Exploring the Potential of Ultra-Low Latency Spiking Neural Network in Remote Sensing Image Classification

With the rapid advancement of artificial intelligence, the depth of artificial neural networks (ANN) has increased, resulting in higher power consumption and resource utilization. To address these concerns, researchers have begun to explore spiking neural networks (SNN) as an alternative. SNN offers the potential for improved energy efficiency while achieving comparable performance. In recent years, there has been a growing interest in applying SNN to remote sensing classification scenarios. However, one major challenge hindering its widespread adoption is the high latency associated with real-time processing. To address this challenge, we adopt a novel hybrid training approach that combines ANN-SNN conversion with backpropagation through time. We evaluate the effectiveness of our approach on three publicly available datasets: MSTAR, AID and UCMerced. Our experiment results demonstrate that the proposed ultra-low latency SNN achieves classification accuracy rates of 99.05%, 91.36%, and 92.62%, respectively, with a single time step. To the best of our knowledge, this is the first time to apply ultra-low latency SNN to remote sensing image classification tasks. Our findings suggest that this innovative approach has significant potential for future research in the deployment of these models for on-board processing.