Methodology

We evaluate OEC systems running a remote sensing application on nanosatellite constellations.

We evaluate an OEC system in which each nanosatellite includes a Jetson TX2 module. Prior work shows that these systems remain effective in the space radiation environment [96]. Each nanosatellite collects data and either downlinks to a ground station or performs onboard machine inference. We use building footprint detection for the remote sensing application. We train the DetectNet [90] CNN on satellite images and ground-truth labels from the SpaceNet [94] dataset, and we evaluate performance on separate test data. The SpaceNet dataset is a collection of 0.3 m/px satellite images with labeled building footprints; we decimate the images to achieve higher GSDs. To evaluate the energy cost of computing on a satellite, we directly measure average power and latency of the inference application running on a Jetson TX2. We measure power with multimeters, recording current and voltage into the Jetson while workloads run from energy stored in a capacitor bank closely resembling our modeled power system. These operating energy values are an input to cote in its model of energy available to a nanosatellite during a deployment.

To quantify the limitations of bent-pipe architectures, we use cote to evaluate the performance of existing and future constellations. Space segments consist of polar (97.3°) orbits with 250 or 1000 satellites. This orbit is identical to one occupied by existing, deployed satellites; we use TLEs from nanosatellites operated by Planet. For each of the two space segments, we consider three constellation configurations, as described in Section 4.2: close-spaced, frame-spaced, and orbit-spaced. These configurations are compared to the current practice, bent-pipe configuration. We consider a polar ground segment consisting of two rings of ground stations, one at 87°N and one at 87°S. Each ring contains the same number of stations spaced evenly longitudinally.

The downlink frequency is centered at 8.15 GHz with a bandwidth of 20.0 MHz. We model nanosatellite patch antennas with a peak gain of 6.0 dB, and we model ground station receiving dishes with a peak gain of 44.1 dB.

我们对在纳米卫星星座上运行遥感应用的OEC（轨道边缘计算）系统进行评估。

我们评估的OEC系统中，每颗纳米卫星都包含一个Jetson TX2模块。先前的工作表明，这些系统在空间辐射环境中能保持有效性[96]。每颗纳米卫星收集数据后，要么下传到地面站，要么执行星上机器推断。我们选用建筑物轮廓检测作为遥感应用。我们使用 SpaceNet[94]数据集中的卫星图像和真值标签来训练DetectNet[90]卷积神经网络（CNN），并在独立的测试数据上评估其性能。

SpaceNet数据集是一组带有已标注建筑物轮廓的0.3米/像素的卫星图像；我们对这些图像进行降采样以获得更高的地面采样距离（GSD）值。为了评估星上计算的能源成本，我们直接测量了在Jetson TX2上运行推断应用的平均功率和延迟。我们使用万用表测量功率，记录当工作负载运行时，从一个与我们建模的电源系统非常相似的电容器组中输入到Jetson的电流和电压。这些运行时的能耗值，将作为cote模型中关于卫星在部署期间可用能量的输入参数。

为了量化“弯管”架构的局限性，我们使用cote来评估现有和未来星座的性能。空间部分由250颗或1000颗卫星组成的极地轨道（97.3°）星座构成。该轨道与现有已部署卫星所处的轨道相同；我们使用了由Planet公司运营的纳米卫星的两行轨道根数（TLEs）。对于这两种规模的空间部分，我们分别考虑了三种星座配置，如第4.2节所述：近距离间隔、帧间隔和轨道间隔。我们将这些配置与当前实践中的“弯管”配置进行比较。我们考虑一个由两个地面站环组成的极地地面部分，一个位于北纬87°，另一个位于南纬87°。每个环包含相同数量且在经度上均匀分布的地面站。

下行链路频率中心为8.15 GHz，带宽为20.0 MHz。我们建模的纳米卫星贴片天线峰值增益为6.0 dB，地面站的接收天线碟盘峰值增益为44.1 dB。