Hyperspectral lithology identification roadway modeling method

Abstract

The invention belongs to the field of intelligent mining and geological exploration of mines, and particularly relates to a hyperspectral lithology recognition roadway modeling method which comprises the following steps: S1, data acquisition, S2, point cloud processing and feature extraction, S3, data restoration and enhancement, S4, lithology preliminary classification, S5, lithology fine interpretation, S6, semantic constraint registration, S7, semantic model generation and S8, model output and application. Compared with single-point measurement and manual station scanning of a traditional fixed spectrometer, the hyperspectral lithologic recognition roadway modeling method has the advantages that a mobile continuous scanning mode is adopted, real-time positioning and splicing are achieved in combination with tight coupling synchronous positioning and a map building algorithm, and automation of the whole process from data collection to model output is achieved. According to the method, the scanning efficiency is improved by more than 20 times, the manual intervention and station setting cost are greatly reduced, the whole process automation from data acquisition to semantic model generation is realized, and the modeling efficiency and the lithology identification precision are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent mining and geological exploration technology, specifically a hyperspectral lithology identification method for tunnel modeling. Background Technology

[0002] Hyperspectral technology is a technique based on a large amount of narrow-band image data, the core of which lies in combining imaging technology with spectral technology. Hyperspectral technology can acquire continuous and detailed spectral information, forming a three-dimensional data cube containing spatial and spectral information. As the main man-made structure, the tunnel is the main channel for coal mining, transportation and ventilation. The three-dimensional tunnel model can more intuitively show the spatial relationship between tunnel engineering and geological bodies.

[0003] Currently, existing hyperspectral systems are bulky, difficult to disassemble and shrink, and inconvenient for outdoor use. Furthermore, the field of view of the imaging spectral units in existing hyperspectral systems is fixed, resulting in poor versatility. Therefore, we propose a hyperspectral lithology identification method for tunnel modeling. Summary of the Invention

[0004] The main objective of this invention is to provide a hyperspectral lithology identification method for tunnel modeling, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention proposes a hyperspectral lithology identification tunnel modeling method, comprising the following steps: S1. Data Acquisition: Hyperspectral image sequences, depth point cloud sequences, and inertial measurement unit data are simultaneously acquired during tunnel travel using a mobile hyperspectral-geometric synchronous acquisition system. S2. Point Cloud Processing and Feature Extraction: Based on the tightly coupled synchronous localization and mapping (SLAM) algorithm, real-time pose estimation is performed on the inertial measurement unit data and the depth point cloud sequence collected in S1. Based on the pose estimation, the point cloud data of each frame in the depth point cloud sequence is transformed to the same coordinate system to complete the initial stitching of the point cloud. At the same time, the feature point cloud generated by the tightly coupled synchronous localization and mapping algorithm during the operation is output as the skeleton point cloud of the alleyway structure. S3. Data Repair and Enhancement: The overall hyperspectral data cube formed after the initial stitching in S2 is preprocessed, and missing or abnormal spectral data in the overall hyperspectral data cube caused by environmental interference is repaired and enhanced based on the Kriging interpolation method. The spatial variation model used in the Kriging interpolation method is an anisotropic model, and its range parameters along the axial and radial directions of the tunnel are adaptively determined by analyzing the spatial distribution range of the tunnel structure skeleton point cloud output by S2 in different main directions. S4. Preliminary lithological classification: Based on the spectral angle mapping algorithm, the similarity between the spectral vector of each pixel in the overall hyperspectral data cube and the reference spectral vectors of various lithologies in the pre-constructed endmember spectral library is calculated to complete the preliminary classification of the lithology of the roadway surrounding rock. S5. Refined Lithology Interpretation: Based on the multi-source heterogeneous tunnel dataset composed of the tunnel structure skeleton point cloud generated in S2 and the overall hyperspectral data cube repaired in S3, a dual-stream 3D convolutional neural network model is constructed and applied to refine the preliminary lithology classification results obtained in S4. The dual-stream 3D convolutional neural network model includes a spectral stream and a geometric stream. The spectral stream takes the local 3D blocks of the overall hyperspectral data cube as input, and the geometric stream takes the local voxel grid converted from the tunnel structure skeleton point cloud output in S2 as input. The point cloud normal vector feature sequence in the local voxel grid is encoded through a gated recurrent unit module. The output features of the spectral stream and the geometric stream are integrated through an attention fusion module, and finally a refined lithology classification map is output. This process realizes AI automatic annotation of the tunnel surrounding rock through a deep learning model, completely replacing the traditional manual visual interpretation. S6. Semantic Constraint Registration: The refined lithology classification map obtained in S5 is used as a semantic label and assigned to the corresponding points in the preliminary stitched point cloud obtained in S2. The non-rigid iterative nearest point algorithm is used to accurately register and fuse the multi-view preliminary stitched point cloud with semantic labels. In the registration process, the semantic consistency cost is calculated for the pair of points to be matched in each iteration. If the lithology labels of the two points are different, a penalty term with a preset weight is added to the matching error of the point pair. S7. Semantic Model Generation: Convert the precisely registered semantic point cloud obtained in S6 into a triangular mesh surface model to generate a three-dimensional semantic tunnel model with lithological properties. S8. Model Output and Application: Convert the 3D semantic tunnel model obtained in S7 into an industrial basic class format for output, add material attribute sets to the mesh patches corresponding to different lithological zones in the triangular mesh surface model, and integrate an application programming interface that automatically calculates and outputs rock quality indicators based on the material attributes.

[0006] Preferably, the mobile hyperspectral-geometric synchronous acquisition system in S1 integrates a lightweight hyperspectral camera, a depth camera, and an inertial measurement unit. The spatial transformation relationship between the hyperspectral camera and the depth camera is determined by hand-eye calibration. The hyperspectral image sequence, the depth point cloud sequence, and the data acquisition of the inertial measurement unit are synchronized in time by a hardware trigger signal. The depth camera is a lidar depth camera based on the time-of-flight principle, and the hyperspectral camera preferably has a wavelength range of 400-2500nm.

[0007] Preferably, in S3, the adaptive determination of the range parameter by analyzing the spatial distribution range of the tunnel structure skeleton point cloud in different principal directions is specifically as follows: principal component analysis is performed on the tunnel structure skeleton point cloud, the first principal component direction is taken as the tunnel axis, and the range of the Kriging interpolation anisotropic model along the tunnel axis and radial direction is determined by calculating the distribution scale of the point cloud in the first principal component direction and the second principal component direction perpendicular to it. This method can automatically and objectively capture the structural characteristics of the narrow space of the tunnel, making the interpolation process more targeted.

[0008] Preferably, the end-member spectral library described in S4 is constructed and verified by collecting spectral data from rock core samples collected at the roadway site and combining them with laboratory X-ray diffraction analysis results.

[0009] Preferably, the point cloud normal vector feature sequence in S5 is a sequence formed by performing principal component analysis on the point cloud within the local voxel grid, calculating the normal vector of each point, calculating the angle between the normal vector of each point and the average normal vector of all points in its neighborhood, and arranging them in spatial order.

[0010] Preferably, the weight coefficient of the penalty term in S6 is a constant, which is determined by grid search and cross-validation on a typical lane dataset.

[0011] Preferably, in the triangular mesh surface model described in S7, the boundary of the point cloud partition with different lithological labels is used as an insurmountable constraint edge during the reconstruction process, and the constrained Delaunay triangulation algorithm is used for meshing.

[0012] Preferably, the automatic calculation of the rock quality index in S8 includes the following steps: generating a series of virtual borehole lines along the tunnel axis in the three-dimensional semantic tunnel model; traversing the grid patches that intersect with the virtual borehole lines and identifying continuous line segments with consistent lithology based on their material properties; calculating the total length of all line segments with a length exceeding 10 cm and comparing it with the total length of the virtual borehole lines to obtain the rock quality index value.

[0013] Preferably, the tightly coupled synchronous positioning and map building algorithm at the visual front end extracts and tracks feature points of the spatial locations of pixels with stable spectral characteristics from the preliminary lithology classification results of S4, and incorporates the reprojection error of these feature points into the pose optimization process.

[0014] Preferably, the mobile acquisition system dynamically adjusts its scanning speed in the tunnel based on the sliding average of the triaxial acceleration sum measured in real time by the inertial measurement unit, and automatically reduces its speed when the sliding average exceeds a set threshold.

[0015] This invention provides a hyperspectral lithology identification method for tunnel modeling. It has the following beneficial effects: (1) Compared with the traditional single-point measurement and manual station scanning of fixed spectrometers, the hyperspectral lithology identification tunnel modeling method of this invention adopts a mobile continuous scanning mode, combined with tightly coupled synchronous positioning and map building algorithms to achieve real-time positioning and stitching, realizing full automation from data acquisition to model output. This method improves scanning efficiency by more than 20 times and significantly reduces manual intervention and station setting costs, making it particularly suitable for rapid modeling of long tunnel networks.

[0016] (2) This hyperspectral lithology identification tunnel modeling method overcomes the drawback of separating lithology identification and 3D modeling in traditional methods. This invention simultaneously acquires spectral and geometric data at the hardware level and deeply integrates spectral features, 3D geometry, and inertial navigation data at the software level, ultimately generating a unified semantic model through semantic constraint registration. This enables the model to possess both accurate geometric morphology and realistic lithological properties, providing a complete information foundation for subsequent analysis.

[0017] (3) This hyperspectral lithology identification tunnel modeling method achieves automatic identification and annotation of the surrounding rock of tunnels through improved spectral angle mapping and a dual-stream 3D-CNN model that enables AI automatic annotation. This replaces the time-consuming and subjective manual geological interpretation, resulting in more objective and repeatable results. At the same time, the generated semantic model is not a "static display model." Its embedded lithological properties can directly drive the automatic calculation of rock mechanics indicators such as RQD, extending the application value of the model from visualization to professional engineering analysis and decision support, greatly expanding the practical depth of the model. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the overall steps of the present invention; Figure 2 This is a flowchart illustrating some of the steps of the present invention. Figure 1 ; Figure 3 This is a flowchart illustrating some of the steps of the present invention. Figure 2 ; Figure 4 This is a flowchart illustrating some of the steps of the present invention. Figure 3 ; Figure 5 This is a flowchart illustrating some of the steps of the present invention. Figure 4 ; Figure 6 This is a flowchart illustrating some of the steps of the present invention. Figure 5 ; Figure 7 This is a flowchart illustrating some of the steps of the present invention. Figure 6 ; Figure 8 This is a flowchart illustrating some of the steps of the present invention. Figure 7 ; Figure 9 This is a flowchart illustrating some of the steps of the present invention. Figure 8 .

[0020] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 - Figure 9 This invention proposes a hyperspectral lithology identification method for tunnel modeling, comprising the following steps: S1. Data Acquisition: A mobile hyperspectral-geometric synchronous acquisition system is used to simultaneously acquire hyperspectral image sequences, depth point cloud sequences, and inertial measurement unit (IMU) data while traversing the tunnel. In this step, the hyperspectral image sequence is used to capture the fine spectral characteristics of the surrounding rock in the tunnel, providing core data for subsequent lithology identification; the depth point cloud sequence is used to obtain the three-dimensional spatial geometric information of the tunnel, establishing its three-dimensional framework; and the IMU data records the device's motion attitude and acceleration in real time, providing key parameters for subsequent point cloud stitching and motion distortion correction. Hardware triggering and precise clock synchronization ensure strict spatiotemporal alignment of the three types of data. This multi-source data synchronous acquisition strategy lays a solid foundation for subsequent data fusion processing, overcoming the limitations of a single sensor's insufficient sensing capability in the complex environment of a tunnel. S2. Point Cloud Processing and Feature Extraction: Based on a tightly coupled synchronous localization and mapping (SMR) algorithm, real-time pose estimation is performed on the inertial measurement unit (IMU) data and depth point cloud sequence acquired in S1. Based on the pose estimation, the point cloud data from each frame in the depth point cloud sequence is transformed to the same coordinate system, completing the initial point cloud stitching. Simultaneously, the feature point cloud generated by the tightly coupled SMR algorithm during operation is output as the tunnel structure skeleton point cloud. This step, through tight coupling, deeply fuses visual observation and inertial measurement data, significantly improving positioning accuracy and robustness under challenging environments such as tunnel illumination variations and texture repetition. The extracted tunnel structure skeleton point cloud not only contains spatial topological information of the tunnel, but its point cloud density and distribution characteristics also directly reflect the macroscopic geometric morphology of the tunnel, providing crucial spatial structural priors for subsequent spectral data restoration. S3. Data Repair and Enhancement: The overall hyperspectral data cube formed after the initial stitching in S2 is preprocessed, and missing or abnormal spectral data caused by environmental interference in the overall hyperspectral data cube is repaired and enhanced based on the Kriging interpolation method. The spatial variation model used in the Kriging interpolation method is an anisotropic model. Its range parameters along the tunnel axis and radial direction are adaptively determined by analyzing the spatial distribution range of the tunnel structure skeleton point cloud output from S2 in different principal directions. This step utilizes the inherent narrow spatial structure characteristics of the tunnel to guide the interpolation process. By analyzing the distribution scale of the structure skeleton point cloud in the principal directions (such as the axial and radial variances calculated through principal component analysis), the range parameters of the Kriging model are dynamically adjusted, making the repaired spectral data more consistent with the actual distribution of geological bodies in the tunnel, significantly improving the physical rationality and accuracy of the repair results. S4. Preliminary Lithology Classification: Based on the spectral angle mapping algorithm, the similarity between the spectral vector of each pixel in the overall hyperspectral data cube and the reference spectral vectors of various lithologies in the pre-constructed endmember spectral library is calculated to complete the preliminary classification of the surrounding rock lithology of the tunnel. The spectral angle mapping algorithm effectively reduces the impact of changes in illumination intensity on the classification results by calculating the generalized angle between spectral vectors, making it particularly suitable for scenarios with uneven lighting in tunnels. This preliminary classification result provides reliable initial labels and attention guidance for subsequent fine interpretation. Furthermore, the stable lithology feature points generated can be fed back to the tightly coupled synchronous localization and map building algorithm process in S2, enhancing the localization capability in areas lacking texture and forming positive synergy between processes. S5. Refined Lithological Interpretation: Based on the multi-source heterogeneous tunnel dataset composed of the tunnel structure skeleton point cloud generated in S2 and the overall hyperspectral data cube repaired in S3, a dual-stream 3D convolutional neural network model is constructed and applied to refine the preliminary lithological classification results obtained in S4. The dual-stream 3D convolutional neural network model includes a spectral stream and a geometric stream. The spectral stream takes local 3D blocks of the overall hyperspectral data cube as input, and the geometric stream takes the local voxel mesh transformed from the tunnel structure skeleton point cloud output in S2 as input. A gated recurrent unit module processes the point cloud in the local voxel mesh. The normal vector feature sequence is encoded, and the output features of the spectral and geometric flows are integrated through an attention fusion module to finally output a refined lithology classification map. This process achieves AI-automated labeling of the surrounding rock of the tunnel through a deep learning model, completely replacing traditional manual visual interpretation. This dual-stream 3D convolutional neural network model is the core of AI-automated lithology labeling. Through end-to-end deep learning, it automatically learns discriminative features from the fused spectral and geometric data to complete the intelligent identification and labeling of lithology at each location on the tunnel surface, achieving deep-level complementarity and enhancement of spectral features and spatial structural features. The GRU module introduced in the geometric flow can effectively capture the sequential spatial patterns of rock mass joints, bedding, and other structural surfaces, while the attention fusion mechanism can adaptively weigh the contribution weights of spectral and geometric information in different contexts, thus demonstrating stronger discriminative ability when facing complex lithology identification problems such as "same material, different spectra" and "different material, same spectrum". S6. Semantic Constraint Registration: The refined lithology classification map obtained in S5 is used as a semantic label and assigned to the corresponding points in the preliminary stitched point cloud obtained in S2. The non-rigid iterative nearest point algorithm is used to accurately register and fuse the preliminary stitched point cloud with semantic labels from multiple perspectives. During the registration process, the semantic consistency cost is calculated for the pair of points to be matched in each iteration. If the lithology labels of the two points are different, a penalty term with a preset weight is added to the matching error of the pair of points. By introducing semantic constraints as a strong prior, erroneous matching between different lithology regions is prohibited from the root of the optimization process. This ensures that while optimizing geometric deformation, the semantic boundary remains clear and consistent, thereby generating a unified model with excellent geometric accuracy and semantic integrity. S7. Semantic Model Generation: The precisely registered semantic point cloud obtained in S6 is converted into a triangular mesh surface model to generate a three-dimensional semantic tunnel model with lithological attributes. During the meshing process, the constrained Delaunay triangulation algorithm is used to treat the lithological boundary as a hard constraint to ensure that the meshes of different lithological regions are correctly stitched at the boundary, avoiding the ambiguity or loss of semantic information during the patching process, and providing a high-quality model foundation for subsequent attribute query and mechanical analysis. S8. Model Output and Application: The 3D semantic tunnel model obtained in S7 is converted into an industrial-grade format for output. Material attribute sets are added to the mesh patches corresponding to different lithological zones in the triangular mesh surface model. Simultaneously, an application programming interface (API) is integrated to automatically calculate and output rock quality indicators based on material attributes. This transforms the model from a 3D visualization asset into a computable and analyzable digital twin. Users can interactively query lithology and spectral curves at any location and automatically obtain key rock mechanics indicators such as RQD through built-in interfaces. This allows geological information to directly serve engineering decisions such as tunnel support design and stability assessment, greatly enhancing the model's practical value.

[0023] In an embodiment of the present invention, the mobile hyperspectral-geometric synchronous acquisition system in S1 integrates a lightweight hyperspectral camera, a depth camera, and an inertial measurement unit. The spatial transformation relationship between the hyperspectral camera and the depth camera is determined by hand-eye calibration. The time synchronization of the hyperspectral image sequence, the depth point cloud sequence, and the data acquisition by the inertial measurement unit is achieved by hardware triggering signals. The depth camera can be a lidar depth camera based on the time-of-flight principle to obtain high-precision depth information. The hyperspectral camera is preferably a device with a wavelength range of 400-2500nm to comprehensively cover the diagnostic spectral characteristics of common diagenetic minerals.

[0024] Furthermore, in S3, the range parameters are adaptively determined by analyzing the spatial distribution range of the tunnel structure skeleton point cloud in different principal directions. Specifically, principal component analysis is performed on the tunnel structure skeleton point cloud, the first principal component direction is taken as the tunnel axis, and the range of the Kriging interpolation anisotropic model along the tunnel axis and radial direction is determined by calculating the distribution scale of the point cloud in the first principal component direction and the second principal component direction perpendicular to it. This method can automatically and objectively capture the structural characteristics of the narrow space of the tunnel, making the interpolation process more targeted.

[0025] Furthermore, the S4 mid-terminal spectral library was constructed and validated by collecting spectral data from rock core samples from the tunnel site and combining them with laboratory X-ray diffraction analysis results.

[0026] Furthermore, in S5, the point cloud normal vector feature sequence is formed by performing principal component analysis on the point cloud within the local voxel grid, calculating the normal vector of each point, and then calculating the angle between the normal vector of each point and the average normal vector of all points in its neighborhood, and arranging them in spatial order. This sequence can effectively characterize the occurrence and development of surface structural planes (such as joints and bedding) on ​​the rock mass, providing key geometric evidence for lithological identification.

[0027] Furthermore, in S6, the weight coefficient of the penalty term is a constant, which is determined by grid search and cross-validation on a typical lane dataset. Through systematic parameter optimization, the best balance between geometric accuracy and semantic consistency can be achieved.

[0028] Furthermore, in the triangular mesh surface model in S7, the boundary of the point cloud partitions of different lithological labels is used as an insurmountable constraint edge during the reconstruction process. The constrained Delaunay triangulation algorithm is used for meshing, which ensures that the boundaries of different lithologies remain clear and sharp after the model is surfaced, avoiding ambiguity of semantic information.

[0029] Furthermore, the automatic calculation of rock quality indicators in S8 involves the following steps: generating a series of virtual borehole lines along the tunnel axis in the three-dimensional semantic tunnel model; traversing the grid patches that intersect with the virtual borehole lines and identifying continuous line segments with consistent lithology based on their material properties; and calculating the total length of all line segments longer than 10 centimeters and comparing it with the total length of the virtual borehole lines to obtain the rock quality indicator value. This digitizes and automates the traditional workflow of geological engineers, greatly improving efficiency and consistency.

[0030] Furthermore, the tightly coupled synchronous localization and mapping algorithm at the visual front end extracts and tracks feature points of the spatial locations of pixels with stable spectral characteristics from the preliminary lithology classification results of S4, and incorporates the reprojection error of these feature points into the pose optimization process. This feedback mechanism can utilize the intermediate results of lithology identification to enhance the robustness and stability of the tightly coupled synchronous localization and mapping algorithm in areas with missing roadway textures.

[0031] Furthermore, the mobile acquisition system dynamically adjusts its scanning speed in the tunnel based on the sliding average of the triaxial acceleration measured in real time by the inertial measurement unit. When the sliding average exceeds a set threshold, the travel speed is automatically reduced, and when the vibration is severe, the speed is automatically reduced. This adaptive mechanism helps to ensure the quality of the acquired data and reduce the impact of factors such as motion blur on subsequent processing.

[0032] It should be noted that the above electrical components are all existing technology products. Those skilled in the art should select, install and complete the circuit debugging work according to the needs of use to ensure that all electrical appliances can work normally. The components are all general standard parts or components known to those skilled in the art. Their structure and principle can be known by those skilled in the art through technical manuals or conventional experimental methods. No specific restrictions are made here.

[0033] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for modeling roadways using hyperspectral lithology identification, characterized in that: Includes the following steps: S1. Data Acquisition: Hyperspectral image sequences, depth point cloud sequences, and inertial measurement unit data are simultaneously acquired while traveling through the tunnel using a mobile hyperspectral-geometric synchronous acquisition system. S2. Point Cloud Processing and Feature Extraction: Based on the tightly coupled synchronous positioning and mapping algorithm, the inertial measurement unit data and the depth point cloud sequence collected in S1 are subjected to real-time pose estimation. Based on the pose estimation, the point cloud data of each frame in the depth point cloud sequence are transformed to the same coordinate system to complete the initial stitching of the point cloud. At the same time, the feature point cloud generated by the tightly coupled synchronous positioning and mapping algorithm during the operation is output as the skeleton point cloud of the alleyway structure. S3. Data Repair and Enhancement: The overall hyperspectral data cube formed after the initial stitching in S2 is preprocessed, and missing or abnormal spectral data in the overall hyperspectral data cube caused by environmental interference is repaired and enhanced based on the Kriging interpolation method. The spatial variation model used in the Kriging interpolation method is an anisotropic model, and its range parameters along the axial and radial directions of the tunnel are adaptively determined by analyzing the spatial distribution range of the tunnel structure skeleton point cloud output by S2 in different main directions. S4. Preliminary lithological classification: Based on the spectral angle mapping algorithm, calculate the similarity between the spectral vector of each pixel in the overall hyperspectral data cube and the reference spectral vectors of various lithologies in the pre-constructed endmember spectral library, and complete the preliminary classification of the lithology of the roadway surrounding rock. S5. Refined Lithology Interpretation: Based on the multi-source heterogeneous tunnel dataset composed of the tunnel structure skeleton point cloud generated in S2 and the overall hyperspectral data cube repaired in S3, a dual-stream 3D convolutional neural network model is constructed and applied to refine the preliminary lithology classification results obtained in S4. The dual-stream 3D convolutional neural network model includes a spectral stream and a geometric stream. The spectral stream takes the local 3D blocks of the overall hyperspectral data cube as input, and the geometric stream takes the local voxel grid converted from the tunnel structure skeleton point cloud output in S2 as input. The point cloud normal vector feature sequence in the local voxel grid is encoded through a gated recurrent unit module. The output features of the spectral stream and the geometric stream are integrated through an attention fusion module to finally output a refined lithology classification map. S6. Semantic Constraint Registration: The refined lithology classification map obtained in S5 is used as a semantic label and assigned to the corresponding points in the preliminary stitched point cloud obtained in S2. The non-rigid iterative nearest point algorithm is used to accurately register and fuse the multi-view preliminary stitched point cloud with semantic labels. In the registration process, the semantic consistency cost is calculated for the pair of points to be matched in each iteration. If the lithology labels of the two points are different, a penalty term with a preset weight is added to the matching error of the point pair. S7. Semantic Model Generation: Convert the precisely registered semantic point cloud obtained in S6 into a triangular mesh surface model to generate a three-dimensional semantic tunnel model with lithological properties. S8. Model Output and Application: Convert the 3D semantic tunnel model obtained in S7 into an industrial basic class format for output, add material attribute sets to the mesh patches corresponding to different lithological zones in the triangular mesh surface model, and integrate an application programming interface that automatically calculates and outputs rock quality indicators based on the material attributes.

2. The method for modeling roadways using hyperspectral lithology identification according to claim 1, characterized in that: The mobile hyperspectral-geometric synchronous acquisition system described in S1 integrates a lightweight hyperspectral camera, a depth camera, and an inertial measurement unit. The spatial transformation relationship between the hyperspectral camera and the depth camera is determined by hand-eye calibration. The time synchronization of the hyperspectral image sequence, the depth point cloud sequence, and the data acquisition by the inertial measurement unit is achieved by hardware triggering signals. The depth camera is a lidar depth camera based on the time-of-flight principle.

3. The method for modeling tunnels using hyperspectral lithology identification according to claim 1, characterized in that: The variable range parameters described in S3 are adaptively determined by analyzing the spatial distribution range of the tunnel structure skeleton point cloud in different principal directions. Specifically, principal component analysis is performed on the tunnel structure skeleton point cloud, the first principal component direction is taken as the tunnel axis, and the variable range of the Kriging interpolation anisotropic model along the tunnel axis and radial direction is determined by calculating the distribution scale of the point cloud in the first principal component direction and the second principal component direction perpendicular to it.

4. The method for modeling roadways using hyperspectral lithology identification according to claim 1, characterized in that: The end-member spectral library described in S4 is constructed and verified by collecting spectral data from rock core samples collected in the tunnel and combining them with laboratory X-ray diffraction analysis results.

5. The method for modeling tunnels using hyperspectral lithology identification according to claim 1, characterized in that: The point cloud normal vector feature sequence described in S5 is a sequence formed by performing principal component analysis on the point cloud within the local voxel grid, calculating the normal vector of each point, calculating the angle between the normal vector of each point and the average normal vector of all points in its neighborhood, and arranging them in spatial order.

6. The method for modeling tunnels using hyperspectral lithology identification according to claim 1, characterized in that: The weight coefficient of the penalty term described in S6 is a constant, which is determined by grid search and cross-validation on a typical lane dataset.

7. The method for modeling tunnels using hyperspectral lithology identification according to claim 1, characterized in that: In the triangular mesh surface model described in S7, the boundary of the point cloud partition with different lithological labels is used as an insurmountable constraint edge during the reconstruction process, and the constrained Delaunay triangulation algorithm is used for meshing.

8. The method for modeling roadways using hyperspectral lithology identification according to claim 1, characterized in that: The automatic calculation of rock quality index described in S8 includes the following steps: generating a series of virtual borehole lines along the tunnel axis in the three-dimensional semantic tunnel model; traversing the grid patches that intersect with the virtual borehole lines and identifying continuous line segments with consistent lithology based on their material properties; calculating the total length of all line segments with a length exceeding 10 cm and comparing it with the total length of the virtual borehole lines to obtain the rock quality index value.

9. The method for modeling roadways using hyperspectral lithology identification according to claim 1, characterized in that: The tightly coupled synchronous localization and map building algorithm extracts and tracks feature points in the spatial location of pixels with stable spectral characteristics from the preliminary lithology classification results of S4 at the visual front end, and incorporates the reprojection error of these feature points into the pose optimization process.

10. The method for modeling roadways using hyperspectral lithology identification according to claim 1, characterized in that: The mobile acquisition system dynamically adjusts its scanning speed in the tunnel based on the sliding average of the triaxial acceleration measured in real time by the inertial measurement unit. When the sliding average exceeds a set threshold, the system automatically reduces its speed.

Images