Abstract <p>Road extraction from remote sensing images plays a crucial role in navigation, traffic management, and urban construction, etc. With the development of deep learning technology, the performance of road extraction has been greatly improved. However, it is still a challenging task to extract narrow road with different directions accurately, and edge information that plays an important role in road extraction has not been sufficiently utilized. To deal with the above problems, we proposed an improved road extraction network Swift-SegEdgeNet based on the lightweight SwiftFormer structure. There are mainly two contributions. Firstly, we proposed a multi-task learning structure by integrating the edge detection task with the semantic segmentation task. With the guidance of edge information, it is beneficial to improve the road location accuracy. Secondly, in order to obtain the road direction more accurately, we propose a dual-branch module, which includes a standard convolution branch and a stripe convolution branch. By introducing four strip convolutions, it is helpful to extract the direction information of the roads. Experiments on the Massachusetts road dataset and the DeepGlobe 2018 road dataset validating that the proposed Swift-SegEdgeNet can effectively improve the accuracy of road extraction with lower computational complexity and few memory usage. And comparison results with the other state-of-the-art methods demonstrated that our method achieved superior performance with the highest mIOU, mDC, F1-score and Recall.</p> Graphical abstract <p></p>