跳转至

Background

In the 1990’s, Iridium, Globestar, and Teledesic [27,30,42] planned constellations of tens of satellites to provide direct connectivity to handheld terminals. Similarly, early imaging constellations, such as NOAA’s series for weather sensing, comprised of a couple of satellites. These early constellations triggered important research in satellite networking [19,21, 34,38,49,52].

More recently, within the last 5-10 years, the emergence of large LEO constellations, comprising of hundreds of satellites, has been driven by lower launch and manufacturing costs of small satellites. For instance Planet’s Dove constellation for Earth imagery is composed of nearly 200 low-cost cubesats (‘shoebox-sized’ satellites) with off-the-shelf components. Our work focuses on these modern constellations for Earth observation. These modern constellations differ from traditional satellite constellations in three ways:

在1990年代，Iridium、Globestar和Teledesic [27,30,42] 公司规划了由数十颗卫星组成的星座，旨在为手持终端提供直接连接。同样，早期的成像星座，如用于气象感知的NOAA系列卫星，也仅由几颗卫星组成。这些早期的星座引发了卫星网络领域的重要研究 [19,21, 34,38,49,52]。

近年来，在过去的5-10年里，随着小型卫星发射和制造成本的降低，由数百颗卫星组成的大型低地球轨道（LEO）星座应运而生。例如，Planet公司的用于地球影像的Dove星座由近200颗低成本立方体卫星（‘鞋盒大小’的卫星）组成，这些卫星使用了现成的商业组件。我们的工作专注于这些用于地球观测的现代星座。这些现代星座与传统卫星星座在三个方面有所不同：

• Constellation size: Modern satellite constellations (e.g., Planet Inc. [26], Spire Inc. [3], etc.), consist of hundreds of satellites as opposed to few satellites in traditional constellations. This allows modern constellations to get more frequent imagery of any part of Earth with revisit frequency of few hours as opposed to a delay of several days from traditional constellations.

• Data volumes: The low orbit of LEO satellites and improved imaging hardware enables high resolution imagery (e.g. 1m 2 per pixel). They capture images of Earth in different parts of the frequency spectrum, e.g., RGB, Radio Waves, Infrared, etc. The multi-spectral imagery, increased satellite number, and high resolution lead to increased data volumes—from few GBs of data per day to TBs of data per day. For instance, Planet’s Dove satellites generates approximately one Terabyte of data per satellite per day.

• Applications: Traditional Earth observation satellites could only support delay-tolerant applications like crop yield estimates, land cover use, etc. Modern constellations gather images more frequently and offer the promise of real-time applications like disaster monitoring, traffic analysis, maritime monitoring, etc.

• Processing pipelines: Modern data processing pipelines increasingly rely on modern Machine Learning (ML) methods, in contrast to (merely) traditional signal processing based approaches. For example, European Space Agency’s Φ-Sat-1 [28] recently demonstrated the ability to perform neural network-based cloud detection on board a satellite.

Network pipeline: Finally, we provide a brief description of satellite network pipeline as context for the rest of the paper. LEO satellites for Earth observation operate in polar orbits and go around the Earth once every 1.5 hours appoximately. During each orbit, they pass over a different part of the Earth due to Earth’s rotation. The data from these satellites is usually downloaded using few dedicated ground stations with Gbps link capacities [25]. Due to a satellite’s orbital motion, it can contact each ground station four to six times a day, with each contact lasting up to ten minutes. To improve download latency and increase the number of contacts, recent work has proposed distributed ground station designs with multiple general-purpose ground stations [60,61]. In the industry, Amazon and Microsoft have launched ground-station-as-a-service platforms [7,46], wherein satellite operators can rent time on existing ground stations to download data from satellites.

  • 星座规模： 现代卫星星座（例如，Planet Inc. [26]，Spire Inc. [3]等）由数百颗卫星组成，而传统星座仅有几颗卫星。这使得现代星座能够更频繁地获取地球上任何地区的影像，重访频率达到几小时，而传统星座则有数天的延迟。

  • 数据量： LEO卫星的低轨道和改进的成像硬件使其能够获得高分辨率影像（例如，每像素1平方米）。它们在不同频谱范围（如RGB、无线电波、红外线等）捕捉地球图像。多光谱影像、增加的卫星数量和高分辨率导致了数据量的激增——从每天几GB增加到每天TB级别。例如，Planet公司的Dove卫星每颗每天大约产生1TB的数据。

  • 应用： 传统的地球观测卫星只能支持延迟容忍型应用，如作物产量估算、土地覆盖利用等。现代星座更频繁地收集图像，为实时应用（如灾害监测、交通分析、海事监控等）提供了可能。

  • 处理流程： 现代数据处理流程越来越依赖于现代机器学习（ML）方法，而不仅仅是传统的基于信号处理的方法。例如，欧洲航天局的Φ-Sat-1 [28] 最近展示了在卫星上（星上）执行基于神经网络的云检测的能力。

网络流程： 最后，我们简要描述卫星网络流程，作为本文其余部分的背景。用于地球观测的LEO卫星在极地轨道上运行，大约每1.5小时绕地球一圈。在每次轨道运行中，由于地球的自转，它们会飞越地球的不同区域。这些卫星的数据通常通过少数几个具有Gbps级别链路容量的专用地面站进行下载。由于卫星的轨道运动，它每天可以与每个地面站接触四到六次，每次接触持续长达十分钟。为了改善下载延迟并增加接触次数，近期的工作提出了包含多个通用地面站的分布式地面站设计 [60,61]。在工业界，亚马逊和微软已经推出了“地面站即服务”平台 [7,46]，卫星运营商可以在现有地面站上租用时间来从卫星下载数据。