WeavingUnet: Enhancing dense road extraction by integrating horizontal and vertical features

Road extraction plays a crucial role in automatically identifying road networks from remote sensing images, which is essential for various applications. However, existing road extraction methods perform poorly when dealing with dense roads. Inspired by the act of ”weaving” in real life, this paper proposes a dense road extraction method named WeavingUnet, aimed at addressing the difficulties in effectively extracting dense roads using existing methods. This model explores a novel approach to road feature extraction by analyzing road features in various directions and simulating the geometric shapes of roads to more accurately perform dense road extraction. The model incorporates innovative components such as Snake Weaving Attention, Contextual Information Weaving Module, Global Information Extraction Module, and Multi-Scale Weaving Decoder, which collectively enhance the extraction of multi-level road information and optimize the road reconstruction process. The model is tested on the public datasets DeepGlobe and Massachusetts, achieving F1 scores of 82.73% and 72.52%, and IoUs of 71.44% and 61.01%, respectively. Compared with the latest SOTA models, the F1 scores increase by 1.53% and 1.22%, and the IoUs improve by 1.76% and 1.87%. This study also conducts a comprehensive analysis of dense road extraction and compared it with SOTA models. In the densest areas, the false negative rates (FNs) in the test images are reduced by 4.93‰ and 1.47‰, respectively, demonstrating the effectiveness of WeavingUnet in handling dense roads. The code and dataset are available at https://github.com/XianZhi-Ma/WeavingUnet . • Find road extraction models underperform in dense areas; verify experimentally. • Utilize striped convolutions in various directions to extract dense road features. • Design WeavingUnet specifically for dense road extraction tasks, achieving SOTA.