Exploration of Deep Neural Networks with Symmetric Simplicial Layers for On-Satellite Earth Observation Processing

Bringing artificial intelligence on-board of space crafts holds significant promise for enhancing the capabilities of space missions. On-board processing can enable responsive missions that are not limited by the latency and bandwidth of
the communication to earth. However, in many cases dedicated solutions are required due to the resource constraint environment of satellites. In this contribution we explore architectures for onboard processing of earth observation imagery based on deep neural networks with Symmetric Simplicial (SymSim) layers. The performance of the networks are assessed for the task of plume and scenery classification in RGB earth observation pictures. We propose an extended SymSim algorithm and show its performance in small variants of ResNet compared to the same architectures with convolutional blocks.