Abstract: Information extraction of rural roads from remote sensing images faces significant challenges associated with adaptation to complex scenarios and the balance between high accuracy and lightweight deployment. To address these issues, this study proposes a lightweight information extraction method that integrates dual-branch feature fusion with attention-based knowledge distillation, aiming to achieve precise detail-topology modeling of rural roads and efficient deployment. First, to overcome the inherent limitations of single-structure models that highlight either local details or global context, a teacher model characterized by the dual-branch fusion integrating ResNet with Vision Transformer (ViT) was established, termed Teachers-Unet. In this model, a bidirectional fusion module (BiFusion) was employed to dynamically fuse the local detail features captured by the convolutional neural network (CNN) model with the global topological features captured by the Transformer. Second, an attention-based knowledge distillation strategy was designed for a balance between high accuracy and lightweight deployment. Specifically, a KDAttention module was introduced into a lightweight student model, enabling the model to focus on key rural roads and inherit structured representations from the teacher model. Finally, ablation and comparative experiments were conducted on the DeepGlobe and WHU-RuR rural road datasets. The results demonstrate that the Teachers-Unet model yielded an intersection-over-union (IoU) value of 75.23% and an F₁ score of 85.86% on the DeepGlobe dataset, increasing by 13.59% and 9.59%, respectively, compared to the optimal single-branch model. The attention-based knowledge-distilled student model involved only 8.3×10⁶ parameters, less than 25% of those contained in the teacher model. The student model yielded an IoU value of 67.14% and an F₁ score of 80.34% on the WHU-RuR dataset, with an accuracy gap of less than 1% compared to the teacher model. Compared to conventional models such as U-Net and D-LinkNet, the proposed method increased the IoU by more than 12% on the rural road dataset. This method can effectively overcome multiple issues induced by strong background interference, a high proportion of narrow roads, and severe occlusions in rural road information extraction. Overall, the method developed in this study achieves a balance between accuracy and lightweight deployment, providing effective technical support for information extraction of rural roads from remote sensing images.