Space Communications Technology | The Grainger College of Engineering

Deepak Vasisht

Taking satellites to the edge

Deepak Vasisht is showing us the best ways to quickly get sensitive data

How do we fight wildfires when the tiniest shift in wind direction or change in humidity can drastically alter their course?

How can we keep track of agricultural fields too vast for people to regularly visit?

How can we monitor environmental developments large and complex as polar ice melting?

The answer to these and other questions may lie in space. A constellation of satellites in low-Earth orbit (LEO) at altitudes between 300 and 600 miles would enable us to study our planet with unprecedented detail.

Grainger Engineering computer science professor Deepak Vasisht explains, “The principle is actually very simple: by moving the camera closer to the object, you get a better photo. Current satellite constellations orbit at 12,000 or 22,000 miles, which allows us to see that there is a wildfire, for example. However, an LEO constellation would show in detail how the fire unfolds in real-time, making a huge difference in evacuation efforts and resource allocation to fight it.”

As the private space industry develops and orbital launch systems become more efficient, deploying Earth observation constellations in LEO is becoming feasible. However, using them for time-sensitive applications like wildfire monitoring is challenging due to networking and communications deficiencies.

Unlike ground-based networks, satellites cannot transmit information instantaneously. They must wait until they pass over a transmitter, where they “dump” as much information as possible before losing contact. Since LEO satellites move so quickly — about 5 miles per second — the contact window with the transmitter is very brief. In the wildfire example, a photograph can take hours to days to reach authorities, during which time the fire can significantly change.

Vasisht wants to cut this down to a minute. An expert in networking, he and his research group are working to reduce the latency in satellite communication through innovations in both satellite hardware and software. They have developed the Serval system to demonstrate how a constellation of LEO satellites could classify, prioritize and transmit images in a matter of minutes.

“We’re taking ideas from terrestrial networking and moving them into space for the first time,” Vasisht said. “One LEO satellite has limited compute capacity, so the idea of Serval is to distribute the load across satellites and even to ground stations. This way, the smaller, specialized satellite hardware – an ‘edge computer’ – is utilized as efficiently as possible.”

Classifying images based on their content and assigning priority is a perfect use for AI technology, but these algorithms require extensive resources to run. To perform such tasks in space, Serval splits them into two parts: one that depends on rapidly changing information, and one that depends on slowly changing information. The slowly changing part can be calculated in advance using ground-based computing resources then transmitted to the satellite to use as a baseline for the rapidly changing part.

For example, the AI algorithm to detect wildfires in California can be decomposed into “detect forests in California” – the slow part – and “detect fire” – the rapid part. The former is pre-computed on the ground and loaded onto the satellite, and the latter is evaluated on the satellite using an onboard GPU or other AI system.

This still leaves the problem of ensuring that the prioritized image is transferred in a timely manner. Serval addresses this by exploiting the predictability of orbital motion.

According to Vasisht, “It’s not like cellular service where a large group of transmitters move unpredictably. A satellite is subject to deterministic laws of motion, and its path can be calculated to very high precision. We can exploit this to create precise download schedules that make the most of the contact window.”

The research group found that the combination of distributed computing and precise network scheduling reduces the 50th percentile of transmission times from 47 hours to 2 minutes, with the 90th percentile reduced from 149 hours to 47 minutes. Moreover, the satellite compute load was reduced by 80%.

“This result is exciting for another reason: our scheme would broaden the scale of the internet of things,” Vasisht said. “Right now, networked devices can cover 5, maybe 10 miles. If they could connect to the cloud via satellites, they could cover 500 miles. Imagine if we had moisture sensors to study crops over entire regions, or sensors to track wildfires over their entire boundaries. We need efficient satellite networking to make this happen, and our work brings us a step closer.”

Graduate students Om Chabra, Ishani Javeja, Maleeha Masood and Bill Tao in Vasisht’s research group as well as computer science professor Indranil Gupta also contributed to the development of Serval.

Vasisht’s group collaborates with the company Planet, Inc. which operates around 200 LEO satellites. The research team used the company’s Planetscope imagery database to test their ideas.

This work was largely supported by Vasisht’s award “Networking and Compute for Next Generation Low-Earth Orbit Satellites” through the National Science Foundation’s CAREER program. Cisco and Microsoft also supported this work.