Artificial intelligence-based topographic surveying and mapping method and system

Abstract

The invention discloses a topographic surveying and mapping method based on artificial intelligence. The topographic surveying and mapping method comprises the steps of (1) multi-source data fusion acquisition; (2) intelligently extracting topographic features; (3) three-dimensional reconstruction optimization; (4) dynamic terrain evolution modeling; the invention also discloses a system for implementing the topographic mapping method based on artificial intelligence, and the system comprises a data acquisition and preprocessing module, a topographic feature intelligent extraction module, a three-dimensional reconstruction optimization module, a dynamic topographic evolution modeling module and an interactive visualization and decision platform. The data acquisition and preprocessing module, the terrain feature intelligent extraction module, the three-dimensional reconstruction optimization module, the dynamic terrain evolution modeling module and the interactive visualization and decision-making platform are connected in sequence, and each module communicates through a gRPC interface and supports independent upgrading.

Description

Technical Field

[0001] The present invention relates to the technical field of topographic surveying and mapping, and more specifically, it relates to a topographic surveying and mapping method and system based on artificial intelligence. Background Art

[0002] With the emergence of photography technology, photogrammetry has begun to be used in topographic surveying and mapping. By taking photos from the air or the ground, a stereoscopic model is created and topographic information is extracted. During the process of engineering construction, topographic map surveying and mapping work is the preparatory work in the early stage. Measuring and collecting relevant data of the geographical space can make the building have a more harmonious relationship with nature, and make the artificial building information become more clear in the standardization process of the geographical space position. Using methods such as numbers, names, and distances to label the space position, clearly labeling the information of the construction project, dividing distances, place names, and various attributes, so that the geographical space position can be clearly shown on the topographic map, facilitating the development of engineering construction. At present, the following technical bottlenecks exist in traditional topographic surveying and mapping methods:

[0003] Low data fusion efficiency: The spatio-temporal benchmarks of multi-source sensor data are inconsistent, and traditional ICP registration is time-consuming and has insufficient accuracy.

[0004] Coarse feature extraction granularity: Networks such as U-Net are difficult to identify sub-meter topographic elements (such as ditches, cracks) due to the loss of feature map resolution (1 / 32).

[0005] Lack of dynamic prediction: Existing 3D reconstruction methods (such as oblique photography) lack physical constraints and cannot simulate the geological mechanics evolution process.

[0006] Lag in decision-making support: Traditional numerical simulation calculations are time-consuming and cannot meet the real-time warning requirements. Summary of the Invention

[0007] Aiming at the deficiencies of the existing technology, the purpose of the present invention is to provide a topographic surveying and mapping method and system based on artificial intelligence.

[0008] To achieve the above purpose, the present invention provides the following technical solutions:

[0009] A topographic surveying and mapping method based on artificial intelligence, comprising the following steps:

[0010] (1) Multi-source data fusion acquisition: Using multiple devices to collect data sources of the terrain to be surveyed and mapped, and performing multi-modal data registration on the collected data sources through artificial intelligence to suppress noise and enhance data.

[0011] (2) Intelligent extraction of terrain features: The terrain features of the collected and fused data sources are extracted. The backbone network uses the HRNet+ architecture to maintain high-resolution features. In the task branch, DeepLabv3+ is used for semantic segmentation to extract information such as water systems, vegetation, and artificial buildings. Mask R-CNN is used for instance segmentation to identify independent terrain units. SuperPoint is used for key point detection to locate feature markers. Slope calculation and curvature analysis are performed on the data source.

[0012] (3) 3D reconstruction optimization: The data source after feature extraction is added to the neural radiation field model to obtain a 3D reconstruction model based on spatial image features, and the error of the formed 3D reconstruction model is corrected;

[0013] (4) Dynamic terrain evolution modeling: Construct a multi-mechanism coupled intelligent evolution model that integrates time and space, integrates physical constraint neural networks, multi-factor graph coupling systems, and adversarial generation prediction frameworks, and realizes multi-scale dynamic simulation and interactive inversion of terrain evolution through online learning and mixed reality deduction platforms to meet the needs of geological disaster warning and long-term strategic planning.

[0014] Furthermore, the specific method of using multiple devices to collect data sources for the terrain to be surveyed in step (1) includes deploying a stereo observation network, integrating satellite multispectral images, drone LiDAR point clouds, oblique photography data from ground mobile surveying and mapping systems, and geological radar profiles to construct a multi-scale data set from centimeters to kilometers.

[0015] Furthermore, the multimodal data registration in step (1) includes adopting a Transformer-based cross-modal registration network to solve the spatiotemporal reference difference through a deformable attention mechanism to achieve high-precision spatial matching of optical images and point cloud data.

[0016] Furthermore, the specific method of suppressing noise and enhancing data in step (1) includes constructing a noise-aware generative adversarial network, separating terrain signals from sensor noise through a spectral normalization discriminator, using the StyleGAN2-ADA algorithm to perform texture enhancement on low-quality areas, and improving the resolution of weak feature areas by more than 4 times. At the same time, a lightweight edge computing module is developed to fuse multi-source data streams in real time and output a standardized terrain feature tensor to support end-to-end feature transfer for subsequent terrain extraction tasks.

[0017] Furthermore, the HRNet+ high-resolution feature preservation in step (2) adopts an improved high-resolution network to maintain spatial accuracy through multi-resolution parallel processing and enhanced feature fusion.

[0018] Further, the multi-resolution parallel processing constructs a resolution sequence with four branches through the formula \(\{(H / 2^{i - 1},W / 2^{i - 1})\}\), where the input image size is \(H\times W\), and calculates the feature exchange executed in each stage through the formula where \(\) is the cross-resolution feature transformation operator, which includes 1×1 convolution and bilinear interpolation, and \(i = 1, 2, 3, 4\). is the cross-resolution feature transformation operator, including 1×1 convolution and bilinear interpolation, \(i = 1, 2, 3, 4\).

[0019] Further, the enhanced feature fusion introduces a coordinate attention module (CA) to optimize feature selection. The processing result of the coordinate attention module on the feature map is calculated through the formula

[0020] where GAP is global average pooling, GMP is global max pooling, and \(\sigma\) is the Sigmoid activation.

[0021] Further, the specific method for 3D model reconstruction and optimization in step (3) is to use NeRF modeling with feature enhancement, fuse feature encoding and spatial coordinates, predict density and color through MLP, generate a predicted image using differentiable rendering, jointly optimize reconstruction and feature consistency, iteratively optimize model details based on geometric constraints and multi-view Figure 1 consistency, prune the low-density regions at the same time to reduce floating noise, convert the point cloud output by NeRF into an explicit mesh, fill in the missing patches, enforce the optical flow of the rendering results of adjacent views to be consistent with the real offset, introduce a discriminator network to distinguish the high-frequency details between real and rendered images

[0022] Further, the dynamic terrain evolution modeling in step (4) specifically includes using a spatio-temporal tensor data structure and a Transformer architecture to process historical and real-time data, capture terrain changes at different time scales, combine numerical simulation and a data-driven dual-branch network, ensure that the prediction conforms to physical laws through a differentiable physics layer, use multi-agent simulation and graph neural network to process the interactions of factors such as erosion and sedimentation, perform multi-resolution prediction using GAN, and update the model online learning, and develop a visualization platform to support various simulations and real-time adjustments.

[0023] Further, an artificial intelligence-based terrain mapping system for implementing the artificial intelligence-based terrain mapping method, the system includes:

[0024] A data acquisition and preprocessing module, including a multi-sensor control unit, a cross-modal registration engine, a noise suppression enhancer, and an edge computing node, for multi-source heterogeneous data acquisition and intelligent fusion;

[0025] The intelligent terrain feature extraction module includes a high-resolution feature backbone, a semantic segmentation unit, an instance segmentation unit, a key point detection unit, and a terrain parameter calculator, and is used for multi-dimensional terrain element analysis;

[0026] The 3D reconstruction optimization module includes a neural radiance field core, an error correction unit, and an LOD generator, and is used for high-precision 3D modeling and correction;

[0027] The dynamic terrain evolution modeling module includes a physical-data dual-driven engine, a multi-factor coupling system, an adversarial prediction framework, and an online learning interface, and is used for spatio-temporal coupled terrain evolution prediction;

[0028] The interactive visualization and decision-making platform includes a mixed reality engine, a dynamic deduction sand table, and intelligent decision-making assistance, and is used for multi-dimensional data fusion display and deduction;

[0029] The data acquisition and preprocessing module, the intelligent terrain feature extraction module, the 3D reconstruction optimization module, the dynamic terrain evolution modeling module, and the interactive visualization and decision-making platform are connected in sequence, and each module communicates through a gRPC interface and supports independent upgrade.

[0030] In summary, the present application includes at least one of the following beneficial technical effects:

[0031] Through the cross-modal intelligent registration and noise suppression mechanism, this method breaks through the limitations of traditional single data sources. At the same time, based on the spatial-spectral joint registration framework of deep features, it effectively overcomes the spatio-temporal reference differences and sensor noise interference of multi-source data, and significantly improves the data consistency of complex terrains. The dynamic adaptive enhancement of this method uses the spectral domain filtering and texture reconstruction technology of the generative adversarial network to achieve autonomous optimization of data quality, especially suitable for feature restoration in harsh acquisition environments such as cloud cover and multiple echoes. The system realizes the holographic analysis of terrain elements through a multi-task collaborative deep learning architecture. The HRNet+ backbone network retains sub-meter terrain details (such as water system tributaries and engineering structure joints) to the greatest extent through multi-scale parallel convolution and feature recombination mechanisms, overcoming the problem of small target information loss caused by downsampling in traditional networks.

[0032] This system realizes physical-data joint drive, embeds terrain parameter calculation (slope / curvature) into the feature extraction process, enhances the geoscientific interpretability of the model through physical prior constraints, and improves the recognition reliability of steep change regions and complex geometric structures. Through the gradient interaction between the numerical simulation branch and the data-driven branch, the deep integration of physical laws and observational data is realized, ensuring that the prediction results not only conform to the principles of geomechanics but also can adapt to the evolution of complex environments. At the same time, based on the model update architecture of incremental learning, new sensor data and on-site monitoring feedback can be dynamically fused to continuously improve the generalization performance of the system in unknown terrain scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A method flow chart of a terrain surveying and mapping method based on artificial intelligence in the present invention;

[0034] Figure 2 The system block diagram of a terrain surveying and mapping system based on artificial intelligence in the present invention. DETAILED DESCRIPTION

[0035] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application may be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0036] It should be noted that, unless otherwise specified, all technical and scientific terms used in this application have the same meanings as commonly understood by ordinary technicians in the technical field to which this application belongs.

[0037] According to one embodiment of the present invention, Figure 1 As shown, a terrain surveying and mapping method based on artificial intelligence comprises the following steps:

[0038] (1) Multi-source data fusion acquisition: Use a variety of devices to collect data sources of the terrain to be surveyed and mapped, and use artificial intelligence to perform multi-modal data registration on the collected data sources to suppress noise and enhance data;

[0039] (2) Intelligent extraction of terrain features: The terrain features of the collected and fused data sources are extracted. The backbone network uses the HRNet+ architecture to maintain high-resolution features. In the task branch, DeepLabv3+ is used for semantic segmentation to extract information such as water systems, vegetation, and artificial buildings. Mask R-CNN is used for instance segmentation to identify independent terrain units. SuperPoint is used for key point detection to locate feature markers. Slope calculation and curvature analysis are performed on the data source.

[0040] (3) 3D reconstruction optimization: The data source after feature extraction is added to the neural radiation field model to obtain a 3D reconstruction model based on spatial image features, and the error of the formed 3D reconstruction model is corrected;

[0041] (4) Dynamic terrain evolution modeling: Construct a multi-mechanism coupled intelligent evolution model that integrates time and space, integrates physical constraint neural networks, multi-factor graph coupling systems, and adversarial generation prediction frameworks, and realizes multi-scale dynamic simulation and interactive inversion of terrain evolution through online learning and mixed reality deduction platforms to meet the needs of geological disaster warning and long-term strategic planning.

[0042] The specific method of collecting data sources of the terrain to be surveyed and mapped in step (1) by using multiple devices includes deploying a stereo observation network, integrating satellite multispectral images, UAV LiDAR point clouds, oblique photography data of a ground mobile mapping system, and geological radar profiles to construct a multi-scale data set from centimeter level to kilometer level; the multi-modal data registration includes using a Transformer-based cross-modal registration network to solve the spatio-temporal reference difference through a deformable attention mechanism to achieve high-precision spatial matching of optical images and point cloud data; the specific method of suppressing noise and enhancing data includes constructing a noise-aware generative adversarial network, separating terrain signals and sensor noise through a spectral normalization discriminator, using the StyleGAN2-ADA algorithm to enhance the texture of low-quality regions, improving the resolution of weak feature regions by more than 4 times, and at the same time developing a lightweight edge computing module to fuse multi-source data streams in real time and output a standardized terrain feature tensor to support end-to-end feature transfer of subsequent terrain extraction tasks.

[0043] In step (2), the HRNet + high-resolution feature preservation uses an improved high-resolution network to maintain spatial accuracy through multi-resolution parallel processing and enhanced feature fusion; the multi-resolution parallel processing constructs a resolution sequence of four branches through the formula \(\{(H / 2^{i - 1},W / 2^{i - 1})\}\), where the input image size is \(H\times W\), and calculates the feature exchange executed in each stage through the formula where \(\mathcal{F}_{i\rightarrow j}\) is a cross-resolution feature transformation operator, including 1×1 convolution and bilinear interpolation, and \(i = 1,2,3,4\); the enhanced feature fusion introduces a coordinate attention module (CA) to optimize feature selection, and the processing result of the coordinate attention module on the feature map is calculated through the formula where GAP is global average pooling, GMP is global maximum pooling, and \(\sigma\) is the Sigmoid activation.

[0044]

[0045] Figure 1 In step (3), the specific method of three-dimensional model reconstruction and optimization is to use NeRF modeling with feature enhancement, fuse feature encoding and spatial coordinates, predict density and color through an MLP, generate a predicted image using differentiable rendering, jointly optimize reconstruction and feature consistency, and iteratively optimize model details based on geometric constraints and multi-view consistency. At the same time, prune the low-density regions to reduce floating noise, convert the point cloud output by NeRF into an explicit mesh, fill in the missing patches, force the optical flow of the rendering results of adjacent views to be consistent with the real offset, and introduce a discriminator network to distinguish the high-frequency details between real and rendered images.

[0046]

[0046] The dynamic terrain evolution modeling in step (4) specifically includes using a spatio-temporal tensor data structure and a Transformer architecture to process historical and real-time data, capturing terrain changes at different time scales, combining numerical simulation and a data-driven dual-branch network, ensuring that predictions conform to physical laws through a differentiable physics layer, using multi-agent simulation and graph neural networks to handle the interactions of factors such as erosion and settlement, performing multi-resolution predictions using GAN, and online learning to update the model, and developing a visualization platform to support various simulations and real-time adjustments.

[0047] According to another embodiment of the present invention, as Figure 2 shown, an artificial intelligence-based terrain mapping system for implementing an artificial intelligence-based terrain mapping method, the system includes a data acquisition and preprocessing module 1 for multi-source heterogeneous data acquisition and intelligent fusion, a terrain feature intelligent extraction module 2 for multi-dimensional terrain element analysis, a three-dimensional reconstruction optimization module 3 for high-precision three-dimensional modeling and correction, a dynamic terrain evolution modeling module 4 for spatio-temporal coupled terrain evolution prediction, and an interactive visualization and decision-making platform 5 for multi-dimensional data fusion display and deduction. Each module is connected in sequence and communicates through a gRPC interface, supporting independent upgrades.

[0048] The data acquisition and preprocessing module 1 includes:

[0049] The multi-sensor control unit 11 includes integrated devices such as satellites, UAV LiDAR, ground MMS, and ground-penetrating radar, and realizes hardware synchronization through the ROS robot operating system, with a timestamp alignment accuracy < 1 ms.

[0050] The cross-modal registration engine 12 is based on the Transformer-based CMR-Net network, supports the registration of optical images - LiDAR - geological section data, and the registration error RMSE < 0.3 pixels.

[0051] The noise suppression enhancer 13 uses a NA-GAN adversarial network to denoise LiDAR point clouds and remove clouds from images.

[0052] The edge computing node 14 deploys a TensorRT-accelerated real-time fusion pipeline through NVIDIA Jetson AGX Xavier and outputs a standardized terrain tensor.

[0053] The terrain feature intelligent extraction module 2 includes:

[0054] The high-resolution feature backbone 21 is based on an improved HRNet+ architecture, outputs a feature map with a resolution of 1 / 4 of the original image, and supports a 4-level feature pyramid.

[0055] The semantic segmentation unit 22 is based on the improved DeepLabv3+ and integrates a terrain prior knowledge base, with the water system boundary recognition accuracy reaching 0.5m.

[0056] The instance segmentation unit 23 is based on the terrain-adapted version of Mask R-CNN and supports the instance segmentation of landslide boundaries.

[0057] The key point detection unit 24 is based on the optimized version of SuperPoint and integrates slope constraints.

[0058] The terrain parameter calculator 25 is based on CUDA-accelerated parallel gradient calculation and supports the second-level generation of slope / curvature of 10km 2 DEM.

[0059] The three-dimensional reconstruction optimization module 3 includes:

[0060] The neural radiance field core 31 is based on the improved NeRF architecture and integrates the HRNet+ features as the input of position encoding, with the training efficiency increased by 40%.

[0061] The error correction unit 32 is jointly optimized based on ICP (Iterative Closest Point) and GAN:

[0062]

[0063] Among them, CD is the Chamfer Distance, λ_photo = 1.0, and λ_geo = 0.6.

[0064] The LOD generator 33 automatically generates LOD1 to LOD4 multi-level of detail models, with a compression ratio of 20:1. Among them, the LOD4 model only requires 5MB / km 2 .

[0065] The dynamic terrain evolution modeling module 4 includes:

[0066] The physics-data dual-driven engine 41, the numerical simulation branch and the 3D ConvLSTM branch are coupled through differentiable layers to achieve gradient backpropagation.

[0067] The multi-factor coupling system 42, the graph attention network (GAT) models the interaction of erosion-settlement-tectonic movement, and the node update formula:

[0068]

[0069] Among them, α_ij is calculated by the attention mechanism.

[0070] The adversarial prediction framework 43 is based on the CycleGAN architecture to achieve multi-resolution prediction. The 1m resolution model is used for landslide warning, and the 10m model is used for urban planning.

[0071] The online learning interface 44 supports real-time update of RT-DInSAR data stream, and the model iteration period is < 2 hours.

[0072] The interactive visualization and decision-making platform 5 includes:

[0073] The mixed reality engine 51 is based on Unity3D + Holographic Remoting and supports real-time loading of TB-level terrain data on the Hololens 2 side.

[0074] The dynamic deduction sand table 52 integrates the Houdini physics engine and can simulate the influence of parameters such as rainfall and earthquake intensity.

[0075] The intelligent decision-making assistance 53 generates a risk heat map based on the SVM classifier, and the PSO algorithm inversely calculates the optimal engineering parameters, with a response delay < 200 ms.

[0076] In summary, the present invention constructs a terrain intelligent mapping system with the ability of autonomous evolution through a technical closed-loop of multi-modal data fusion, multi-granularity feature analysis, and multi-mechanism coupling modeling. This method and system not only reflect the breakthrough of a single technology, but more importantly, establish a "data-model-decision" collaborative enhancement paradigm, providing an extensible technical base for emerging fields such as digital twin geology and smart cities.

[0077] Taking the above-mentioned ideal embodiments of the present invention as an inspiration, through the above description, relevant staff can completely make various changes and modifications without departing from the technical idea of this invention. The technical scope of this invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. An artificial intelligence-based topographic surveying method, characterized in that: The following steps are involved: (1) Multi-source data fusion acquisition: Use a variety of devices to collect data sources of the terrain to be surveyed and mapped, and use artificial intelligence to perform multi-modal data registration on the collected data sources to suppress noise and enhance data; (2) Intelligent extraction of terrain features: The terrain features of the collected and fused data sources are extracted. The backbone network uses the HRNet+ architecture to maintain high-resolution features. In the task branch, DeepLabv3+ is used for semantic segmentation to extract information such as water systems, vegetation, and artificial buildings. Mask R-CNN is used for instance segmentation to identify independent terrain units. SuperPoint is used for key point detection to locate feature markers. Slope calculation and curvature analysis are performed on the data source. (3) 3D reconstruction optimization: The data source after feature extraction is added to the neural radiation field model to obtain a 3D reconstruction model based on spatial image features, and the error of the formed 3D reconstruction model is corrected; (4) Dynamic terrain evolution modeling: Construct a multi-mechanism coupled intelligent evolution model that integrates time and space, integrates physical constraint neural networks, multi-factor graph coupling systems, and adversarial generation prediction frameworks, and realizes multi-scale dynamic simulation and interactive inversion of terrain evolution through online learning and mixed reality deduction platforms to meet the needs of geological disaster warning and long-term strategic planning.

2. The topographic mapping method based on artificial intelligence according to claim 1, characterized in that: The specific method of using multiple devices to collect data sources for the terrain to be surveyed in step (1) includes deploying a stereo observation network, integrating satellite multispectral images, UAV LiDAR point clouds, oblique photography data from ground mobile surveying and mapping systems, and geological radar profiles to construct a multi-scale data set from centimeters to kilometers.

3. The topographic mapping method based on artificial intelligence according to claim 1, characterized in that: The multimodal data registration in step (1) includes adopting a Transformer-based cross-modal registration network, solving the spatiotemporal reference difference through a deformable attention mechanism, and achieving high-precision spatial matching of optical images and point cloud data.

4. The topographic mapping method based on artificial intelligence according to claim 1, characterized in that: The specific method of suppressing noise and enhancing data in step (1) includes constructing a noise-aware generative adversarial network, separating terrain signals from sensor noise through a spectral normalization discriminator, and using the StyleGAN2-ADA algorithm to perform texture enhancement on low-quality areas, thereby increasing the resolution of weak feature areas by more than 4 times. At the same time, a lightweight edge computing module is developed to fuse multi-source data streams in real time and output a standardized terrain feature tensor to support end-to-end feature transfer for subsequent terrain extraction tasks.

5. The topographic mapping method based on artificial intelligence according to claim 4, characterized in that: The HRNet+ high-resolution feature preservation in step (2) adopts an improved high-resolution network and maintains spatial accuracy through multi-resolution parallel processing and enhanced feature fusion.

6. The topographic mapping method based on artificial intelligence according to claim 5, characterized in that: The multi-resolution parallel processing constructs a resolution sequence with four branches through the formula \(\{(H / 2^{i - 1},W / 2^{i - 1})\}\), where the input image size is \(H\times W\), and calculates the feature exchange performed at each stage through the formula where \(\lt0000019\gt\) is a cross-resolution feature transformation operator including \(1\times1\) convolution and bilinear interpolation, and \(i = 1, 2, 3, 4\). is a cross-resolution feature transformation operator, including \(1\times1\) convolution and bilinear interpolation, \(i = 1, 2, 3, 4\).

7. The topographic surveying method based on artificial intelligence according to claim 5, characterized in that: The enhanced feature fusion introduces a coordinate attention module (CA) to optimize feature selection. The processing result of the coordinate attention module on the feature map is obtained through the formula where GAP is global average pooling, GMP is global max pooling, and σ is the Sigmoid activation.

8. The topographic mapping method based on artificial intelligence according to claim 2, characterized in that: The specific method for 3D model reconstruction and optimization in step (3) is to use NeRF modeling with feature enhancement, fuse feature encoding with spatial coordinates, predict density and color through MLP, generate predicted images using differentiable rendering, jointly optimize reconstruction and feature consistency, iteratively optimize model details based on geometric constraints and multi-view consistency. At the same time, prune low-density regions to reduce floating noise, convert the point cloud output by NeRF into an explicit mesh, complete missing patches, enforce the optical flow of the rendering results of adjacent views to be consistent with the real offset, introduce a discriminator network to distinguish high-frequency details between real and rendered images.

9. The topographic mapping method based on artificial intelligence according to claim 8, characterized in that: The dynamic terrain evolution modeling in step (4) specifically includes using a spatio-temporal tensor data structure and a Transformer architecture to process historical and real-time data, capture terrain changes at different time scales, combine numerical simulation and a data-driven dual-branch network, ensure that predictions conform to physical laws through a differentiable physics layer, use multi-agent simulation and graph neural networks to handle the interactions of factors such as erosion and settlement, perform multi-resolution predictions using GAN, and update the model through online learning, and develop a visualization platform to support various simulations and real-time adjustments.

10. An artificial intelligence-based topographic mapping system for implementing the artificial intelligence-based topographic mapping method according to any one of claims 1-9, characterized in that, The system includes: A data acquisition and preprocessing module, including a multi-sensor control unit, a cross-modal registration engine, a noise suppression enhancer, and an edge computing node, for multi-source heterogeneous data acquisition and intelligent fusion; A terrain feature intelligent extraction module, including a high-resolution feature backbone, a semantic segmentation unit, an instance segmentation unit, a key point detection unit, and a terrain parameter calculator, for multi-dimensional terrain element analysis; A 3D reconstruction and optimization module, including a neural radiance field core, an error correction unit, and an LOD generator, for high-precision 3D modeling and correction; A dynamic terrain evolution modeling module, including a physics-data dual-driven engine, a multi-factor coupling system, an adversarial prediction framework, and an online learning interface, for spatio-temporal coupled terrain evolution prediction; An interactive visualization and decision-making platform, including a mixed reality engine, a dynamic deduction sand table, and intelligent decision-making assistance, for multi-dimensional data fusion display and deduction; The data acquisition and preprocessing module, the terrain feature intelligent extraction module, the 3D reconstruction and optimization module, the dynamic terrain evolution modeling module, and the interactive visualization and decision-making platform are connected in sequence, and each module communicates through a gRPC interface, supporting independent upgrades.

Images