Detecting Small Space Objects for Space Situational Awareness with Collaborative Edge AI

Xingyu Qu, Peng HuarXiv 2026 · arXiv

TL;DR

This paper introduces a collaborative Edge AI framework that allows clusters of satellites to work together to detect small space objects. By benchmarking various deep learning models and data fusion strategies, the authors demonstrate that while high-resolution imagery is vital for accuracy, "image-level fusion" offers the best balance of detection precision and low latency for time-sensitive space situational awareness.

What Problem Does This Paper Address?

As Earth's orbit becomes increasingly crowded, the risk of the "Kessler syndrome"—a cascade of satellite collisions—grows. To prevent this, satellites need robust Space Object Detection (SOD) systems. However, detecting small debris or distant satellites is incredibly difficult because they often appear as mere pixels against a vast background.

Current onboard systems face a "triple threat" of challenges:

  1. Data Scarcity: There are very few high-quality, labeled datasets of objects in space.
  2. Resolution vs. Speed: High-resolution images are needed to see small objects, but they take longer to process and transmit.
  3. Communication Bottlenecks: While using multiple satellites (multiview) improves accuracy, sending high-def images over Inter-Satellite Links (ISLs) can create significant delays.

Methods

The authors propose a cluster-based collaborative Edge AI system. In this setup, one "primary" satellite acts as a central node, receiving observations from several "support" satellites.

They evaluated three distinct fusion strategies to combine these multiple viewpoints:

  • Image-Level Fusion: Concatenating raw images from different satellites before feeding them into the AI model.
  • Feature-Level Fusion: Each satellite's view is processed through a neural network "backbone" separately, and the resulting data maps are merged.
  • Attention-Based Fusion: Using a "Query-Key-Value" mechanism (similar to Transformers) to let the model decide which parts of the support views are most relevant to the primary view.

Crucially, the paper also introduces a Communication Cost Model that accounts for ISL bandwidth, the speed of light (propagation delay), and "waiting times" that occur when a primary satellite's receiver is overwhelmed by too many incoming data streams at once.

Datasets

The researchers developed and released a new synthetic dataset called SODv3.

  • Source: Generated using the Unity engine to simulate realistic space environments.
  • Scale: 155,663 images organized into 32,538 synchronized multiview groups.
  • Quality: Features ultra-high-resolution (4K) images, which is rare for SOD datasets.
  • Utility: It includes metadata like observer positions and target distances, supporting both single-view and collaborative multiview research.

Key Insights and Ideas

  • Resolution is Non-Negotiable: For small objects, downsampling is a "feature killer." Increasing resolution from 1920p to 3840p improved detection accuracy (mAP50) by a staggering 138.2%.
  • The "More is Less" Paradox: Adding more viewpoints doesn't always help. For image-level fusion, using only 1–3 support views actually yielded better accuracy than using all eight available views. This suggests that "noisy" or redundant viewpoints can actually confuse the model.
  • The Communication-Computation Balance Point: There is a specific ISL bandwidth threshold where the time spent transmitting data equals the time spent running the AI. Below this point, the network is the bottleneck; above it, the onboard processor is the limit.
  • Reception Capacity Matters: Latency doesn't just scale with data size; it spikes sharply if the number of support satellites exceeds the primary satellite's concurrent reception capacity ($n_r$), forcing data to be sent in "rounds."

Key Results

  • Top Performer: Among the models tested (YOLOv9, v11, and v13), YOLOv11-S achieved the best balance, with an mAP50 of 69.47% at 4K resolution.
  • Efficiency: Image-level fusion was the clear winner for speed, reducing combined communication and inference latency by roughly 30–33% compared to feature-level or attention-based methods.
  • Latency: Under standard ISL constraints, multiview collaboration adds minimal overhead as long as the satellite's hardware can handle the concurrent data streams.

Limitations and Future Work

While the results are promising, the authors acknowledge a few hurdles:

  • Synthetic vs. Real: The models were trained on synthetic data. Real-world space conditions (sensor noise, unpredictable lighting) might degrade performance.
  • Bandwidth Savings: The authors suggest that future work should explore transmitting compressed feature maps instead of full images to further reduce the communication burden.
  • Dynamic Routing: The current model assumes a relatively stable link; future iterations need to account for the highly dynamic nature of laser-based ISLs in moving satellite constellations.

Summary updated

Abstract

With the rapid growth of space activities, near-Earth space is becoming increasingly congested, leading to a higher risk of spacecraft collisions and space debris generation. Recent close-proximity encounters and collision incidents necessitate efficient and cost-effective onboard systems for space object detection (SOD) and collision avoidance. While deep learning (DL) enabled Edge AI using low-cost vision sensors on single or multiple satellites has shown promises for SOD, significant challenges remain, including limited datasets, small SOD, and the lack of comprehensive performance assessments. This paper presents a collaborative Edge AI framework for multiview SOD and systematically benchmarks DL models and fusion strategies. Results show that higher image resolution significantly improves small SOD. For multiple views, image-level fusion reduces the combined communication and inference latency by 30.13% and 32.80% on average compared with feature-level and attention-based fusion across the evaluated inter-satellite link (ISL) bandwidth. For larger view sets, feature-level and attention-based fusion achieve better detection accuracy at the cost of analysis in the sensing-communication pipeline, revealing that multiview collaboration introduces minimal latency overhead under ISL constraints when reception capacity is not exceeded. Our work provides valuable design insights for future collaborative satellite network design, supporting the emerging space situational awareness and space traffic management requirements.
BibTeX
@misc{qu2026detecting,
  author = {Xingyu Qu and Peng Hu},
  title = {{Detecting Small Space Objects for Space Situational Awareness with Collaborative Edge AI}},
  year = {2026},
  eprint = {2611.19394},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.19394},
}
Figure