Home Blogs Using Machine Learning and Neural Networks for advanced space solutions

Using Machine Learning and Neural Networks for advanced space solutions

KP Labs, a NewSpace company based in Poland, specializes in machine learning applications for the space industry and hyperspectral imagery acquisition and processing. The company focus on accelerating space exploration by advancing autonomous spacecraft operation and robotic technology. Its flagship product is the Intuition-1 satellite mission. It is 6U satellite, equipped with a hyperspectral optical instrument and an on-board computer “Leopard” that advanced data processing through convolutional neural networks.

In downstream, the company concentrates on the use of hyperspectral images for industrial and agricultural purposes. For this application, it has developed a lightweight, hyperspectral imager “Zebra”. 

A few months back KP Labs signed a contract with ACC Clyde Space for the delivery of a satellite bus and placing the satellite into low Earth orbit. The launch of Intuition-1 is expected by 2023. The mission of the Intuition-1 is to perform Earth observation using a hyperspectral instrument and advanced data processing based on deep neural networks (AI) on-board the satellite.

KP Labs is also involved with the European Space Agency in projects like HYPERNET, which is about HYPER-spectral image segmentation using deep neural networks, and SUPER RESOLUTION – Super-resolution reconstruction (SRR) with deep neural networks. The aim of SRR techniques is to improve the quality and increase the resolution of images (upscale) while restoring as many details as possible from the source image or images.

“By processing the data on board of Intuition-1, it will be possible to quickly assess the condition of plants and forests, to forecast crop yield, says Michał Zachara, Vice President, KP Labs, in an exclusive interview with Geospatial World.

Michał Zachara, Vice President, KP Labs

Intuition-1 processes hyperspectral images in orbit using neural networks. What are some of the major benefits of this approach?

The supercomputer “Leopard” on-board the Intuition-1 will enable for segmentation of hyperspectral imagery right in Earth’s orbit.

Thanks to in-orbit processing of the collected imagery, at least 100 times less data must be transferred to the ground station. The end-users will thus have lower access time to any information gathered by the satellite that they find relevant. Images will be captured daily, which will allow for continuous monitoring during e.g. a flood.

The main payload is an innovative, high-performance data processing unit, processing high-resolution hyperspectral imagery. The payload will be complemented with a high-resolution specialized optical instrument operating within the range of visible light and near-infrared. Dividing that spectrum into 150 bands will greatly increase the amount of data provided when compared to already existing instruments.

Uses Deep Neural Networks to process data on-board and therefore only sends the most important and valuable insights to the ground. By reducing the time and cost of data transfer and processing, it enables you to focus on a rapid response to any detected phenomena.

How do you think AI coupled with Machine Learning would lead to a paradigm shift in the space sector?

Thanks to an increasing number of research projects related to AI in space but also due to first commercial products becoming available on the market, like the Leopard AI data processing unit (DPU) by KP Labs, it is a very valid question. Nearly all space missions rely on significant spacecraft oversight by ground control operators. This approach obviously introduces considerable delays in decision making (especially for e.g. Low Earth Orbit – LEO) posing significant risks for mission designers in terms of critical situation response. Time is the essence in such situations and AI might be the answer here.

Various ML technologies could lead to the development of autonomous systems on board of satellites capable of responding to critical situations automatically, just as good as a human operator could. As you can imagine, this could lead to a significant reduction in costs of ground segment and potential improvement in spacecraft reliability further leading to driving overall space mission costs down.

The above example is something that we are nearly capable of doing as of today. But the paradigm shift to come would mean much more to the space sector. Thanks to AI and Machine Learning we will be capable of dealing with new, unpredictable environments out there e.g. operating a mining expedition on the surface of an asteroid. We should also witness more autonomous collision avoidance systems for our planet’s orbit opening the way for new mega-constellations (which will come rather sooner than later). All those inventions and AI technologies will pave the way for space exploration, where autonomy will allow humans to deal with the increasing complexity of scenarios and mysterious worlds to discover.

Also Read: Singapore startup to build world’s first open-source satellite network

How do you ensure data segmentation?

Our segmentation technologies are based on Machine Learning algorithms. What we have developed during collaboration with ESA (Hypernet project) allowed us to reach state-of-the-art supervised segmentation accuracy but also, opened the way to a more data agnostic unsupervised segmentation approach. These automated (or unsupervised) segmentation algorithms allow any imagery from regular RGB, through multispectral Sentinel-2 to hyperspectral imagery with hundredths of data bands to be segmented nearly without human intervention and tuning. This, in turn, allows identifying new object classes when working with imagery from regions that weren’t previously classified by humans.

These unsupervised segmentation mechanisms will be capable of segmenting data even directly in orbit. Our Intiution-1 satellite mission together with its hyperspectral instrument and AI data processing unit will demonstrate (among other) this capability.

Where do you think the future of Earth observation is heading?

Looking at current trends, we are probably heading towards further improvements in revisit times for Earth observation satellites (through constellations) maintaining medium- to high-resolution imaging capacity. This means the technology will open ways for new applications where revisit times of less than 1 hour are critical. Monitoring e.g. sea vessels or various calamities could significantly benefit from this capability, but the truth is that the time will show which applications will be turned into sustainable beneficial business models. It could be that they weren’t identified yet and are waiting for a brand-new New Space start-up.

Another trend that will be enabled by more extensive use of AI in space will be autonomous tasking of satellites and their capability to decide which areas need to be captured. Imagine a constellation that autonomously detects various anomalies, reorganizing its formation to focus on particular events (e.g. floods, eruptions, oil leaks) and then, through inter-satellite communication optimized data transfer down to Earth delivering information as quickly and as efficiently as possible.

Above all, what is certain, is that we will need to deal with the increasing number of artificial satellites orbiting our planet. This will deliver new capabilities mentioned above, but at the same time will require new technologies, like autonomous decision making and collision avoidance to be developed. Without that, it seems that small satellite development might grind to a near halt, at some point due to the ever-increasing likelihood of orbital object collision.

With increasing cloud capabilities, refined AI-analytics and emerging business models like Groundstation-as-a-service, do you foresee a disruption in the data analytics industry?

It is inevitable. Data analytics industry will face the challenge of emerging markets looking for near and real-time data to make decisions faster and more accurate. Our AI solutions support such companies enabling the possibility to pre-process data on-orbit and downlink only necessary information to bring data to end clients right on time. For example, algorithms on Leopard DPU will search for dedicated scenarios and alert when the potentially dangerous scenario is found. Such information might be crucial in case of emergency situations. It might also improve data-driven companies offering precision agriculture or IoT. Along with new software opportunities, hardware will also influence the market and help processing data much faster.

Data processing on-orbit is definitely a game-changer. There are NewSpace companies looking for innovations, developing cutting edge technologies that perfectly respond to the current and future needs. An increasing number of constellations and CubeSats, spacecraft autonomy and optical navigation in space will influence the market and force data processing on orbit.

Also Read: Geospark Analytics deploys AI for risk assessment and effective decision-making