Space Use-Case: Onboard Satellite Image Classification

Satellite imagery is the most important sector of space industry, as around 38% of satellites are fully dedicated to earth observation. Then, for each sample, the error gradient is propagated backward in the network, and the weights are adjusted to minimise this error. However, the bandwidth between the satellite and its ground control centre is very narrow. Furthermore, the quality of photographs can be altered by a wide variety of factors including clouds, fumes, shadows, and planes. Thus, one might want to avoid congesting the already limited bandwidth with such useless images. The main idea is to pre-process the images onboard the satellite, automatically deciding whether a photograph is exploitable and worth sending to the ground. To do so, Neural Network applications can be deployed on a low-power FPGA, such as the low-power image processing oriented Tulipp platform. In this use-case, a Hybrid Neural Network architecture developed for satellite applications is adapted to the Tulipp EMC2-ZU3EG board, to serve as a use-case for the Tulipp project and demonstrate the possibilities of the board in real-world applications.