Background and Motivation

In this work, we will focus on Low-Earth orbit (LEO) Earth observation satellites — satellites with orbital periods of < 128 min and with low eccentricity (i.e., near-circular orbits), resulting in altitude < 2000 km. We focus on LEO EO satellites because a) EO satellites are often placed in LEO orbit in order to improve the spatial resolution of the generated imagery, and b) The number and size of LEO EO satellite constellations has been increasing [48], in large part due to significant decreases in LEO satellite launch costs [58, 83] as well as due to emergence of new EO applications (Section 5).

Unlike the EO data generation rate, which is increasing rapidly, there is limited opportunity to increase RF downlink capacity [125]. As such, downlink data rates have increased less than data generation rates (Fig. 3). Several approaches have been proposed to deal with this downlink deficit. Lossless and high quality lossy compression can be used to decrease the number of bits needed to represent each pixel downlinked. More aggressively, data can be discarded — either not downlinked or not even generated. This is done commonly in practice (e.g., Dove does not image the ocean); prior work [54] also propose to do it via image processing (e.g., detect and discard images occluded by clouds). Our work does not focus on reducing the amount of EO data to be sent to the applications running on Earth; we move the applications themselves to space.

The closest related work is the deployment of HPE’s SpaceBorne and SpaceBorne-2 computers to the International Space Station (ISS). These computers have been used to compute on data generated in space which had historically been slow to downlink. For example, astronauts have used these computers to monitor their DNA for mutation due to radiation exposure. This decreased the amount of time needed to analyze astronaut DNA from 12 h (mostly in downlink time) to 6 min [144]. Unlike our work, the HPE ISS computers do not process EO data from EO satellites.

Another closely related work is by Orbits Edge [110], a start-up that is trying to build frames to send servers to outer space. Limited information is available about their design.

To the best of our knowledge, no prior work makes a quantitative case for SµDCs. Ours is also the first work to analyze the computation requirements for a SµDC, the associated communication bottlenecks, and the SµDC-communication co-design approaches to address the bottlenecks.