A lunar lava tube 3D modeling method and system

Abstract

This invention belongs to the field of lunar exploration and in-situ resource utilization, and specifically discloses a method and system for 3D modeling of lunar lava tubes. The method comprises: fusing multimodal data of lunar lava tubes to obtain a basic 3D model; the multimodal data includes point cloud data collected by a lidar, point cloud data collected by an orbiter penetrating radar, 3D images collected by a stereo camera, and thermal images collected by an infrared thermal imager; inputting the complete regions of the basic 3D model into generative models based on CGAN and VAE, respectively, to obtain two corresponding 3D completion models; and performing a weighted fusion of the two 3D completion models to obtain a complete 3D model of the lunar lava tubes. This method can improve the integrity and accuracy of 3D models of lunar lava tubes, providing technical support for in-depth exploration of lunar lava tubes.

Description

Technical Field

[0001] The present invention belongs to the field of lunar exploration and in-situ resource utilization, and more specifically, relates to a three-dimensional modeling method and system for lunar lava tubes. Background Art

[0002] In recent years, as lunar exploration missions have advanced, lunar lava tubes have gradually garnered widespread attention from the scientific community. Lunar lava tubes are unique underground caverns formed by ancient volcanic activity on the Moon. As the lava cooled and solidified, the difference in velocity created a solid insulating layer on the outside, while the lava continued to flow inside, ultimately forming a hollow tube structure. Due to the Moon's low gravity, lunar lava tubes are much larger than those on Earth, reaching diameters of hundreds or even thousands of meters, making them ideal sites for lunar bases. These lava tubes not only serve as natural barriers, effectively blocking cosmic radiation and reducing the impact of lunar micrometeorite impacts and lunar dust on equipment and personnel, but they also potentially store a wealth of geological information, providing valuable data for studying the Moon's internal structure.

[0003] However, lunar lava tube exploration faces multiple technical challenges. First, because the underground environment of the lava tubes is unknown, the detection equipment cannot rely on prior maps for navigation, and the pathfinding and modeling speeds are greatly reduced. Second, the harsh environment on the lunar surface, such as lunar dust and radiation, interferes with wireless communication signals, making it difficult to maintain real-time data transmission and command interaction. In addition, the lack of satellite navigation signals such as GPS in lunar lava tubes increases the difficulty of equipment navigation and positioning. Finally, the extremely limited lighting inside lunar lava tubes makes it difficult for visual sensors to capture clear images, significantly reducing perception and positioning accuracy, and affecting the equipment's autonomous detection capabilities.

[0004] To address these challenges, multimodal data fusion technology has shown great potential in 3D modeling of lunar lava tubes. By combining multiple sensors, such as lidar, orbiter-penetrating radar, stereo cameras, and infrared thermal imagers, it can more comprehensively capture complex terrain details and improve 3D modeling accuracy. However, in practical applications, the effective integration and synchronous processing of data from these different sensors, as well as the inference of unscanned areas, remain technical challenges that need to be addressed. Summary of the Invention

[0005] In response to the above-mentioned defects or improvement needs of the prior art, the present invention provides a method and system for three-dimensional modeling of lunar lava tubes, the purpose of which is to improve the integrity and accuracy of the three-dimensional modeling of lunar lava tubes.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for three-dimensional modeling of lunar lava tubes is proposed, comprising the following steps:

[0007] fusing multimodal data of lunar lava tubes to obtain a basic three-dimensional model; the multimodal data includes point cloud data collected by a lidar, point cloud data collected by an orbiter penetrating radar, three-dimensional images collected by a stereo vision camera, and thermal images collected by an infrared thermal imager;

[0008] The complete areas in the basic 3D model are input into the generative models based on CGAN and VAE respectively, and two corresponding 3D completion models are obtained; the two 3D completion models are weightedly fused to obtain a complete 3D model of the lunar lava tube.

[0009] As a further preferred method, the method of obtaining multimodal data is specifically as follows:

[0010] In the lunar lava tubes, the lidar uses short-wave infrared laser technology to collect point cloud data; the orbiter penetrating radar collects point cloud data through 30Hz to 100Hz VHF signals; the stereo vision camera uses digital light enhancement technology to generate three-dimensional images; and an infrared thermal imager with a sensitivity of 0.05℃ to 0.1℃ and a band of 3μm to 14μm is used to collect thermal images.

[0011] As a further preference, when acquiring multimodal data, auxiliary positioning is performed through an inertial navigation system.

[0012] As a further preferred method, the multimodal data of the lunar lava tubes are fused to obtain a basic three-dimensional model, including the following steps:

[0013] The point cloud data collected by the lidar, the point cloud data collected by the surround penetrating radar, and the images collected by the stereo vision camera are reconstructed in three dimensions using SLAM technology to obtain three corresponding three-dimensional reconstruction models.

[0014] Feature points are extracted from the three 3D reconstruction models and thermal images respectively, and then based on the feature points, model matching is performed using the ICP algorithm to obtain a basic 3D model.

[0015] As a further preferred method, model matching is performed using an ICP algorithm, specifically:

[0016] In each iteration of the ICP algorithm, when the error function value is less than the error threshold, the iteration is stopped to obtain the optimal transformation matrix and achieve model matching; the error function adopts the mean square error MSE;

[0017] The error threshold gradually decreases as the iteration proceeds:

[0018]

[0019] Among them, T nis the error threshold at the nth iteration, T0 is the initial error threshold, error(0) is the initial error, error(i) is the error at the i-th iteration, and β is the parameter for adjusting the decay rate.

[0020] As a further preferred method, the method for constructing a generative model based on VAE includes:

[0021] Perform data enhancement on the basic 3D model to obtain a preliminary 3D model; extract several complete parts from the preliminary 3D model to construct a training set;

[0022] VAE consists of an encoder and a decoder. VAE is trained through a training set: the encoder takes the complete part as input to obtain latent space variables; the decoder decodes the latent space variables to generate a three-dimensional completion model; the loss function L used in training VAE as follows:

[0023] L VAE =D KL (q(z|x)||p(z))-E q(z|x) [p(x|z)]

[0024] Among them, D KL represents KL divergence, q(z|x) represents the posterior distribution of the latent variable z generated by the encoder according to the input data x, p(z) represents the prior distribution of the latent variable z, and || represents the distribution difference operator in KL divergence; E q(z|x) represents the expectation of the latent variable z, and p(x|z) represents the three-dimensional completion model generated by the decoder;

[0025] The trained VAE is the VAE-based generative model.

[0026] As a further preferred method, the method for constructing a generative model based on CGAN includes:

[0027] CGAN consists of a generator and a discriminator. CGAN is trained using a training set: the generator takes the complete part as input to generate a 3D completion model; the discriminator discriminates the authenticity of the complete part and the 3D completion model generated by the generator; the loss function L used in training is CGAN as follows:

[0028] L CGAN =E[logD(x)]+E[log(1-D(G(z))]

[0029] Where E represents the expectation of all samples, G(z) represents the three-dimensional completion model generated by the generator, D is the discriminator, which determines the probability of the data being true, D(x) represents the discrimination result of the input data x, and D(G(z)) represents the discrimination result of the three-dimensional completion model G(z);

[0030] The trained generator is the CGAN-based generation model.

[0031] As a further preferred embodiment, data enhancement is performed on the basic three-dimensional model, including: performing data enhancement on the basic three-dimensional model by adding noise, local cropping, and local amplification.

[0032] As a further preferred method, the two three-dimensional completion models are weightedly fused, including: using a multilinear interpolation method to weightedly fuse the two three-dimensional completion models; then correcting the geometric consistency of the fused model through a global optimization algorithm to obtain a complete three-dimensional model of the lunar lava tube.

[0033] According to another aspect of the present invention, a lunar lava tube 3D modeling system is provided, comprising a processor configured to execute the above-mentioned lunar lava tube 3D modeling method.

[0034] In general, the above technical solutions conceived by the present invention have the following technical advantages compared with the existing technology:

[0035] 1. This invention combines multimodal data fusion with a generative model inference and completion approach to 3D modeling, inferring and completing the structure of unscanned areas of lunar lava tubes. This improves the integrity and accuracy of the constructed 3D model of lunar lava tubes and provides important technical reserves and scientific basis for future lunar lava tube exploration.

[0036] 2. The present invention optimizes the design of multi-modal sensor data acquisition to address the challenges of detecting lunar lava tubes without light, complex terrain, and deep structures. The lidar uses short-wave infrared laser technology to reduce interference from lunar dust reflection and light scattering, thereby enhancing the ability to detect the surface and deep structures of the tube. The surround penetrating radar uses the low VHF band and has strong penetrating detection capabilities. It can obtain the overall direction and hidden deep structure of the lunar lava tube, and provide accurate pipeline model positioning and direction information. The stereo vision camera uses digital light enhancement technology, which is suitable for the weak light environment of the lunar lava tube and can generate high-precision stereo images in a light-limited environment. The infrared thermal imager can capture local small temperature differences in a large temperature difference environment through high sensitivity adjustment, and identify cracks, rock layer changes and hidden structures inside the tube. The collaborative fusion of multiple sensors ensures that the system can achieve high-precision detection and three-dimensional modeling in the complex environment of the moon.

[0037] 3. The present invention integrates an inertial navigation system (INS), which can achieve high-precision positioning even in a signal-free environment, ensuring the accuracy of navigation and modeling of the equipment in lunar lava tubes.

[0038] 4. When dealing with complex terrain or obstructions, a single sensor's limited perception capabilities make it difficult to provide comprehensive data support. This invention achieves precise modeling of the complex internal structure of lunar lava tubes through multi-sensor data fusion. LiDAR and orbiter penetrating radar provide three-dimensional point clouds and information on the overall direction of the tubes, stereo vision cameras generate three-dimensional depth maps, and infrared thermal imagers supplement temperature change data. SLAM technology is then used to generate a three-dimensional point cloud model of the scanned area. Based on the ICP algorithm, the multimodal data of the lunar lava tubes is temporally and spatially aligned to improve the robustness of data fusion. This fusion technology overcomes the limitations of single sensors in complex terrain, enhances data comprehensiveness, improves model accuracy and stability, and meets the requirements for three-dimensional modeling of lunar lava tubes.

[0039] 5. In lunar lava tubes, due to narrow structures and occlusions, some areas are difficult to fully scan, resulting in missing models. The present invention uses a pre-trained generative large model to infer and complete the unscanned areas of lunar lava tubes. Specifically, a generative large model combining variational autoencoders (VAE) and conditional generative adversarial networks (CGAN) is used to infer and complete the three-dimensional structure of unscanned areas. VAE encodes the three-dimensional point cloud of the scanned area, extracts complex geometric features, and reconstructs the spatial information of the missing area through decoding; CGAN infers the three-dimensional structure of the unscanned area from the scanned data through adversarial training of the generator and the discriminator, and improves the performance of the generated three-dimensional model in terms of details and overall consistency through multiple iterative training and dynamic parameter challenges. This model completion method further improves the modeling accuracy of unscanned areas through multi-source data fusion, ensures the integrity of the three-dimensional model in complex environments, and meets the modeling requirements of lunar lava tubes. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a schematic diagram of the process of a three-dimensional modeling method for a lunar lava tube according to an embodiment of the present invention;

[0041] Figure 2 A schematic diagram of the process of constructing a basic three-dimensional model according to an embodiment of the present invention;

[0042] Figure 3 Schematic diagram of the three-dimensional model completion and optimization process according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0044] The embodiment of the present invention provides a method for three-dimensional modeling of lunar lava tubes, such as Figure 1 As shown, the following steps are included:

[0045] S1. Multimodal data acquisition and preliminary processing: Multimodal data acquisition of lunar lava tubes is performed using lidar, orbiter penetrating radar, stereo vision camera, and infrared thermal imager.

[0046] Furthermore, the lidar uses shortwave infrared lasers to generate high-precision three-dimensional point cloud data, reducing interference from lunar dust and reflected light, ensuring accurate detection of complex terrain within lunar lava tubes, including ridges, depressions, and intersections. The orbiter-penetrating radar uses low-VHF signals (preferably 30Hz-100Hz) to penetrate the lunar surface, capturing the direction, structure, and deep hidden areas of lunar lava tubes, providing rough three-dimensional topography. The stereo vision camera utilizes digital illumination enhancement technology, actively supplementing illumination and pre-processing low-light images to enhance visual quality, ensuring high-resolution imagery captures geometric and textural details of the lava tube surface in low or no light conditions. A high-sensitivity infrared thermal imager (preferably with a sensitivity of 0.05°C-0.1°C and a wavelength range of 3μm-14μm) is used to capture temperature distribution within the tubes. This allows for the detection of subtle temperature variations within lunar lava tubes despite extreme temperature differences, detecting cracks, rock layer changes, and other discontinuities within the lava tubes, and identifying thermal radiation signatures of different geological layers. After synchronous data collection by multiple sensors, the data is denoised and invalid data is eliminated to ensure data quality.

[0047] Specifically, the acquisition operation is carried out according to the following steps: The orbiter penetrating radar is located on a satellite or orbiter platform, using low VHF signals to penetrate the lunar surface and detect the direction and deep structure of the lunar lava tube. At the same time, the detection equipment is mounted on a mobile platform and slowly advances along the lava tube, triggering multi-modal sensors to collect data synchronously at fixed intervals or at key geological structure locations. The lidar collects point cloud data in real time and uses a built-in filtering system to remove interference from lunar dust and reflected light to ensure the clarity of the detection data. The binocular stereo images generated by the stereo vision camera can clearly reflect the details and geometric changes on the surface of the lava tube. The infrared thermal imager captures the temperature difference distribution inside the lava tube, identifies areas with large temperature differences and local hotspots, and helps locate areas with large terrain changes within the tube.

[0048] The collected multimodal data is wirelessly transmitted to the platform's computing unit, where it undergoes format conversion and noise filtering. It is then aligned in time and space based on the collected location information, timestamp, and trajectory of the mobile platform. During this alignment process, the platform's computing unit adjusts the sensor data in real time to ensure precise alignment of the multimodal data within the same time and spatial coordinate system. Simultaneously, the data processing unit monitors the quality of each sensor's data in real time, triggering a resampling mechanism when missing data or significant errors are detected to ensure data integrity.

[0049] Furthermore, the acquisition equipment has been optimized to adapt to the complex environment of the lunar lava tubes, and can operate stably under conditions such as no communication signals, no positioning navigation, and no light. The data of each sensor is assisted in positioning by the inertial navigation system (INS), which enables the equipment to achieve high-precision positioning and modeling operations through INS in an environment without navigation signals. The equipment has local data storage and automatic synchronization functions to ensure that the equipment can continuously collect and store multimodal data in the absence of communication signals; after the communication signal is restored, the system will automatically upload the stored data to achieve seamless data synchronization and ensure data integrity and continuity. In addition, the internal processing unit of the equipment has been optimized to ensure that data processing and fusion are completed independently without external communication, ensuring the smooth progress of the exploration mission.

[0050] S2. Multimodal data alignment and fusion: Based on the collected multimodal data, SLAM technology is used for 3D reconstruction. Then, the ICP algorithm is used to match and fuse the multimodal 3D models to generate a basic 3D model of the lava tube.

[0051] like Figure 2 As shown, the specific steps include:

[0052] S21. Use SLAM technology to perform three-dimensional reconstruction on the point cloud data collected by the lidar, the point cloud data collected by the surround penetrating radar, and the images collected by the stereo vision camera, and obtain three corresponding three-dimensional reconstruction models.

[0053] Specifically, the laser radar data is processed by LIDAR-SLAM to generate a high-precision three-dimensional point cloud model; the orbiter penetrating radar data is processed by GPR-SLAM to obtain the overall direction and deep structure of the lava tube; the three-dimensional images of the stereo vision camera are processed by Stereo Visual SLAM to generate a three-dimensional surface structure map, which improves the texture information of the three-dimensional modeling of the lunar lava tube.

[0054] S22, extracting feature points from the three 3D reconstructed models and the thermal image respectively, and then performing model matching based on the feature points using the ICP algorithm to obtain a basic 3D model. Specifically, this includes:

[0055] First, feature points are extracted from each model. Feature point extraction is based on normal vector and curvature analysis, selecting key points representing the model's geometric structure.

[0056] Then, the improved ICP algorithm is used to register the feature point set, and the registration error is calculated in the iterative process. The error calculation optimization indicator is the mean square error (MSE), which is expressed as follows:

[0057]

[0058] Among them, p i is the source point cloud feature point, q i is the target point cloud feature point, N is the number of feature points, R is the rotation matrix, and t is the translation vector. This error function describes the matching error between point clouds, and the optimization goal is to minimize this error by adjusting the rotation and translation matrices.

[0059] The above indicators are used to evaluate whether the model meets the set threshold standards. When the MSE drops to the error threshold, the transformation matrix at this time is output to complete the model alignment.

[0060] Each model was matched using the aforementioned method. For example, the point cloud data collected by the lidar was first matched with the point cloud data collected by the orbiter's penetrating radar. The resulting 3D model was then aligned with the 3D image captured by the stereo camera. Finally, the aligned 3D model was matched with the thermal image data collected by the infrared thermal imager, ultimately forming a basic 3D model of the lunar lava tubes. This model incorporates features such as the lunar lava tube's topographic relief, interior wall contours, and height differences.

[0061] Furthermore, while the standard ICP algorithm uses a fixed threshold for point cloud matching, the improved ICP algorithm of this invention introduces an adaptive threshold adjustment mechanism based on error accumulation. Specifically, a larger error threshold is set at the beginning of the registration process to achieve a rough match and quickly reduce multimodal data errors. As the iterations proceed, the error threshold is gradually reduced based on the error changes at each iteration, thereby improving registration accuracy. The error threshold update formula is as follows:

[0062]

[0063] Among them, T n is the error threshold at the nth iteration, T0 is the initial error threshold, error(0) is the initial error, error(i) is the error at the i-th iteration, and β is the parameter for adjusting the decay rate.

[0064] S3. Training and inferential completion of the generative large model: Due to occlusion and other factors during data collection, the basic 3D model is missing. Based on the generative model, the basic 3D model is optimized and completed to obtain a complete 3D model of the lunar lava tubes.

[0065] like Figure 3 As shown, the specific steps include:

[0066] S31. Pre-trained generative model.

[0067] First, a training set is constructed: data augmentation is performed on the base 3D model. Specifically, data augmentation techniques such as adding noise, local cropping, and local magnification are used to increase the diversity of the training set and improve the model's generalization ability in inferring unknown areas, resulting in a preliminary 3D model. The complete 3D structure is then extracted from the preliminary 3D model to construct a training set. In some embodiments, samples of Earth's lava tubes can also be used to create a training set for model training.

[0068] Then, VAE and CGAN are trained separately using the training set to obtain generation models based on VAE and CGAN.

[0069] Furthermore, the VAE model includes an encoder and decoder structure. The encoder takes the complete part of the base 3D model as input, extracts latent space variables, and focuses on reconstructing the missing areas. The decoder decodes the latent space variables, infers the geometry of the missing areas, and generates a completed 3D model. The VAE loss function during training is as follows:

[0070] L VAE =D KL (q(z|x)||p(z))-E q(z|x) [p(x|z)]

[0071] Among them, D KL represents the KL divergence, q(z|x) represents the posterior distribution of the latent variable z generated by the encoder, p(z) represents the prior distribution of the latent variable z, and || represents the distribution difference operator in the KL divergence; E q(z|x) represents the expectation of the latent variable z, p(x|z) represents the 3D completion model generated by the decoder; z represents the latent variable, x represents the input data or 3D model;

[0072] Furthermore, the CGAN model includes a generator and a discriminator structure. The generator takes the complete part of the base model as input and generates a 3D completion model. The discriminator distinguishes between the complete part and the 3D completion model, evaluates the authenticity of the generated result, and optimizes the generator performance through a loss function. The CGAN loss function during training is:

[0073] L CGAN=E[logD(x)]+E[log(1-D(G(z)))]

[0074] Among them, L CGAN is the CGAN loss function, E represents the expectation of all samples, G(z) represents the three-dimensional completion model generated by the generator, D is the discriminator, which judges the probability of the data being true, D(x) represents the discrimination result of the input data x, and D(G(z)) represents the discrimination result of the three-dimensional completion model G(z)

[0075] After training, the VAE and CGAN generators can complete the three-dimensional model respectively. The VAE generates the geometric shape of the missing area through latent space variables, and the CGAN generator generates complete three-dimensional data.

[0076] S32. Input the complete area in the basic 3D model into the generative models based on CGAN and VAE respectively, and obtain two corresponding 3D completion models, complete the occluded or unscanned areas, and perform weighted fusion on the two 3D completion models to obtain a complete 3D model of the lunar lava tube.

[0077] Furthermore, the two 3D completion models are weightedly fused, including: using multilinear interpolation to weightedly fuse the overlapping areas of the two models; then, using a global optimization algorithm to correct the geometric consistency of the fused models, to obtain a complete 3D model. This complete 3D model is continuous and consistent, covering known areas and inferring unknown areas.

[0078] Specifically, weighted fusion assigns different weights based on the quality and confidence of the data, ensuring that high-quality data has a greater impact on the final model; the weighting process uses multiple linear interpolation to ensure smooth transitions in boundary areas. After splicing, a global optimization algorithm is used to correct geometric consistency and correct errors and discontinuities at the splicing points; the global error function is optimized by minimizing the following formula:

[0079]

[0080] Among them, p i With q i The corresponding point of the boundary area, T(q i ) is the geometric transformation of the boundary area.

[0081] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A three-dimensional modeling method for lunar lava tubes, characterized in that: The steps include: fusing multimodal data of lunar lava tubes to obtain a basic three-dimensional model; the multimodal data includes point cloud data collected by a lidar, point cloud data collected by an orbiter penetrating radar, three-dimensional images collected by a stereo vision camera, and thermal images collected by an infrared thermal imager; The complete areas in the basic 3D model are input into the generative models based on CGAN and VAE respectively, and two corresponding 3D completion models are obtained; the two 3D completion models are weightedly fused to obtain a complete 3D model of the lunar lava tube.

2. The method for three-dimensional modeling of lunar lava tubes according to claim 1, wherein: The specific method for obtaining multimodal data is: In the lunar lava tubes, the lidar uses short-wave infrared laser technology to collect point cloud data; the orbiter penetrating radar collects point cloud data through 30Hz to 100Hz VHF signals; the stereo vision camera uses digital light enhancement technology to generate three-dimensional images; and an infrared thermal imager with a sensitivity of 0.05℃ to 0.1℃ and a band of 3μm to 14μm is used to collect thermal images.

3. The three-dimensional modeling method of lunar lava tubes according to claim 1, wherein: When acquiring multimodal data, auxiliary positioning is performed through the inertial navigation system.

4. The method for three-dimensional modeling of lunar lava tubes according to claim 1, wherein: The multimodal data of lunar lava tubes are fused to obtain a basic 3D model, which includes the following steps: The point cloud data collected by the lidar, the point cloud data collected by the surround penetrating radar, and the images collected by the stereo vision camera are reconstructed in three dimensions using SLAM technology to obtain three corresponding three-dimensional reconstruction models. Feature points are extracted from the three 3D reconstruction models and thermal images respectively, and then based on the feature points, model matching is performed using the ICP algorithm to obtain a basic 3D model.

5. The three-dimensional modeling method of lunar lava tubes according to claim 4, characterized in that: Model matching is performed using the ICP algorithm, specifically: In each iteration of the ICP algorithm, when the error function value is less than the error threshold, the iteration is stopped to obtain the optimal transformation matrix and achieve model matching; the error function adopts the mean square error MSE; The error threshold gradually decreases as the iteration proceeds: Among them, T n is the error threshold at the nth iteration, T0 is the initial error threshold, error(0) is the initial error, error(i) is the error at the i-th iteration, and β is the parameter for adjusting the decay rate.

6. The method for three-dimensional modeling of lunar lava tubes according to claim 1, wherein: Methods for constructing VAE-based generative models include: Perform data enhancement on the basic 3D model to obtain a preliminary 3D model; extract several complete parts from the preliminary 3D model to construct a training set; VAE consists of an encoder and a decoder. VAE is trained through a training set: the encoder takes the complete part as input to obtain latent space variables; the decoder decodes the latent space variables to generate a three-dimensional completion model; the loss function L used in training VAE as follows: L VAE =D KL (q(z|x)||p(z))-E q(z|x) [p(x|z)] Among them, D KL represents KL divergence, q(z|x) represents the posterior distribution of the latent variable z generated by the encoder according to the input data x, p(z) represents the prior distribution of the latent variable z, and || represents the distribution difference operator in KL divergence; E q(z|x) represents the expectation of the latent variable z, and p(x|z) represents the three-dimensional completion model generated by the decoder; The trained VAE is the VAE-based generative model.

7. The method for three-dimensional modeling of lunar lava tubes according to claim 6, wherein: The construction method of the CGAN-based generative model includes: CGAN consists of a generator and a discriminator. CGAN is trained using a training set: the generator takes the complete part as input to generate a 3D completion model; the discriminator discriminates the authenticity of the complete part and the 3D completion model generated by the generator; the loss function L used in training is CGAN as follows: L CGAN =E[logD(x)]+E[log(1-D(G(z))] Where E represents the expectation of all samples, G(z) represents the three-dimensional completion model generated by the generator, D is the discriminator, which determines the probability of the data being true, D(x) represents the discrimination result of the input data x, and D(G(z)) represents the discrimination result of the three-dimensional completion model G(z); The trained generator is the CGAN-based generation model.

8. The method for three-dimensional modeling of lunar lava tubes according to claim 7, wherein: The basic 3D model is enhanced by data enhancement, including adding noise, local cropping and local amplification.

9. The method for three-dimensional modeling of a lunar lava tube according to any one of claims 1 to 8, wherein: The two 3D completion models are weightedly fused, including: using multilinear interpolation to weightedly fuse the two 3D completion models; then the geometric consistency of the fused models is corrected through a global optimization algorithm to obtain a complete 3D model of the lunar lava tube.

10. A lunar lava tube 3D modeling system, characterized in that: It includes a processor, which is used to execute the lunar lava tube three-dimensional modeling method as described in any one of claims 1-8.

Images