A method for capturing DEM elevation data resources for smart rail transit

Abstract

This invention discloses a method for capturing DEM elevation data resources for intelligent rail transit, relating to the field of intelligent rail transit technology. This invention utilizes the cross-modal attention mechanism in the AIGC multimodal feature extraction model to deeply mine the semantic relationships of terrain between different data sources, making the generated results more accurate in representing the terrain structure. In the diffusion model generative reconstruction process, geometric boundary constraints of the rail transit line corridor are introduced to ensure that the generated DEM elevation data is strictly limited to the line corridor, avoiding boundary anomalies caused by unconstrained generation. Furthermore, by combining multi-scale spatial consistency checks with local outlier identification, it can finely distinguish between abrupt and gradual anomalies, and gradually converge to a preset accuracy through a neighborhood-weighted iterative correction mechanism. This eliminates anomalies while preserving the true terrain details to the greatest extent, ultimately outputting high-quality DEM elevation data resources that meet the requirements of rail transit engineering applications.

Description

Technical Field

[0001] This invention relates to the field of intelligent rail transit technology, specifically to a method for capturing intelligent rail transit DEM elevation data resources. Background Technology

[0002] As the backbone of modern urban public transportation, rail transit has a strong demand for high-precision digital elevation model (DEM) data along its routes at all stages of planning, design, construction, operation, and maintenance. DEM data provides fundamental topographic support for key business scenarios such as longitudinal profile design, flood control and drainage analysis, determination of the spatial relationship between surrounding buildings and the line, and emergency response decision-making. However, rail transit lines typically stretch for tens to hundreds of kilometers, traversing diverse landforms including urban built-up areas, suburban farmland, and hilly terrain, presenting complex and varied topographic conditions. This places extremely high demands on the accuracy, completeness of coverage, and timeliness of the acquired DEM data.

[0003] Currently, the acquisition of DEM data along rail transit lines mainly relies on the following technical approaches: The first is based on traditional photogrammetry, using aerial photographs to generate DEMs through stereo pair matching. This method works well in textured areas such as urban built-up areas, but is prone to matching failures in densely vegetated areas and areas with weak texture such as water bodies, leading to holes or anomalies in the elevation data. The second is based on airborne LiDAR to directly acquire 3D point clouds of the ground surface and interpolate to generate DEMs. This method has high accuracy and is less affected by ground cover, but the coverage area of ​​a single flight is limited, requiring multiple flights for large-scale lines, resulting in high data acquisition costs. Furthermore, data occlusion problems still exist in steep mountains and densely built-up areas. The third is based on spaceborne synthetic aperture radar interferometry (InSAR) technology to acquire DEMs. This method has the advantages of all-weather and large-area coverage, but its spatial resolution and absolute elevation accuracy are usually lower than LiDAR methods due to radar wavelength limitations, and it is prone to volume scattering bias in vegetated areas. The fourth is to directly utilize existing publicly available DEM products (such as SRTM, ASTER GDEM, and Copernicus). While DEMs and other publicly available products can be used for cropping, their spatial resolution is mostly 30 meters or 90 meters, which is insufficient to meet the requirements of rail transit engineering for sub-meter to meter-level elevation accuracy. Furthermore, the data is often not up-to-date and cannot reflect the elevation changes brought about by urban construction and topographical changes in recent years. Summary of the Invention

[0004] To address the aforementioned technical problems, a method for capturing DEM elevation data resources for intelligent rail transit is provided. This technical solution resolves the issues raised in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] In a first aspect of the present invention, a method for capturing DEM elevation data resources for smart rail transit is provided, comprising:

[0007] Acquire multi-source remote sensing data within a preset width range along the rail transit line. The multi-source remote sensing data includes optical satellite imagery data, airborne lidar point cloud data, and spaceborne synthetic aperture radar interferometric data. At the same time, acquire the three-dimensional coordinate sequence of the design centerline of the rail transit line and existing DEM reference data.

[0008] Radiometric correction, geometric fine correction and point cloud denoising are performed sequentially on the multi-source remote sensing data, and the processed data from each source are uniformly converted to the same geographic coordinate system and the same spatial resolution to form a registered multi-source dataset.

[0009] The registered multi-source dataset is input into the pre-trained AIGC multimodal feature extraction model. The terrain semantic information from different data sources is fused through a cross-modal attention mechanism to generate a multimodal feature tensor containing elevation gradient distribution, surface roughness features and terrain structure semantics.

[0010] Guided by multimodal feature tensors, generative reconstruction of DEM elevation grid data is performed using a diffusion model. Geometric boundary constraints of rail transit corridors are introduced during the reconstruction process, and initial DEM elevation grid data is output.

[0011] Multi-scale spatial consistency checks and local outlier identification are performed on the initial DEM elevation grid data. Elevation deviation statistics are calculated in combination with existing DEM reference data to generate data quality assessment labels.

[0012] Based on the data quality assessment labels, an iterative correction process based on neighborhood weighting is performed on the abnormal areas in the initial DEM elevation grid data until the elevation deviation statistics converge to the preset threshold range, and the final rail transit DEM elevation data resources are output.

[0013] Preferably, the step of sequentially performing radiometric correction, geometric fine correction, and point cloud denoising on the multi-source remote sensing data, and then uniformly converting the processed source data to the same geographic coordinate system and the same spatial resolution to form a registered multi-source dataset, specifically includes:

[0014] Radiometric correction based on the sensor's radiometric response function is performed on the optical satellite imagery data, converting the raw digital quantization values ​​into surface reflectance values. The conversion relationship is as follows:

[0015] ;

[0016] in, For band The surface reflectance value, The original numerical quantization value, This represents the radiation shift in that band. This represents the solar irradiance at the top of the atmosphere in this band. The zenith angle of the sun. This is the Earth-Sun distance correction factor;

[0017] The radiation drift is calculated using the overlapping area of ​​adjacent temporal images, and the reflectivity value is compensated for the drift.

[0018] The airborne lidar point cloud data undergoes point-by-point intensity normalization to eliminate intensity differences between different flight sorties. The normalization calculation method is as follows:

[0019] ;

[0020] in, The normalized intensity value. This is the original strength value. This represents the average intensity at all points during that flight sortie. The standard deviation of the intensity of all points for this flight sortie;

[0021] The statistical outlier removal algorithm is used to identify and remove isolated noise points in a point cloud whose distance from surrounding point groups exceeds a preset multiple of the local standard deviation. The judgment criteria are as follows:

[0022] ;

[0023] in, For the first a point to its The average distance between nearest neighbors, For all points The mean of the average distance to nearest neighbors. For all points Standard deviation of the nearest neighbor mean distance A preset multiple threshold is used to identify points that meet the condition as noise and remove them.

[0024] Perform flat-ground phase removal and terrain phase unwrapping processing on spaceborne synthetic aperture radar interferometric data based on reference DEM;

[0025] The radiometrically corrected optical image, the denoised lidar point cloud, and the unwrapped interferometric data were geometrically refined based on ground control points. The geographic coordinates of all source data were uniformly transformed to the CGCS2000 coordinate system, and the spatial resolution of all source data was uniformly adjusted to a 2-meter grid spacing using a bilinear resampling method. The bilinear interpolation calculation method is as follows:

[0026] ;

[0027] in, For target grid points Resampled value at that point, , , , For the four known grid points surrounding the target point, The original values ​​corresponding to the known grid points are used to form a registered multi-source dataset.

[0028] Preferably, the step of inputting the registered multi-source dataset into the pre-trained AIGC multimodal feature extraction model and fusing terrain semantic information from different data sources through a cross-modal attention mechanism specifically includes:

[0029] The registered optical image data, lidar point cloud data, and interferometric data are input into the corresponding single-mode encoders to extract the local terrain feature vectors of each source data.

[0030] A cross-modal attention query matrix is ​​constructed, using the optical image feature vector as the query vector and the lidar point cloud feature vector as the key vector. The attention weight matrix between the optical image and the point cloud is calculated as follows:

[0031] ;

[0032] in, The optical image feature vector matrix, The feature vector matrix of the lidar point cloud. The feature vector matrix of the lidar point cloud. Given the dimension of the key vector, the weight matrix is ​​weighted and summed with the point cloud feature vector to obtain the optical-point cloud cross-modal fusion feature;

[0033] Then, using the interferometric data feature vector as the query vector and the optical-point cloud fusion feature as the key vector, the attention weight matrix between the interferometric data and the fusion feature is calculated, and a second weighted sum is performed to obtain the three-modal fusion feature vector.

[0034] Max pooling and average pooling operations are performed on the three-modal fusion feature vector along the spatial dimension. After concatenating the pooling results, a multimodal feature tensor containing elevation gradient distribution features, surface roughness features, and terrain structure semantic features is generated through a fully connected layer mapping.

[0035] Preferably, the generation of a multimodal feature tensor comprising elevation gradient distribution, surface roughness features, and terrain structure semantics specifically includes:

[0036] The elevation values ​​of all points within each grid cell are calculated for the registered lidar point cloud data. The elevation differences between adjacent grid cells are calculated along the east-west and north-south directions, respectively, to obtain the east-west and north-south elevation gradient components. The east-west elevation gradient is calculated as follows:

[0037] ;

[0038] in, For the first The east-west elevation gradient components of the location This represents the elevation value of the adjacent grid cell on the east side. This represents the elevation value of the current grid cell;

[0039] The method for calculating the north-south elevation gradient is as follows:

[0040] ;

[0041] in, For the first The north-south elevation gradient components of the location, The elevation value of the adjacent grid cell on the north side;

[0042] The two components are stitched together according to their grid positions to form an elevation gradient distribution feature matrix; for the registered optical image data, the local standard deviation of the pixel values ​​within each grid cell is calculated. The local standard deviation is calculated as follows:

[0043] ;

[0044] in, For the first Local standard deviation of pixel values ​​within the location window For The central window area, The total number of pixels within the window. For the first in the window The pixel value of the location, The mean of all pixel values ​​within the window is used to arrange the local standard deviations according to the grid position, forming a surface roughness feature matrix;

[0045] For the registered interferometric data, the interferometric phase value of each grid cell is extracted, the phase change rate is calculated along multiple directions, and the phase change rates in each direction are arranged according to the grid position to form a terrain structure semantic feature matrix;

[0046] The elevation gradient distribution feature matrix, surface roughness feature matrix, and terrain structure semantic feature matrix are concatenated along the channel dimension to obtain a multimodal feature tensor.

[0047] Preferably, the generative reconstruction of DEM elevation grid data, guided by multimodal feature tensors and using a diffusion model, incorporates geometric boundary constraints of the rail transit corridor during the reconstruction process to output initial DEM elevation grid data. Specifically, this includes:

[0048] The multimodal feature tensor is mapped to a conditional vector corresponding to the time step of the diffusion model through a conditional embedding layer;

[0049] Using the pure noise tensor as the initial state, in the reverse denoising process of the diffusion model, at each time step, the condition vector is concatenated with the current noise tensor and then input into the denoising network to predict the noise residual at the current time step.

[0050] A geometric boundary mask for the rail transit corridor is applied after the output layer of the denoising network. This mask is generated by expanding a preset width to both sides according to the three-dimensional coordinate sequence of the design centerline. The denoising results within the mask are retained, while the denoising results outside the mask are set to zero. The mask processing method is as follows:

[0051] ;

[0052] in, For the first time after masking Elevation values ​​of grid cells The output of the denoising network is the first Elevation values ​​of grid cells For the geometric boundary mask in the first Binary identifier of location, This indicates that the location is within the corridor of the rail transit line. This indicates that the location is outside the designated area;

[0053] The denoising results after masking are accumulated step by step until the diffusion process ends, and the initial DEM elevation grid data is output.

[0054] Preferably, the introduction of geometric boundary constraints of the rail transit line corridor during the reconstruction process specifically includes:

[0055] The three-dimensional coordinate sequence of the design centerline of the rail transit line is read, and linear interpolation is performed between adjacent three-dimensional coordinate points with an interpolation interval of 1 meter to generate a dense set of centerline points. The linear interpolation calculation method is as follows:

[0056] ;

[0057] in, The centerline point obtained by interpolation, and For two adjacent original 3D coordinate points, The interpolation parameters are taken uniformly at 1-meter intervals;

[0058] Using each point on the centerline point set as a reference, extend half the preset width to both sides along the normal direction of the centerline at that point to obtain the boundary points on the left and right sides.

[0059] Connect all centerline points and their corresponding left and right boundary points in sequence to form a closed polygonal boundary of the rail transit corridor.

[0060] The polygon boundary is rasterized into a binary mask with the same resolution as the DEM elevation grid data. The area inside the mask is assigned a value of 1, and the area outside the mask is assigned a value of 0.

[0061] In the first half of the time step of the diffusion model's reverse denoising process, the boundary region of the binary mask is extended outward by two mesh cells to form an extended mask. A linearly decreasing weight decay process is then applied to the denoising result outside the extended mask region. The weight decay is calculated as follows:

[0062] ;

[0063] in, For the first Location decay weight, This represents the shortest grid distance from this location to the mask boundary. To maximize the expansion distance, two grid cells are used here;

[0064] In the latter half of the time step of the diffusion model's reverse denoising process, the mask is restored to the original binary mask, and the denoising result outside the mask range within the original mask range is directly set to zero.

[0065] Preferably, the step of performing multi-scale spatial consistency checks and local outlier identification on the initial DEM elevation grid data, calculating elevation deviation statistics in conjunction with existing DEM reference data, and generating data quality assessment labels specifically includes:

[0066] Sliding window mean filtering was performed on the initial DEM elevation grid data using three window sizes: 3×3, 5×5, and 7×7, respectively, to obtain local mean surfaces at three scales. The local mean was calculated as follows:

[0067] ;

[0068] in, Window size Next Local mean of location For Centered on, with dimensions of Sliding window, This represents the total number of grid cells within the window. For the first Initial DEM elevation values ​​for the location;

[0069] The elevation differences between the initial DEM elevation grid data and the local mean surfaces at three scales are calculated respectively to obtain the residual surfaces at the three scales;

[0070] For the residual surfaces at three scales, corresponding elevation difference thresholds are set respectively, and grid cells with elevation differences exceeding the threshold at any scale are marked as suspicious anomalous cells;

[0071] For each suspected anomaly cell, the elevation values ​​of all grid cells within its eight-neighbor range are statistically analyzed. The deviation between the elevation value of the suspected cell and the average elevation of its eight-neighbor range is calculated. The deviation is calculated as follows:

[0072] ;

[0073] in, For the first Elevation deviation values ​​of suspiciously anomalous units. The elevation mean of all grid cells within the eight-neighborhood of this cell is used, and cells whose deviation exceeds a preset multiple of the neighborhood standard deviation are identified as local outlier cells.

[0074] The location coordinates and elevation values ​​of all local outlier cells are recorded as an outlier area dataset. Existing DEM reference data is cropped to the same grid range and spatial resolution as the initial DEM elevation grid data to obtain a reference DEM subset. The initial DEM elevation grid data is then subtracted from the reference DEM subset grid by grid to obtain the elevation deviation grid. The elevation deviation is calculated as follows:

[0075] ;

[0076] in, For the first The elevation deviation of the location, The initial DEM elevation value at this location. The elevation value of the DEM at this location is used as a reference.

[0077] Calculate the mean and standard deviation of all deviation values ​​in the elevation deviation grid, and separately calculate the mean of positive deviations and the mean of negative deviations. The method for calculating the mean of positive deviations is as follows:

[0078] ;

[0079] in, The mean of the positive deviation. This is the set of all grid cells in the elevation deviation grid that have a deviation value greater than zero. This represents the number of grid cells in the set.

[0080] The mean of negative deviation is calculated as follows:

[0081] ;

[0082] in, It is the absolute value of the negative deviation mean. This is the set of all grid cells in the elevation deviation grid where the deviation value is less than zero. This represents the number of grid cells in the set.

[0083] The data quality assessment label is formed by using the absolute value of the mean, standard deviation, mean of positive deviation, and mean of negative deviation as four statistical dimensions. Each of the four statistical dimensions is compared with four preset thresholds. The comparison result of each statistical dimension generates a binary label. The four binary labels are combined to form a data quality assessment label.

[0084] Preferably, the step of performing neighborhood-weighted iterative correction processing on abnormal areas in the initial DEM elevation grid data based on data quality assessment labels until the elevation deviation statistics converge to a preset threshold range, and outputting the final rail transit DEM elevation data resource, specifically includes:

[0085] Extract the spatial coordinates of each local outlier cell in the outlier region dataset, and extract the elevation values ​​of all valid grid cells within a circular neighborhood with a radius of 5 grid cells centered on that cell.

[0086] Calculate the Euclidean distance from each effective grid cell in the neighborhood to the central anomaly cell. The Euclidean distance is calculated as follows:

[0087] ;

[0088] in, For the neighboring region Location of valid grid cell to center anomaly cell The Euclidean distance;

[0089] Using the reciprocal of the distance as the weight, the elevation values ​​of all valid grid cells in the neighborhood are weighted and summed, then divided by the sum of the weights to obtain the corrected elevation value of the anomalous cell in the first iteration. The correction calculation method is as follows:

[0090] ;

[0091] in, The corrected elevation value for the first iteration of the central anomaly unit. It is the set of all valid grid cells within a circular neighborhood with a radius of 5. For the neighboring region The original elevation values ​​of the effective grid cells at the location;

[0092] Replace the original abnormal elevation values ​​with the corrected elevation values ​​from the first iteration, and update the corresponding elevation values ​​in the initial DEM elevation grid data.

[0093] The multi-scale spatial consistency check process is re-executed with the updated DEM data to re-identify local outlier units. The neighborhood weighted correction process described above is repeated for the newly identified outlier units.

[0094] The process of correction and re-verification is repeated until the number of newly identified local outlier units in two consecutive iterations is less than the preset number, and the elevation deviation statistics corresponding to all retained outlier units fall within the preset threshold range, and the final rail transit DEM elevation data resource is output.

[0095] Preferably, the acquisition of multi-source remote sensing data within a preset width range along the rail transit line, and the simultaneous acquisition of the three-dimensional coordinate sequence of the design centerline of the rail transit line and existing DEM reference data, specifically includes:

[0096] Using each coordinate point on the three-dimensional coordinate sequence of the design centerline of the rail transit line as the center, a buffer zone is generated with a radius of half the preset width. The buffer zone radius is calculated as follows:

[0097] ;

[0098] in, Where is the radius of the buffer zone. The preset width for acquiring data along the rail transit line;

[0099] All buffers are merged along the centerline to obtain the data acquisition range polygon along the rail transit line;

[0100] Spatial intersection calculation is performed between the polygonal area and the coverage area of ​​the optical satellite image, and the optical satellite image data within the intersection area is extracted as the optical satellite image data source;

[0101] Spatial intersection calculation is performed between the polygon range and the coverage area of ​​the airborne lidar point cloud, and the airborne lidar point cloud data within the intersection area is extracted as the point cloud data source;

[0102] Spatial intersection calculation is performed between the polygonal range and the coverage area of ​​the spaceborne synthetic aperture radar interferometry data, and the interferometric data within the intersection area is extracted as the interferometric data source;

[0103] Simultaneously, the existing DEM reference data within the polygon area is read, and the existing DEM reference data is cropped to a rectangular grid area consistent with the polygon area, which is used as the benchmark data for subsequent elevation deviation calculation.

[0104] In a second aspect of the invention, a smart rail transit DEM elevation data resource acquisition system under the AIGC framework is also provided, comprising:

[0105] Acquisition Module: Acquires multi-source remote sensing data within a preset width range along the rail transit line. The multi-source remote sensing data includes optical satellite imagery data, airborne lidar point cloud data, and spaceborne synthetic aperture radar interferometric data. At the same time, it acquires the three-dimensional coordinate sequence of the design centerline of the rail transit line and existing DEM reference data.

[0106] Data processing module: Performs radiometric correction, geometric fine correction and point cloud denoising on multi-source remote sensing data in sequence, and converts the processed data from each source to the same geographic coordinate system and the same spatial resolution to form a registered multi-source dataset;

[0107] The first generation module inputs the registered multi-source dataset into the pre-trained AIGC multimodal feature extraction model, and fuses the terrain semantic information from different data sources through a cross-modal attention mechanism to generate a multimodal feature tensor containing elevation gradient distribution, surface roughness features and terrain structure semantics.

[0108] First output module: Guided by multimodal feature tensors, it performs generative reconstruction of DEM elevation grid data using a diffusion model. During the reconstruction process, it introduces geometric boundary constraints of rail transit corridors and outputs initial DEM elevation grid data.

[0109] The second generation module performs multi-scale spatial consistency checks and local outlier identification on the initial DEM elevation grid data, calculates elevation deviation statistics by combining existing DEM reference data, and generates data quality assessment labels.

[0110] The second output module performs neighborhood-weighted iterative correction processing on abnormal areas in the initial DEM elevation grid data based on data quality assessment labels until the elevation deviation statistics converge to a preset threshold range, and outputs the final rail transit DEM elevation data resources.

[0111] Compared with existing technologies, this invention provides a method for capturing intelligent rail transit DEM elevation data resources, which has the following beneficial effects:

[0112] This invention effectively overcomes the limitations of single-source data in obtaining elevation data in areas lacking texture, vegetation cover, and data occlusion by fusing three complementary data sources: optical satellite imagery, airborne lidar point clouds, and spaceborne InSAR interferometric data. This significantly improves the coverage integrity and elevation accuracy of DEM data along rail transit lines. Simultaneously, by utilizing the cross-modal attention mechanism in the AIGC multimodal feature extraction model, it deeply mines the semantic relationships between different data sources, generating a multimodal feature tensor containing elevation gradient, surface roughness, and terrain structure, making the generated results more accurate in representing terrain structure. Furthermore, by introducing geometric boundary constraints of the rail transit corridor during the diffusion model generative reconstruction process, it ensures that the generated DEM elevation data is strictly limited to the corridor, avoiding boundary anomalies caused by unconstrained generation. Moreover, this invention combines multi-scale spatial consistency testing with local outlier identification to finely distinguish between abrupt and gradual anomalies, and gradually converges to a preset accuracy through a neighborhood-weighted iterative correction mechanism. This eliminates anomalies while preserving true terrain details to the greatest extent, ultimately outputting high-quality DEM elevation data resources that meet the requirements of rail transit engineering applications. Attached Figure Description

[0113] Figure 1 This is a schematic diagram of the method flow of S101-S106 in this invention;

[0114] Figure 2 This is a schematic diagram of the method flow for S201-S206 in this invention;

[0115] Figure 3 This is a schematic diagram of the method flow for S301-S304 in this invention;

[0116] Figure 4 This is a schematic diagram of the system module framework of the present invention. Detailed Implementation

[0117] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0118] Example 1

[0119] Please refer to Figure 1 As shown, in a first aspect of the present invention, a method for capturing intelligent rail transit DEM elevation data resources is provided, comprising:

[0120] S101. Acquire multi-source remote sensing data within a preset width range along the rail transit line. The multi-source remote sensing data includes optical satellite imagery data, airborne lidar point cloud data, and spaceborne synthetic aperture radar interferometric data. At the same time, acquire the three-dimensional coordinate sequence of the design centerline of the rail transit line and existing DEM reference data.

[0121] S102. Perform radiometric correction, geometric fine correction and point cloud denoising on the multi-source remote sensing data in sequence, and convert the processed source data to the same geographic coordinate system and the same spatial resolution to form a registered multi-source dataset.

[0122] S103. Input the registered multi-source dataset into the pre-trained AIGC multimodal feature extraction model, and fuse the terrain semantic information from different data sources through a cross-modal attention mechanism to generate a multimodal feature tensor containing elevation gradient distribution, surface roughness features and terrain structure semantics.

[0123] S104. Guided by multimodal feature tensors, generative reconstruction of DEM elevation grid data is performed using a diffusion model. Geometric boundary constraints of rail transit corridors are introduced during the reconstruction process, and initial DEM elevation grid data is output.

[0124] S105. Perform multi-scale spatial consistency checks and local outlier identification on the initial DEM elevation grid data, calculate elevation deviation statistics in combination with existing DEM reference data, and generate data quality assessment labels.

[0125] S106. Perform neighborhood-weighted iterative correction processing on abnormal areas in the initial DEM elevation grid data according to the data quality assessment label until the elevation deviation statistics converge to the preset threshold range, and output the final rail transit DEM elevation data resources.

[0126] Those skilled in the art will understand that this invention, by fusing three complementary data sources—optical satellite imagery, airborne lidar point clouds, and spaceborne InSAR interferometric data—effectively overcomes the limitations of single data sources in acquiring elevation data in areas lacking texture, vegetation cover, and data occlusion, significantly improving the coverage integrity and elevation accuracy of DEM data along rail transit lines. Simultaneously, by utilizing the cross-modal attention mechanism in the AIGC multimodal feature extraction model, it deeply mines the semantic relationships between different data sources, generating a multimodal feature tensor containing elevation gradient, surface roughness, and terrain structure, thus enhancing the terrain structure representation of the generated results. More accurate; in addition, the geometric boundary constraints of the rail transit line corridor are introduced into the generative reconstruction process of the diffusion model to ensure that the generated DEM elevation data is strictly limited to the line corridor, avoiding the boundary anomaly problem caused by unconstrained generation; furthermore, the present invention can finely distinguish between abrupt anomalies and gradual anomalies by combining multi-scale spatial consistency test and local anomaly identification, and gradually converges to the preset accuracy through a neighborhood weighted iterative correction mechanism, eliminating anomalies while preserving the real terrain details to the greatest extent, and finally outputting high-quality DEM elevation data resources that meet the requirements of rail transit engineering applications.

[0127] Please refer to Figure 2 As shown, radiometric correction, geometric fine correction, and point cloud denoising are sequentially performed on the multi-source remote sensing data. The processed data from each source are then uniformly transformed to the same geographic coordinate system and spatial resolution to form a registered multi-source dataset, specifically including:

[0128] S201. Perform radiometric correction on the optical satellite image data based on the sensor radiometric response function, converting the original digital quantization values ​​into surface reflectance values. The conversion relationship is as follows:

[0129] ;

[0130] in, For band The surface reflectance value, The original numerical quantization value, This represents the radiation shift in that band. This represents the solar irradiance at the top of the atmosphere in this band. The zenith angle of the sun. This is the Earth-Sun distance correction factor;

[0131] S202. Calculate the radiation drift using the overlapping area of ​​adjacent temporal images and compensate for the drift in reflectance values.

[0132] S203. Perform point-by-point intensity normalization on the airborne lidar point cloud data to eliminate intensity differences between different flight sorties. The normalization calculation method is as follows:

[0133] ;

[0134] in, The normalized intensity value. This is the original strength value. This represents the average intensity at all points during that flight sortie. The standard deviation of the intensity of all points for this flight sortie;

[0135] S204. Use the statistical outlier removal algorithm to identify and remove isolated noise points in the point cloud whose distance from surrounding point groups exceeds a preset multiple of the local standard deviation. The judgment criteria are:

[0136] ;

[0137] in, For the first a point to its The average distance between nearest neighbors, For all points The mean of the average distance to nearest neighbors. For all points Standard deviation of the nearest neighbor mean distance A preset multiple threshold is used to identify points that meet the condition as noise and remove them.

[0138] S205. Perform flat-ground phase removal and terrain phase unwrapping processing on the spaceborne synthetic aperture radar interferometric data based on the reference DEM;

[0139] S206. The radiometrically corrected optical image, the denoised lidar point cloud, and the unwrapped interferometric data are geometrically refined based on ground control points. The geographic coordinates of each source data are uniformly transformed to the CGCS2000 coordinate system, and the spatial resolution of each source data is uniformly adjusted to a 2-meter grid spacing using a bilinear resampling method. The bilinear interpolation calculation method is as follows:

[0140] ;

[0141] in, For target grid points Resampled value at that point, , , , For the four known grid points surrounding the target point, The original values ​​corresponding to the known grid points are used to form a registered multi-source dataset.

[0142] Please refer to Figure 3 As shown, the registered multi-source dataset is input into the pre-trained AIGC multimodal feature extraction model, and terrain semantic information from different data sources is fused through a cross-modal attention mechanism, specifically including:

[0143] S301. Input the registered optical image data, lidar point cloud data and interferometric data into the corresponding single-mode encoder to extract the local terrain feature vectors of each source data.

[0144] S302. Construct a cross-modal attention query matrix, using the optical image feature vector as the query vector and the lidar point cloud feature vector as the key vector, and calculate the attention weight matrix between the optical image and the point cloud. The attention weight is calculated as follows:

[0145] ;

[0146] in, The optical image feature vector matrix, The feature vector matrix of the lidar point cloud. The feature vector matrix of the lidar point cloud. Given the dimension of the key vector, the weight matrix is ​​weighted and summed with the point cloud feature vector to obtain the optical-point cloud cross-modal fusion feature;

[0147] S303. Then, using the interferometric data feature vector as the query vector and the optical-point cloud fusion feature as the key vector, calculate the attention weight matrix between the interferometric data and the fusion feature, perform a second weighted summation, and obtain the three-modal fusion feature vector.

[0148] S304. Perform max pooling and average pooling operations on the three-modal fusion feature vector along the spatial dimension, and after concatenating the pooling results, map them through a fully connected layer to generate a multimodal feature tensor containing elevation gradient distribution features, surface roughness features, and terrain structure semantic features.

[0149] Generate a multimodal feature tensor containing elevation gradient distribution, surface roughness features, and terrain structure semantics, specifically including:

[0150] The elevation values ​​of all points within each grid cell are calculated for the registered lidar point cloud data. The elevation differences between adjacent grid cells are calculated along the east-west and north-south directions, respectively, to obtain the east-west and north-south elevation gradient components. The east-west elevation gradient is calculated as follows:

[0151] ;

[0152] in, For the first The east-west elevation gradient components of the location This represents the elevation value of the adjacent grid cell on the east side. This represents the elevation value of the current grid cell;

[0153] The method for calculating the north-south elevation gradient is as follows:

[0154] ;

[0155] in, For the first The north-south elevation gradient components of the location, The elevation value of the adjacent grid cell on the north side;

[0156] The two components are stitched together according to their grid positions to form an elevation gradient distribution feature matrix; for the registered optical image data, the local standard deviation of the pixel values ​​within each grid cell is calculated. The local standard deviation is calculated as follows:

[0157] ;

[0158] in, For the first Local standard deviation of pixel values ​​within the location window For The central window area, The total number of pixels within the window. For the first in the window The pixel value of the location, The mean of all pixel values ​​within the window is used to arrange the local standard deviations according to the grid position, forming a surface roughness feature matrix;

[0159] For the registered interferometric data, the interferometric phase value of each grid cell is extracted, the phase change rate is calculated along multiple directions, and the phase change rates in each direction are arranged according to the grid position to form a terrain structure semantic feature matrix;

[0160] The elevation gradient distribution feature matrix, surface roughness feature matrix, and terrain structure semantic feature matrix are concatenated along the channel dimension to obtain a multimodal feature tensor.

[0161] Guided by multimodal feature tensors, generative reconstruction of DEM elevation grid data is performed using a diffusion model. Geometric boundary constraints of rail transit corridors are introduced during the reconstruction process, and initial DEM elevation grid data is output, specifically including:

[0162] The multimodal feature tensor is mapped to a conditional vector corresponding to the time step of the diffusion model through a conditional embedding layer;

[0163] Using the pure noise tensor as the initial state, in the reverse denoising process of the diffusion model, at each time step, the condition vector is concatenated with the current noise tensor and then input into the denoising network to predict the noise residual at the current time step.

[0164] A geometric boundary mask for the rail transit corridor is applied after the output layer of the denoising network. This mask is generated by expanding a preset width to both sides according to the three-dimensional coordinate sequence of the design centerline. The denoising results within the mask are retained, while the denoising results outside the mask are set to zero. The mask processing method is as follows:

[0165] ;

[0166] in, For the first time after masking Elevation values ​​of grid cells The output of the denoising network is the first Elevation values ​​of grid cells For the geometric boundary mask in the first Binary identifier of location, This indicates that the location is within the corridor of the rail transit line. This indicates that the location is outside the designated area;

[0167] The denoising results after masking are accumulated step by step until the diffusion process ends, and the initial DEM elevation grid data is output.

[0168] The reconstruction process incorporates geometric boundary constraints for the rail transit corridor, specifically including:

[0169] The three-dimensional coordinate sequence of the design centerline of the rail transit line is read, and linear interpolation is performed between adjacent three-dimensional coordinate points with an interpolation interval of 1 meter to generate a dense set of centerline points. The linear interpolation calculation method is as follows:

[0170] ;

[0171] in, The centerline point obtained by interpolation, and For two adjacent original 3D coordinate points, The interpolation parameters are taken uniformly at 1-meter intervals;

[0172] Using each point on the centerline point set as a reference, extend half the preset width to both sides along the normal direction of the centerline at that point to obtain the boundary points on the left and right sides.

[0173] Connect all centerline points and their corresponding left and right boundary points in sequence to form a closed polygonal boundary of the rail transit corridor.

[0174] The polygon boundary is rasterized into a binary mask with the same resolution as the DEM elevation grid data. The area inside the mask is assigned a value of 1, and the area outside the mask is assigned a value of 0.

[0175] In the first half of the time step of the diffusion model's reverse denoising process, the boundary region of the binary mask is extended outward by two mesh cells to form an extended mask. A linearly decreasing weight decay process is then applied to the denoising result outside the extended mask region. The weight decay is calculated as follows:

[0176] ;

[0177] in, For the first Location decay weight, This represents the shortest grid distance from this location to the mask boundary. To maximize the expansion distance, two grid cells are used here;

[0178] In the latter half of the time step of the diffusion model's reverse denoising process, the mask is restored to the original binary mask, and the denoising result outside the mask range within the original mask range is directly set to zero.

[0179] Multi-scale spatial consistency checks and local outlier identification are performed on the initial DEM elevation grid data. Elevation deviation statistics are calculated using existing DEM reference data to generate data quality assessment labels, specifically including:

[0180] Sliding window mean filtering was performed on the initial DEM elevation grid data using three window sizes: 3×3, 5×5, and 7×7, respectively, to obtain local mean surfaces at three scales. The local mean was calculated as follows:

[0181] ;

[0182] in, Window size Next Local mean of location For Centered on, with dimensions of Sliding window, This represents the total number of grid cells within the window. For the first Initial DEM elevation values ​​for the location;

[0183] The elevation differences between the initial DEM elevation grid data and the local mean surfaces at three scales are calculated respectively to obtain the residual surfaces at the three scales;

[0184] For the residual surfaces at three scales, corresponding elevation difference thresholds are set respectively, and grid cells with elevation differences exceeding the threshold at any scale are marked as suspicious anomalous cells;

[0185] For each suspected anomaly cell, the elevation values ​​of all grid cells within its eight-neighbor range are statistically analyzed. The deviation between the elevation value of the suspected cell and the average elevation of its eight-neighbor range is calculated. The deviation is calculated as follows:

[0186] ;

[0187] in, For the first Elevation deviation values ​​of suspiciously anomalous units. The elevation mean of all grid cells within the eight-neighborhood of this cell is used, and cells whose deviation exceeds a preset multiple of the neighborhood standard deviation are identified as local outlier cells.

[0188] The location coordinates and elevation values ​​of all local outlier cells are recorded as an outlier area dataset. Existing DEM reference data is cropped to the same grid range and spatial resolution as the initial DEM elevation grid data to obtain a reference DEM subset. The initial DEM elevation grid data is then subtracted from the reference DEM subset grid by grid to obtain the elevation deviation grid. The elevation deviation is calculated as follows:

[0189] ;

[0190] in, For the first The elevation deviation of the location, The initial DEM elevation value at this location. The elevation value of the DEM at this location is used as a reference.

[0191] Calculate the mean and standard deviation of all deviation values ​​in the elevation deviation grid, and separately calculate the mean of positive deviations and the mean of negative deviations. The method for calculating the mean of positive deviations is as follows:

[0192] ;

[0193] in, The mean of the positive deviation. This is the set of all grid cells in the elevation deviation grid that have a deviation value greater than zero. This represents the number of grid cells in the set.

[0194] The mean of negative deviation is calculated as follows:

[0195] ;

[0196] in, It is the absolute value of the negative deviation mean. This is the set of all grid cells in the elevation deviation grid where the deviation value is less than zero. This represents the number of grid cells in the set.

[0197] The data quality assessment label is formed by using the absolute value of the mean, standard deviation, mean of positive deviation, and mean of negative deviation as four statistical dimensions. Each of the four statistical dimensions is compared with four preset thresholds. The comparison result of each statistical dimension generates a binary label. The four binary labels are combined to form a data quality assessment label.

[0198] Based on the data quality assessment labels, an iterative correction process based on neighborhood weighting is performed on the abnormal areas in the initial DEM elevation grid data until the elevation deviation statistics converge to a preset threshold range. The final rail transit DEM elevation data resource is then output, specifically including:

[0199] Extract the spatial coordinates of each local outlier cell in the outlier region dataset, and extract the elevation values ​​of all valid grid cells within a circular neighborhood with a radius of 5 grid cells centered on that cell.

[0200] Calculate the Euclidean distance from each effective grid cell in the neighborhood to the central anomaly cell. The Euclidean distance is calculated as follows:

[0201] ;

[0202] in, For the neighboring region Location of valid grid cell to center anomaly cell The Euclidean distance;

[0203] Using the reciprocal of the distance as the weight, the elevation values ​​of all valid grid cells in the neighborhood are weighted and summed, then divided by the sum of the weights to obtain the corrected elevation value of the anomalous cell in the first iteration. The correction calculation method is as follows:

[0204] ;

[0205] in, The corrected elevation value for the first iteration of the central anomaly unit. It is the set of all valid grid cells within a circular neighborhood with a radius of 5. For the neighboring region The original elevation value of the effective grid cell at the location;

[0206] Replace the original abnormal elevation values ​​with the corrected elevation values ​​from the first iteration, and update the corresponding elevation values ​​in the initial DEM elevation grid data.

[0207] The multi-scale spatial consistency check process is re-executed with the updated DEM data to re-identify local outlier units. The neighborhood weighted correction process described above is repeated for the newly identified outlier units.

[0208] The process of correction and re-verification is repeated until the number of newly identified local outlier units in two consecutive iterations is less than the preset number, and the elevation deviation statistics corresponding to all retained outlier units fall within the preset threshold range, and the final rail transit DEM elevation data resource is output.

[0209] Acquire multi-source remote sensing data within a predetermined width range along the rail transit line, and simultaneously acquire the three-dimensional coordinate sequence of the rail transit line's design centerline and existing DEM reference data, specifically including:

[0210] Using each coordinate point on the three-dimensional coordinate sequence of the design centerline of the rail transit line as the center, a buffer zone is generated with a radius of half the preset width. The buffer zone radius is calculated as follows:

[0211] ;

[0212] in, Where is the radius of the buffer zone. The preset width for acquiring data along the rail transit line;

[0213] All buffers are merged along the centerline to obtain the data acquisition range polygon along the rail transit line;

[0214] Spatial intersection calculation is performed between the polygonal area and the coverage area of ​​the optical satellite image, and the optical satellite image data within the intersection area is extracted as the optical satellite image data source;

[0215] Spatial intersection calculation is performed between the polygonal range and the coverage area of ​​the airborne lidar point cloud, and the airborne lidar point cloud data within the intersection area is extracted as the point cloud data source;

[0216] Spatial intersection calculation is performed between the polygonal range and the coverage area of ​​the spaceborne synthetic aperture radar interferometry data, and the interferometric data within the intersection area is extracted as the interferometric data source;

[0217] Simultaneously, the existing DEM reference data within the polygon area is read, and the existing DEM reference data is cropped to a rectangular grid area consistent with the polygon area, which is used as the benchmark data for subsequent elevation deviation calculation.

[0218] Please refer to Figure 4 As shown, in a second aspect of the present invention, a smart rail transit DEM elevation data resource acquisition system under the AIGC framework is also provided, comprising:

[0219] Acquisition Module: Acquires multi-source remote sensing data within a preset width range along the rail transit line. The multi-source remote sensing data includes optical satellite imagery data, airborne lidar point cloud data, and spaceborne synthetic aperture radar interferometric data. At the same time, it acquires the three-dimensional coordinate sequence of the design centerline of the rail transit line and existing DEM reference data.

[0220] Data processing module: Performs radiometric correction, geometric fine correction and point cloud denoising on multi-source remote sensing data in sequence, and converts the processed data from each source to the same geographic coordinate system and the same spatial resolution to form a registered multi-source dataset;

[0221] The first generation module inputs the registered multi-source dataset into the pre-trained AIGC multimodal feature extraction model, and fuses the terrain semantic information from different data sources through a cross-modal attention mechanism to generate a multimodal feature tensor containing elevation gradient distribution, surface roughness features and terrain structure semantics.

[0222] First output module: Guided by multimodal feature tensors, it performs generative reconstruction of DEM elevation grid data using a diffusion model. During the reconstruction process, it introduces geometric boundary constraints of rail transit corridors and outputs initial DEM elevation grid data.

[0223] The second generation module performs multi-scale spatial consistency checks and local outlier identification on the initial DEM elevation grid data, calculates elevation deviation statistics by combining existing DEM reference data, and generates data quality assessment labels.

[0224] The second output module performs neighborhood-weighted iterative correction processing on abnormal areas in the initial DEM elevation grid data based on data quality assessment labels until the elevation deviation statistics converge to a preset threshold range, and outputs the final rail transit DEM elevation data resources.

[0225] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for capturing DEM elevation data resources of intelligent rail transit, characterized in that, include: Acquire multi-source remote sensing data within a preset width range along the rail transit line. The multi-source remote sensing data includes optical satellite imagery data, airborne lidar point cloud data, and spaceborne synthetic aperture radar interferometric data. At the same time, acquire the three-dimensional coordinate sequence of the design centerline of the rail transit line and existing DEM reference data. Radiometric correction, geometric fine correction and point cloud denoising are performed sequentially on the multi-source remote sensing data, and the processed data from each source are uniformly converted to the same geographic coordinate system and the same spatial resolution to form a registered multi-source dataset. The registered multi-source dataset is input into the pre-trained AIGC multimodal feature extraction model. The terrain semantic information from different data sources is fused through a cross-modal attention mechanism to generate a multimodal feature tensor containing elevation gradient distribution, surface roughness features and terrain structure semantics. Guided by multimodal feature tensors, generative reconstruction of DEM elevation grid data is performed using a diffusion model. Geometric boundary constraints of rail transit corridors are introduced during the reconstruction process, and initial DEM elevation grid data is output. Multi-scale spatial consistency checks and local outlier identification are performed on the initial DEM elevation grid data. Elevation deviation statistics are calculated in combination with existing DEM reference data to generate data quality assessment labels. Based on the data quality assessment labels, an iterative correction process based on neighborhood weighting is performed on the abnormal areas in the initial DEM elevation grid data until the elevation deviation statistics converge to the preset threshold range, and the final rail transit DEM elevation data resources are output.

2. The method according to claim 1, wherein, The process involves sequentially performing radiometric correction, geometric fine correction, and point cloud denoising on the multi-source remote sensing data, and then uniformly transforming the processed data from each source to the same geographic coordinate system and spatial resolution to form a registered multi-source dataset. Specifically, this includes: Radiometric correction based on the sensor's radiometric response function is performed on the optical satellite imagery data, converting the raw digital quantization values ​​into surface reflectance values. The conversion relationship is as follows: ； wherein, is a surface reflectance value for the waveband , is an original digital quantization value, is a radiometric offset value for the waveband, is a top-of-atmosphere solar irradiance value for the waveband, is a solar zenith angle, is a correction factor for the Earth-Sun distance; The radiation drift is calculated using the overlapping area of ​​adjacent temporal images, and the reflectivity value is compensated for the drift. The airborne lidar point cloud data undergoes point-by-point intensity normalization to eliminate intensity differences between different flight sorties. The normalization calculation method is as follows: ； wherein, is the normalized intensity value, is the original intensity value, is the mean value of all point intensities of the flight pass, is the standard deviation of all point intensities of the flight pass; The statistical outlier removal algorithm is used to identify and remove isolated noise points in a point cloud whose distance from surrounding point groups exceeds a preset multiple of the local standard deviation. The judgment criteria are as follows: ； in, For the first a point to its The average distance between nearest neighbors, For all points The mean of the average distance to nearest neighbors. For all points Standard deviation of the nearest neighbor mean distance A preset multiple threshold is used to identify points that meet the condition as noise and remove them. Perform flat-ground phase removal and terrain phase unwrapping processing on spaceborne synthetic aperture radar interferometric data based on reference DEM; The radiometrically corrected optical image, the denoised lidar point cloud, and the unwrapped interferometric data were geometrically refined based on ground control points. The geographic coordinates of all source data were uniformly transformed to the CGCS2000 coordinate system, and the spatial resolution of all source data was uniformly adjusted to a 2-meter grid spacing using a bilinear resampling method. The bilinear interpolation calculation method is as follows: ； in, For target grid points Resampled value at that point, , , , For the four known grid points surrounding the target point, The original values ​​corresponding to the known grid points are used to form a registered multi-source dataset.

3. The method for capturing intelligent rail transit DEM elevation data resources according to claim 2, characterized in that, The process of inputting the registered multi-source dataset into the pre-trained AIGC multimodal feature extraction model and fusing terrain semantic information from different data sources through a cross-modal attention mechanism specifically includes: The registered optical image data, lidar point cloud data, and interferometric data are input into the corresponding single-mode encoders to extract the local terrain feature vectors of each source data. A cross-modal attention query matrix is ​​constructed, using the optical image feature vector as the query vector and the lidar point cloud feature vector as the key vector. The attention weight matrix between the optical image and the point cloud is calculated as follows: ； in, The optical image feature vector matrix, The feature vector matrix of the lidar point cloud. The feature vector matrix of the lidar point cloud. Given the dimension of the key vector, the weight matrix is ​​weighted and summed with the point cloud feature vector to obtain the optical-point cloud cross-modal fusion feature; Then, using the interferometric data feature vector as the query vector and the optical-point cloud fusion feature as the key vector, the attention weight matrix between the interferometric data and the fusion feature is calculated, and a second weighted sum is performed to obtain the three-modal fusion feature vector. Max pooling and average pooling operations are performed on the three-modal fusion feature vector along the spatial dimension. After concatenating the pooling results, a multimodal feature tensor containing elevation gradient distribution features, surface roughness features, and terrain structure semantic features is generated through a fully connected layer mapping.

4. The method for capturing intelligent rail transit DEM elevation data resources according to claim 3, characterized in that, The generation of a multimodal feature tensor, which includes elevation gradient distribution, surface roughness features, and terrain structure semantics, specifically includes: The elevation values ​​of all points within each grid cell are calculated for the registered lidar point cloud data. The elevation differences between adjacent grid cells are calculated along the east-west and north-south directions, respectively, to obtain the east-west and north-south elevation gradient components. The east-west elevation gradient is calculated as follows: ； in, For the first The east-west elevation gradient components of the location This represents the elevation value of the adjacent grid cell on the east side. This represents the elevation value of the current grid cell; The method for calculating the north-south elevation gradient is as follows: ； in, For the first The north-south elevation gradient components of the location, The elevation value of the adjacent grid cell on the north side; The two components are stitched together according to their grid positions to form an elevation gradient distribution feature matrix; for the registered optical image data, the local standard deviation of the pixel values ​​within each grid cell is calculated. The local standard deviation is calculated as follows: ； in, For the first Local standard deviation of pixel values ​​within the location window For The central window area, The total number of pixels within the window. For the first in the window The pixel value of the location, The mean of all pixel values ​​within the window is used to arrange the local standard deviations according to the grid position, forming a surface roughness feature matrix; For the registered interferometric data, the interferometric phase value of each grid cell is extracted, the phase change rate is calculated along multiple directions, and the phase change rates in each direction are arranged according to the grid position to form a terrain structure semantic feature matrix; The elevation gradient distribution feature matrix, surface roughness feature matrix, and terrain structure semantic feature matrix are concatenated along the channel dimension to obtain a multimodal feature tensor.

5. A method for capturing DEM elevation data resources for intelligent rail transit according to claim 4, characterized in that, The generative reconstruction of DEM elevation grid data, guided by multimodal feature tensors and utilizing a diffusion model, incorporates geometric boundary constraints of rail transit corridors during the reconstruction process, outputting initial DEM elevation grid data. Specifically, this includes: The multimodal feature tensor is mapped to a conditional vector corresponding to the time step of the diffusion model through a conditional embedding layer; Using the pure noise tensor as the initial state, in the reverse denoising process of the diffusion model, at each time step, the condition vector is concatenated with the current noise tensor and then input into the denoising network to predict the noise residual at the current time step. A geometric boundary mask for the rail transit corridor is applied after the output layer of the denoising network. This mask is generated by expanding a preset width to both sides according to the three-dimensional coordinate sequence of the design centerline. The denoising results within the mask are retained, while the denoising results outside the mask are set to zero. The mask processing method is as follows: ； in, For the first time after masking Elevation values ​​of grid cells The output of the denoising network is the first Elevation values ​​of grid cells For the geometric boundary mask in the first Binary identifier of location, This indicates that the location is within the corridor of the rail transit line. This indicates that the location is outside the designated area; The denoising results after masking are accumulated step by step until the diffusion process ends, and the initial DEM elevation grid data is output.

6. A method for capturing DEM elevation data resources for intelligent rail transit according to claim 5, characterized in that, The introduction of geometric boundary constraints for the rail transit corridor during the reconstruction process specifically includes: The three-dimensional coordinate sequence of the design centerline of the rail transit line is read, and linear interpolation is performed between adjacent three-dimensional coordinate points with an interpolation interval of 1 meter to generate a dense set of centerline points. The linear interpolation calculation method is as follows: ； in, The centerline point obtained by interpolation, and For two adjacent original 3D coordinate points, The interpolation parameters are taken uniformly at 1-meter intervals; Using each point on the centerline point set as a reference, extend half the preset width to both sides along the normal direction of the centerline at that point to obtain the boundary points on the left and right sides. Connect all centerline points and their corresponding left and right boundary points in sequence to form a closed polygonal boundary of the rail transit corridor. The polygon boundary is rasterized into a binary mask with the same resolution as the DEM elevation grid data. The area inside the mask is assigned a value of 1, and the area outside the mask is assigned a value of 0. In the first half of the time step of the diffusion model's reverse denoising process, the boundary region of the binary mask is extended outward by two mesh cells to form an extended mask. A linearly decreasing weight decay process is then applied to the denoising result outside the extended mask region. The weight decay is calculated as follows: ； in, For the first Location decay weight, This represents the shortest grid distance from this location to the mask boundary. To maximize the expansion distance, two grid cells are used here; In the latter half of the time step of the diffusion model's reverse denoising process, the mask is restored to the original binary mask, and the denoising result outside the mask range within the original mask range is directly set to zero.

7. A method for capturing intelligent rail transit DEM elevation data resources according to claim 6, characterized in that, The process of performing multi-scale spatial consistency checks and local outlier identification on the initial DEM elevation grid data, calculating elevation deviation statistics in conjunction with existing DEM reference data, and generating data quality assessment labels specifically includes: Sliding window mean filtering was performed on the initial DEM elevation grid data using three window sizes: 3×3, 5×5, and 7×7, respectively, to obtain local mean surfaces at three scales. The local mean was calculated as follows: ； in, Window size Next Local mean of location For Centered on, with dimensions of Sliding window, This represents the total number of grid cells within the window. For the first Initial DEM elevation values ​​for the location; The elevation differences between the initial DEM elevation grid data and the local mean surfaces at three scales are calculated respectively to obtain the residual surfaces at the three scales; For the residual surfaces at three scales, corresponding elevation difference thresholds are set respectively, and grid cells with elevation differences exceeding the threshold at any scale are marked as suspicious anomalous cells; For each suspected anomaly cell, the elevation values ​​of all grid cells within its eight-neighbor range are statistically analyzed. The deviation between the elevation value of the suspected cell and the average elevation of its eight-neighbor range is calculated. The deviation is calculated as follows: ； in, For the first Elevation deviation values ​​of suspiciously anomalous units. The elevation mean of all grid cells within the eight-neighborhood of this cell is used, and cells whose deviation exceeds a preset multiple of the neighborhood standard deviation are identified as local outlier cells. The location coordinates and elevation values ​​of all local outlier cells are recorded as an outlier area dataset. Existing DEM reference data is cropped to the same grid range and spatial resolution as the initial DEM elevation grid data to obtain a reference DEM subset. The initial DEM elevation grid data is then subtracted from the reference DEM subset grid by grid to obtain the elevation deviation grid. The elevation deviation is calculated as follows: ； in, For the first The elevation deviation of the location, The initial DEM elevation value at this location. The elevation value of the DEM at this location is used as a reference. Calculate the mean and standard deviation of all deviation values ​​in the elevation deviation grid, and separately calculate the mean of positive deviations and the mean of negative deviations. The method for calculating the mean of positive deviations is as follows: ； in, The mean of the positive deviation. This is the set of all grid cells in the elevation deviation grid that have a deviation value greater than zero. This represents the number of grid cells in the set. The mean of negative deviation is calculated as follows: ； in, It is the absolute value of the negative deviation mean. This is the set of all grid cells in the elevation deviation grid where the deviation value is less than zero. This represents the number of grid cells in the set. The data quality assessment label is formed by using the absolute value of the mean, standard deviation, mean of positive deviation, and mean of negative deviation as four statistical dimensions. Each of the four statistical dimensions is compared with four preset thresholds. The comparison result of each statistical dimension generates a binary label. The four binary labels are combined to form a data quality assessment label.

8. A method for capturing DEM elevation data resources for intelligent rail transit according to claim 7, characterized in that, The process involves performing neighborhood-weighted iterative correction on outlier areas in the initial DEM elevation grid data based on data quality assessment labels until the elevation deviation statistics converge to a preset threshold range, ultimately outputting the final rail transit DEM elevation data resource. Specifically, this includes: Extract the spatial coordinates of each local outlier cell in the outlier region dataset, and extract the elevation values ​​of all valid grid cells within a circular neighborhood with a radius of 5 grid cells centered on that cell. Calculate the Euclidean distance from each effective grid cell in the neighborhood to the central anomaly cell. The Euclidean distance is calculated as follows: ； in, For the neighboring region Location of valid grid cell to center anomaly cell The Euclidean distance; Using the reciprocal of the distance as the weight, the elevation values ​​of all valid grid cells in the neighborhood are weighted and summed, then divided by the sum of the weights to obtain the corrected elevation value of the anomalous cell in the first iteration. The correction calculation method is as follows: ； in, The corrected elevation value for the first iteration of the central anomaly unit. It is the set of all valid grid cells within a circular neighborhood with a radius of 5. For the neighboring region The original elevation value of the effective grid cell at the location; Replace the original abnormal elevation values ​​with the corrected elevation values ​​from the first iteration, and update the corresponding elevation values ​​in the initial DEM elevation grid data. The multi-scale spatial consistency check process is re-executed with the updated DEM data to re-identify local outlier units. The neighborhood weighted correction process described above is repeated for the newly identified outlier units. The process of correction and re-verification is repeated until the number of newly identified local outlier units in two consecutive iterations is less than the preset number, and the elevation deviation statistics corresponding to all retained outlier units fall within the preset threshold range, and the final rail transit DEM elevation data resource is output.

9. A method for capturing DEM elevation data resources for intelligent rail transit according to claim 8, characterized in that, The acquisition of multi-source remote sensing data within a predetermined width range along the rail transit line, and the acquisition of the three-dimensional coordinate sequence of the design centerline of the rail transit line and existing DEM reference data, specifically includes: Using each coordinate point on the three-dimensional coordinate sequence of the design centerline of the rail transit line as the center, a buffer zone is generated with a radius of half the preset width. The buffer zone radius is calculated as follows: ； in, Where is the radius of the buffer zone. The preset width for acquiring data along the rail transit line; All buffers are merged along the centerline to obtain the data acquisition range polygon along the rail transit line; Spatial intersection calculation is performed between the polygonal area and the coverage area of ​​the optical satellite image, and the optical satellite image data within the intersection area is extracted as the optical satellite image data source; Spatial intersection calculation is performed between the polygonal range and the coverage area of ​​the airborne lidar point cloud, and the airborne lidar point cloud data within the intersection area is extracted as the point cloud data source; Spatial intersection calculation is performed between the polygonal range and the coverage area of ​​the spaceborne synthetic aperture radar interferometry data, and the interferometric data within the intersection area is extracted as the interferometric data source; Simultaneously, the existing DEM reference data within the polygon area is read, and the existing DEM reference data is cropped to a rectangular grid area consistent with the polygon area, which is used as the benchmark data for subsequent elevation deviation calculation.

10. A smart rail transit DEM elevation data resource acquisition system under the AIGC framework, used to implement a smart rail transit DEM elevation data resource acquisition method as described in any one of claims 1-9, characterized in that, include: Acquisition Module: Acquires multi-source remote sensing data within a preset width range along the rail transit line. The multi-source remote sensing data includes optical satellite imagery data, airborne lidar point cloud data, and spaceborne synthetic aperture radar interferometric data. At the same time, it acquires the three-dimensional coordinate sequence of the design centerline of the rail transit line and existing DEM reference data. Data processing module: Performs radiometric correction, geometric fine correction and point cloud denoising on multi-source remote sensing data in sequence, and converts the processed data from each source to the same geographic coordinate system and the same spatial resolution to form a registered multi-source dataset; The first generation module inputs the registered multi-source dataset into the pre-trained AIGC multimodal feature extraction model, and fuses the terrain semantic information from different data sources through a cross-modal attention mechanism to generate a multimodal feature tensor containing elevation gradient distribution, surface roughness features and terrain structure semantics. First output module: Guided by multimodal feature tensors, it performs generative reconstruction of DEM elevation grid data using a diffusion model. During the reconstruction process, it introduces geometric boundary constraints of rail transit corridors and outputs initial DEM elevation grid data. The second generation module performs multi-scale spatial consistency checks and local outlier identification on the initial DEM elevation grid data, calculates elevation deviation statistics by combining existing DEM reference data, and generates data quality assessment labels. The second output module performs neighborhood-weighted iterative correction processing on abnormal areas in the initial DEM elevation grid data based on data quality assessment labels until the elevation deviation statistics converge to a preset threshold range, and outputs the final rail transit DEM elevation data resources.

Images