AI Inference Technology toward a Space Datacenter

1. Introduction

Space development has been accelerating, and the number of satellites—particularly those for Earth observation and communications—continues to increase dramatically. Satellites are equipped not only with advanced sensors for Earth observation, communication terminals, and solar panels but also onboard computers used for functions such as satellite control and communications. Specialized embedded computers with radiation resistance have been developed and installed specifically for space use. However, it has recently become common to adapt commercial off-the-shelf computers with radiation protection measures for use in space. Thus, onboard computers are now being used for a wider range of applications. In particular, Earth observation satellites equipped with high-resolution sensors are expected to use onboard computers for tasks such as data compression and advanced analysis using artificial intelligence (AI).

Low Earth orbit (LEO) satellites—which are commonly used for Earth observation—are becoming smaller and have limited power supply, resulting in constraints on computational capacity and processing time. Therefore, conducting advanced AI analysis using high-precision AI models such as vision and language models running on satellites—which are sometimes called an ultimate edge environment—is not straightforward. Similarly, conducting comparative analysis onboard using data captured with other observation satellites or metadata obtained from the ground also presents significant challenges.

2. Toward a space datacenter business

Space Compass—a joint company between NTT and SKY Perfect JSAT—aims to launch a space datacenter business using geostationary orbit (GEO) satellites and is taking on the challenge of building a new infrastructure [1].

Compared with LEO satellites, GEO satellites are expected to offer rich power and computing resources. They can provide data storage for preserving observation data and enable high-capacity, quasi-real-time data transmission between LEO satellites and ground stations.

By uploading observation data from multiple LEO satellites to a GEO satellite for aggregation and conducting comparative analysis with past archives saved at the GEO data storage, GEO satellites are expected to achieve more advanced analysis of Earth observation data.

California Institute of Technology/Jet Propulsion Laboratory held a workshop in 2019, where they proposed the concept Nebulae, which envisions the paradigm shift from traditional space missions aimed at individual scientific goals to an infrastructure-based approach that provides computing, data storage, networking, and cloud services to enable efficient and sustainable services [2]. Efforts to build infrastructures in space are accelerating worldwide, spanning both public and private sectors, particularly in the context of defense applications [3]. The NTT Group is also working toward building a sustainable society by integrating terrestrial and non-terrestrial infrastructure through space-integrated computing [1].

Space Compass plans to launch multiple GEO satellites and begin its space datacenter business. By deploying three or more GEO satellites, it will become possible to offer constant connectivity with most LEO satellites and ground stations. This space datacenter business consists of two main components: optical data relay and space edge computing. Optical data relay is a high-speed network service that connects LEO satellites, which orbit the Earth every 90 minutes and cannot keep constant connection with ground stations, with ground stations via GEO satellites using optical communication. This enables secure and real-time utilization of satellite data. However, since observation data tends to be large in size, an increase in the volume of such data can cause the communication link between GEO satellites and ground stations to become a bottleneck, potentially delaying data analysis on the ground. To address this, space edge computing, which leverages onboard computing resources on GEO satellites, can be used to analyze observation data in orbit. By selecting only the most important data for download to the ground, this approach is expected to reduce the volume of data that needs to be transmitted and enhance the real-time analysis of critical information [4].

GEO satellites that can offer constant connectivity with observation satellites (typically LEO satellites) are also ideal as data hubs for observation satellites. Figure 1 illustrates three representative use cases using space datacenters.

Fig. 1. Representative use cases using space datacenters.

3. Event-driven AI inference technology

Figure 2 shows an example data processing pipeline of event-driven AI inference applied to SAR data. Modern SAR sensors have high resolving ability and enable us to produce high-definition images, resulting in extremely large data sizes—even after imaging and compression, a single file can exceed several hundred megabytes. Since downloading all SAR data captured with fast-orbiting LEO satellites to the ground is both costly and time-consuming, a two-phase inference approach is used: the onboard computer executes a first round of coarse inference, and only the data deemed to require more detailed analysis are downloaded for a second round of inference on the ground. In our experiment, we used a relatively lightweight object detection model as the first-phase AI model. For the onboard computing platform, we conducted validation using the Unibap ix10, which has a strong track record in LEO satellite deployments, and the Intel Myriad X vision processing unit (VPU) [6] as the AI chip. While the results depend on how frequently the target objects appear, experiments using a publicly available ship dataset [8] showed that on-orbit data volume was reduced by 58.7 to 80% [5]. Using the VPU significantly reduced both power consumption and execution time compared with the central processing unit, confirming that AI analysis is feasible on satellite onboard computers. We also developed a ship detection application that runs on the ix10 and demonstrated it at events such as NTT R&D Forum 2024 (Fig. 3).

Fig. 2. Two-phase AI inference for onboard SAR data processing.

Fig. 3. Ship detection demo running on a space-grade onboard computer.

4. Lightweight change detection technology

The idea of event-driven AI cannot be applied unless we know what kind of event will happen beforehand. One method to overcome this limitation is change detection, which identifies differences from past data and focuses processing only on the areas where changes have occurred. By executing change detection, it becomes possible to store only the changed regions—reducing data volume—and to limit the processing scope, thereby lowering computational costs. As shown in Fig. 4, the task is defined to output a change map indicating the changed locations between two input images. Change detection has been studied for many years in research communities, and numerous methods have been proposed. We developed a model for lightweight change detection on satellite onboard computers [7]. This model enhances existing Transformer-based change detection algorithms through two improvements: early exit, and lightweight decoders. Early exit is an idea in which, if the neural network is confident in its prediction for the input data, it can output results before reaching the final layer, thus saving computation. In our case, early exit was added and trained on the basis of metric-learning approaches.

Fig. 4. Proposed lightweight change detection model.

5. Conclusion

We introduced three envisioned use cases for a space datacenter using GEO satellites along with two key technologies that enable these use cases and contribute to reducing communication volume with the ground and improving the real-time performance of data analysis: event-driven AI inference for SAR data and lightweight change detection technology. The NTT Software Innovation Center will continue research, development, and testing toward on-orbit demonstrations with LEO satellites while advancing R&D toward distributed space computing for satellite constellations evolving from standalone onboard computers.

References

[1]	Press release, “NTT and SKY Perfect JSAT conclude collaboration agreement on new space enterprise to aid realization of a sustainable society,” May 20, 2021. https://group.ntt/en/newsrelease/2021/05/20/210520a.html
[2]	IEEE SMC-IT/SCC 2023: Explainable AI and Space Clouds, https://sites.astro.caltech.edu/xaisc/
[3]	Press release, “Leonardo: Kick Off for the Project of the First Space Cloud System for Defense,” Feb. 19, 2024. https://www.leonardo.com/documents/15646808/28143462/ComLDO_SpaceCloud_ENG.pdf
[4]	T. Eda, A. Yamanaka, K. Tabata, and I. Yamasaki, “Case Study: Two-Phase AI Prediction Techniques for Space Edge Computing,” 2023 IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2023), Pasadena, CA, USA, July 2023. https://doi.org/10.1109/IGARSS52108.2023.10282932
[5]	M. R. Busto, T. Eda, T. Udagawa, and T. Sekine, “Case-Study: Two-Phase AI Prediction for Onboard Synthetic Aperture Radar (SAR) Data Processing,” 2024 IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2024), Athens, Greece, July 2024. https://doi.org/10.1109/IGARSS53475.2024.10642771
[6]	T. Eda, M. Busto, T. Udagawa, N. Ishihama, K. Tabata, Y. Matsuo, and I. Yamasaki, “Technical Challenges for AI in Space Data Centers,” IGARSS 2024, Athens, Greece, July 2024. https://doi.org/10.1109/IGARSS53475.2024.10642138
[7]	M. R. Busto, S. Nouri, and T. Eda, “Parameter and Data Efficient Framework for Lightweight Change Detection,” IGARSS 2024, Athens, Greece, July 2024. https://doi.org/10.1109/IGARSS53475.2024.10641816
[8]	T. Zhang, X. Zhang, X. Ke, X. Zhan, J. Shi, S. Wei, D. Pan, J. Li, H. Su, Y. Zhou, and D. Kumar, “LS-SSDD-v1.0: A Deep Learning Dataset Dedicated to Small Ship Detection from Large-Scale Sentinel-1 SAR Images,” Remote Sensing, Vol. 12, No. 18, p. 2997, 2020. https://doi.org/10.3390/rs12182997