Road Extraction with Weak Features and Complex Backgrounds Based on Atrous–Strip–UNet

TL;DR

A new neural network called ASUNet improves road extraction from satellite images by handling weak road features and complex backgrounds.

Contribution

ASUNet integrates atrous and strip convolution modules into an encoder-decoder architecture for better road extraction in remote sensing.

Findings

ASUNet achieves F1 scores of 0.7292 on the Zhouqu Road Dataset and 0.7134 on the DeepGlobe Dataset.

The model performs better in scenarios with weak road features or complex backgrounds compared to existing methods like BiSeNet and LinkNet.

Abstract

What are the main findings?
An atrous–strip–unet (ASUNet) is proposed, which integrates atrous convolution and strip convolution modules into an encoder–decoder architecture to address road extraction challenges from remote sensing images (e.g., weak road features, complex backgrounds, and occlusions).On the self-compiled Zhouqu Road Dataset (covering rural/suburban/urban roads of western Chinese counties) and the public DeepGlobe Dataset, ASUNet achieves F1 scores of 0.7292 and 0.7134, respectively. The proposed algorithm demonstrates high accuracy and effectiveness for road extraction, constituting a valuable addition to the road extraction toolkit.

An atrous–strip–unet (ASUNet) is proposed, which integrates atrous convolution and strip convolution modules into an encoder–decoder architecture to address road extraction challenges from remote sensing images (e.g., weak road features, complex backgrounds, and occlusions).

On the self-compiled Zhouqu Road Dataset (covering rural/suburban/urban roads of western Chinese counties) and the public DeepGlobe Dataset, ASUNet achieves F1 scores of 0.7292 and 0.7134, respectively. The proposed algorithm demonstrates high accuracy and effectiveness for road extraction, constituting a valuable addition to the road extraction toolkit.

What are the implications of the main findings?
ASUNet provides a reliable approach for accurate road extraction in remote sensing scenarios with weak features or complex backgrounds, supporting applications like transportation planning and disaster emergency response.The Zhouqu Road Dataset complements the lack of representative county-level road data in western China, offering a valuable resource for subsequent remote sensing road extraction research.

ASUNet provides a reliable approach for accurate road extraction in remote sensing scenarios with weak features or complex backgrounds, supporting applications like transportation planning and disaster emergency response.

The Zhouqu Road Dataset complements the lack of representative county-level road data in western China, offering a valuable resource for subsequent remote sensing road extraction research.

With the continuous improvement of remote sensing image resolution, accurately extracting road information from complex backgrounds remains challenging. This is because roads present diverse morphological characteristics across regions and scales, and their spectral features are highly similar to those of surrounding objects, such as buildings and bare soil, making them hard to distinguish. Occlusion by buildings and trees leads to incomplete road extraction. To solve the above problems, this paper proposed the atrous–strip–Unet (ASUNet), an encoder–decoder network into which atrous and strip convolution modules are inserted to extract roads with weak features and complex backgrounds from high-resolution remote sensing images. In this study, we construct the Zhouqu Road Dataset from high-resolution aerial imagery, covering representative road types (rural, suburban, and urban) characteristic of county-level settlements in western China. By comparing several advanced algorithms with excellent learning performance—including BiSeNet and LinkNet—on both the Zhouqu Road and DeepGlobe Datasets, the improved and optimized model presented in this paper demonstrates better extraction accuracy and effectiveness; it achieves F1 scores of 0.7292 and 0.7134 on the two datasets, respectively. It is particularly worth mentioning that our proposed algorithm shows better performance in scenarios where road features are weak or backgrounds are complex.

Figures