Towards Energy-Efficient Satellite-Onboard Landcover Classification with Spiking Neural Networks

Abstract

While artificial neural networks significantly boosted remote sensing analytics with unsupervised, semi-supervised, and supervised deep learning techniques in the past decade, utilization of such data science models implies notable demand in energy resrouces. While graphical processing units such as NVIDIA's A100 run at power consumptions of up to ~300W, our brain operates at about 20W to master tasks such as image analysis. Spiking Neural Networks (SNN) model brain neurons with their ability to accumulate signals to transmit a unit signal to the next neuron when a given threshold is passed. This sparse, energy-efficient propagation and processing of SNN has potential to get implemented in dedicated chips for AI-applications at the edge.