MSCR-HRNetV2: High-Resolution Remote Sensing Image Segmentation for Railway Scenes

Remote sensing image semantic segmentation not only enables the acquisition of valuable category information regarding surface objects but also provides crucial spatial information. This foundational data, with its high research value, plays a pivotal role in advancing the understanding and utilization of spatial environmental resources. To construct a suitable space for railway transportation and improve the channel information mismatch problem in the multi-scale fusion unit of HRNetV2, in terms of data, a high-resolution remote sensing image dataset of two railway sections of Kaha Railway and Ganquan Railway was collected, and annotated, including 210 original remote sensing images with a resolution of
$4800 \times 2986$
, a viewing altitude of 2km, and a spatial resolution of 0.24m. In terms of methods, an embedded self-attention mechanism module was introduced and a multi-scale consistency regularization (MSCR) method was introduced to propose MSCR-HRNetV2, achieving pixel-level semantic segmentation representation of high-resolution remote sensing images. The experimental results show that the MIoU results of MSCR-HRNetV2 on the self-built railway dataset and the public Potsdam remote sensing image dataset reached 0.8137 and 0.7596 respectively, which were 0.0313 and 0.0201 higher than the original HRNetV2. Compared with other typical remote sensing image semantic segmentation methods such as Deeplabv3+, the segmentation accuracy has also been improved by 0.0386 and 0.0219. The high-resolution remote sensing image dataset constructed in this study provides an important data source for studying the segmentation problem of high-resolution remote sensing images around railways. The proposed MSCR-HRNetV2 benchmark method lays the foundation for the full development and utilization of railway transportation space resources.