Onboard Neuromorphic Split Computing via Optical Links for LEO Remote Sensing (2507.08490v1)

Abstract: Low Earth orbit (LEO) satellite constellations increasingly require onboard intelligence under strict power and communication constraints. This paper proposes a neuromorphic split computing framework tailored for hierarchical LEO systems, where edge satellites perform event-driven sensing using dynamic vision sensors (DVS) and lightweight spiking neural network (SNN) encoders, while core satellites conduct inference using a powerful SNN decoder. A learned spike mapping scheme enables direct transmission over optical inter-satellite links (OISLs) without conventional modulation overhead. Experimental results on synthetic aerial scene classification demonstrate that the proposed architecture achieves accuracy on par with modern large vision-based pipelines, while offering energy efficiency comparable to that of existing lightweight implementations. These findings highlight the potential of neuromorphic computing for energy-efficient inter-satellite split computing in LEO remote sensing missions.

• The paper introduces a neuromorphic split computing framework for LEO satellites, integrating event-driven sensors with optical links for efficient data transmission.
• It employs dynamic vision sensors, spiking neural networks, and a learned time-hopping mechanism, achieving up to a 10× reduction in encoder and 2.5× in decoder energy consumption.
• The architecture integrates Spiking Vision Transformers at core nodes, ensuring robust, high-accuracy classification even under pointing error variations.

Onboard Neuromorphic Split Computing via Optical Links for LEO Remote Sensing

Introduction to Neuromorphic Split Computing in LEO Constellations

The paper presents a novel framework for deploying neuromorphic split computing within low Earth orbit (LEO) satellite systems. This approach seeks to optimize the onboard processing capabilities of satellite constellations by employing neuromorphic computing paradigms, which mimic biological neural systems to enhance the energy efficiency and processing speed of networked satellite nodes.

Figure 1: Conceptual illustration of the LEO hierarchical satellite architecture. Edge nodes are responsible for sensing and preliminary processing, while core nodes aggregate data via optical inter-satellite links (OISLs), perform higher-level tasks, and manage downlink to ground stations.

Neuromorphic Architecture and Its Components

The proposed architecture is tailored for hierarchical satellite systems. Edge satellites conduct event-driven sensing using dynamic vision sensors (DVS) and encode information with lightweight spiking neural networks (SNNs). The core satellites, equipped with more robust computational resources, utilize a powerful SNN decoder for inference tasks. The system capitalizes on optical inter-satellite links (OISLs) for direct neuromorphic signal transmission, negating the need for conventional modulation processes.

Figure 2: Comparison of conventional (a) and neuromorphic (b) remote sensing pipelines in a hierarchical satellite system.

The edge-to-core signal flow utilizes a unique learned spike mapping strategy designed for efficient transmission across optical channels, facilitating energy efficiency in data communication. A highlighted advantage of this system is its ability to match the classification accuracy of modern vision pipelines while offering up to a 10$\times$ reduction in encoder energy consumption and a 2.5$\times$ reduction in decoder energy consumption.

Spatiotemporal Event Processing and Encoding

The satellites' edge nodes utilize DVS, generating spatiotemporal event sequences represented as sparse, spike-based outputs suitable for SNN processing. These temporal event frames are processed using an architecture based on the Spiking Token Mixer (STMixer), which is designed to accommodate the event-driven characteristics of DVS data.

Figure 3: Spatiotemporal event streams generated from representative satellite scenes using a DVS-based sensor. Each column shows a region of interest, with corresponding event frames over five time steps ($t=1$ to $t=4$).

This processing enables high-dimensional input to be encoded into temporally expanded sequences optimized through a learned time-hopping (LTH) mechanism, allowing direct signal transmission. The STMixer framework allows for energy-efficient encoding, contributing to the system's overall power efficiency.

Optical Inter-Satellite Communication

A key innovation in this paper is the integration of a learned temporal hopping (LTH) encoding mechanism to optimize spike transmission over optical links. By embedding temporal patterns within the spike sequences, the system achieves compatible and efficient optical channel communication, even under conditions of high noise variance and pointing errors, a common challenge in free-space optical systems.

Figure 4: Illustration of the learned time-hooping (LTH) encoding process in the proposed neuromorphic split computing system.

Core Node Processing with Spiking Vision Transformers

The core satellite nodes deploy a Spiking Vision Transformer (SViT) architecture, which supports nuanced signal decoding and high-accuracy inference. SViT's architecture leverages event-driven attention mechanisms to maintain energy-efficient processing while ensuring rapid response times.

Figure 5: Network architecture of (a) the edge encoder based on STMixer with LTH output, and (b) the core decoder based on SViT.

This design is particularly advantageous given its compatibility with hybrid analog-digital platforms, facilitating the practical deployment of machine learning models in power-constrained satellite environments.

Numerical Evaluation and Results

The paper demonstrates the proposed architecture's performance under simulated scenarios using aerial scene classification tasks. Results indicate that the neuromorphic framework effectively reduces energy consumption while maintaining competitive accuracy relative to conventional large-scale neural networks. Particularly noteworthy is the architecture's ability to handle variations in radial pointing error with minimal degradation in performance.

Conclusion

This work advances the field of satellite-based remote sensing by proposing a neuromorphic split computing framework optimized for low-power and high-efficiency operation. It aligns with the growing need for onboard intelligence in LEO satellite systems and offers compelling evidence for the practicality of neuromorphic computing in space. Future research directions may explore further optimization strategies and broader application domains within satellite constellations.