The First Verification Test of Space-Ground Collaborative Intelligence via Cloud-Native Satellites

Shangguang Wang, Qiyang Zhang, Ruolin Xing, Fei Qi, Mengwei Xu

TL;DR

This paper presents the first real-world deployment and verification of a cloud-native, space-ground collaborative AI inference system on Low Earth Orbit (LEO) satellites. By leveraging lightweight container orchestration, the proposed framework achieves a ~50% boost in object detection accuracy while slashing downlink data transmission by 90%.

What Problem Does This Paper Address?

Earth Observation (EO) satellites capture massive amounts of high-resolution imagery, but traditional "bent-pipe" satellite architectures must downlink all raw data to ground stations for processing. This creates a massive bottleneck due to limited downlink bandwidth, high latency, and communication link instability. Furthermore, much of this transmitted data is redundant or useless—for instance, up to 90% of raw imagery in certain regions is completely obscured by cloud cover.

While Orbital Edge Computing (OEC) aims to solve this by processing data directly on the satellite, existing onboard architectures are inflexible, difficult to update in orbit, and limited by the low accuracy of lightweight models that can fit on constrained satellite hardware.

Methods

The authors introduce a cloud-native satellite design that treats the satellite as an edge node integrated with ground-based cloud computing.

Figure 2. The design of the cloud-native satellite. The com- munication, computing, and storage capabilities are inte- grated into the cloud-native satellite.

The architecture is built on two open-source frameworks adapted for space:

  • KubeEdge: Provides lightweight container orchestration (EdgeCore) to manage applications on the satellite, ensuring offline autonomy and reliable data sync despite intermittent satellite-to-ground connections.
  • Sedna: An AI edge-cloud synergy suite that manages collaborative training and inference tasks.

The Collaborative Inference Workflow

Instead of relying on a single "one-size-fits-all" model, the authors implement a two-stage shallow-deep cooperative inference mechanism:

  1. Onboard Preprocessing: The satellite runs a lightweight, low-precision "shallow" model (YOLOv3-tiny) to detect features of interest and filter out redundant data (like clouds).
  2. Conditional Offloading: If the onboard model's prediction confidence is high, the results are sent directly to the ground. If the confidence is low, the satellite transmits the image fragments to the ground station, where a heavy, high-precision "deep" model (YOLOv3) performs exact detection.
Figure 5. The workflow of satellite-ground collaborative inference system architecture.

Additionally, because satellites cannot natively process massive high-resolution images due to memory constraints, the authors utilize an image-splitting strategy to break large remote sensing images into smaller fragments before running onboard inference.

Datasets

The framework was evaluated using DOTA (v1.0 and v1.5), a widely recognized large-scale dataset for object detection in aerial and remote sensing imagery. No new datasets were introduced.

Key Insights and Ideas

  • The Viability of Cloud-Native Space Software: The paper successfully demonstrates that standard cloud-native technologies (containers, microservices, and Kubernetes-style orchestration) can run reliably on resource-constrained CubeSats, paving the way for agile DevOps and continuous software updates in space.
  • Dynamic Resource Allocation: Rather than trying to force massive deep learning models onto low-power satellite hardware, the "shallow-deep" split efficiently balances the satellite's limited compute power against its limited downlink bandwidth.
  • Real-World Energy Quantified: The authors provide a rare look at actual in-orbit power consumption. Computing on the Baoyun satellite accounts for roughly 17% of its total energy budget (and 33% of its payload energy), proving that onboard AI is as much an energy management challenge as it is a computational one.

Key Results

The system was flight-tested on two operational LEO satellites (Baoyun and Chuangxingleishen) within the Tiansuan constellation.

  • Accuracy Improvements: Compared to purely performing inference in orbit, the collaborative system improved mean Average Precision (mAP) by 44% on DOTA v1.0 and 52% on DOTA v1.5 (averaging a ~50% overall improvement).
  • Bandwidth Savings: By filtering out redundant data and only downlinking low-confidence images, the system achieved a 90% reduction in data transmission volume.
Data filter rate across different image sizes on the DOTA dataset.
  • Power Metrics: Telemetry data from the Baoyun satellite showed that the onboard computing unit (a Raspberry Pi system) consumed 8.78W out of the satellite's total 51.07W power draw.

Limitations and Future Work

While this work successfully verifies a single satellite-to-ground collaborative link, it does not yet address the complexities of a fully distributed, multi-satellite constellation network where satellites must communicate peer-to-peer (inter-satellite links).

The authors note that future work will focus on deploying this framework in batches across the remaining satellites of the Tiansuan constellation to establish a true collaborative space computing network. They also aim to test the system's capabilities in specific verticals like precision agriculture, infrastructure monitoring, and humanitarian assistance.

Summary updated

Abstract

Recent advancements in satellite technologies and the declining cost of access to space have led to the emergence of large satellite constellations in Low Earth Orbit. However, these constellations often rely on bent-pipe architecture, resulting in high communication costs. Existing onboard inference architectures suffer from limitations in terms of low accuracy and inflexibility in the deployment and management of in-orbit applications. To address these challenges, we propose a cloud-native-based satellite design specifically tailored for Earth Observation tasks, enabling diverse computing paradigms. In this work, we present a case study of a satellite-ground collaborative inference system deployed in the Tiansuan constellation, demonstrating a remarkable 50\% accuracy improvement and a substantial 90\% data reduction. Our work sheds light on in-orbit energy, where in-orbit computing accounts for 17\% of the total onboard energy consumption. Our approach represents a significant advancement of cloud-native satellite, aiming to enhance the accuracy of in-orbit computing while simultaneously reducing communication cost.
BibTeX
@misc{wang2023first,
  author = {Shangguang Wang and Qiyang Zhang and Ruolin Xing and Fei Qi and Mengwei Xu},
  title = {{The First Verification Test of Space-Ground Collaborative Intelligence via Cloud-Native Satellites}},
  year = {2023},
  eprint = {2311.06078},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2311.06078},
}
Figure