Methodology

We evaluate Kodan with multiple space data processing applications and on multiple hardware platforms. For input, we use a Sentinel dataset [14] with classification vector labels and per-pixel masks. This dataset contains 48% high-value (i.e., non-cloudy) data and 52% low-value (i.e., cloudy) data. For test applications, we use publicallyavailable semantic segmentation neural networks [43] customized to generate a per-pixel mask for each data sample. Table 1 summarizes these applications. We reserve a subset of the representative dataset for model validation. During training, we apply data augmentation to improve accuracy and avoid over-fitting.

We deploy each application to multiple hardware platforms. We run applications on an NVIDIA Jetson AGX Orin Tegra embedded GPU in its 15 W power mode — near the maximum reasonable power draw for a 3U cubesat subsystem. We also run applications on a Core i7-7800X CPU containing 12 cores clocked at 3.5 GHz and drawing around 140 W of power, and on a GeForce GTX 1070 Ti GPU drawing around 180 W of power. Both devices represent forward-looking computational hardware for the space edge.

Throughout our evaluation, we model orbital mechanics, data collection, and communication using the cote space computing simulation software [8]. We model the Landsat 8 orbit, camera sensor, and data frames by extending cote to import the Landsat World Reference System (WRS), and we log the image frame captures as the satellite passes over its ground track. We model the positions (latitude and longitude) and communication characteristics of the Landsat ground segment. Using cote, we compute the frame deadline for each satellite deployment based on its orbit characteristics.

我们在多个硬件平台上，使用多种空间数据处理应用对 Kodan 进行评估。输入方面，我们使用了一个带有分类向量标签和逐像素掩码的 Sentinel 数据集[14]。该数据集包含 48% 的高价值数据（即无云数据）和 52% 的低价值数据（即有云数据）。测试应用方面，我们使用了公开可用的语义分割神经网络[43]，并对其进行定制，使其能为每个数据样本生成一个逐像素掩码。表 1 总结了这些应用。我们保留了代表性数据集的一个子集用于模型验证。在训练过程中，我们应用了数据增强（data augmentation）以提高准确度并避免过拟合。

我们将每个应用部署到多个硬件平台上。我们在 NVIDIA Jetson AGX Orin Tegra 嵌入式 GPU 上以 15W 的功耗模式运行这些应用——该功耗接近一个 3U 立方星子系统所能承受的合理功耗上限。我们还在一个拥有 12 核、主频为 3.5 GHz、功耗约 140W 的 Core i7-7800X CPU 上，以及一个功耗约 180W 的 GeForce GTX 1070 Ti GPU 上运行了这些应用。这两种设备代表了未来空间边缘计算硬件的发展方向。

在整个评估过程中，我们使用 cote 空间计算模拟软件[8]来对轨道力学、数据收集和通信进行建模。我们通过扩展 cote 以导入 Landsat 世界参考系统（WRS），对 Landsat 8 的轨道、相机传感器和数据帧进行建模，并在卫星经过其星下点轨迹时记录图像帧的捕获。我们对 Landsat 地面段的位置（经纬度）和通信特性进行了建模。利用 cote，我们根据每个卫星部署的轨道特性，计算出其帧处理的截止时间（frame deadline）。