Abstract:
Although the cost of development is cheap, cube satellites are limited in power, size, and downlink capabilities. By optimizing algorithms and the hardware these algorith...Show MoreMetadata
Abstract:
Although the cost of development is cheap, cube satellites are limited in power, size, and downlink capabilities. By optimizing algorithms and the hardware these algorithms run, one overcomes these limitations, thus, allowing more missions to run and more data to be collected from it. Images, for example, are relatively big in size and if the satellite were able to know which image data to downlink, it could save a lot of time and resources. For this purpose, a cloud detection algorithm based on the U-net architecture was developed using the TensorFlow library. This model will be trained using a dataset of 15,263 images taken from the Landsat 8 satellite while the SPARCS cloud assessment dataset was used to evaluate the model on images it was not trained on. To limit the size of the input data, only the RGB band was used. After optimizing the model's parameters, the model shows that it achieved an overall accuracy of ∼85%. Furthermore, testing of the same model on images of lower resolution taken from CubeSats showed that it still was fairly accurate and would manage to work in most CubeSats that would only be able to take low resolution images. The model was then quantized and was then converted to a C code 8 bytes array using the TensorFlow Lite library to reduce its size and operation. It is then implemented inside a STM32F746BGT6 microcontroller which can then be used by cube satellites to detect clouds from the images it would take. This module is the Image Classification Unit (ICU). As a proof of concept, this ICU will be implemented inside a 3U CubeSats mission developed at the National Space Science and Technology Center, UAE.
Date of Conference: 11-16 July 2021
Date Added to IEEE Xplore: 12 October 2021
ISBN Information: