Vancouver-based Metaspectral is sending a hyperspectral imaging payload to the International Space Station in early 2023. Working with HySpeed Computing and Nanoracks, they’ll be there to test and improve hyperspectral data analysis using deep learning algorithms and neural networks.
They’re also there to prove that Metaspectral’s compression tech can solve a big problem with hyperspectral imaging: the pictures are enormous—but the bandwidth is not.
In an interview with SpaceQ, Metaspectral Co-Founder and CTO Migel Tissera talked about Metaspectral, hyperspectral imaging, and what his company is bringing to the table.
Hyperspectral imaging
Hyperspectral imaging is a hot topic in space circles, and there’s at least one prominent Canadian space startup, Wyvern Space, that is focused on the subject.
It’s a kind of imaging that goes far beyond traditional images. Traditional color images will have three spectral bands for each pixel, corresponding to “red” “green” and “blue.” Hyperspectral images have complete spectrums of hundreds of different wavelengths per pixel. Each hyperspectral image isn’t just a two-dimensional picture, it’s a data cube: X dimension, Y dimension, and a third “dimension” that contains the hundreds of different wavelengths captured by the pixel.
When properly analyzed, these images are tremendously useful, able to pick up things like a material’s composition, radiation, and thermal properties. So satellites that employ hyperspectral imaging have a wide variety of use cases. In defence and rescue, they can be used to ferret out images of people and equipment from a relatively small number of pixels in poor conditions. They can also be used in agriculture to monitor crops and soil conditions, in mining and forestry to find promising resources, and in meteorology to track weather and climate change.
That last topic is especially salient to people at the moment, considering the growing climate change crisis and the extreme temperatures being seen in the Northern hemisphere this summer. It’s also one of the main things that Metaspectral is focused on.
Metaspectral’s hyperspectral compression solution
The problem is these images’ tremendous size. Tissera said that a normal 4K image can range up to ten megabytes in size, but a single hyperspectral data cube can go up to four gigabytes in size. A satellite pass could produce an equally huge number of these images, and most LEO hyperspectral imaging is done by whole constellations. It’s a truly awe-inspiring amount of data.
Analyzing them is already challenging. Metaspectral uses machine learning algorithms and deep learning networks for analysis. An arguably bigger problem, though, is transport. “Edge” computing—adding processing to the satellites and sensors where it’s collected—does help somewhat, but most hyperspectral data analysis needs to be done in a datacenter, and so you need to get those terabytes of data there somehow. While this is no great challenge on Earth in 2022, the bandwidth available on a typical LEO satellite is greatly limited.
That’s where Metaspectral comes in. Their forte is compression. They do more than that—Tissera said that they operate on the “data layer” as opposed to the hardware layer—but compression is one of their big competitive advantages. Their data compression is based on an implementation of a standard created by the international Consultative Committee of Space Data Systems.
Tissera believes that they’re the only ones who’ve actually implemented the standard, and even he’s not sure why; possibly because it’s so novel, only finalized in 2019. Whatever the reason, it works. Their implementation of the standard can reduce data losslessly to up to 40% of its original size, and reduce it near-losslessly up to 10% of its original size. He said that “that gives us the flexibility to transmit that data, shuttle that data back and forth in real time.”
It’s computationally intensive, requiring custom Metaspectral FPGA hardware at both the sending and receiving ends of the transmission. Even GPUs aren’t powerful enough. Once it’s there, though, it can take these gargantuan images and get them to an Earth-side data center, where Metaspectral uses deep learning to quickly process and prepare the data. He said that their software platform “makes the data analysis-ready within fifteen minutes of downlinking.”
ISS, climate, and space-proofing
Hyperspectral compression has a lot of terrestrial applications, and so Metaspectral does work with terrestrial clients, including work for the Canadian Department of National Defence. But even though they graduated from the Creative Destruction Lab’s General Stream, not its Space Stream, they’re still focusing hard on space. Space provides an ideal use case for their compression technology due to the hard constraints on data transmission.
That led them to their recent greenhouse gas monitoring project with the Canadian Space Agency, where they will be working on a method to use hyperspectral data from the total carbon column observing network (TC CON), along with data from several satellites, to “systematically and methodically quantify the CO2 levels present at ground elevation.” Tissera said that it’s going well; that their neural nets are able to predict CO2 levels within the datasets “within a 3% error margin now.” The goal is to track the effects of different kinds of carbon sequestration from agriculture.
Their big new project, though, is the OPTICA (Onboard Programmable Technology for Image Classification and Analysis) mission. It’s an Earth observation payload destined for the ISS that will be pairing an off-the-shelf hyperspectral camera with Metaspectral’s compression and analysis technology. Tissera said that they would be “providing the hardware for the payload, the software for the payload, and providing the software for processing the data on the ground.”
It’s going to be serving as a proof-of-concept for Metaspectral’s tech; as Tissera puts it, it will allow them to “space-rate our technology stack…we’re basically taking the edge device and putting it in the space station.” This will give them room to market their solution more directly to hyperspectral imaging companies, as well as other organizations that may want to take advantage of these powerful tools but balk at the difficulty of transporting and preparing the data.
They have two partners on the project: HySpeed Computing and Nanoracks. Nanoracks will be handling all aspects of getting the payload into space, installing and hosting it in their Nanoracks External Platform (NREP), and providing the data connection from the ISS to Earth. HySpeed Computing will be (according to the release) responsible for “creating the necessary data processing pipeline and analysis tools.” They have, according to Tissera, a “very rich history of analyzing hyperspectral data,” and they’ll be analyzing the data that Metaspectral’s pipeline provides them. Metaspectral will be doing their own analysis as well.
As for the mission timing, Tissera said that while delays are possible, they expect that their payload will be going up with SpaceX’s Cargo Resupply Mission 27 in late Q1, early Q2 next year.
Metaspectral’s searching for talent
Tissera couldn’t talk about their investment or customer situation beyond the already-available contract information, only saying that they’re receiving interest on all sides.
Being located in Vancouver has been helpful, though it’s had its downsides. The weather is great, they like the city, and Vancouver is close enough to west-coast technology hubs like San Francisco and Silicon Valley that they have comparatively easy access to the customer and investor pools there. He acknowledged that there’s a downturn happening, but expressed confidence in Metaspectral’s ability to weather it, especially considering their proven terrestrial successes.
Tissera said, however, that “I think one thing that has kind of affected us is Vancouver’s talent.” What’s there is high-quality, but it’s being snapped up by big American firms, especially as they can promise larger salaries. He admitted that it’s a “challenge” to find top-tier talent, and one of Metaspectral’s primary focuses right now is getting tech talent: front end and back end developers, deep learning engineers, and deep learning scientists. He said that “we’re doing it, we’re growing and we’re doing it.”
At a time when the big US tech firms are letting go of people, this may end up providing an opportunity for Metaspectral in the competition for top talent.