Ensuring the safety of the external environment is a critical aspect of railway operation safety. Traditional manual inspections for identifying safety hazards along high-speed railway (HSR) lines are labor-intensive, costly, and inefficient due to geographical and weather conditions. Leveraging deep learning techniques for the intelligent extraction of hazard targets from remote sensing imagery can significantly improve data utilization and processing efficiency, becoming an essential trend for the intelligent development of HSR systems. However, the development of this technology still faces challenges at three levels: data quality, model performance, and application interaction. At the data level, the primary challenge lies in the lack of high-quality training datasets. High-resolution color remote sensing images are typically generated by fusing multispectral (MS) images with panchromatic (PAN) images, and the effectiveness of fusion algorithms directly influences the fidelity of ground target information. Although existing methods have made progress in preserving spatial and spectral fidelity, challenges such as spatial-spectral misalignment and information loss remain unresolved, particularly as satellite capabilities improve for observing small targets. In addition, the absence of dedicated datasets for extracting hazards along HSR lines restricts the development and training of intelligent extraction models. At the model level, challenges arise from significant variations in target scales, as well as the spectral and spatial inconsistencies of targets (e.g., identical objects appearing differently and different objects sharing similar spectral characteristics). As the resolution of remote sensing images improves, non-target background information increasingly contributes to noise, complicating the discrimination and learning of target features. At the application level, existing intelligent detection methods based on single-modal information (images) limit the understanding of image semantics and require a high level of professional expertise from users. With the new interactive experiences brought by language models to various fields, exploring how to combine language input for more flexible and efficient human-machine interaction is a key issue that needs to be addressed. In this paper, we aim to systematically investigate intelligent detection technology for hazards along high-speed rail lines based on remote sensing images, focusing on the three aforementioned aspects to promote the intelligent development of high-speed rail safety assurance.

We explore the application of optical remote sensing imagery in railway safety. It investigates intelligent hazard detection techniques for HSR by addressing two core aspects: constructing high-quality datasets of HSR hazard targets and enhancing the performance of remote sensing image extraction models. To improve image fusion quality, we analyze the pansharpening task from a frequency domain perspective. By exploring the commonalities between the multiscale decomposition capability of wavelet transforms and the multiscale structures of deep learning models, we propose a multiscale spatial-frequency domain dynamic fusion algorithm, BiDFNet. BiDFNet employs a bidirectional subnet structure: the backward subnet extracts high-frequency components of the PAN image via wavelet transform, while the forward subnet reconstructs high-frequency information from the MS image progressively through a wavelet-based adaptive fusion module. Furthermore, a dual-domain dynamic filtering module enhances the algorithm's generalizability with a parameter-efficient design. We create an HSR hazard dataset targeting typical hazard types using high-quality fused data. To address issues such as significant target scale variations, small-target omission, and severe background noise interference, we develop a deep-learning-based adaptive deformable fitting method, DFEANet, for remote sensing target extraction. DFEANet employs an encoder-decoder architecture to extract multiscale features. Deformable convolution is introduced to adaptively fit the receptive field to target shapes, while edge alignment between adjacent feature levels is enhanced using optical flow concepts, improving edge extraction precision. In addition, a gating mechanism is adopted to regulate information flow during feature fusion, effectively suppressing background noise. To enhance the model's ability to understand advanced semantics, we construct a multimodal segmentation dataset for hazards along high-speed rail lines based on the aforementioned hazard dataset. We also conduct both qualitative and quantitative comparisons of several multimodal segmentation algorithms, represented by LAVT, on the hazard multimodal segmentation task, exploring intelligent detection technology along high-speed rail lines by integrating textual data.

The proposed pansharpening method is compared with traditional and deep learning-based methods using imagery from the GaoFen-2, SuperView-1, and WorldView-III satellites. Quantitative experimental results (Table 4) show that the proposed method achieves optimal values across five metrics in reduced resolution experiments, with minimal differences between fusion results and reference images. In full-resolution experiments, the proposed method demonstrates excellent generalization performance, especially for small objects such as vehicles. As shown in Fig. 13, it can accurately reconstruct spatial structure and spectral information. Using the established HSR hazard dataset, the proposed hazard extraction model is tested. Experimental results (Table 5 and Fig. 14) indicate that the proposed method successfully achieves complete extraction of hazard masks across different scales and effectively learns the spectral characteristics of color-coated steel sheet (CCSS) roof buildings. Both quantitative analyses and qualitative comparisons demonstrate that it outperforms other benchmark algorithms, achieving the best performance. In addition, this is the first to explore multimodal segmentation technology for hazard detection along high-speed rail lines, providing a reference for future research on more flexible and efficient intelligent hazard detection techniques. By integrating the proposed algorithm with railway geoinformation, statistical analysis and risk-level classification of hazard information are enabled. Combined with electronic maps, the visualization of hazard data is facilitated, effectively supporting the inspection of external railway safety hazards. Field verification of the algorithm's results shows that, among 20 sampled hazard sites, 17 are correctly identified, with 3 missed detections.

In this study, we focus on the intelligent safety hazard detection task for HSR based on optical remote sensing images. High-quality data form the foundation of intelligent hazard detection. To address the challenges of small-target reconstruction introduced by the enhanced resolution of remote sensing imagery, we integrate traditional multiresolution analysis concepts with deep learning models. By progressively extracting and processing high-frequency components from PAN images, the method mitigates issues of small-target information blending or loss during MS image fusion, achieving optimal spatial and spectral fidelity compared to deep learning methods operating purely in the spatial domain. Based on the constructed HSR hazard dataset, the proposed hazard extraction model demonstrates superior accuracy in extracting CCSS roof buildings compared to state-of-the-art deep learning methods, significantly improving the efficiency of hazard inspection along HSR lines. By incorporating textual data, we construct a multimodal segmentation dataset for hazards along high-speed rail lines. Test results of various multimodal segmentation algorithms on this dataset show that the introduction of textual data enhances the flexibility of intelligent hazard detection. Users are expected to selectively detect hazards based on their relative position to the railway line, thus building a more intelligent human-machine interaction model. Considering practical application needs and technological development trends, we discuss the future development requirements of HSR safety hazard detection technology based on optical remote sensing images. We also conduct a systematic investigation of intelligent hazard detection along HSR lines, starting from data selection and progressing to application deployment. Given the limited amount of related research in this field, the findings of this paper provide valuable references for future studies.