SpaceExit: Enabling Efficient Adaptive Computing in Space with Early Exits

Advances in satellite technology and reduced launch costs have led to a proliferation of Earth observation (EO) satellites in low-Earth orbit (LEO). These satellites generate massive high-resolution imagery, creating a significant downlink bottleneck due to limited satellite-to-ground communication bandwidth. While orbit edge computing (OEC) can reduce data volume, existing static approaches fail to adapt to the varying complexity of satellite imagery, resulting in limited system performance and inefficient resource utilization.

We therefore propose SpaceExit, an integrated system for efficient adaptive computing on satellites. SpaceExit introduces three key components: (1) a geospatial-contextual adaptive detector that leverages both visual semantics and geospatial context to adjust processing complexity for each image, (2) a complexity-driven adaptive task scheduler that partitions images into tiles and allocates inference tasks across onboard devices based on content complexity and device capabilities, and (3) a satellite resource adaptive controller that ensures safe and efficient execution under changing conditions. Evaluations of diverse satellite settings and hardware platforms demonstrate that SpaceExit increases the performance by 5.2%-37.6% compared with the SoTA design.

卫星技术的进步和发射成本的降低导致了近地轨道（LEO）地球观测（EO）卫星的激增。这些卫星产生大量高分辨率影像，但由于星地通信带宽有限，造成了严重的下行链路瓶颈。

虽然在轨边缘计算（OEC）可以减少数据量，但现有的静态方法无法适应卫星影像复杂度的变化，导致系统性能受限和资源利用效率低下。

因此，我们提出了SpaceExit，一个用于卫星高效自适应计算的集成系统。SpaceExit引入了三个关键组件:

（1）地理空间上下文自适应探测器: 该探测器利用视觉语义和地理空间上下文来为每个图像调整处理复杂度

（2）复杂度驱动的自适应任务调度器: 该调度器根据内容复杂度和设备能力将图像划分为图块，并在星上设备间分配推理任务

（3）卫星资源自适应控制器: 该控制器确保在变化的条件下安全高效地执行任务

对不同卫星设置和硬件平台的评估表明，与最先进的（SoTA）设计相比，SpaceExit将性能提升了5.2%-37.6%

Introduction

Recent advances in satellite technology and reduced launch costs have led to a significant increase in Earth observation (EO) satellites in low-Earth orbit (LEO) [37]. These advanced satellites, equipped with high-resolution cameras, can generate terabytes of imagery data per day, providing valuable insights for applications such as precision agriculture [4,29], disaster response [33,35], and environmental monitoring [28]. However, this massive data generation creates unprecedented challenges for satellite-based computing systems that differ fundamentally from terrestrial computing environments in three critical aspects.

First, satellite computing occurs in an extremely constrained and hostile environment. Satellites face severe thermal variations (e.g., -22°C to +77°C [34]) as they move between sunlight and Earth’s shadow during their orbital periods. Such extreme conditions can significantly impact onboard chips and other equipment. As a result, effective satellite computing requires careful thermal management strategies that extend beyond conventional resource optimization. For instance, existing analysis of operational LEO satellites shows that thermal constraints alone can reduce computation performance by up to 10% [39].

Second, satellite resources exhibit unique dynamic characteristics tied to orbital mechanics. As micro-datacenters in space, satellites face network and power constraints that differ fundamentally from terrestrial systems. Satellite-to-ground communication bandwidth varies dramatically (e.g., from 0 to 220 Mbps) based on orbital position and ground station visibility [1]. Similarly, available power fluctuates with solar panel orientation and eclipse periods. Traditional "bent pipe" approaches that simply downlink raw imagery become impractical under these constraints [10].

Third, satellite observation tasks present distinct workload patterns that challenge conventional computing paradigms. Continuous Earth monitoring generates spatially and temporally correlated image sequences with varying complexity from simple ocean scenes to intricate urban landscapes [37]. These images often contain redundant information across overlapping observation swaths, yet may also capture rapidly evolving phenomena requiring immediate analysis.

近年来，卫星技术的进步和发射成本的降低导致近地轨道（LEO）的地球观测（EO）卫星数量显著增加[37]。这些配备了高分辨率相机的先进卫星每天可产生数太字节（TB）的影像数据，为精准农业[4,29]、灾害响应[33,35]和环境监测[28]等应用提供了宝贵的见解。然而，海量数据的生成给星上计算系统带来了前所未有的挑战，这些挑战在三个关键方面与地面计算环境有着根本的不同。

首先，卫星计算在一个极其受限和恶劣的环境中进行。卫星在其轨道周期内穿梭于日光和地球阴影之间，面临着剧烈的温度变化（例如，-22°C至+77°C [34]）。这种极端条件会严重影响星上芯片和其他设备。因此，有效的卫星计算需要超越传统资源优化的精细热管理策略。例如，对在轨LEO卫星的现有分析表明，仅热约束就可能使计算性能降低高达10% [39]。

其次，卫星资源表现出与轨道力学相关的独特动态特性。作为太空中的微型数据中心，卫星面临的网络和电力约束与地面系统根本不同。星地通信带宽根据轨道位置和地面站可见性而急剧变化（例如，从0到220 Mbps）[1]。同样，可用功率也随着太阳能电池板的朝向和日食周期而波动。在这种约束下，简单地将原始影像下传的传统“弯管”（bent pipe）方法变得不切实际[10]。

第三，卫星观测任务 呈现出独特的工作负载模式, 对传统计算范式构成了挑战。连续的地球监测会生成空间和时间上相关的图像序列，其复杂度各不相同，从简单的海洋场景到错综复杂的城市景观[37]。这些图像在重叠的观测条带中通常包含冗余信息，但也可能捕捉到需要立即分析的快速演变现象。

While recent orbital edge computing (OEC) approaches [37] attempt to address these by performing onboard analysis, existing solutions rely largely on static processing pipelines that fail to adapt to the dynamic nature of both satellite operations and EO data [9–11]. They typically attempt to reduce computational overhead by selectively discarding imagery or employing varying models for different scenes [9,10]. Despite these methods leveraging prior knowledge about images to implement differentiated processing, they ultimately rely on static models. This inherent rigidity presents significant limitations in EO contexts, where scene complexity, illumination conditions, and orbital parameters undergo constant flux. Moreover, these systems cannot adaptively respond to runtime variations in both computational resources (like chips) and environmental factors which is a critical requirement for resource-constrained satellite platforms. As illustrated in Figure 1, an adaptive processing pipeline that adaptively modulates computational intensity based on real-time scene complexity and resource can achieve superior efficiency by enabling early termination for simpler inputs while maintaining high-fidelity processing for challenging scenarios.

虽然近期的在轨边缘计算（OEC）方法[37]试图通过执行星上分析来解决这些问题，但 现有解决方案主要依赖于静态处理流水线，无法适应卫星操作和EO数据的动态特性[9–11]。它们通常试图通过选择性地丢弃影像或对不同场景采用不同模型来减少计算开销[9,10]。

尽管 这些方法利用了关于图像的先验知识来实现差异化处理，但它们最终依赖于静态模型。

在场景复杂度、光照条件和轨道参数不断变化的EO背景下，这种固有的僵化性带来了显著的局限性。

此外，这些系统无法自适应地响应计算资源（如芯片）和环境因素的运行时变化，而这对于资源受限的卫星平台来说是一项关键要求。

如图1所示，一个能根据实时场景复杂度和资源状况自适应调节计算强度的自适应处理流水线，可以通过对简单输入实现提前终止，同时对复杂场景保持高保真度处理，从而实现卓越的效率:

We therefore propose SpaceExit, an integrated system that addresses all three satellite-specific characteristics through algorithm-system co-design. Unlike previous methods that merely simulate adaptive behavior through predefined rules, SpaceExit continuously adjusts its processing based on current input and system conditions, offering a truly flexible and efficient solution to the growing demands of EO missions.

At its core, SpaceExit features a geospatial-contextual adaptive detector that serves as the foundation for adaptive computation. This module incorporates a multi-scale feature extraction pipeline with an innovative exiting mechanism, adaptively adjusting processing complexity for each image region based on visual semantics and geospatial context. The system’s architecture enables efficient handling of both fine-grained details and large-scale patterns while allowing early termination for easily recognizable images. To enhance adaptability, a geospatial context component integrates auxiliary information such as geographical location, time of day, and historical data to guide exit decisions and optimize resource allocation. For example, it can prioritize more complex processing in urban areas while enabling quicker exits in open-water regions, thus optimizing resource allocation based on contextual information.

Complementing this adaptive processing capability, SpaceExit features a complexity-driven adaptive task scheduler designed to optimize onboard computing resource utilization, which is crucial for long-term stable execution of in-orbit computing tasks [39]. It employs an adaptive tiling strategy, partitioning input images into regions of varying sizes based on their estimated complexity. Then it implements an adaptive allocation scheme that distributes these tiles across available onboard devices, considering factors such as tile complexity, device capabilities, and current workload. Additionally, to ensure robust operation in the challenging space environment, SpaceExit incorporates a satellite resource adaptive controller. This component continuously monitors system parameters including computing load, bandwidth utilization, temperature, and energy consumption. It adaptively adjusts scheduling and exiting strategies to maintain safe operating conditions while maximizing processing throughput.

因此，我们提出了SpaceExit，一个通过算法-系统协同设计来解决所有三个卫星特有挑战的集成系统。与以往仅通过预定义规则模拟自适应行为的方法不同，SpaceExit能根据当前的输入和系统条件持续调整其处理过程，为日益增长的EO任务需求提供了一个真正灵活高效的解决方案。

(1) 地理空间上下文自适应探测器

SpaceExit的核心是一个作为自适应计算基础的地理空间上下文自适应探测器。该模块集成了 一个具有创新性退出机制的多尺度特征提取流水线，能根据视觉语义和地理空间上下文，自适应地调整每个图像区域的处理复杂度。

该系统架构能够高效处理细粒度细节和大规模模式，同时允许对易于识别的图像进行提前终止。

为增强适应性，一个地理空间上下文组件融合了地理位置、时间、历史数据等辅助信息，以指导退出决策并优化资源分配。例如，它可以在城市区域优先进行更复杂的处理，而在开阔水域则能更快地退出，从而根据上下文信息优化资源分配。

(2) 复杂度驱动的自适应任务调度器

为补充这种自适应处理能力，SpaceExit还配备了一个复杂度驱动的自适应任务调度器，旨在优化星上计算资源的利用率，这对于在轨计算任务的长期稳定执行至关重要[39]。它采用一种自适应分块策略，根据输入图像的预估复杂度将其 划分为不同大小的区域。 然后，它实施一种自适应分配方案，综合考虑图块复杂度、设备能力和当前工作负载等因素， 将这些图块分配到可用的星上设备上。

(3) 卫星资源自适应控制器

此外，为确保在充满挑战的太空环境中稳健运行，SpaceExit集成了一个卫星资源自适应控制器。该组件持续监控计算负载、带宽利用率、温度和能耗等系统参数。它能 自适应地调整调度和退出策略，以在最大化处理吞吐量的同时维持安全的操作条件。

Our comprehensive experiments simulate various scenarios and hardware configurations with real-world datasets. The results demonstrate that SpaceExit consistently outperforms existing static detection pipelines, increasing the performance by 24.3% on average compared with the SoTA design.

The contributions of this work are multifaceted. To the best of our knowledge, SpaceExit represents the first integrated system for adaptive onboard earth observation that is optimized across both algorithm and system levels for satellite deployment. The novel geospatial-contextual adaptive detector pushes the boundaries of efficient inference in resourceconstrained environments. The complexity-driven adaptive task scheduler offers a new paradigm for managing heterogeneous computing resources in space, while the satellite resource adaptive controller ensures reliable operation under the unique constraints of orbital platforms.

我们通过全面的实验，使用真实世界的数据集模拟了各种场景和硬件配置。结果表明，SpaceExit的性能始终优于现有的静态检测流水线，与最先进的（SoTA）设计相比，平均性能提升了24.3%

这项工作的贡献是多方面的。据我们所知，SpaceExit是首个在算法和系统层面都为卫星部署进行优化的集成式自适应星上地球观测系统。新颖的地理空间上下文自适应探测器推动了资源受限环境下高效推理的边界。复杂度驱动的自适应任务调度器为管理太空中的异构计算资源提供了新的范式，而卫星资源自适应控制器则确保了在轨道平台的独特约束下的可靠运行

Background and Challenges

2.1 Static OEC Systems

To address the limitations of classic "bent-pipe" approaches, OEC is proposed to process data onboard satellites before transmission (Figure 2). Existing OEC systems mainly rely on static methods for data processing. TargetFuse [40] deploys a DNN model serving as a counter, while Kodan [9] provides several context-based models to yield higher accuracy in different scenes. These static approaches also adopt selective discarding techniques or data deduplication, achieving significant data reduction. However, these methods struggle to adapt to the diverse and dynamic nature of Earth observation data, often leading to suboptimal results when confronted with unexpected or complex scenes.

A major limitation of multi-model systems for satellitebased image processing is the substantial overhead associated with model loading and storage. Satellites have limited onboard storage capacity, constraining the number and complexity of models that can be deployed [3]. Our analysis, as shown in Figure 3a, reveals that the time spent loading different object detection models is often much longer than their computation time. This overhead significantly impacts system efficiency and responsiveness. Furthermore, the typical approach of applying simple models first and escalating to more complex ones if necessary can result in significant delays and computational waste. If a simple model proves inadequate for a given scene, the resources expended on its application and the time spent loading subsequent models are essentially lost. These limitations force a trade-off between model diversity and processing capability, highlighting the need for a more efficient and adaptive onboard processing approach.

Another critical drawback of static methods is their reliance on prior knowledge, which may be inaccurate or unavailable in the dynamic domain of Earth observation [7]. Pre-defined rules for scene classification or object detection may fail when confronted with unexpected phenomena or rapidly changing environments. Our analysis, illustrated in Figure 3b, demonstrates this vulnerability. We trained two models specifically designed for Building and Transportation detection. When the prior knowledge aligns with the input data, these models achieve high performance. However, when confronted with unexpected scenes or objects, their performance drops significantly. This inflexibility leads to suboptimal resource utilization across heterogeneous scenes, as the system cannot adaptively adjust its processing strategy based on the current input and operational context.

为解决传统“弯管”（bent-pipe）方法的局限性，学界提出了在轨边缘计算（OEC），即在数据传输前先在卫星上进行处理（图2）。现有的OEC系统主要依赖静态方法进行数据处理。例如，TargetFuse [40] 部署一个用作计数器的DNN模型，而Kodan [9] 则提供多个基于上下文的模型，以在不同场景中获得更高的准确性。这些静态方法也采用选择性丢弃或数据去重技术，实现了显著的数据量削减。然而，这些方法难以适应地球观测数据多样化和动态化的特性，在面对意料之外或复杂的场景时，其结果往往不是最优的。

对于星上图像处理，多模型系统的一个主要限制是与模型加载和存储相关的巨大开销。卫星的星上存储容量有限，这限制了可部署模型的数量和复杂性[3]。

我们的分析（如图3a所示）揭示， 加载不同目标检测模型所花费的时间通常远超其计算时间。

这种开销严重影响了系统的效率和响应速度。此外， 那种先应用简单模型，必要时再升级到更复杂模型的典型方法，可能导致严重的延迟和计算浪费。如果一个简单模型被证明不适用于某个给定场景，那么花在其应用上的资源以及加载后续模型所耗费的时间基本上都被浪费了。 这些限制迫使我们在模型多样性和处理能力之间做出权衡，凸显了开发一种更高效、更具自适应性的星上处理方法的必要性。

静态方法的另一个关键缺点是它们对先验知识的依赖，而在地球观测这一动态领域，先验知识可能不准确或无法获得[7]。 为场景分类或目标检测预定义的规则在遇到意外现象或快速变化的环境时可能会失效。

我们的分析（如图3b所示）证明了这一脆弱性。我们训练了两个专门用于检测“建筑”和“交通工具”的模型。当先验知识与输入数据吻合时，这些模型表现出色。然而，当面对意料之外的场景或物体时，它们的性能会急剧下降。这种僵化性导致了在异构场景下资源的次优利用，因为系统无法根据当前的输入和运行环境自适应地调整其处理策略。

2.2 Challenges of Adaptive OEC

While implementing a truly adaptive method for Earth observation processing holds great potential, implementing such a system presents several significant challenges:

Challenge-1: Adapting multi-scale information and geospatial context for object detection in OEC. Dynamic object detectors have shown great promise in non-satellite scenarios, where they can adjust computation based on the complexity of the input data. However, applying these methods to satellite imagery introduces additional challenges. Satellite imagery requires the processing of multi-scale image information and incorporation of geospatial context, which is not typically considered in traditional methods. Furthermore, the constraints of the satellite environment, such as limited computational resources and the need for real-time processing, make dynamic methods in satellites more energy-sensitive compared to ground-based systems. Developing an adaptive approach that can adjust computation cost without sacrificing accuracy while leveraging both visual and geospatial information is a complex task requiring careful algorithm design.

Challenge-2: Scheduling dynamic task across heterogeneous onboard devices. Unlike static methods where processing requirements can be predicted in advance, adaptive approaches must continuously adjust resource allocation based on current input and system state. This is particularly challenging in the satellite environment where multiple heterogeneous computing devices with varying capabilities are often deployed. Designing a scheduling algorithm that can efficiently distribute workload across these devices while adapting to dynamic processing requirements is non-trivial.

Challenge-3: Runtime resource management under varying onboard constraints. The ever-changing space environment and varying mission requirements pose unique difficulties in runtime management. Satellites face severe constraints such as bandwidth, power, and thermal budget, which fluctuate based on orbital position and mission phase. Additionally, onboard resources like power and memory are limited and must be carefully managed. Prior work [39] has shown that these constraints can significantly impact the stablitity and performance of onboard processing systems. Ensuring safe and efficient execution under these dynamic conditions requires careful design of the resource management system.

虽然为地球观测处理实现一种真正的自适应方法潜力巨大，但部署这样的系统也面临若干重大挑战：

挑战一：在OEC中为目标检测适配多尺度信息和地理空间上下文 动态目标探测器在非卫星场景中已展现出巨大潜力，它们可以根据输入数据的复杂度来调整计算量。然而，将这些方法应用于卫星影像会引入额外的挑战。卫星影像需要处理多尺度图像信息并融合地理空间上下文，而传统方法通常不考虑这些因素。此外，卫星环境的约束，如有限的计算资源和实时处理的需求，使得卫星上的动态方法比地面系统对能耗更为敏感。因此，开发一种既能利用视觉和地理空间信息，又能在不牺牲精度的前提下调整计算成本的自适应方法，是一项需要精心进行算法设计的复杂任务。

挑战二：在星上异构设备间调度动态任务 与静态方法中可以预先预测处理需求不同，自适应方法必须根据当前的输入和系统状态持续调整资源分配。这在卫星环境中尤其具有挑战性，因为卫星上通常部署了多个能力各异的异构计算设备。设计一种能够高效地在这些设备间分配工作负载，同时适应动态处理需求的调度算法，绝非易事。

挑战三：在多变的星上约束下进行运行时资源管理 不断变化的空间环境和多样的任务需求给运行时管理带来了独特的困难。卫星面临着带宽、功率和热预算等严格约束，这些约束会根据轨道位置和任务阶段而波动。此外，星上的功率和内存等资源有限，必须得到精细管理。先前的工作[39]已经表明，这些约束会严重影响星上处理系统的稳定性和性能。要确保在这些动态条件下安全高效地执行任务，需要对资源管理系统进行周密的设计。

SpaceExit System Design

The SpaceExit system architecture is designed to address the key challenges of onboard object detection on satellites holistically. As shown in Figure 4, SpaceExit takes the high-resolution satellite images as input and processes them with an efficient adaptive object detector on heterogeneous onboard computing devices.

SpaceExit的系统架构旨在整体性地解决卫星上目标检测的关键挑战。如图4所示，SpaceExit接收高分辨率卫星图像作为输入，并通过一个高效的自适应目标探测器在星上异构计算设备上对它们进行处理。

Specifically, SpaceExit contains three main modules that synergistically optimize the detection pipeline across various dimensions of algorithm-system co-design:

• Geospatial-Contextual Adaptive Detector (GCAD, Section 3.1): This module forms the core of SpaceExit’s adaptive processing capability to address Challenge-1. It employs a multi-scale feature backbone with a novel exiting mechanism that leverages both visual semantics and geospatial context. The GCAD can adaptively adjust the processing complexity for each tile, allowing for efficient early exits on simpler areas while dedicating more resources to complex regions. This approach enables SpaceExit to balance detection accuracy and computational efficiency across diverse satellite imagery.

• Complexity-Driven Adaptive Task Scheduler (CATS, Section 3.2): The CATS optimizes the utilization of onboard computing resources across multiple devices to address Challenge-2. It implements a complexity-aware scheduling algorithm that partitions input images into tiles based on their estimated complexity. These tiles are then allocated across available onboard devices, considering factors such as tile complexity, device capabilities, and current workload. This adaptive approach ensures efficient load balancing and maximizes overall system throughput.

• Satellite Resource Adaptive Controller (SRAC, Section 3.3): The SRAC is responsible for maintaining safe and efficient system operation under the dynamic constraints of the space environment to address Challenge-3. It continuously monitors system parameters including computing load, temperature, and energy consumption. Based on these inputs, the SRAC adaptively adjusts system configurations and exiting strategies to optimize resource utilization while ensuring the system operates within safe limits.

These three modules work together to enable SpaceExit’s efficient and adaptive approach to onboard object detection. The GCAD provides content-aware processing, the CATS ensures optimal resource utilization across devices, and the SRAC maintains system stability and efficiency in changing environmental conditions.

具体而言，SpaceExit包含三个主要模块，它们通过算法-系统协同设计，在多个维度上协同优化检测流水线：

• 地理空间上下文自适应探测器 (Geospatial-Contextual Adaptive Detector, GCAD, 3.1节): 该模块是SpaceExit自适应处理能力的核心，旨在解决挑战一

  • 它采用一个带有创新性退出机制的多尺度特征主干网络，该机制同时利用了视觉语义和地理空间上下文
  • GCAD能为每个图像图块自适应地调整处理复杂度，从而在较简单的区域实现高效的提前退出（early exits），同时将更多资源投入到复杂区域
  • 这种方法使SpaceExit能够在多样化的卫星影像中平衡检测精度和计算效率

• 复杂度驱动的自适应任务调度器 (Complexity-Driven Adaptive Task Scheduler, CATS, 3.2节): CATS旨在优化星上多设备计算资源的利用率，以解决挑战二

  • 它实现了一种复杂度感知的调度算法，该算法根据预估的复杂度将输入图像划分为不同的图块
  • 然后，综合考虑图块复杂度、设备能力和当前工作负载等因素，将这些图块分配到可用的星上设备上
  • 这种自适应方法确保了高效的负载均衡，并最大化了整体系统吞吐量

• 卫星资源自适应控制器 (Satellite Resource Adaptive Controller, SRAC, 3.3节): SRAC负责在太空环境的动态约束下维持系统安全高效的运行，以解决挑战三

  • 它持续监控包括计算负载、温度和能耗在内的系统参数
  • 基于这些输入，SRAC自适应地调整系统配置和退出策略，以优化资源利用，同时确保系统在安全阈值内运行

这三个模块协同工作，共同实现了SpaceExit高效且自适应的星上目标检测方法。GCAD提供内容感知的处理，CATS确保跨设备的最优资源利用，而SRAC则在变化的环境条件下维持系统的稳定性和效率。

3.1 ~ 3.3

建模与算法细节, omitted here

Experimental Setup

4.1 Computational Hardware

We design and implement a heterogeneous computing testbed that faithfully emulates the resource-constrained environment of 3U CubeSat onboard computers. Our experimental platform incorporates two primary computing units that represent typical space-grade hardware configurations, carefully selected to balance performance and power constraints.

Firstly, we employ the NVIDIA Jetson Nano, which is equipped with a Quad-core ARM Cortex-A57 MPCore processor clocked at 1.43 GHz, paired with a 128-core Maxwell GPU that delivers 472 GFLOPS of compute performance. The system is supported by 4 GB of LPDDR4x memory, offering 25.6 GB/s bandwidth. Operating in a 10W TDP configuration, the Jetson Nano supports both FP16 and FP32 precision computations, making it suitable for efficient neural network inference within a compact 70 mm × 45 mm form factor [27]. Additionally, we utilize the NVIDIA Jetson Xavier NX, which offers higher performance in a similar size to the Jetson Nano. It features a 6-core NVIDIA Carmel ARMv8.2 64-bit CPU, a 384-core Volta GPU with 48 Tensor Cores, and 8 GB of LPDDR4x memory with 59.7 GB/s bandwidth. It delivers up to 21 TOPS of AI performance and supports TDP configurations of 10W, 15W and 20W, making it ideal for demanding AI applications in space-constrained environments.

To emulate a 3U CubeSat onboard computer with heterogeneous payloads, we setup an NVIDIA Jetson Nano with a 10W power mode connected to NVIDIA Jetson Xavier NX with 20W power mode. This configuration provides substantial AI computing power within the 4kg mass and 30W power budgets of a typical 3U bus [11], making it faithfully model the real onboard setup.

我们设计并实现了一个异构计算测试平台，该平台忠实地模拟了3U立方星（CubeSat）星上计算机的资源受限环境。我们的实验平台集成了两个主要的计算单元，它们代表了典型的航天级硬件配置，并经过精心选择以平衡性能和功耗限制。

首先，我们采用NVIDIA Jetson Nano。它配备了一个主频为1.43 GHz的四核ARM Cortex-A57 MPCore处理器，并搭配一个128核的Maxwell GPU，可提供472 GFLOPS的计算性能。

该系统由4 GB的LPDDR4x内存支持，带宽为25.6 GB/s。Jetson Nano在10W的热设计功耗（TDP）配置下运行，支持FP16和FP32精度计算，使其非常适合在70 mm × 45 mm的紧凑尺寸内进行高效的神经网络推理[27]。

此外，我们还使用了NVIDIA Jetson Xavier NX，它在与Jetson Nano相似的尺寸下提供了更高的性能。它配备一个6核NVIDIA Carmel ARMv8.2 64位CPU，一个带有48个张量核心（Tensor Cores）的384核Volta GPU，以及8 GB的LPDDR4x内存，带宽为59.7 GB/s。它能提供高达21 TOPS的人工智能性能，并支持10W、15W和20W的TDP配置，是空间受限环境中要求苛刻的AI应用的理想选择。

为了模拟一个带有异构载荷的3U立方星星上计算机，我们将一个设置为10W功耗模式的NVIDIA Jetson Nano与一个设置为20W功耗模式的NVIDIA Jetson Xavier NX连接。这种配置在典型3U卫星平台4公斤的质量预算和30W的功率预算内，提供了强大的人工智能计算能力[11]，从而忠实地模拟了真实的星上部署环境。

4.2 Satellite dataset

Following previous research [40], we evaluate SpaceExit using the DOTA (Dataset of Object deTection in Aerial images) dataset [38], a comprehensive satellite imagery dataset designed specifically for object detection tasks. The dataset contains 403k annotated instances across 15 object categories, including various vehicles, ships and infrastructures, with objects captured at arbitrary orientations to reflect real-world conditions. We preprocess to resemble the conditions encountered in EO applications closely.

我们遵循先前的研究[40]，使用DOTA（Dataset of Object deTection in Aerial images）数据集[38]来评估SpaceExit。DOTA是一个专为目标检测任务设计的综合性卫星影像数据集。该数据集包含15个物体类别下的40.3万个标注实例，涵盖了各种车辆、船只和基础设施，并且物体以任意方向被捕捉，以反映真实世界的条件。我们对数据进行了预处理，以使其更贴近地球观测（EO）应用中遇到的实际情况。

4.3 SpaceExit Implementation

In SpaceExit, the adaptive router takes the terrain-type embedding (one-hot vector) and four visual embeddings as input. The selected scale then passes through a feature fusion module to combine the coarse semantics and fine details. The terrain type is predicted by a pre-trained MobileNet classifier fine-tuned on the DOTA images with ground-truth land cover labels. We use 2 categories: land and ocean, which are visually distinctive. This router requires only about 130KB of memory (approximately 2.4% of the full model’s memory footprint), which is negligible.

We construct our geospatial embeddings by concatenating terrain type, land cover, and pre-processed POI data. Each geospatial embedding requires at most 32B of storage, resulting in a total database size under 100MB to cover the entire Earth. For single-orbit operations, this requirement decreases significantly to less than 100 KB. Given the inherent spatial locality of satellite imagery, our system retrieves 50 geospatial embeddings simultaneously. This efficient batching strategy means database access is required only once per minute, rendering the access time negligible in our overall processing pipeline. As for the complexity estimation model, we experimented with both a simple two-layer CNN and a color variation-based method. While both performed well, the latter proved significantly more efficient, leading us to adopt it as our default model.

The model is trained end-to-end with Adam optimizer at 1e-4 learning rate for 100 epochs. We use 0.1 weight decay and a cosine learning rate schedule with 5 epochs of warmup. The input size is 2048×2048 pixels, randomly cropped from the full images. The batch size is 8 per GPU (32 total) with synchronized BN. For augmentation, we apply HSV jittering, flipping, translation, and multi-scale training.

在SpaceExit中，自适应路由器的输入包括地形类型嵌入（独热向量）和四个视觉嵌入。被选中的尺度随后会通过一个特征融合模块，以结合粗粒度的语义和精细的细节。地形类型由一个预训练的MobileNet分类器预测，该分类器在带有真实地表覆盖标签的DOTA图像上进行了微调。我们使用“陆地”和“海洋”这两个视觉上易于区分的类别。该路由器的内存需求仅约130KB（大约占完整模型内存占用的2.4%），可以忽略不计。

我们通过拼接地形类型、地表覆盖和预处理的兴趣点（POI）数据来构建地理空间嵌入。每个地理空间嵌入最多需要32B的存储空间，这意味着覆盖整个地球的总数据库大小低于100MB。对于单轨道操作，这一需求显著降低至小于100KB。考虑到卫星影像固有的空间局部性，我们的系统可以一次性检索50个地理空间嵌入。这种高效的批处理策略意味着每分钟仅需访问一次数据库，使得访问时间在我们的整体处理流水线中可以忽略不计。至于复杂度估计模型，我们实验了一种简单的双层CNN方法和一种基于颜色变化的方法。虽然两者表现都很好，但后者被证明效率要高得多，因此我们采纳其为我们的默认模型。

模型使用Adam优化器进行端到端的训练，学习率为1e-4，共训练100个周期。我们使用0.1的权重衰减和一个带有5个周期预热（warmup）的余弦学习率调度策略。输入尺寸为2048×2048像素，从完整图像中随机裁剪。每个GPU的批处理大小为8（总计32），并采用同步批归一化（BN）。在数据增强方面，我们应用了HSV抖动、翻转、平移和多尺度训练。

4.4 Baselines

We evaluate SpaceExit against four satellite image processing systems that represent the state-of-the-art system in OEC:

• BentPipe [11] implements the traditional ground-centric approach, where satellites downlink raw imagery to ground stations for processing. This system maximizes processing capability but is severely limited by communication bandwidth constraints.

• SpaceOnly [11] performs all object detection computations onboard using deployed neural networks, transmitting only the detection results to ground stations. While this minimizes bandwidth usage, it faces significant onboard computational resource constraints.

• Kodan [9] leverages specialized machine learning models selected based on prior knowledge about observation targets. This approach improves efficiency through targeted processing but is heavily dependent on the accuracy of prior information.

• TargetFuse [40] employs a combination of adaptive tiling, object clustering, and bandwidth-aware transmission scheduling to minimize detection errors under resource constraints. This hybrid approach balances onboard processing with efficient data transmission.

These baselines, ranging from fully ground-based to completely onboard processing, provide comprehensive reference points for evaluating SpaceExit ’s performance.

我们将SpaceExit与四个代表了OEC领域最先进水平的卫星图像处理系统进行评估对比：

• BentPipe [11]：该系统实现了传统的以地面为中心的方法，即卫星将原始影像下传至地面站进行处理
  • 这种系统最大化了处理能力，但受到通信带宽的严重限制
• SpaceOnly [11]：该系统使用部署好的神经网络在星上执行所有目标检测计算，仅将检测结果传输到地面站
  • 虽然这最大限度地减少了带宽使用，但它面临着严峻的星上计算资源限制
• Kodan [9]：该系统利用根据观测目标的先验知识选择的专用机器学习模型
  • 这种方法通过 目标明确的处理 提高了效率，但严重依赖于先验信息的准确性
• TargetFuse [40]：该系统采用自适应分块、目标聚类和带宽感知的传输调度相结合的方法，以在资源受限的情况下最小化检测错误
  • 这种混合方法在星上处理和高效数据传输之间取得了平衡

这些基准系统涵盖了从完全地面处理到完全星上处理的各种范式，为评估SpaceExit的性能提供了全面的参考点

4.5 Metrics

For detection performance, we adopt the standard MS COCO evaluation protocol, reporting mean Average Precision (mAP) at different Intersections over Union (IoU) thresholds. Specifically, we report the standard MS COCO metrics [24], including AP@IoU=0.5 and AP@IoU=0.5:0.95 for each category.

For efficiency, we measure the end-to-end processing time and total charge on the satellite system. This includes the time taken for taking photos, inference, and the communication of information from the satellite to the ground station. We use the total amount of information transmitted from the satellite to the ground station as the benchmark for evaluating the system’s processing speed.

To evaluate the overall performance of our system, we use goodput which combines both accuracy and throughput. The goodput of a system is defined as the correct information produced per unit time. Goodput takes into account not only the amount of data processed but also the quality and relevance of the processed data, ensuring that our system meets the stringent requirements of satellite-based Earth observation.

在检测性能方面，我们采用标准的MS COCO评估协议，报告在不同交并比（Intersection over Union, IoU）阈值下的平均精度均值（mean Average Precision, mAP）。具体而言，我们报告了每个类别的标准MS COCO指标[24]，包括\(AP@IoU=0.5\)和\(AP@IoU=0.5:0.95\)

在效率方面，我们测量了端到端的处理时间以及卫星系统的总能耗。这包括了拍摄照片、执行推理以及信息从卫星到地面站的通信所花费的时间。我们使用从卫星传输到地面站的总信息量作为评估系统处理速度的基准

为了评估我们系统的综合性能，我们使用有效吞吐量（goodput）这一指标，它结合了准确度和吞吐量。 一个系统的有效吞吐量被定义为单位时间内产生的正确信息量。有效吞吐量不仅考虑了处理的数据量，还考虑了已处理数据的质量和相关性， 从而确保我们的系统能满足星基地球观测的严苛要求

Related Work

6.1 Orbitital Edge Computing

Orbital Edge Computing (OEC) has emerged as a promising approach to address the growing challenges of Earth observation data processing and transmission [9,11,37,39]. The foundational concept and architecture of OEC were introduced by Denby and Lucia, who demonstrated that processing remote sensing images onboard satellites can significantly alleviate the communication bottleneck between space and ground segments [11]. Subsequent research has expanded the scope of OEC, with numerous systems built to process Earth observation tasks [8,9,40]. For instance, Kodan maximizes the utility of saturated satellite downlinks while mitigating the computational bottleneck [9], while EagleEye proposes a mixed-resolution, leader-follower constellation design [8]. Recent proposals have even discussed the possibility of building microdatacenters with 4KW power in space [5], further expanding the potential capabilities of OEC.

While these OEC approaches have made significant progress in enabling onboard data processing and reducing transmission volumes, they generally rely on fixed processing pipelines or pre-defined decision criteria. This limits their adaptability to the diverse and dynamic EO tasks.

在轨边缘计算（Orbital Edge Computing, OEC）已成为一种应对日益增长的地球观测数据处理与传输挑战的有效方法 [9,11,37,39]。OEC的基础概念和架构由Denby和Lucia提出，他们证明了在卫星上处理遥感图像可以显著缓解天地通信链路的瓶颈 [11]。后续研究扩展了OEC的范围，并构建了众多系统来处理地球观测任务 [8,9,40]。例如，Kodan系统在缓解计算瓶颈的同时，最大化了饱和卫星下行链路的效用 [9]，而EagleEye则提出了一种混合分辨率的“领导者-跟随者”星座设计 [8]。近期的提案甚至讨论了在太空中建造功率达4KW的微型数据中心的可能性 [5]，进一步扩展了OEC的潜在能力。

尽管这些OEC方法在实现星上数据处理和减少传输数据量方面取得了显著进展，但它们通常依赖于固定的处理流水线或预定义的决策标准。这限制了它们对多样化和动态化的地球观测任务的适应能力。

6.2 Resource Management in Satellite Systems

Resource management in satellite systems, particularly in OEC contexts, faces fundamental challenges in balancing limited bandwidth, computation, and energy resources [7, 12]. Research has explored various approaches to optimize computation offloading and resource allocation, including deep reinforcement learning and game theory [14,41]. Energy management remains critical for solar-powered satellite operations [42], with optimization frameworks employing techniques like mixed-integer programming to balance latency, energy consumption, and computational throughput [6].

Despite these advancements, current resource management strategies for satellite systems often lack close integration with the specific requirements of dynamic onboard image processing and object detection tasks.

卫星系统中的资源管理，尤其是在OEC背景下，面临着在有限的带宽、计算和能源资源之间取得平衡的根本性挑战 [7, 12]。研究人员已探索了多种方法来优化计算卸载和资源分配，包括深度强化学习和博弈论 [14,41]。对于太阳能供电的卫星操作而言，能源管理仍然至关重要 [42]，相关优化框架采用混合整数规划等技术来平衡延迟、能耗和计算吞吐量 [6]。

尽管取得了这些进展，但当前卫星系统的资源管理策略通常缺乏与动态星上图像处理和目标检测任务的具体需求进行紧密集成。

6.3 Dynamic Neural Network Inference

Dynamic neural network inference has emerged as a promising approach to balance accuracy and efficiency in various applications [15,18,30,31]. Early exit networks, which dynamically terminate inference at earlier layers, represent a significant advancement in this field [22]. Several methods have been proposed in this direction, including BranchyNet [32], SkipNet [36], MSDNet [21], LoRAExit [26] and MMexit [17]. These approaches typically add early exit points to neural networks, enabling dynamic inference times for classification tasks and reducing average computation time for easily classifiable inputs. Few works have extended dynamic inference to object detection. For instance, DynamicDet employs an adaptive router to analyze multi-scale information and automatically determine the inference route [25]. The computational efficiency of these methods has led to their adoption in resource-constrained environments [16,19].

Despite these advancements, current dynamic inference methods primarily target classification tasks and general-purpose computing environments. They have yet to fully address the unique challenges posed by object detection in satellite imagery or the specific constraints of orbital edge computing systems.

动态神经网络推理已成为在各种应用中平衡准确性与效率的一种很有前景的方法 [15,18,30,31]。 能够动态地在较早的网络层终止推理的提前退出网络（Early exit networks）是该领域的一大进步 [22]。

沿着这个方向已提出了多种方法，包括BranchyNet [32]、SkipNet [36]、MSDNet [21]、LoRAExit [26]和MMexit [17]。这些方法 通常在神经网络中添加提前退出点，从而为分类任务实现动态的推理时间，并减少对易于分类的输入的平均计算时间。

少数工作已将动态推理扩展到目标检测领域。例如，DynamicDet采用一个自适应路由器来分析多尺度信息并自动确定推理路径 [25]。这些方法的计算效率使其在资源受限的环境中得到了应用 [16,19]。

尽管取得了这些进展，但当前的动态推理方法主要针对分类任务和通用计算环境。它们尚未完全解决卫星影像目标检测所带来的独特挑战，也未充分考虑在轨边缘计算系统的特定约束。

Conclusion

In this paper, we presented SpaceExit, an integrated system that enables efficient adaptive Earth observation on satellites. By leveraging a novel multi-scale feature backbone with adaptive exiting, a unified adaptive scheduling framework, and a runtime resource management module, SpaceExit addresses the unique challenges of orbital edge computing. Our comprehensive evaluations across diverse satellite settings and hardware platforms demonstrate that SpaceExit significantly outperforms existing methods. By enabling adaptive, efficient onboard processing, SpaceExit paves the way for a new generation of autonomous and capable Earth observation satellites. It demonstrates the potential to maximize the value of limited downlink bandwidth, opening new possibilities for space-based data analytics and real-time global monitoring.

在本文中，我们提出了SpaceExit，一个能够在卫星上实现高效自适应地球观测的集成系统。通过利用一种新颖的带有自适应退出的多尺度特征主干网络、一个统一的自适应调度框架以及一个运行时资源管理模块，SpaceExit解决了在轨边缘计算的独特挑战。我们在多样化的卫星设置和硬件平台上进行的综合评估表明，SpaceExit的性能显著优于现有方法。通过实现自适应、高效的星上处理，SpaceExit为新一代自主、功能强大的地球观测卫星铺平了道路。它展示了最大化有限下行链路带宽价值的潜力，为天基数据分析和实时全球监测开辟了新的可能性。