能覆盖所有经度。近地轨道的高度为数百公里，其轨道周期约为 90 分钟。

历史上，对地观测卫星体积庞大且为单体结构。近年来，廉价的纳米卫星已大量普及。诸如 Worldview [13]、地球观测1号 (EO-1) [33] 及陆地卫星 (Landsat) [27] 等单体卫星，每颗成本高达数亿美元（例如，855,000,000 美元 [20]）。现在，许多任务采用立方体卫星 (cubesats) [32]、芯片卫星 (chipsats) [42] 和口袋方块卫星 (pocketqubes) [9, 36]，以提高硬件更新频率并避免单体卫星的高昂成本。成本的降低催生了由数百个设备组成的对地观测星座 [5, 26]。

轨道力学决定了卫星传感器数据的可及性与质量。对于图像而言，卫星沿着其星下点轨迹 (ground track) 捕获一帧 (frame) 图像。一帧图像覆盖了广阔的地理区域；地理空间应用通常将图像帧分割成许多更小的图块 (tiles) 以便分析。卫星图像质量由地面采样距离 (Ground Sample Distance, GSD) 来表征——即相邻像素间的地理距离，其范围可从公里/像素 (km/px) 到厘米/像素 (cm/px) [13, 27]，具体取决于轨道高度和相机特性。图1阐释了这些概念。

The bent pipe: Today, most Earth-observation satellite operators manually task their devices to sense and downlink raw observations to a datacenter for processing, i.e., a bent pipe [12, 25]. Communication opportunities for high-velocity, LEO satellites last only for a few minutes while the device is near a ground station and may occur infrequently depending on the orbit. State-of-the-art communication systems downlink a total data quantity of MBs or GBs per pass. This downlink bottleneck constrains observation rate because not all data can be sent (we quantify this bottleneck in Section 2.1).

“弯管”模式： 如今，大多数对地观测卫星的运营商通过人工指令来驱动其设备，令其感知并下传原始观测数据至数据中心进行处理，即所谓的“弯管” (bent pipe) 模式 [12, 25]。 对于高速运行的近地轨道卫星而言，通信机会仅限于设备飞近地面站的几分钟内，并且根据轨道的不同，通信窗口可能非常稀少。 最先进的通信系统在单次过境期间总共能下传 MB 或 GB 量级的数据。这种下行链路瓶颈限制了观测速率，因为并非所有数据都能被发送回地面（我们将在 2.1 节中量化此瓶颈）。

Orbital edge computing: Recent work proposes orbital edge computing (OEC) [7, 8], in which satellites process data at the space edge. OEC distributes computation across a constellation where each satellite contains highly-capable, commercial, off-the-shelf (COTS) compute hardware rather than low-performance, radiationhardened, space CPUs. An OEC satellite tiles each image frame and processes tiles to identify interesting data to transmit. Especially when interesting features are rare, OEC addresses the downlink bottleneck with edge computing by triaging sensor data before transmission. However, OEC must process all frame tiles before a new frame enters the sensor view, creating a frame processing deadline usually between 1 − 30 s. Failing to meet the deadline is a computational bottleneck.

在轨边缘计算： 近期的研究提出了在轨边缘计算 (Orbital Edge Computing, OEC) [7, 8]，即卫星在太空边缘处理数据。OEC 将计算任务分布于一个星座，其中每颗卫星都搭载了高性能的商用现成品 (Commercial, Off-The-Shelf, COTS) 计算硬件，而非性能较低的抗辐射 (radiation-hardened) 航天级 CPU。

一颗采用 OEC 模式的卫星会对每帧图像进行分块，并处理这些图块以识别值得传输的趣味数据 。尤其是在趣味特征稀少的情况下，OEC 通过在传输前对传感器数据进行分类筛选，有效解决了下行链路瓶颈 。然而，OEC 必须在下一帧图像进入传感器视野之前处理完当前帧的所有图块，这便产生了一个通常在 1−30 秒之间的帧处理时限 。未能满足此时限即构成计算瓶颈。

Challenges at the Orbital Edge

We highlight two orbital edge challenges: a downlink bottleneck prevents sending all raw data, and a computational bottleneck prevents processing all data on orbit. We quantify these bottlenecks using the cote [8] simulator to model existing satellites, including orbital dynamics, sensing, communication, and the ground segment. We validate our analysis with publicly-available satellite and ground segment performance metrics [27]. In this section, we address three key questions:

1. What limitations arise from today’s downlink bottleneck, and what will its impact be in the future (Section 2.1.1)?

2. To what extent could edge computing address the downlink bottleneck (Section 2.1.2)?

3. How much of this potential improvement can be realized by directly deploying data processing applications to the space edge, and to what extent does the computational bottleneck limit these improvements (Section 2.1.3)?

我们着重指出在轨边缘面临的两个挑战：下行链路瓶颈导致无法发送所有原始数据，而计算瓶颈则导致无法在轨处理所有数据 。我们使用 cote [8] 模拟器来量化这些瓶颈，该模拟器对现有卫星进行建模，涵盖了轨道动力学、传感、通信及地面部分。我们通过公开的卫星和地面部分性能指标 [27] 对我们的分析进行了验证。在本节中，我们旨在回答三个关键问题：

• 当今的下行链路瓶颈带来了哪些限制，其在未来的影响又将如何（第 2.1.1 节）？
• 边缘计算能在多大程度上解决下行链路瓶颈（第 2.1.2 节）？
• 通过将数据处理应用直接部署到太空边缘，这种潜在的改进能实现多少？计算瓶颈在多大程度上限制了这些改进（第 2.1.3 节）？

2.1.1 The Downlink Bottleneck. Downlink capacity cannot support existing satellite sensor datarates, and the gap grows with sensor fidelity and constellation population. We quantify this gap for Landsat 8 in Figure 2. Over one orbit revolution, the ground segment supports reception of just 2% of the available observations of hyperspectral, 10K image frames. When satellite count in the same orbit plane increases from one to 16, downlinked data increases from 5 frames during one period to 60 frames during the same period; this improvement stems from claiming previously idle ground station time. When the Landsat ground segment serves one satellite, stations sit idle while the satellite is out of range (i.e., most of the time). Additional satellites, when not contending with each other, claim idle time and increase total downlinked data. However, as constellation population increases, the space segment eventually saturates the downlink. Adding satellites beyond this population count increases the ability to observe but not downlink additional data, widening the downlink gap.

Why are large constellations desirable? Larger constellations increase sensor coverage of Earth. Figure 3 shows the satellite count required for daily global coverage, i.e. the opportunity to observe all Landsat frames each day (see Section 2.1.1 for discussion on whether such observations could be downlinked). We add support to cote for the Landsat Path/Row World Reference System (WRS) [38] and import the WRS-2 scene boundary shapefiles to produce this plot. Reaching global daily coverage requires a constellation population of at least 40. Constellations in different orbits or with different sensors, like the Spire “Lemur” or Planet “Dove” satellites, deploy even greater numbers of devices.

2.1.1 下行链路瓶颈

下行链路容量无法支持现有卫星传感器的数据率，并且随着传感器保真度和星座规模的增长，这一差距日益扩大。我们在图 2 中量化了陆地卫星 8 号 (Landsat 8) 的这一差距。在一次轨道周期内，地面部分仅能支持接收高光谱、10K 分辨率图像帧的可用观测数据中的 2%。当同一轨道平面内的卫星数量从一颗增加到 16 颗时，一个周期内下传的数据量从 5 帧增加到 60 帧；这一提升源于利用了先前空闲的地面站时间。当陆地卫星的地面网络服务于单颗卫星时，地面站在卫星超出通信范围时处于空闲状态（即大部分时间）。增加的卫星在不相互争用资源的情况下，可以利用这些空闲时间，从而增加总下传数据量。然而，随着星座规模的增加，空间段最终会使下行链路饱和。在此之后继续增加卫星数量，只会增强观测能力，却无法下传更多数据，从而扩大了下行链路的差距。

为何需要大型星座？ 更大的星座可以提升对地球的传感器覆盖范围。图 3 展示了实现每日全球覆盖所需的卫星数量，即每天有机会观测到所有陆地卫星图像帧（关于这些观测数据能否被下传的讨论，请参见第 2.1.1 节）。我们为 cote 添加了对陆地卫星全球参考系统 (WRS) 的路径/行号 [38] 的支持，并导入了 WRS-2 的场景边界 shapefile 文件来生成此图。要实现全球每日覆盖，需要一个规模至少为 40 颗卫星的星座。采用不同轨道或不同传感器的星座，如 Spire 公司的“狐猴”(Lemur) 卫星或 Planet 公司的“鸽子”(Dove) 卫星，部署的设备数量甚至更多。

2.1.2 Addressing the Saturated Downlink. Bent-pipe satellites waste limited downlink capacity by indiscriminately sending observations containing both high-value and low-value data. To demonstrate this fact, we examine a cloud-filtering application. On average, 67% of satellite images are obscured by clouds [23] and are low-value to most customers. For Landsat 8, Figure 4 (left column) shows that, during one day, just 1/3 of the data from nearly 3600 observable frames is high-value (i.e., not cloudy). Accounting for the downlink bottleneck (middle), less than 21% of observable high-value data is downlinked with a bent-pipe. Ideal edge filtering (100% accuracy and zero execution time) delivers over 3× more high-value data — 63% of the total, observable high-value data, and the maximum possible under the downlink bottleneck (right column). Increasing the ratio of high-value to low-value data increases the data value density of the saturated downlink: the fraction of downlinked data composed of high-value bits.

“弯管”模式的卫星不加选择地发送包含高价值和低价值数据的观测结果，从而浪费了有限的下行链路容量 。为了证明这一点，我们研究了一个云过滤应用。平均而言，67% 的卫星图像被云层遮挡 [23]，对大多数客户而言价值很低。对于陆地卫星 8 号，图 4（左栏）显示，在一天之内，从近 3600 个可观测帧中产生的数据只有 1/3 是高价值的（即无云）。考虑到下行链路瓶颈（中栏），采用“弯管”模式下传的高价值数据不足可观测高价值数据的 21%。 理想的边缘过滤（准确率 100% 且执行时间为零）能够多提供超过 3 倍的高价值数据 —— 占总可观测高价值数据的 63%，这也是下行链路瓶颈下可能达到的最大值（右栏）。提高高价值数据与低价值数据的比例，可以增加饱和下行链路的数据价值密度：即下传数据中高价值比特所占的比例。

2.1.3 The Computational Bottleneck. While ideal cloud filtering offers a potential 3× improvement in valuable data downlinked, the space edge must contend with the inelastic and limited computational resources on satellites. Volume, mass, energy, and cost constraints at the space edge prevent deployment of unlimited computational resources inside a satellite [8]. Unless a filtering application completes within the frame deadline, a satellite cannot process all frames, creating a computational bottleneck that limits the ability of OEC to address the downlink bottleneck. We quantify this effect for a real cloud filter application [31]. Figure 5 shows the fraction of high-value data downlinked with and without edge filtering for a range of constellation sizes. Direct-deployment of cloud filtering improves high-value data downlinked by just 9%, far short of the potential 3×. The shortfall stems from the 98 s frame processing time, which exceeds the 22 s frame deadline; only a fraction of captured frames can be filtered. Therefore, while OEC has potential to mitigate the downlink bottleneck, the computational bottleneck limits its benefit.

Limitations of parallel, distributed computation: Prior OEC work addresses computational needs by distributing work across a pipeline of satellites. While effective at reducing per-satellite compute time to meet full ground track coverage, pipeline populations must be very large (e.g., > 100 devices per application). This approach is costly and designed for vertically-integrated constellations deployed for a single purpose. In the future, we expect the emergence of constellations acting as a platform for customer applications, i.e., a constellation-as-a-service. As a result, the naturally inelastic space edge has pressure to operate at its computational limit in order to maximize platform value. Prior OEC work provides no technique to reduce per-satellite computational load without increasing constellation population; this shortcoming is a key motivation for Kodan.

尽管理想的云过滤有望将下传的有价值数据量提升 3 倍，但太空边缘必须应对卫星上缺乏弹性且有限的计算资源。太空边缘在体积、质量、功耗和成本上的限制，使得在卫星内部署无限的计算资源成为不可能 [8]。除非一个过滤应用能在帧时限 (frame deadline) 内完成，否则卫星无法处理所有图像帧，从而产生一个计算瓶颈，该瓶颈限制了在轨边缘计算 (OEC) 解决下行链路瓶颈的能力。我们针对一个真实的云过滤应用 [31] 量化了这种影响。图 5 显示了在不同星座规模下，有无边缘过滤时下传的高价值数据比例。 直接部署云过滤应用仅将下传的高价值数据提升了 9%，远低于 3 倍的潜力。这一差距源于其 98 秒的帧处理时间，远超 22 秒的帧时限；因此只有一小部分捕获的帧能被过滤。 所以，尽管 OEC 有潜力缓解下行链路瓶颈，但计算瓶颈限制了其效益。

并行分布式计算的局限性： 以往的 OEC 工作通过将任务分布到一条卫星流水线 (pipeline of satellites) 上来满足计算需求。虽然这种方法能有效降低单颗卫星的计算时间以实现完整的星下点轨迹覆盖，但流水线中的卫星数量必须非常庞大（例如，每个应用需要 >100 个设备）。这种方法成本高昂，且是为单一目的部署的垂直整合星座设计的。未来，我们预期会出现作为客户应用平台的星座，即 “星座即服务” (constellation-as-a-service)。因此，天然缺乏弹性的太空边缘面临着在其计算极限下运行的压力，以最大化平台价值。 以往的 OEC 工作没有提供在不增加星座规模的情况下降低单星计算负载的技术；这一不足是 Kodan 的一个关键动机。