Pixalytics In-Orbit Demonstration Results
We’ve recently been involved in an exciting project to see if we could run one of our Machine Learning (ML) land cover classification algorithms directly on satellite images in space onboard the ION SCV004 craft around 500 kilometres above the planet.
This activity was an In-Orbit Demonstration as part of the inaugural Cloud Computing Service Call for Ideas, organised by D-Orbit in collaboration with Unibap AB (publ) and Trillium Technologies, with the support of the European Space Agency – ESA Φ-lab, and we were lucky enough to be selected as a participant in the programme.
The ION SCV004 is a free-flying CubeSat deployer and technology demonstrator developed by D-Orbit, which was launched on the 13th January 2022 into a low-Earth sun-synchronous orbit. The ION stands for ‘InOrbit NOW’, and the SCV stands for ‘Space Carrier Vessel’.
We chose our waste plastic land classification algorithm that was developed as part of the Marlisat project, which was one ESA’s Open Space Innovation Platform (OSIP) portfolio of ideas to track how marine plastic litter move around the world’s ocean and other water bodies of the world and was led by the French organisation CLS, Collecte Localization Satellites.
For the In-Orbit Demonstration, the task of the project was to streamline our algorithm to meet the constraints of the project to allow it to run using Unibap’s SpaceCloud® onboard D-Orbit’s orbital transfer vehicle; these were that the algorithm needed to run within 5 minutes and have a maximum memory usage of less than 10 Mb on top of the base container.
A test bed was created, and we undertook several iterative processes where we tried to meet the constraints. The approach required significantly simplifying the ML algorithm regarding inputs and Neural Network model structure. So, instead of having both Sentinel-1 and Sentine-2 as inputs, we used a simplified Sentinel-2 dataset with only four bands that another participant in the In-Orbit Demonstration generated. Also, we reduced the land cover classification to a few broad categories and used the 2015 ESA Climate Change Initiative Landcover product for training. The focus was on getting the overall broad-scale classification for anywhere on the planet approximately correct rather than expecting an accurate pixel-based classification. We were delighted when it passed the on-ground testing, and we then waited for deployment onto the spacecraft. This deployment happened earlier this year in mid-March when our code was successfully deployed and ran in orbit!
This is a demonstrator project, but it has shown that running a ML land cover classification algorithm is possible directly on satellite images in orbit. By undertaking classifications in-orbit, this will, hopefully, reduce the amount of data end-users have to download significantly and speed up the data transfer and the overall efficiency of the process. In the longer term, this classification approach could allow onward decisions, e.g., whether this or another satellite should be tasked to make further acquisitions or on-ground activities should be initiated.
Processing satellite images in space is another of the bucket list tasks – Pixalytics has been to space (possibly twice!), gone over 200 miles per hour in an F1 car, and crossed the Atlantic in an ROV! We’re delighted to have taken part in this project, and we are grateful for the help and support from D-Orbit, Unibap AB, Trillium Technologies, and ESA for making this project happen. It will be exciting to see where this might lead.
