DGRDet: Dynamic Gaussian Receptive Field Encoding-Based Spiking Neural Networks for Remote Sensing Object Detection

Remote sensing object detection aims to identify and localize specific targets in satellite or aerial imagery. Spiking Neural Networks (SNNs), benefiting from their implicit feedback-based and event-driven brain-inspired dynamics, offer a promising solution to alleviate the high energy consumption of conventional ANN-based detection models. However, existing SNN-based approaches for remote sensing object detection—particularly for small, arbitrarily rotated objects—are still in their infancy and suffer from a substantial performance gap compared with ANN counterparts. In this work, we draw inspiration from the hierarchical sparse perception mechanisms of biological vision and integrate dynamic receptive field modulation into the encoding stage, proposing a high-precision spiking object detection framework tailored for remote sensing image. Specifically, we design a Hierarchical Feedback-based Gaussian Encoding (HFG) scheme, in which the parameters of Gaussian kernels are dynamically adjusted through spike-triggered top-down feedback connections. This mechanism enables the encoding process to adaptively respond to complex geometric variations of remote sensing objects, including rotation and scale changes. Based on the proposed encoding strategy, we develop DGRDet (Dynamic Gaussian Receptive Field Encoding-based Spiking Neural Networks for Remote Sensing Object Detection), a directly trained deep SNN detector for remote sensing image. Extensive evaluations on the large-scale public DOTA dataset demonstrate that DGRDet achieves competitive detection accuracy, outperforming existing SNN-based object detection methods. Moreover, compared with ANN models of comparable detection performance, DGRDet reduces spike activity by 81.31% and requires only 0.12% of the inference energy consumption, achieving a favorable balance between detection accuracy, efficiency, and energy efficiency.