EagleEye: Nanosatellite constellation design for high-coverage, high-resolution sensing

Advances in nanosatellite technology and low launch costs have led to more Earth-observation satellites in low-Earth orbit. Prior work shows that satellite images are useful for geospatial analysis applications (e.g., ship detection, lake monitoring, and oil tank volume estimation). To maximize its value, a satellite constellation should achieve high coverage and provide high-resolution images of the targets. Existing homogeneous constellation designs cannot meet both requirements: a constellation with low-resolution cameras provides high coverage but only delivers low-resolution images; a constellation with high-resolution cameras images smaller geographic areas. We develop EagleEye, a novel mixed-resolution, leader-follower constellation design. The leader satellite has a low-resolution, high-coverage camera to detect targets with onboard image processing. The follower satellite(s), equipped with a high-resolution camera, receive commands from the leader and take high-resolution images of the targets. The leader must consider actuation time constraints when scheduling follower target acquisitions. Additionally, the leader must complete both target detection and follower scheduling in a limited time. We propose an ILP-based algorithm to schedule follower satellite target acquisition, based on positions identified by a leader satellite. We evaluate on four datasets and show that Eagle-Eye achieves 11–194% more coverage compared to existing solutions.

纳米卫星技术的进步与发射成本的降低促使更多的地球观测卫星被部署在低地球轨道。已有研究表明，卫星图像在地理空间分析应用中（如船舶检测、湖泊监测及油罐体积估算）具有重要价值。为实现其价值最大化，卫星星座需同时达成对目标的高覆盖率和高分辨率成像。

然而，现有的同构星座设计无法同时满足这两项要求：配备低分辨率相机的星座虽能提供高覆盖率，但图像分辨率较低；而配备高分辨率相机的星座虽能提供高质量图像，但其成像的地理范围则相对有限。

为此，我们开发了 “鹰眼”（EagleEye），这是一种新颖的混合分辨率、“引领者-跟随者”式星座设计 。

1. 在该设计中，引领者卫星搭载低分辨率、高覆盖率的相机，通过星上图像处理技术来探测目标。
2. 跟随者卫星则配备高分辨率相机，它接收来自引领者卫星的指令，对指定目标进行高分辨率成像。

在调度跟随者卫星的目标捕获任务时，引领者卫星必须考虑执行时间约束。此外，引领者卫星必须在有限的时间内完成目标检测和对跟随者的任务调度。

我们提出了一种基于整数线性规划（ILP）的算法，该算法能根据引领者卫星识别出的目标位置，为跟随者卫星调度成像任务。通过在四个数据集上的评估，结果显示，“鹰眼”方案与现有解决方案相比，覆盖率提升了11%–194%

Introduction

The emergence of small, cheap nanosatellites — e.g., chipsats, pocketqubes, and cubesats [31, 46, 54, 64] — and the maturation of commercial space launch services [34] bring space within reach for a wide range of valuable, new space-based cyberphysical systems applications. These applications leverage the unique vantage point of a low-Earth orbit (LEO) satellite to capture spectral (i.e., visual, infrared) sensor data that are inaccessible on Earth. Low-cost, high-cadence launches allow the deployment of constellations: groups of satellites that work together to implement an application. Applications transform this data into valuable insights, such as optimizing transport [20, 26], enhancing agriculture [59], and supporting disaster relief efforts [36]. Applications typically sense a target area on Earth and process the sensor data using machine learning. Historically, all computation happens on Earth. Data centers process information downlinked by satellites that were individually and manually tasked with capturing and downlinking particular observations. This outdated operational model is a fundamental barrier to increasing the capability of future satellite systems, posing physical limitations and limiting operational autonomy.

LEO nanosatellites are limited by physical and operational constraints. Satellites are physically limited by the low communication bandwidth to the ground and the limited energy. LEO satellites are also constrained in their acquisition of sensor data. Applications want sensor coverage in large geographic areas to maximize the reach of their applications. Applications also want high-resolution data to maximize the quality of data delivered to analysts. Satellite cameras meet only one of these constraints: wide-area sensors capture lowresolution data (e.g., 100km per pixel) and high-resolution sensors capture data with a narrow-area view (e.g., a few square kilometers per image). These constraints present an unsatisfying design choice for satellite application developers, forcing them to choose coverage or resolution, but not both. Today’s satellite constellations are also limited by a lack of autonomy and are relying on manual tasking. A human operator identifies the targets and sends pointing commands. The inability to detect and prioritize tasks precludes dynamic and reactive applications.

小型、廉价的纳米卫星 —— 例如，芯卫星（chipsats）、袖珍卫星（pocketqubes）和立方星（cubesats）[31, 46, 54, 64]——的兴起，以及商业航天发射服务 [34] 的日趋成熟，使得各类极具价值的新型天基信息物理系统应用得以进入太空。这些应用利用低地球轨道（LEO）卫星的独特视角，来捕获在地球上无法获取的光谱（即视觉、红外）传感器数据。低成本、高频次的发射使得部署卫星星座（constellations）成为可能——即由一组协同工作的卫星来共同执行应用任务。这些应用将采集到的数据转化为宝贵的洞察，例如优化交通运输 [20, 26]、提升农业生产力 [59] 以及支持救灾工作 [36]。应用通常需要遥感地球上的某个目标区域，并使用机器学习技术处理传感器数据。在过去，所有计算都在地面数据中心完成，卫星仅根据人工下达的指令单独进行观测和数据下传。这种过时的运行模式对未来卫星系统能力的提升构成了根本性障碍，不仅带来了物理上的限制，也制约了其运行自主性。

低地球轨道（LEO）纳米卫星受到物理和运行上的双重约束。在物理层面，卫星受限于对地通信的低带宽和有限的能源。在数据采集方面，低轨卫星同样面临限制。一方面，应用需要大范围的传感器覆盖以扩大其服务范围；另一方面，应用也需要高分辨率数据以最大化提供给分析人员的数据质量。然而，卫星相机通常只能满足其中一个要求：广域传感器能捕捉大范围但分辨率低的数据（例如，每像素代表100公里），而高分辨率传感器则视场极窄，只能拍摄小范围的图像（例如，每张图像几平方公里）。这些限制给卫星应用开发者带来了一个两难的设计抉择，迫使他们在覆盖范围和分辨率之间二选一，而无法兼得。此外，当今的卫星星座还受限于缺乏自主性，高度依赖人工任务分配。操作员需要手动识别目标并发送指向指令。这种无法自主检测目标和划分任务优先级的模式，阻碍了动态和响应式应用的发展。

Getting the human out of the loop and autonomously, dynamically identifying sensing targets from orbit is challenging. An autonomous constellation of satellites must detect targets using orbital edge computing [28–30]. Onboard computing requires careful system design balancing energy consumption with computational capability to avoid being bottlenecked by compute or energy collection time. Second, satellites in the constellation must collectively identify events of sufficient interest with low time and energy overhead. Today’s constellations operate with little or no autonomy making collective decision-making about points of interest infeasible. Third, after identifying interesting events, satellites in the constellation must plan actuation actions to point their sensors at targets.

要实现“人在回路外”（getting the human out of the loop），即 从轨道上自主、动态地识别遥感目标，是极具挑战性的。

首先，一个自主的卫星星座必须 利用星上边缘计算（orbital edge computing）[28–30] 来检测目标 。星上计算需要精心的系统设计，以平衡能耗与计算能力，从而避免受限于计算或能量收集时间，形成瓶颈。

其次，星座中的卫星必须 以低时间和能量开销，协同识别出足够有价值的事件 。目前的星座几乎没有自主性，这使得针对兴趣点进行集体决策变得不可行。

第三，在识别出感兴趣的事件后，星座中的卫星必须 规划执行动作 （actuation actions），将其传感器指向目标。

We develop EagleEye, a new nanosatellite constellation operating model that enables capturing high-resolution data with high spatial coverage and a high degree of autonomy. The goal of EagleEye is to identify points of interest with high coverage and to sense those points of interest with high resolution.

为此，我们开发了 “鹰眼”（EagleEye） ，一种新型的纳米卫星星座运行模型，它能够在实现高空间覆盖率和高度自主性的同时，捕获高分辨率数据。“鹰眼”的目标是以高覆盖率识别兴趣点，并以高分辨率对这些兴趣点进行遥感。

EagleEye organizes a constellation of satellites into a mixedresolution, leader-follower as Fig. 1a illustrates. In the leaderfollower model, a low-resolution, high-coverage leader satellite identifies points of interest by continuously processing each image using orbital edge computing [28, 30]. A coorbital follower satellite trails the leader by a small distance and points its high-resolution (but low-coverage) sensor at the identified points of interest and captures them. EagleEye resolves the tension between sensor coverage and image resolution: instead of choosing between high coverage and high-resolution, the leader-follower model provides both with little increase in cost. Fig. 1b shows the impact of EagleEye’s design, with both high coverage and high-resolution with far fewer satellites than existing high-resolution solutions.

“鹰眼”将卫星星座组织成一种混合分辨率的 “引领者-跟随者”模式 ，如图1a所示。在该模式中:

1. 一个搭载低分辨率、高覆盖率相机的引领者卫星通过持续使用星上边缘计算 [28, 30] 处理每幅图像来识别兴趣点
2. 一个共轨的跟随者卫星在引领者后方不远处飞行，并将其 高分辨率（但低覆盖率）的传感器指向已识别的兴趣点 并进行拍摄

“鹰眼”解决了传感器覆盖范围与图像分辨率之间的矛盾：它不再需要在高覆盖率和高分辨率之间做出选择，而是通过“引领者-跟随者”模式以微小的成本增量同时提供两者。

图1b展示了“鹰眼”设计的效果：它能以远少于现有高分辨率方案的卫星数量，同时实现高覆盖率和高分辨率。

leader && follower

(1) leader: 范围matters，识别“兴趣点”

(2) follower: 只把 “精度” 对准 “兴趣点”

The design of EagleEye requires solving two main problems in computational nanosatellite constellation design. The first problem is designing a leader satellite that uses orbital edge computing to identify targets in low-resolution data without introducing a new time or energy bottleneck compared to existing low-resolution systems. Moreover, the leader must perform inference on low-resolution data with high precision to avoid sending false sensing cues to followers. The second problem is designing a constellation in which leaders and followers collectively perform actuation-aware scheduling to point and capture targets. Actuation-aware scheduling is challenging because the leader must calculate a feasible actuation plan for followers that covers targets on a pointing trajectory. The scheduler must have a low overhead and account for followers’ time and energy.

“鹰眼”的设计需要解决计算型纳米卫星星座设计中的两个主要问题:

第一个问题是设计引领者卫星，使其能够利用星上边缘计算识别低分辨率数据中的目标，同时相对于现有的低分辨率系统， 不会引入新的时间或能源瓶颈。

此外，引领者必须高精度地对低分辨率数据进行推断，以 避免向跟随者发送错误的遥感指令。

第二个问题是 设计一种星座，其中引领者和跟随者能够协同执行 “感知作动的调度”（actuation-aware scheduling），以实现对目标的指向和捕获。

这种调度方式之所以具有挑战性，是因为 引领者必须为跟随者计算出一条可行的执行规划，使其指向轨迹能够覆盖多个目标 。该调度器必须开销低，并考虑到跟随者的时间和能源限制。

难点所在

1. leader 要正确识别图像
2. leader 要正确推理出“兴趣点”并传递
3. 要设计 “可以满足这种scheduling” 模式的 constellation

We develop and evaluate a prototype of a EagleEye constellation design. The system identifies targets using several ML object detection models. The leader runs actuation-aware scheduling for followers using a low-cost integer linear programming (ILP) formulation that models pointing time. We evaluate EagleEye for several applications — airplane tracking, ship tracking, and lake algal bloom detection — using an orbital edge computing simulator from prior work [30], and four publicly available datasets. Our evaluation shows that EagleEye achieves high-coverage and high-resolution with as much as 4.3× reduction in required constellations size.

To summarize, the main contributions of this work are:

• EagleEye, a new leader-follower mixed-resolution contellation organization and operating model for computational nanosatellite constellations.

• Actuation-aware scheduling and target clustering, which leverage onboard computing on low-resolution data to coordinate satellites to point at targets autonomously to capture high-resolution target data.

• An EagleEye prototype implementation including an ILP formulation of scheduling and several ML target detectors.

• A comprehensive evaluation showing that for four realworld Earth observation tasks, EagleEye improves coverage and image resolution, while reducing constellation cost and complexity.

我们开发并评估了一个“鹰眼”星座设计的原型系统。该系统使用多种机器学习目标检测模型来识别目标。引领者卫星通过一个低开销的 整数线性规划（ILP） 公式来为跟随者进行感知作动的调度，该公式能够对卫星的指向时间进行建模。我们使用了先前工作 [30] 中的一个星上边缘计算模拟器和四个公开可用的数据集，针对飞机追踪、船舶追踪和湖泊水华检测等多个应用对“鹰眼”进行了评估。我们的评估表明，“鹰眼”在实现高覆盖率和高分辨率的同时，所需星座规模最多可减少4.3倍。

总结而言，本文的主要贡献如下：

• “鹰眼”（EagleEye）：一种用于计算型纳米卫星星座的、新型的“引领者-跟随者”混合分辨率组织与运行模型
• 感知作动的调度与目标聚类：该技术利用对低分辨率数据的星上计算能力，自主协调卫星指向目标，以捕获高分辨率的目标数据
• 一个“鹰眼”原型系统的实现：包括一个用于调度的整数线性规划（ILP）公式和多个机器学习目标检测器
• 一项全面的评估：评估结果表明，在四个真实世界的地球观测任务中，“鹰眼”提升了覆盖率和图像分辨率，同时降低了星座的成本和复杂性