Acta Optica Sinica, Volume. 45, Issue 7, 0728004(2025)
Intelligent Detection of Safety Hazards Along High-Speed Railway Lines Based on Optical Remote Sensing Images
Figures & Tables(24)
Fig. 2. Typical external environmental hazards of railways. (a) Color-coated steel sheet (CCSS) roof buildings; (b) greenhouses
Fig. 3. Frequency-domain analysis of pansharpening. (a) Original images; (b) spectral residual maps between PAN and HRMS images after FFT; (c) spectral residual maps between LMS and HRMS images after FFT
Fig. 4. Overall architecture of BiDFNet (consisting of two components: the HRMS image reconstruction subnet and the PAN image decomposition subnet)
Fig. 5. Structural diagrams of WAF and D3F modules. (a) Structural diagram of WAF; (b) structural diagram of D3F
Fig. 6. Samples of high-speed railway external hazard dataset. (a) Samples of CCSS roof buildings and corresponding labels; (b) samples of greenhouses and corresponding labels
Fig. 8. Structural diagram of DFEM (containing a spatial branch and a channel branch)
Fig. 10. Comparison of model structures of LAVT, CrossVLT, RefSegFormer, and RMSIN
Fig. 11. Qualitative comparisons of reduced resolution experiment on reduced resolution testing set of GaoFen-2 datasets [the first row shows the results of all compared methods, with the rectangular box highlighting the zoomed-in region of interest; the second row presents the residual maps (after normalization) between the fusion results and the reference image, with the color bar on the left indicating the mapping of residual values to colors]
Fig. 12. Qualitative comparisons of the reduced resolution experiment on the reduced resolution testing set of SuperView-1 datasets [the first row shows the results of all compared methods, with the rectangular box highlighting the zoomed-in region of interest; the second row presents the residual maps (after normalization) between the fusion results and the fusion reference image, with the color bar on the left indicating the mapping of residual values to colors]
Fig. 13. Qualitative comparisons of the reduced resolution experiment on the reduced resolution testing set of WorldView-Ⅲ datasets [the first row shows the results of all compared methods, with the rectangular box highlighting the zoomed-in region of interest, and the second row presents the residual maps (after normalization) between the fusion results and the reference image, with the color bar on the left indicating the mapping of residual values to colors]
Fig. 14. Qualitative comparisons of the full resolution experiment on the full resolution testing set of WorldView-Ⅲ datasets [the first row displays the fusion results, with rectangular boxes marking the key regions of interest; the second row shows the magnified view of these selected areaes]
Fig. 15. Qualitative results of the comparison algorithms for CCSS roof building extraction on the testing set (“sample” refers to the original image, “label” refers to the annotated labels, and the rectangular box highlights the regions of interest)
Fig. 16. Multimodal segmentation dataset of hazards along the high-speed rail line (the text description is below the image, with the red annotations representing the target masks corresponding to the text)
Fig. 17. Qualitative comparison of referring image segmentation results (with the segmentation results of each method highlighted in red in the image)
Fig. 18. Examples of on-site verification of external hazard extraction results along the Beijing-Zhangjiakou high-speed railway images (①‒⑥ are on-site verification photographs of hazard targets, with arrows indicating the correspondence between the hazard extraction results in the remote sensing images and the on-site targets). (a) Extraction results around the Nant Grand Bridge in Zhangjiakou City; (b) extraction results around the No. 1 Grand Bridge at Xuanhua Station
Table 1. Sensor parameters of all satellite data sources covered in this article
View tableTable 1. Sensor parameters of all satellite data sources covered in this article
Satellite SuperView-1 GaoFen-2 WorldView-III Ground sample
distance （GSD）
PAN： 0.5 m PAN： 0.8 m PAN： 0.31 m MS： 2 m MS： 3.24 m MS： 1.24 m Sensor band PAN： 450-890 nm PAN： 450-900 nm PAN： 450-800nm Blue： 450-520 nm Blue： 450-520 nm Coastal blue： 400-450 nm Red： 630-690 nm Green： 520-590 nm Green： 520-590 nm Blue： 450-510 nm Red edge： 705-745 nm Red： 630-690 nm Red： 630-690 nm Green： 510-580 nm Near-IR1： 770-895 nm Near-IR： 770-890 nm Near-IR： 770-890 nm Yellow： 585-625 nm Near-IR2： 860-1040 nm Revisit time 2 days by twin satellites Within 5 days 1 day at 1-metre GSD resolution
4.5 days at 20° off-nadir （0.59 m GSD）
Dynamic range 11 bit 10 bit 11 bit
Table 2. Sample number and statistical distribution of hazard targets in self-constructed high-speed railway hazard dataset
View tableTable 2. Sample number and statistical distribution of hazard targets in self-constructed high-speed railway hazard dataset
Satellite Hazard type Sample
number
Label
number
Minimum /m2 Maximum /m2 Sum /m2 Mean /m2 Standard
deviation /m2
SuperView-1 CCSS roof building 12156 25711 1.143800 29692.8 9781844 380.4537 932.3558 Greenhouse 994 7889 0.121742 10920.7 3786571 479.9811 448.2920
Table 3. Number of samples and image resolution in pansharpening experimental datasets
View tableTable 3. Number of samples and image resolution in pansharpening experimental datasets
Satellite
sensor
Sample number of training set Image
size of
training
set
Sample number of validation set Image
size of
validation
set
Sample number
of reduced
resolution
testing set
Image size
of reduced
resolution
testing set
Sample
number of
full resolution testing set
Image
size of full
resolution
testing set
SuperView-1 18126 PAN：
128×128×1
MS：
32×32×4
2015 PAN：
128×128×1
MS：
32×32×4
328 PAN：
512×512×1
MS：
128×128×4
— — GaoFen-2 19809 PAN：
64×64×1
MS：
16×16×4
2201 PAN：
64×64×1
MS：
16×16×4
20 PAN：
256×256×1
MS：
64×64×4
— — WorldView-Ⅲ 9714 PAN：
64×64×1
MS：
16×16×8
1080 PAN：
64×64×1
MS：
16×16×8
20 PAN：
256×256×1
MS：
64×64×8
20 PAN：
512×512×1
MS：
128×128×8
Table 4. Quantitative experimental results of pansharpening (optimal values are in bold)
View tableTable 4. Quantitative experimental results of pansharpening (optimal values are in bold)
Dataset WorldView-Ⅲ Method SSIM SAM ERGAS SCC Q2n RQNR Wavelet 0.8201 0.1206 5.8221 0.4447 0.6852 0.1917 0.0695 0.5403 SFIM 0.8670 0.0932 4.9804 0.5648 0.8089 0.1853 0.0584 0.5594 BDPN 0.9552 0.0685 2.9075 0.6890 0.8901 0.0393 0.0531 0.8439 LPPN 0.9672 0.0597 2.4814 0.7070 0.8933 0.0199 0.0672 0.8689 FAFNet 0.9671 0.0564 2.4436 0.7108 0.9081 0.0259 0.0760 0.8455 Ours 0.9722 0.0518 2.2103 0.7393 0.9173 0.0251 0.0402 0.8904
Table 5. Quantitative results of the comparison experiment for CCSS roof building extraction on the testing set (optimal values are in bold)
View tableTable 5. Quantitative results of the comparison experiment for CCSS roof building extraction on the testing set (optimal values are in bold)
Method IoU Recall Precision F1-score PSPNet 70.88 77.00 83.47 80.10 FLANet 77.63 84.40 87.33 85.84 SegNet 81.78 87.91 90.12 89.00 HRNet v2 82.80 87.40 92.35 89.81 U-Net 84.24 87.09 95.19 90.96 DeepLab v3+ 84.64 87.98 94.59 91.16 Ours 86.48 91.46 93.05 92.25
Table 6. Quantitative comparison of hazard intelligent detection using LAVT and its derivative methods on the self-built multimodal hazard dataset along the high-speed rail line (with the optimal values highlighted in bold)
View tableTable 6. Quantitative comparison of hazard intelligent detection using LAVT and its derivative methods on the self-built multimodal hazard dataset along the high-speed rail line (with the optimal values highlighted in bold)
Method Dataset Pre@0.5 Pre@0.6 Pre@0.7 Pre@0.8 Pre@0.9 oIoU mIoU LAVT Test 90.45 86.08 79.62 68.2 40.95 81.790 79.980 Validation 89.69 84.87 78.88 67.37 38.74 80.210 79.050 RefSegFormer Test 90.22 86.54 80.36 70.33 45.71 82.338 80.479 Validation 89.39 84.98 79.14 68.86 42.78 81.273 79.394 RMSIN Test 90.61 86.49 79.78 68.17 40.57 81.580 80.140 Validation 89.85 85.14 78.45 67.11 39.65 80.320 79.350 CrossVLT Test 90.15 85.95 79.78 69.88 43.79 82.300 80.290 Validation 88.94 84.28 78.77 67.60 41.44 80.420 79.330
Tools
Get Citation
Copy Citation Text
Yingjie Li, Dongsheng Zuo, Weiqi Jin, Su Qiu. Intelligent Detection of Safety Hazards Along High-Speed Railway Lines Based on Optical Remote Sensing Images[J]. Acta Optica Sinica, 2025, 45(7): 0728004
Paper Information
Category: Remote Sensing and Sensors
Received: Dec. 14, 2024
Accepted: Jan. 22, 2025
Published Online: Apr. 27, 2025
The Author Email: Weiqi Jin (jinwq@bit.edu.cn)
CSTR:32393.14.AOS241893
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20