Background and Motivation

We provide background on tasking in nanosatellite constellations and highlight key shortcomings of today’s systems in dynamically capturing geo-distributed targets with high coverage and high resolution. We then quantitatively motivate EagleEye by showing the promise of autonomous, mixed-resolution, leader-follower constellations.

Achieving high-coverage, high-resolution sensing with a satellite constellation

More launches to low Earth orbit and the decreasing costs of nanosatellites foster a “new space race:” many satellite constellations monitor the planet for a multitude of Earthobservation tasks. Satellites collect Earth images for geospatial analysis, such as environmental and ecological monitoring, meteorology, and agriculture. Use cases for Earth observation abound. For example, images containing ships could be used to detect illegal fishing [8, 16], oil spills and bilge dumping [42, 50]. High-resolution images containing lakes could help detect algae blooms [20, 52]. Images containing oil storage tanks could be used to estimate total oil reserve volumes [10].

A constellation is a collection of satellites that work together to support an application. The scope of this work is nanosatellite constellations that have development and launch costs that are orders of magnitude lower than larger, “exquisite” [45] satellite designs. Designing a constellation to support an Earth-observation application requires defining an organization and operating model. A constellation organization entails defining the hardware and software composition of each satellite, the number of satellites in the constellation, and the mix of capabilities across a heterogeneous constellation. The operating model involves when and how satellites sense and process Earth-observation signals, and how satellites communicate with one another and with internet-connected receivers in the “ground segment.” The organization also involves defining the orbit altitude and inclination into which satellites deploy.

前往低地球轨道的发射活动日益增多，加之纳米卫星的成本不断下降，共同催生了一场“新太空竞赛”：众多卫星星座为执行大量的对地观测任务而持续监测着我们的星球。卫星采集地球图像用于地理空间分析，例如环境与生态监测、气象学和农业等。对地观测的应用场景十分丰富。例如，包含船舶的图像可用于检测非法捕鱼 [8, 16]、石油泄漏和舱底水倾倒 [42, 50]。包含湖泊的高分辨率图像有助于检测藻类水华 [20, 52]。包含储油罐的图像则可用于估算全球石油储备总量 [10]。

星座是指一组协同工作以支持某个应用的卫星集合。本研究的范围是纳米卫星星座，其研发和发射成本相比于那些更大型、更“精巧”[45]的卫星设计要低数个数量级。设计一个支持对地观测应用的星座，需要定义其组织结构和运行模式。星座组织结构涉及定义每颗卫星的软硬件构成、星座中的卫星数量，以及在一个异构星座中不同能力的组合。运行模式则涉及卫星在何时以及如何感知和处理对地观测信号，以及卫星之间、卫星与地面段（ground segment）的互联网接收器之间如何通信。组织结构还包括定义卫星部署的轨道高度和倾角。

Constellation Organization. Nanosatellite constellations provide a low-cost, low-complexity option for deploying large numbers of satellites. A cubesat uses commercial, off-the-shelf (e.g., Planet [23, 32], NASA [4]) components for electronics and structure and has a small size (e.g., 10 cm × 10 cm × 10 cm for a “1U” cubesat) with masses around 1 − 10 kg. Cubesats often deploy to LEO. Several commercial operators have deployed numerous LEO cubesat constellations [12, 15]. LEO spans an altitude less than 2,000 km and often around 400−700 km. Building and launching a cubesat is relatively inexpensive, with a cost of tens of thousands dollars [6]. The low cost of nanosatellites enables launching constellations of tens or hundreds with a similar capability while costing far less than monolithic satellites [45].

Satellite hardware includes several components. Satellites include an attitude determination and control system (ADACS) with actuators (e.g., reaction wheels) to enable precise pointing at rates between 1 and 10 degrees per second. Onboard imagers capture electromagnetic spectral data inclusive of the visual domain and possibly other spectra, such as RF, near-infrared (NIR), and short-wave infrared (SWIR). In this work, we primarily assume visual spectrum data sensing, the sensors for which are common and have low cost; EagleEye applies generally to arbitrary spectrum data. The ground coverage area and ground sample distance or “GSD” (meters per pixel) of an image produced by a sensor are intrinsically defined by the camera system and the orbital altitude. A fundamental tension between coverage and GSD makes a key trade-off in constellation design at the heart of EagleEye. Existing LEO satellites with COTS imagers capture images of Earth with GSD of tens of centimeters to tens of kilometers per pixel. Fig. 2 illustrates the relationship between a nanosatellite’s ground track, its camera’s swath width, and its camera’s GSD. A GPS/GNSS receiver [40] provides Earth-relative position information, allowing a satellite to perform precise geo-registration of captured sensor data. Recent work [30] showed that it is possible to deploy commodity computing devices (such as the NVidia Jetson/Orin mobile GPU) in a cubesat. High-performance computing hardware equips a nanosatellite to run sophisticated computations, such as image classification, object detection, and pixel segmentation.

星座组织结构 (Constellation Organization)

纳米卫星星座为部署大量卫星提供了一种低成本、低复杂度的选择。 立方星（cubesat）采用商用现成品（COTS），例如来自Planet [23, 32]或NASA [4]的电子和结构部件，其尺寸很小（例如，“1U”立方星的尺寸为10厘米×10厘米×10厘米），质量约为1-10公斤。立方星通常部署于低地球轨道（LEO）。 已有数家商业运营商部署了大量的LEO立方星星座 [12, 15]。LEO的高度通常在2000公里以下，常见于400-700公里范围。建造和发射一颗立方星的成本相对低廉，约为数万美元 [6]。纳米卫星的低成本使得发射由数十甚至数百颗卫星组成的星座成为可能，其能力与单颗大型卫星相当，但成本却远低于后者 [45]。

卫星硬件包括几个关键组件。卫星配有姿态确定与控制系统（ADACS），其执行器（如反作用轮）能以每秒1到10度的速率进行精确定向。星上成像仪可捕获电磁光谱数据，包括可见光波段，也可能包括其他光谱，如射频（RF）、近红外（NIR）和短波红外（SWIR）。在本文中，我们主要假设使用可见光光谱数据进行遥感，因为其传感器普遍且成本低廉；但“鹰眼”的设计普适于任意光谱数据。由传感器生成的图像的地面覆盖区域和地面采样距离（GSD）（米/像素）由相机系统和轨道高度内在地决定。覆盖范围和GSD之间存在的根本性矛盾，构成了星座设计的关键权衡，这也是“鹰眼”系统设计的核心所在。 现有的、搭载COTS成像仪的LEO卫星所拍摄的地球图像GSD范围从几十厘米到几十公里每像素。图2阐释了纳米卫星的星下点轨迹、其相机的刈幅宽度及其相机的GSD之间的关系。GPS/GNSS接收器 [40] 提供相对于地球的位置信息，使卫星能对其捕获的传感器数据进行精确的地理配准。近期的研究 [30] 表明，将商用计算设备（如: 英伟达Jetson/Orin移动GPU）部署在立方星上是可行的。高性能计算硬件使纳米卫星能够运行复杂的计算任务，如图像分类、目标检测和像素分割。

Operating Model. Today, a vast majority of satellites operate with no autonomy. In this operating model, a human operator sends commands to each satellite in the constellation from a ground terminal. On receiving a command, a satellite senses based on GPS coordinates, may process data using orbital edge computing, and may transmit interesting data to Earth.

Recent constellation research proposes on-orbit computing to improve autonomy [28, 30]. In these operating models, a constellation aims to cover its complete ground track. Covering the ground track is challenging with high-resolution data because an individual satellite may be unable to process an entire high-resolution frame before observing the next frame. Nanosatellite pipelining [30] covers the ground track by statically distributing the work among satellites and processing frame data in parallel. Kodan [28] uses ML model specialization, frame tiling, and tile processing elision to reduce the number of satellites required to cover a ground track (often to one satellite). These operating models improve over the command-oriented, human-in-the-loop model, but they do not address heterogeneous camera constellations.

运行模式 (Operating Model)

目前， 绝大多数卫星在无自主性的模式下运行 。在这种运行模式中， 地面终端的人类操作员向星座中的每颗卫星发送指令。卫星在接收到指令后，根据GPS坐标进行遥感，可能会利用星上边缘计算处理数据，并将感兴趣的数据传输回地球。

近期的星座研究提出了通过星上计算来提高自主性的方案 [28, 30]。在这些运行模式中，星座旨在覆盖其完整的星下点轨迹。然而，使用高分辨率数据来覆盖整个轨迹是具有挑战性的，因为单颗卫星可能无法在观测下一帧图像之前处理完当前的高分辨率帧。纳米卫星流水线处理技术（Nanosatellite pipelining）[30]通过在卫星间静态分配工作，并行处理帧数据，从而实现对地面轨迹的覆盖。Kodan [28] 则利用机器学习模型特化、帧分块和分块处理省略等技术，减少了覆盖地面轨迹所需的卫星数量（通常减至一颗）。这些运行模式相较于指令驱动、有人在回路中的模型有所改进，但它们并未解决异构相机星座的问题。

Requirements

A satellite constellation design needs to meet several requirements: high image resolution (i.e., low GSD), high revisit rate (i.e., a small time interval between consecutive views of the same location), high coverage (i.e., a constellation covers a high fraction of earth area), and low cost (i.e., the total satellite count and orbit count is reasonable). Many of these design factors are coupled, and some are naturally negatively correlated. They all influence the cost and feasibility of a constellation deployment.

一个卫星星座设计需要满足若干要求：高图像分辨率（即低GSD）、高重访率（即对同一地点的连续观测时间间隔短）、高覆盖率（即星座覆盖地球面积的比例高）以及低成本（即卫星总数和轨道总数合理）。这些设计因素多是相互耦合的，其中一些甚至天然地呈负相关。它们共同影响着星座部署的成本和可行性。

Data analytics have image resolution thresholds. Some applications require high-resolution images to produce accurate results. We characterize the impact of image resolution with a visual oil volume estimation task [11, 61]. The task consists of two stages: (1) detecting the oil tanks from the images; (2) estimating oil volumes based on the shadow size on the tank lids. Fig. 3 shows the result of running the task on image data at different GSD levels ranging from 0.7 to 11.5 m/px. The results show that the low-resolution image is sufficient to correctly detect the oil tank, but does not contain enough detail to accurately estimate the volume of the oil tank. The key point is that higher resolution images are needed for some applications.

Geo-distributed targets require high geospatial coverage. An application’s sensing targets may be highly geographically distributed, and in motion (e.g., airplanes or ships). To detect such targets, a fixed-size constellation’s satellites’ sensors must capture images of large geographic regions. With a fixed image sensor size (i.e., total sensed pixel count) a larger area per image entails a larger geographic area per pixel, and higher resolution images cover a smaller geographic area. Camera focal length and orbit altitude largely determine these parameters and cameras that can feasibly be deployed to space at low cost are restricted to a single operating point. An operator must then choose: high-resolution data or high-coverage?

数据分析对图像分辨率有阈值要求

某些应用需要高分辨率图像才能产生准确的结果。我们通过一个视觉化的石油储量估算任务 [11, 61] 来描述图像分辨率的影响。该任务包括两个阶段：

（1）从图像中检测储油罐；

（2）根据罐顶的阴影大小估算石油储量

图3展示了在GSD从0.7到11.5米/像素的不同分辨率图像上运行该任务的结果。结果表明， 低分辨率图像足以正确检测到储油罐，但其细节不足以精确估算油罐的储量 。关键在于，某些应用确实需要更高分辨率的图像。

地理分散的目标需要高地理空间覆盖率

应用的遥感目标可能在地理上高度分散，并且可能在移动（如飞机或船舶）。为了探测这类目标，一个固定规模星座的卫星传感器必须能够拍摄大范围的地理区域。 在图像传感器尺寸（即总像素数）固定的情况下，每张图像覆盖的面积越大，意味着每个像素代表的地理区域也越大，而高分辨率图像所覆盖的地理区域则更小 。相机的焦距和轨道高度在很大程度上决定了这些参数，而那些能以低成本部署到太空的相机通常被限制在单一的工作点。因此，运营商必须做出选择：是要高分辨率数据，还是要高覆盖率？

We characterized this unsatisfying tradeoff between data quality and coverage. Fig. 2 shows image GSD versus swath (i.e., ground track width). A large swath operating point uses a short focal length or high altitude to cover a larger area, but at low resolution. A high resolution operating point uses a long focal length or low altitude that captures low-GSD (high-resolution) images, but only at a narrow swath. Fig. 4 (Left) contrasts several existing cubesat cameras (Planet [12], Dragonfly [3], Simera Sense [13]). Fig. 4 (Right) shows the fraction of targets that a constellation captures over a oneday period using different camera swath-widths (we describe our experimental setup in detail in §5.2). Using a large 100 km swath requires only 20 satellites to capture all targets, although only at an unacceptably low resolution. Using a small 10 km swath provides acceptably high-resolution images, but unfortunately, even 40 satellites capture only 41% of the targets.

我们对这种两难的权衡关系进行了量化分析。图2展示了图像GSD与刈幅（即地面轨迹宽度）的关系。一个大刈幅的工作点使用短焦距或高轨道来覆盖更广的区域，但分辨率较低。一个高分辨率的工作点则使用长焦距或低轨道来捕获低GSD（高分辨率）的图像，但刈幅很窄。图4（左）对比了几款现有的立方星相机（Planet [12], Dragonfly [3], Simera Sense [13]）。图4（右）展示了在一个为期一天的周期内，使用不同相机刈幅宽度的星座所捕获的目标比例（我们的实验设置详见§5.2）。使用100公里的大刈幅仅需20颗卫星即可捕获所有目标，但分辨率却低得无法接受。而使用10公里的窄刈幅虽能提供可接受的高分辨率图像，但不幸的是，即使部署40颗卫星也仅能捕获41%的目标。

Constellation size defines total cost. The material cost of a computational nanosatellite constellation scales directly with the satellite count. Satellite cost has several components. Material cost for a nanosatellite is low, especially if equipped with only COTS components (e.g., a COTS GPU is around $2k). Launch costs are by far the largest cost associated with a nanosatellite (e.g., around $50-100k for a 3U cubesat) and must be amortized across many satellites. Moreover, launch costs manifest as a non-linearity as cost varies with constellation size: if the addition of a satellite requires an additional launch, the cost of the second launch amortizes poorly unless still more satellites are added. Operations costs are moderate and scale with satellite count. The primary operation cost is the ground station receiver operation cost, which is being commoditized [1, 21], but still scales with constellation size and data payload.

星座规模决定总成本

一个计算型纳米卫星星座的物料成本与卫星数量成正比。卫星成本包括几个部分。纳米卫星的物料成本很低，特别是如果只装备COTS组件（例如，一个COTS GPU约2000美元）。发射成本是迄今为止与纳米卫星相关的最大开销（例如，一颗3U立方星约需5-10万美元），且必须在多颗卫星间分摊。此外，发射成本会随着星座规模的变化呈现出非线性特征：例如，如果增加一颗卫星就需要一次额外的发射，那么除非后续增加更多卫星，否则第二次发射的成本摊销效率会很低。运营成本适中，并随卫星数量增加而增加。主要的运营成本是地面接收站的运营成本，尽管这一成本正在商品化 [1, 21]，但仍与星座规模和数据量成正比。

Existing Solutions

Tip and Cue. One existing solution uses a “tip and cue” operating model [7, 18, 58], where satellites from different missions with different cameras are used. This approach utilizes a low-resolution camera satellite for target detection and a high-resolution camera satellite for capturing high-resolution images of the targets.

However, this solution has several limitations. First, it requires operators from different missions (entities) to share compute, communication, and sensing resources, posing practical challenges with respect to economics and regulations. Second, the satellites fly in different orbits, leading to long delays (around 12 hours for one deployment [18]) between target identification and high-resolution imaging, which prevents imaging moving targets (e.g., airplanes and ships). Third, in our best reading of the somewhat scant details of these systems, they lack operational autonomy: the target detection satellite sends images to Earth for processing, and the ground segment relays imaging commands to the high-resolution satellite.

AB&B [27] solves the limitations by proposing a bi-satellite cluster where two satellites (a low-resolution camera leader satellite and a high-resolution camera follower satellite) fly in the same orbit, with a separation of 100 s. This addresses the challenges related to resource sharing and image capture delays. To achieve autonomous operation, it runs the target identification and the high-resolution image capture schedule on the leader satellite.

However, AB&B still has several limitations. First, they use a custom branch-and-bound algorithm to schedule high-resolution image capture, which exhibits a high runtime. As shown in §6, AB&B takes more than 15 s to schedule just 19 targets. This leads to difficulties in meeting frame deadlines and result in lower coverage. Second, they neglect satellite energy constraints, whereas cubesats have very limited energy. Their scheduler’s extended runtime exacerbates the energy insufficiency concerns. Third, they only design a bisatellite cluster and evaluate the coverage over a 500 km × 2000 km area, without considering how the constellation size affects the coverage in a larger area (e.g., the entire Earth area is around 510 million km 2 ). Fourth, they only consider a single follower satellite. Although this suffices for some target densities, we show in §6 that multiple followers provide higher coverage for a high target density.

“提示与引导”（Tip and Cue）

一种现有的解决方案采用“提示与引导”的运行模式 [7, 18, 58]，即使用来自不同任务、搭载不同相机的卫星。该方法利用一颗低分辨率相机卫星进行目标探测，再由另一颗高分辨率相机卫星对目标进行高分辨率成像。

然而，该方案有几个局限性。首先，它需要来自不同任务（实体）的运营商共享计算、通信和遥感资源，这在经济和法规方面带来了实际挑战。其次，这些卫星飞行在不同轨道上，导致从目标识别到高分辨率成像之间存在长时间延迟（在一个部署案例中延迟约12小时 [18]），这使得对移动目标（如飞机和船舶）的成像变得不可能。第三，根据我们对这些系统有限公开细节的最佳解读，它们缺乏运行自主性：目标探测卫星将图像发送到地面进行处理，再由地面段向高分辨率卫星转发成像指令。

AB&B [27] 方案通过提出一个双星簇来解决上述局限，其中两颗卫星（一颗搭载低分辨率相机的引领者卫星和一颗搭载高分辨率相机的跟随者卫星）在同一轨道上飞行，间隔为100秒。这解决了资源共享和图像捕获延迟的挑战。为了实现自主运行，它在引领者卫星上运行目标识别和高分辨率图像捕获的调度程序。

然而，AB&B仍存在几个局限。首先，他们使用一个定制的分支定界算法来调度高分辨率图像的捕获，该算法的运行时间很长。如第6节所示，AB&B在调度仅19个目标时就需要超过15秒。这导致难以满足帧处理的时限要求，进而降低了覆盖率。其次，他们忽略了卫星的能源约束，而立方星的能源非常有限。其调度器过长的运行时间加剧了能源不足的问题。第三，他们仅设计了一个双星簇，并在一个500公里×2000公里的区域内评估覆盖率，而没有考虑星座规模如何影响更大区域的覆盖（例如，整个地球表面积约为5.1亿平方公里）。第四，他们只考虑了单个跟随者卫星。尽管这对于某些目标密度是足够的，但我们将在第6节中证明，对于高目标密度的情况，多个跟随者能提供更高的覆盖率。