G2L2Net: A Road Extraction Method for Remote Sensing Images via Gated Global–Local Linear Attention

Road extraction from remote sensing imagery plays a pivotal role in a wide range of geospatial and urban applications. Nevertheless, this task remains inherently challenging due to the intricate morphological variations of roads and frequent occlusions or interference caused by complex background environments. To address these challenges, we propose a road extraction network based on gated global–local linear attention (G $^2$ L $^2$ Attention). First, we introduce a linear deformable convolution and design a linear input-dependent deformable convolution (LID2Conv), which adaptively modulates convolution offsets and weights in a content-aware manner. In addition, we design a top-K-based sparse gated weight (TGW). We use this gated mechanism as a shared weight to multiply with local and global information to achieve G2L2Attention. Local information is obtained by LID2Conv, and we gain global information by introducing 2-D selective scan (SS2D). These two pathways are integrated through the proposed G2L2Attention, enabling an efficient and consistent fusion of hierarchical spatial features. The extracted features are passed to the decoder. This approach improves road detail representation and provides accurate contextual information. Experiments conducted on three public road datasets demonstrate that G2L2Net outperforms the existing methods in various evaluation metrics. Our source code is available at https://github.com/ZehuaChenLab