EAS-Unet: Edge-Aware Spiking Unet for Real-Time Runway Segmentation in Complex SAR Scenes

Airport runway segmentation in spaceborne SAR imagery is critical for aircraft localization and holds strategic importance in military reconnaissance, aviation safety, and emergency response. However, existing semantic segmentation methods based on artificial neural networks (ANNs) usually have a large number of parameters and high energy consumption, failing to meet the strict storage and power constraints of spaceborne platforms. Therefore, this paper proposes an edge-aware spiking Unet (EAS-Unet) based on spiking neural networks (SNNs), which fully utilizes SNNs' advantages in real-time processing and low-power computation. Furthermore, aiming to resolve the scale variability of runways and the complex structures (e.g., taxiways and aprons) which pose challenges to precisely modeling runway edges, we design a dynamic multi-scale feature extraction module with parallel adaptive dilated convolutions for runway scale adaptability, an edge-aware guidance module that enhances contour details using shallow edge features and a local-global context collaborative fusion module integrating hierarchical semantic context. Experiments on GF-3 dataset demonstrate that EAS-Unet outperforms existing ANN segmentation models, achieving better performance and lower energy consumption within a short time step.