Embedded cloud segmentation using AI : Back on years of experiments in orbit on OPS-SAT
Résumé
Since 2019, IRT Saint Exupéry has been researching embedded cloud segmentation and has conducted several experiments on board the OPS-SAT satellite from the European Space Agency. Using FPGA implementations that efficiently execute artificial neural networks, an image with dimensions 2048×1944×3 is inferred in less than 126 ms while consuming less than 2W of power. These neural network inferences in the programmable logic part of a FPGA are, to our knowledge, a
first in orbit.
In this paper, we summarize the work and main results obtained by IRT Saint Exupéry during the whole OPS-SAT mission. We start by describing and comparing our main neural network topologies for cloud segmentation together with a ZGP formula, an ultra-light mathematical equation. Keeping only the best model, we then train it on various evolutions of databases built over the years. Deploying this network on OPS-SAT and on a set of other hardware targets of interest (Google Coral
/ Intel Neural Compute Stick 2), we finally demonstrate that the processing throughput on FPGA is 10 to 36 times faster than on manufacturer-specific ASICs with an equivalent power consumption and better overall algorithmic performance.
This paper also contains original material. In particular, it details the construction of a generic VHDL library developed by IRT Saint Exupéry, and the other tools and methods we used in the context of the CIAR project.