Lightweight Remote Sensing Object Detection Model Based on Spiking Neural Networks

Remote sensing target detection harnesses aerial and satellite imagery to drive critical insights for environmental monitoring, disaster response, and strategic planning. However, the successful implementation of high-performance remote sensing object detection currently relies on artificial neural networks, which need substantial computational resources, posing significant challenges for edge computing scenarios. In this study, we capitalize on the energy efficiency of spiking neural networks to design an innovative remote-sensing object detection model. Compared to equivalent ANN models, our model reduces computational resource requirements significantly. Furthermore, to minimize memory requirements, we introduce a dual quantization strategy for membrane potentials and synaptic weights. It quantifies both membrane potentials and synaptic weights to lower bit levels. Extensive experiments demonstrate that our lightweight remote sensing object detection model achieves competitive accuracy on the NWPU VHR-10 and SSDD datasets with an 80% reduction in memory requirements, respectively. Our solution provides a consistently effective strategy on edge devices with limited resources.