On-board image classification payload for a 3U CubeSat using machine learning for on-orbit cloud detection

Abstract

CubeSats are giving the opportunity for educational institutes to participate in the space industry, develop new technologies and test out new ideas in outer space. CubeSat missions are developed to perform scientific research and demonstrate new space technologies with relatively cheap cost and limited resources. This category of satellites has many limitations such as the short development time, the power consumption and the limited time and capability of data downlink. Earth Observation from a Low Earth Orbit is one of the most appealing m applications of CubeSats developed by students or non-space faring countries. Investigating new technologies to improve image quality and studying ways to increase acquisition adequacy is very promising. This paper aims to introduce a mission hardware design and machine learning-based algorithm used within an Earth Observation (EO) CubeSat. The case study of this paper is Alainsat-1 project which is a 3U CubeSat developed with the support of IEEE Geo-science and Remote Sensing Society (GRSS) at the National Space Science and Technology Center, UAE. The satellite is planned to be launched by 2022. A low-resolution Commercial off-the-shelf (COTS) camera for EO is developed as a primary mission in this
 CubeSat. The compatible hardware design and software algorithm proposed is responsible for classifying the images captured by the camera into different categories based on cloud intensity detected in these images before downloading them to the ground station. A microcontroller-based architecture is developed for controlling the mission board; it is responsible for accessing the memory, reading the images, and running the cloud detection algorithm. The cloud detection algorithm is based on a U-net architecture while the algorithm is developed using a Tensor-flow library. This model is trained using a dataset of images taken from the Landsat 8 satellite project. Moreover, the SPARCS cloud assessment dataset is used to evaluate the developed model on a new set of images. The overall accuracy achieved by the model is around 85% in addition to the acceptable performance of the model observed on a set of low-resolution images. The plan is to make the design modular and optimize its
 performance to be used on-board CubeSats fulfilling the size constraint and overall power consumption limitation of an add-on module to a camera mission.