Experimental Setup

We evaluate L2D2 using a combination of real-world link quality measurements and simulations as we describe below.

Link Quality Measurements

To evaluate L2D2’s capability to predict the quality of individual links, we collect real-world data from 16 ground station-satellite pairs operating in the X-band and one ground station-satellite pair operating in the Ka-band (see Table 1). Recall, X-band is the most popular downlink band for Earth imagery satellites today. The ground stations used to obtain this dataset cover a large geographical spread: Wisconsin, Hawaii, Antarctica, Guam, and Florida. These ground stations collect data from 5 satellite links (X-band: NOAA20/JPSS1, AQUA, TERRA, SNPP & Ka-band: NOAA-20/JPSS1) and measure the signal strength and SNR of the received signal in dBm, along with the elevation and azimuth angles of the satellite. In total, we compile data for 30 days worth of passes for each X-band ground station-satellite pair and 6 days worth of passes for the Ka-band ground station-satellite pair 2 . We augment our dataset using weather data – precipitation intensity, precipitation probability, and cloud cover obtained using the Dark Sky weather API [16].

链路质量测量 (Link Quality Measurements)

为了评估 L2D2 预测单条链路质量的能力，我们从16个工作在 X波段 的“地面站-卫星”对和一个工作在 Ka波段 的“地面站-卫星”对收集了真实世界的数据（见表1）。需要说明的是，X波段是当今地球影像卫星最常用的下行频段。用于获取该数据集的地面站覆盖了广阔的地理范围：威斯康星州、夏威夷、南极洲、关岛和佛罗里达州。这些地面站从5条卫星链路（X波段：NOAA-20/JPSS-1, AQUA, TERRA, SNPP；Ka波段：NOAA-20/JPSS-1）收集数据，并测量接收信号的信号强度和信噪比 (SNR)（单位：dBm），以及卫星的仰角和方位角。我们总共为每个X波段的“地面站-卫星”对汇编了30天的卫星过境数据，为Ka波段的“地面站-卫星”对汇编了6天的过境数据²。我们还使用了 通过 Dark Sky 天气API [16] 获取的天气数据 ——包括降水强度、降水概率和云量——来增强我们的数据集。

Large Scale Simulation

Since the deployment of X-band ground stations today is limited, we rely on amateur ground stations deployed in lower frequencies to evaluate the scheduling aspects of our design. Specifically, we evaluate L2D2 using data collected from deployments of the open-source SatNOGS ground stations [41]. SatNOGS is deployed by independent amateur radio enthusiasts using software-defined radios. The ground stations listen to low-bandwidth satellite broadcast signals primarily from government and academic satellites e.g. from NOAA weather satellites. The observation data is logged in a public database. We select the ground stations that are operational and have made at least 1k observations. In the filtered dataset, we have 173 ground stations (Fig. 5) and 259 satellites. We download the data from all ground station-satellite links for a month-long period. Then, we model L2D2 ground stations to be positioned at the same location and interacting with the same set of satellites. Since our simulations are based on positional and orbital information from real-world ground stations and satellites, our evaluation accurately models the geographical distribution of a network that has independently evolved over time.

A majority of SatNOGS ground stations operate in the sub-500 MHz frequency bands, and some (approx. 20%) support the L-band (1.5 to 1.75 GHz). Since Earth Observation satellites use the X-band (>8 GHz) [19, 34] to download their data, we cannot use the data from the SatNOGS database to get the SNR for satellite-ground station links. Therefore, we simulate data download behavior for each L2D2 ground station in the X-band. For SNR estimation, we use models trained in Sec. 3.2 to predict the SNR at each ground station. Specifically, each L2D2 ground station emulates the behavior (packet loss rates, SNR variation, reflections, etc.) of a randomly chosen ground station in Table 1.

For simulating data transfer, each satellite generates 100 GB of data per day. Since our ground stations are low-complexity, L2D2 ground stations do not use large dishes (5 m or more) typically used by commercial ground stations [34, 52]. We simulate our ground stations to have small, 1m diameter, dish antennas that can be deployed on rooftops or backyards and are similar to the setup deployed by SatNOGS operators. This reduces the SNR of each station by 6 dB. Furthermore, our ground stations use a singlechannel receiver, as opposed to six-channel receivers in centralized designs [19]. Finally, L2D2 computes the data download plan at the granularity of a day. Finer granularity than a day is possible but we haven’t explored this in L2D2.

大规模仿真 (Large Scale Simulation)

由于目前X波段地面站的部署有限，我们依赖于部署在较低频率的业余地面站来评估我们设计的调度方面。具体来说，我们使用 从开源的 SatNOGS 地面站[41]部署中收集的数据来评估 L2D2 。SatNOGS 由独立的业余无线电爱好者使用软件定义无线电 (SDR) 进行部署。这些地面站主要监听来自政府和学术卫星（例如NOAA气象卫星）的低带宽广播信号。其观测数据被记录在一个公开的数据库中。我们筛选出那些正在运行且已进行至少1000次观测的地面站。在筛选后的数据集中，我们有173个地面站（图5）和259颗卫星。我们下载了所有“地面站-卫星”链路长达一个月的数据。然后，我们将 L2D2 地面站建模为位于相同位置并与同一组卫星进行交互。由于我们的仿真是基于真实世界地面站和卫星的位置与轨道信息，因此我们的评估能够准确地模拟一个随时间独立演化形成的网络的地理分布。

大多数 SatNOGS 地面站工作在 500 MHz 以下的频段，部分（约20%）支持L波段（1.5至1.75 GHz）。由于对地观测卫星使用X波段（>8 GHz）[19, 34]来下载数据，我们不能直接使用 SatNOGS 数据库中的数据来获取“卫星-地面站”链路的信噪比。因此，我们为每个 L2D2 地面站仿真了在X波段的数据下载行为。在信噪比估计方面，我们使用在第3.2节中训练的模型来预测每个地面站的信噪比。具体而言，每个 L2D2 地面站会模拟（emulate）表1中随机选择的一个地面站的行为（如丢包率、信噪比变化、信号反射等）。

在数据传输仿真中，每颗卫星每天生成100 GB的数据。由于我们的地面站是低复杂度的，L2D2 地面站不像商业地面站那样使用大型碟形天线（5米或更大）[34, 52]。我们仿真中的地面站配备了可以部署在屋顶或后院的1米直径小型碟形天线，这与 SatNOGS 运营商部署的设备类似。这将每个站点的信噪比降低了 6 dB。此外，我们的地面站使用单信道接收器，而不是中心化设计中的六信道接收器[19]。最后，L2D2 以天为粒度计算数据下载计划。虽然可以实现更细的粒度，但在 L2D2 的当前工作中我们未进行探索。

Model Details

To train the link quality model, we use the last 25% of each dataset as our evaluation set. For X-band satellite, we use 3-weeks of training data and the last week as the test set, unless specified otherwise. We implement the deep learning component of our model in Keras and train our model using the Adam optimizer for 40 epochs in all trials. The model trains on an off-the-shelf Macbook in around 40 minutes and performs a prediction in sub- millisecond (without GPU). We convert estimated SNR to data rate using the 64K blocksize thresholds in a standard satellite data receiver [61].

模型细节 (Model Details)

在训练链路质量模型时，我们将每个数据集的后25%用作我们的评估集。对于X波段卫星，除非另有说明，我们使用3周的训练数据和最后1周作为测试集。我们在 Keras 中实现了模型的深度学习部分，并在所有试验中使用 Adam 优化器训练模型40个轮次 (epochs)。模型在一台商用现成的 Macbook 上训练约需40分钟，进行一次预测的时间在亚毫秒级别（未使用GPU）。我们使用一个标准卫星数据接收器[61]中 64K 块大小的阈值，将估计的信噪比转换为数据速率。

Baselines

Our primary baseline to compare L2D2’s performance is a centralized architecture that deploys the state-of-the-art ground stations described in [19]. This method uses 6 parallel channels as well as high-end receivers with 4m diameter dish antennas. As in [19], we model 5 such high-end ground stations across the planet. In contrast, each L2D2 ground station uses off-the-shelf components – net link capacity is 6𝑑𝐵 lower per link and we do not use any parallel channels. Therefore, each baseline ground station achieves 10x the median throughput achieved by a L2D2 node. Furthermore, L2D2 ground stations experience packet loss due to errors in link prediction. In contrast, the baseline stations are transmit-capable and do not experience such loss.

Our evaluation for link estimation compares against a wellstudied statistical model from the International Telecommunication Union (ITU) [31–33]. This model uses the distance between satellite and ground station, satellite elevation, and precipitation to mathematically model the link quality.

我们用于比较 L2D2 性能的主要基准是一个中心化架构，该架构部署了在[19]中描述的最先进 (state-of-the-art) 地面站。该方案使用6个并行信道以及配备4米直径碟形天线的高端接收器。与[19]中一样，我们在全球范围内建模了5个这样的高端地面站。相比之下，每个 L2D2 地面站使用商用现成组件 (off-the-shelf components)——每条链路的净容量低6dB，且我们不使用任何并行信道。因此，每个基准地面站能达到的吞吐量中位数是 L2D2 节点的10倍。此外，L2D2 地面站会因链路预测错误而产生丢包。相比之下，基准方案中的地面站具备发射能力，不会出现此类损失。

我们对链路估计的评估，是与一个被广泛研究的国际电信联盟 (ITU) 的统计模型[31-33]进行比较。该模型使用卫星与地面站之间的距离、卫星仰角和降水量来数学建模链路质量。