SwinR-UNet: A Hybrid Architecture for High-Precision Recognition and Width Calculation of Rural Highways

This is the authors' abstract. We don't add key points for this paper.

The precise measurement of rural highway widths is essential for the digital administration and maintenance of transportation infrastructure. However, when applied to high-resolution remote sensing (RS) imagery, conventional deep learning models frequently suffer from local feature loss and topological discontinuities. These issues primarily stem from the complex backgrounds, narrow dimensions, and diverse surface materials characteristic of rural highways. To address these limitations, we propose SwinR-UNet, a novel hybrid architecture that integrates U-Net with a Swin-Transformer, specifically designed for RS-assisted recognition and width-attribute updating of rural highways in complex environments. In this architecture, a spatial and channel reconstruction convolution module is introduced into the shallow layers of the U-Net encoder to refine local edge features and suppress background noise. Furthermore, a Swin-Transformer is deeply embedded within both the bottleneck layer and the decoder, leveraging its shifted-window self-attention mechanism to precisely capture the global, long-range linear dependences of the highway networks. In addition, we develop an improved inverse distance weighting interpolation algorithm that incorporates multidimensional feature matching and edge confidence, supporting the physical quantification and interpolation-based updating of highway widths. Extensive experiments on a public dataset (WHU-RUR) and a large-scale custom dataset (CRRD) show that SwinR-UNet outperforms the evaluated baseline models in segmentation accuracy, yielding an F1-score of 79.93% and an intersection over union of 66.57% on the CRRD dataset. Notably, in highly challenging scenarios involving extremely narrow or unpaved highways, our approach maintains a comprehensive mean absolute error of approximately 0.33 m for width calculation, demonstrating improved robustness. Consequently, this study provides an RS-assisted framework for updating and quantifying rural highway width attributes, particularly when existing vector road data are incomplete, inaccurate, or lack reliable width information.