Session
Weekend Session 3: Science/Mission Payloads - Research & Academia I
Location
Utah State University, Logan, UT
Abstract
We discuss the deployment of image processing algorithms developed for BeaverCube-2, a project under development between the MIT Space Telecommunications, Astronomy, Radiation (STAR) Lab and the Northrop Grumman Corporation. The algorithms were uploaded to and executed on OPS-SAT, a 3U CubeSat owned and operated by ESA with a processing payload that allows rapid prototyping, testing, and validation of software and firmware experiments in space at no cost to the experimenter. Testing these algorithms onboard OPS-SAT significantly reduces risk for future on-orbit image processing missions such as BeaverCube-2. We focus on four image processing algorithms used for cloud detection: a luminosity-thresholding method, a random forest method, an U-Net based deep learning method — all developed by STAR Lab for BeaverCube-2 — and a k-means clustering deep learning method implemented by the OPS-SAT Flight Control Team (FCT). We evaluate each method in terms of in terms of overall accuracy, power draw, and temperature rise on-orbit, and discuss the challenges of implementing these methods on embedded hardware and the lessons learned for BeaverCube-2.
Machine Learning Image Processing Algorithms Onboard OPS-SAT
Utah State University, Logan, UT
We discuss the deployment of image processing algorithms developed for BeaverCube-2, a project under development between the MIT Space Telecommunications, Astronomy, Radiation (STAR) Lab and the Northrop Grumman Corporation. The algorithms were uploaded to and executed on OPS-SAT, a 3U CubeSat owned and operated by ESA with a processing payload that allows rapid prototyping, testing, and validation of software and firmware experiments in space at no cost to the experimenter. Testing these algorithms onboard OPS-SAT significantly reduces risk for future on-orbit image processing missions such as BeaverCube-2. We focus on four image processing algorithms used for cloud detection: a luminosity-thresholding method, a random forest method, an U-Net based deep learning method — all developed by STAR Lab for BeaverCube-2 — and a k-means clustering deep learning method implemented by the OPS-SAT Flight Control Team (FCT). We evaluate each method in terms of in terms of overall accuracy, power draw, and temperature rise on-orbit, and discuss the challenges of implementing these methods on embedded hardware and the lessons learned for BeaverCube-2.
