# Energy efficiency analysis of Spiking Neural Networks for space applications

**Authors:** Paolo Lunghi, Stefano Silvestrini, Dominik Dold, Gabriele Meoni, Alexander Hadjiivanov, Dario Izzo  
**arXiv:** [2505.11418](https://arxiv.org/abs/2505.11418)  
**Date:** 2025-05-16  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2505.11418v1/

## Key points

- The exponential growth of the space sector and demand for higher spacecraft autonomy have led to the development of AI-assisted space systems.
- Spiking Neural Networks (SNN) offer a promising solution for energy efficiency in space applications due to their sparse activity induced by binary spikes.
- Practical demonstration of SNN efficiency on real-world tasks remains a challenge despite its potential.
- A study conducted a numerical analysis and comparison of different SNN techniques for scene classification using the EuroSAT dataset, crucial for space applications.
- Models based on temporal coding aim to maximize network efficiency by leveraging sparsity properties inherent in SNN.
- Development of a reliable metric that can compare different architectures in a hardware-agnostic manner was a key aspect of the research.
- The study applied this novel method to predict energy consumption onboard spacecraft, showcasing flexibility in describing specific hardware platforms like BrainChip Akida AKD1000 neuromorphic processor.
- Findings demonstrate potential benefits of utilizing SNN for space applications, particularly in terms of energy efficiency and optimizing performance for spacecraft operations.

## Summary

The exponential growth of the space sector and the demand for higher spacecraft autonomy have led to the development of AI-assisted space systems. However, the low availability of power and energy typical of space applications has hindered progress in this area. Spiking Neural Networks (SNN) offer a promising solution due to their theoretically superior energy efficiency, characterized by sparse activity induced by neurons communicating through binary spikes. Despite this potential, the practical demonstration of SNN efficiency on real-world tasks remains a challenge. To assess the feasibility of implementing SNN onboard spacecraft, a numerical analysis and comparison of different SNN techniques were conducted for scene classification using the EuroSAT dataset. This dataset is crucial for space applications as it represents land use classification using multispectral and RGB imagery from the Sentinel-2A satellite. The study focused on models based on temporal coding, where essential information is encoded in the timing of neuron spikes. These models aim to maximize network efficiency by leveraging the sparsity properties inherent in SNN. A key aspect of this research was the development of a reliable metric that can compare different architectures in a hardware-agnostic manner. This metric establishes a clear theoretical relationship between architecture parameters and expected energy consumption onboard spacecraft. The study also applied this novel method to predict the energy consumption of a BrainChip Akida AKD1000 neuromorphic processor, showcasing its flexibility in describing specific hardware platforms. The findings from this analysis demonstrate the potential benefits of utilizing SNN for space applications, particularly in terms of energy efficiency. By focusing on scene classification tasks with real-world implications, this study contributes valuable insights into optimizing SNN performance for spacecraft operations. Overall, these results highlight the advantages of SNN and pave the way for further advancements in AI-assisted systems within the space sector.