A full-attention network with an open dataset for large-scale building semantic segmentation along long-span high-speed rail lines

Buildings adjacent to high-speed rail (HSR) tracks are one of the key concerns in the environmental hazard monitoring of HSR lines. The rapid and accurate extraction of building information for real-time inspection through high-resolution remote sensing (HRRS) images is a crucial and effective method to ensure the safe operation of high-speed trains. However, a significant challenge associated with this task emerges from the extremely intricate scenes and diverse buildings within HRRS images that cover long-span HSR lines. To surmount this challenge, this paper presents a high-quality HSR dataset and a novel full-attention semantic segmentation network (referred to as FANet) for building detection along HSR lines. The dataset provided contains rich HRRS image data and high-quality pixel-level building annotations. The proposed FANet is designed to enhance model performance by paying comprehensive attention to multi-level, multi-scale, and global relational information. Specifically, after analyzing the special requirements, a multi-scale global attention (MSGA) module and an affinity attention fusion (AAF) module have been devised and incorporated into FANet. The MSGA module is utilized to ensure the integrity and consistency of building extractions across various shapes and sizes, while the AAF module aids in the fine-grained identification of building outline boundaries. Extensive experiments conducted using the introduced HSR dataset, along with two publicly available building datasets–the Massachusetts building dataset and the Inria aerial image dataset–demonstrate the superior performance of the proposed approach. Moreover, the paper comprehensively discusses various applications that the HSR dataset can support for future research endeavors. The HSR dataset and method code will be made publicly accessible at https://github.com/QiaoWenfan/Building_Dataset.