Spatiotemporal AI Digital Base 2D / 3D Integrated Modeling Method and System

Abstract

This invention belongs to the field of digital modeling technology and discloses a spatiotemporal AI digital base integrated 2D and 3D modeling method and system. It includes constructing a five-dimensional spatial hierarchy model and a multi-scale data pyramid to determine the basic visible data range and predicted line-of-sight range, obtaining a pre-loaded candidate tile set, and constructing an intelligent dynamic scheduling mechanism. It acquires the geometric complexity metric and camera distance parameters of the data tiles to be rendered, determines the comprehensive LOD switching threshold of the data tiles to be rendered, and constructs an adaptive LOD transition mechanism. Under a preset coordinate system, it performs spatiotemporal fusion 2D and 3D integrated rendering modeling of physical entities and observation data. The 2D vector elements of this invention can dynamically fit into the 3D terrain, achieving seamless overlay of 2D planar maps and 3D scenes. This satisfies both the needs of professional surveyors for planar accuracy and the needs of decision-makers for 3D situational awareness, greatly improving the maintainability and scalability of the system.

Description

Technical Field

[0001] This invention relates to the field of digital modeling technology, and more specifically, to a spatiotemporal AI digital base integrated two-dimensional and three-dimensional modeling method and system. Background Technology

[0002] With the rapid development of smart cities, digital twin cities, and the demand for comprehensive spatial governance, traditional geographic information systems face numerous technical bottlenecks in dealing with the visualization and interactive analysis of massive, multi-source, and heterogeneous spatial data.

[0003] First, existing GIS platforms typically use fixed-scale 2D map services or independent 3D scene rendering, resulting in a severe data gap between macroscopic maps and detailed microscopic models. When users zoom continuously or switch perspectives, they often encounter data discontinuities, model jumps, or information overload, making it difficult to achieve an immersive, continuous browsing experience from tens of thousands of meters above the ground to the interior of facilities down to the centimeter level.

[0004] Secondly, when faced with massive spatial data at the city level or even the entire region level (including vector elements, remote sensing images, elevation and terrain, UAV oblique photogrammetry models and real-time monitoring data), the traditional full-load mode results in excessive memory usage and network bandwidth pressure, making it difficult for the browser to load and render smoothly, which seriously restricts the real-time interactive performance in big data scenarios.

[0005] Furthermore, the spatial registration accuracy between the UAV oblique photogrammetry model and the terrain base map is insufficient, often resulting in the model appearing suspended or sunken underground, requiring extensive manual intervention for model repair and leading to low production efficiency. Simultaneously, there is a lack of effective dynamic bonding mechanisms between two-dimensional vector elements (such as land parcel boundaries, road networks, and water systems) and the three-dimensional terrain surface, making it difficult to achieve true integrated two-dimensional and three-dimensional visualization.

[0006] In addition, existing LOD (Level of Detail) switching technologies are mostly based on a single geometric complexity or distance threshold for judgment, lacking consideration of the semantic importance of ground features. This results in important facilities and secondary ground features being simplified indiscriminately during scaling, with serious loss of key information, and visual jumps and stuttering are common during level switching.

[0007] Finally, there is a lack of spatiotemporal correlation and fusion rendering mechanism between physical entities (such as buildings, plots of land, and monitoring stations) and dynamic monitoring data (such as real-time observation data of atmosphere, water quality, and soil). Static 3D models and dynamic monitoring data are in a state of separation, making it difficult to support the leap from "seeing" to "managing" and failing to fully reflect the core value of digital twin virtual-real integration.

[0008] In view of this, the present invention proposes a spatiotemporal AI digital base two-dimensional and three-dimensional integrated modeling method and system to solve the above problems. Summary of the Invention

[0009] To overcome the aforementioned deficiencies of the prior art and achieve the above objectives, the present invention provides the following technical solution: A spatiotemporal AI digital foundation integrated 2D / 3D modeling method, the method includes: We acquire multi-source heterogeneous spatial data and divide it into five spatial levels from macro to micro: global, regional, local, object, and observation data. We also configure scale thresholds, data accuracy standards, and semantic labeling systems for each level to construct a five-dimensional spatial hierarchy model. Furthermore, we construct multi-scale data pyramids for different types of multi-source heterogeneous spatial data. The system obtains the current real-time status parameters of the camera, determines the basic visible data range, obtains historical browsing trajectory data to determine the predicted line of sight range, and obtains a set of preloaded candidate tiles based on the basic visible data range and the predicted line of sight range. The system then performs priority sorting and preloading on the data tiles in the preloaded candidate tile set to build an intelligent dynamic scheduling mechanism. The data tile to be rendered is denoted as the data tile to be rendered. The geometric complexity metric and camera distance parameter of the data tile to be rendered are obtained. The comprehensive LOD switching threshold of the data tile to be rendered is determined. Based on the comprehensive LOD switching threshold of all data tiles to be rendered, an adaptive LOD transition mechanism is constructed. Under the preset coordinate system, based on the adaptive LOD transition mechanism and the intelligent dynamic scheduling mechanism, the spatiotemporal fusion of physical entities and observation data is performed to create a unified 2D and 3D rendering model.

[0010] Furthermore, multi-source heterogeneous spatial data includes five data types: vector data, image data, terrain data, and UAV oblique photogrammetry model data. Multi-scale data pyramids include vector pyramids, remote sensing image pyramids, elevation tile pyramids, and model pyramids. Based on vector data, convolutional neural networks are used to perform semantic segmentation and classification labeling on each vector element of the vector data to obtain land cover type labels. Basic importance scores are assigned to different land cover type labels according to preset business rules. The spatial distribution density, topological connectivity and user historical access frequency of vector elements are obtained. Combined with the basic importance scores, the semantic importance weight of each land cover type label is calculated. Based on the semantic importance weight, each vector element is differentiated, simplified, aggregated and symbolized to construct a vector pyramid. Based on image data, a remote sensing image pyramid containing multiple resolution levels was constructed using cloud-optimized GeoTIFF format; Based on the terrain data, the terrain elevation data is extracted from the terrain data, and the terrain elevation data is encoded into RGB image tiles using the Terrarium format to construct a multi-level elevation tile pyramid. Based on UAV oblique photogrammetry model data, the original OSGB format data is converted and lightweighted to generate a 3DTiles format model pyramid containing multiple levels of detail.

[0011] Furthermore, the method for determining the basic visible data range is as follows: Obtain the real-time state parameters of the current camera; construct a hexahedral view frustum composed of the near plane, far plane, left plane, right plane, top plane, and bottom plane based on the real-time state parameters, and calculate the standard plane equation for each plane; Traverse the axis-aligned bounding boxes of all data tiles to be loaded in the multi-scale data pyramid. For each axis-aligned bounding box, substitute the eight vertices into the six standard plane equations and perform sign judgment. If all vertices are located outside any plane, the data tile to be loaded is determined to be completely invisible and excluded; if all vertices are located inside all planes, the data tile to be loaded is determined to be completely visible; if some vertices are located inside the plane and some are located outside, the data tile to be loaded is determined to be partially visible. Calculate the Euclidean distance from the center point of the bounding box of each fully visible and partially visible data tile to be loaded to the camera position, and use it as the view distance value of the data tile to be loaded. Compare the view distance value with the maximum visible distance threshold of the corresponding level of the five-dimensional spatial hierarchy model, and remove the data tiles to be loaded whose view distance value is greater than the maximum visible distance threshold. The set of the remaining fully visible and partially visible data tiles to be loaded is determined as the basic visible data range.

[0012] Furthermore, the method for constructing an intelligent dynamic scheduling mechanism is as follows: Collect and preprocess user historical browsing trajectory data, and train a gaze prediction model based on a long short-term memory network using the preprocessed historical browsing trajectory data. The gaze prediction model takes the camera state sequence of M consecutive frames and the interaction context as input, and outputs the predicted camera state and the corresponding predicted gaze range within the next T millisecond time window. Calculate the difference between the predicted gaze range and the basic visible data range to obtain a preloaded candidate tile set. For each data tile in the preload candidate tile set, a comprehensive preload priority score is calculated. The data tiles are then sorted in descending order of the comprehensive preload priority score to obtain the preload queue, thereby constructing an intelligent dynamic scheduling mechanism.

[0013] Furthermore, the method for constructing an adaptive LOD transition mechanism is as follows: The range of the overall LOD switching threshold is divided into N intervals, each interval corresponding to a LOD rendering level. The smaller the interval value, the higher the precision of the corresponding LOD rendering level. The overall LOD switching threshold of each data tile to be rendered is calculated in real time. When the overall LOD switching threshold of the data tile to be rendered crosses from one interval to an adjacent interval, the LOD level switching is triggered. When LOD level switching is triggered, the transition method is selected according to the switching direction. When switching from a low-precision level to a high-precision level, geometric morphing is used for transition, and texture gradient transition is used when switching from a high-precision level to a low-precision level. The transition duration is dynamically set for each transition method, and the transition duration is inversely proportional to the current camera movement speed. When the current camera movement speed is stationary, the transition duration is set to the preset maximum value; in this way, an adaptive LOD transition mechanism is constructed.

[0014] Furthermore, the method for determining the overall LOD switching threshold of the data tiles to be rendered is as follows: Obtain the geometric information of the data tile to be rendered, calculate the geometric complexity metric and camera distance parameter. The geometric complexity metric is the ratio of the vertex density of the data tile to be rendered to the maximum allowed vertex density of the corresponding data tile level. The camera distance parameter is the ratio of the Euclidean distance from the center point of the axis-aligned bounding box of the data tile to the current camera position to the maximum visible distance threshold of the corresponding level of the current five-dimensional spatial hierarchy model. Obtain the semantic importance weight of the data tile to be rendered, preset corresponding weight coefficients for the semantic importance weight, geometric complexity measure and camera distance parameter respectively, multiply the semantic importance weight, geometric complexity measure and camera distance parameter by their respective weight coefficients and sum them to obtain the comprehensive LOD switching threshold of the data tile to be rendered.

[0015] Furthermore, the method for performing spatiotemporal fusion of physical entities and observation data into a unified 2D / 3D rendering model is as follows: Initialize a rendering environment with a unified preset coordinate system, establish a complete transformation pipeline from geographic coordinates to screen pixel coordinates, and uniformly transform multi-source heterogeneous spatial data and real-time monitoring observation data to the preset coordinate system for spatiotemporal alignment. The camera's current viewpoint is detected in real time. When the viewpoint is detected to be a two-dimensional top view, the terrain elevation is set to zero and two-dimensional vector tile data is loaded for planar map rendering. When the viewpoint is detected to be a three-dimensional tilt view, a non-linear mapping relationship is constructed to dynamically calculate the terrain elevation and the corresponding data tiles are loaded based on an intelligent dynamic scheduling mechanism. Based on the adaptive LOD transition mechanism, various data tiles in the multi-scale data pyramid are rendered in layers to obtain a multi-scale integrated 2D and 3D rendering scene. A spatiotemporal mapping table between physical entities and observation data is established, and spatiotemporal completion is performed on the real-time monitored observation data to obtain completed observation data. The completed observation data is then mapped to the surface or surrounding area of ​​the corresponding physical entity according to the spatiotemporal mapping table, and the corresponding rendering method is selected based on the data type to realize the spatiotemporal fusion rendering modeling of physical entities and observation data.

[0016] Furthermore, the method for dynamically calculating terrain lift height by constructing a nonlinear mapping relationship is as follows: The nonlinear mapping relationship is the nonlinear mapping relationship between terrain lift height, viewing angle tilt, and camera height, expressed as:

[0017] in, To raise the altitude of the terrain, The preset maximum terrain elevation. The visual sensitivity coefficient, For the angle of view tilt, For camera height, This is a decay function that decreases with camera height.

[0018] Furthermore, the method for obtaining complete observation data is as follows: Physical entities include UAV oblique photography models, land parcels, buildings, infrastructure, and monitoring stations. The spatiotemporal correlation mapping table records the unique identifier, spatial range, associated sensor ID, and the corresponding level of each physical entity's data update cycle. Establish a spatiotemporal index structure for the observation data, recording the acquisition timestamp, spatial coordinates, sensor type, data confidence level, and unique identifier of the associated physical entity for each observation data point; In response to the user's preset timeline selection command, the system determines the target time specified by the user, retrieves the observation data for the target time based on the spatiotemporal index structure, performs spatiotemporal completion of the missing observation data for the target time using the Kriging interpolation algorithm, and renders the completed observation data as a particle system, dynamic heat map, or time series label on the surface of the 3D model of the corresponding associated physical entity level, and adjusts the rendering transparency and color intensity according to the data confidence level.

[0019] The spatiotemporal AI digital foundation 2D / 3D integrated modeling system includes: The spatial data organization module is used to acquire multi-source heterogeneous spatial data, divide the multi-source heterogeneous spatial data into five spatial levels from macro to micro: global, regional, local, object, and observation data level, and configure scale thresholds, data accuracy standards and semantic labeling systems for each level to construct a five-dimensional spatial hierarchy model; and construct multi-scale data pyramids for different types of multi-source heterogeneous spatial data. The scheduling mechanism construction module is used to obtain the current real-time status parameters of the camera, determine the basic visible data range, obtain historical browsing trajectory data to determine the predicted line of sight range, obtain a set of preloaded candidate tiles based on the basic visible data range and the predicted line of sight range, perform priority sorting and preloading on the data tiles in the preloaded candidate tile set, and build an intelligent dynamic scheduling mechanism. The transition mechanism construction module is used to record the data tiles to be rendered as data tiles to be rendered, obtain the geometric complexity metric and camera distance parameters of the data tiles to be rendered, determine the comprehensive LOD switching threshold of the data tiles to be rendered, and construct an adaptive LOD transition mechanism based on the comprehensive LOD switching threshold of all data tiles to be rendered. The rendering and modeling module is used to perform spatiotemporal fusion two-dimensional and three-dimensional integrated rendering and modeling of physical entities and observation data in a preset coordinate system, based on an adaptive LOD transition mechanism and an intelligent dynamic scheduling mechanism.

[0020] The technical effects and advantages of the spatiotemporal AI digital base integrated 2D / 3D modeling method and system of this invention are as follows: This invention breaks the problem of separation between macro maps and micro models in traditional GIS by defining a five-dimensional spatial hierarchy. Users can smoothly zoom from 100,000 meters above the ground to the interior of facilities at the centimeter level. The visual experience is consistent and adopts LOD multi-level detail transition technology and vector pyramid technology, which effectively avoids the "jump" phenomenon of sudden disappearance or deformation of ground objects during zooming, and ensures the smoothness of interaction.

[0021] This invention employs a dynamic scheduling mechanism based on view frustum clipping and view distance judgment to ensure that the system only loads data that is "visible" and "needs to be seen," greatly reducing GPU memory usage and network bandwidth pressure. This makes it possible to smoothly load massive amounts of city-level data on the browser side. By utilizing GPU instantiation technology to process large-scale drone models and vector elements, the rendering frame rate is significantly improved, enabling the display of high-concurrency, large-data-volume scenes. Furthermore, by combining a view prediction model based on LSTM for preloading, it achieves an intelligent scheduling effect that becomes "smoother the more you use it."

[0022] This invention utilizes ground control points to accurately register the oblique photogrammetry model, solving the problem of mismatch between the 3D model and the base map. For issues such as models being suspended in mid-air or embedded underground, an automatic correction algorithm based on matrix transformation and terrain sampling is employed to significantly reduce the cost of manual model repair and improve production efficiency. Furthermore, by combining the static 3D model with dynamic monitoring data (atmosphere, water, soil, etc.), it achieves a leap from "seeing" to "managing," truly demonstrating the value of digital twins.

[0023] The two-dimensional vector elements of this invention can dynamically fit the three-dimensional terrain, achieving seamless overlay of 2D planar maps and 3D scenes. This satisfies both the needs of professional surveyors for planar accuracy and the needs of decision-makers for three-dimensional situational awareness. By controlling the rendering style through JSON configuration files, the system can adjust the map style or business rules without recompiling the code, greatly improving the maintainability and scalability of the system. Attached Figure Description

[0024] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system module block diagram of the present invention; Figure 3 This is an overall framework diagram of the present invention. Detailed Implementation

[0025] 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.

[0026] In one embodiment, a method for integrated 2D and 3D modeling of a spatiotemporal AI digital base is disclosed, such as... Figure 1 As shown, the method includes: We acquire multi-source heterogeneous spatial data and divide it into five spatial levels from macro to micro: global, regional, local, object, and observation data. We also configure scale thresholds, data accuracy standards, and semantic labeling systems for each level to construct a five-dimensional spatial hierarchy model. Furthermore, we construct multi-scale data pyramids for different types of multi-source heterogeneous spatial data.

[0027] Specifically, the overall spatial map covers the entire area of ​​the base, with a scale of approximately 1:500,000 to 1:200,000, primarily showcasing the regional topography, major water systems, backbone road networks, administrative boundaries, and macro-ecological patterns for overall situational awareness; the regional spatial map mainly covers natural areas and administrative units, with a scale of 1:100,000 to 1:50,000, used to display detailed provincial, municipal, and county boundaries; the local spatial map mainly covers functional zones and township boundaries, with a scale of 1:25,000 to 1:10,000, presenting contour lines. The first level of space includes functional area boundaries, small watershed ditches, and nature reserve patches, which facilitates refined management. The second level of space targets individual plots, buildings, monitoring stations, and other independent entities, with a scale greater than 1:5000, to clearly identify the outlines, area, and internal structure of land features, and to support attribute queries and interactions. The third level of space overlays real-time or historical observation data (such as atmosphere, water, vegetation, soil, and rocks) on top of the object level, and dynamically presents them in the form of point elements, charts, and labels, realizing the integration of physical entities and numerical data.

[0028] Multi-source heterogeneous spatial data includes five data types: vector data, image data, terrain data, and UAV oblique photogrammetry model data. Correspondingly, multi-scale data pyramids are constructed, including vector pyramids, remote sensing image pyramids, elevation tile pyramids, and model pyramids. In the construction of the vector pyramid, convolutional neural networks are used to perform semantic segmentation and classification labeling on each vector element of the vector data, obtaining land feature type labels (such as buildings, roads, water systems, land parcels, nature reserves, etc.). Based on preset business rules (pre-set according to the experience of personnel in this field and actual needs), basic importance scores are assigned to different land feature type labels. For example, the basic scores for important facilities, backbone road networks, and nature reserves are set higher than those for ordinary land parcels and secondary roads. Further, the spatial distribution density, topological connectivity, and user historical access frequency of vector elements are obtained. Combined with the basic importance scores, the semantic importance weight of each land feature type label is calculated. Spatial distribution density reflects the degree of aggregation of a certain type of land feature in geographic space. Higher density usually indicates more frequent human activity or more important ecological functions in the area. It can be determined based on the vector element type (point, line, area). Using multi-scale data pyramid tiles or preset grids as statistical units, the number of point features, total length of line features, or total area of ​​polygon features within a unit area are calculated, and then linearly normalized to the [0,1] interval. Topological connectivity reflects the pivotal position and connectivity of a feature in the spatial network. The higher the connectivity, the more critical the feature is to the operation of the entire network. For linear network vector features such as road networks and water systems, the betweenness centrality is calculated by converting them into graph structures. For polygon features such as land parcels and buildings, the number of other directly adjacent polygon features is counted, and the calculation results are normalized to the [0,1] interval to obtain the corresponding topological connectivity. The user's historical access frequency reflects the actual usage demand of the corresponding vector feature. The higher the access frequency, the higher the user's attention to the feature. It can be obtained by using a cold start default value combined with a dynamic iterative update mechanism during runtime. When the system builds the data model for the first time and there are no historical user behavior logs, a standardized initial access frequency value is assigned to different land feature types according to preset business rules. For example, the initial value for important infrastructure, medical and educational institutions, and government agencies is set to 0.8, the initial value for residential areas and commercial areas is set to 0.5, and the initial value for undeveloped land and ordinary forest land is set to 0.2. After the system is put into operation, the number of clicks, dwell time, number of times zoomed to the element level, and number of queries within a specified time window are extracted from the system user behavior logs. After weighted summation and application of a time decay function, the data is normalized to the [0,1] interval to obtain the real-time access frequency, which is automatically updated every preset period, gradually replacing the initial default value. Then, based on the experience of those in the field, a weight is preset for spatial distribution density, topological connectivity, and user historical access frequency. After weighted accumulation, the semantic importance weight is calculated by averaging the basic importance scores. The higher the semantic importance weight, the higher the priority of the vector element in the business system, and the more computing and rendering resources will be allocated to it. Finally, based on the semantic importance weight, the vector elements are differentiated, simplified, aggregated, and symbolized: vector elements with high semantic importance weight retain more geometric details and attribute fields, while vector elements with low semantic importance weight undergo a higher degree of aggregation and downsampling, thereby constructing a vector pyramid that balances visual clarity and rendering efficiency.

[0029] In the process of constructing the remote sensing image pyramid, the image data is organized using the Cloud Optimized GeoTIFF (COG) format to construct a pyramid structure containing multiple resolution levels. Each resolution level matches the scale of the corresponding spatial level to ensure fast loading and smooth transition. During the construction of the elevation tile pyramid, the terrain elevation data is extracted from the terrain data and encoded into RGB image tiles using the Terrarium format to achieve multi-level hierarchical storage and support real-time terrain rendering and mountain shadow analysis. During the construction of the model pyramid, the original OSGB format UAV oblique photogrammetry model data is converted and lightweighted to generate 3DTiles format data containing multiple levels of detail (LOD). The model point data is then encapsulated in a JSON file for direct rendering by the front end.

[0030] The system acquires the current real-time status parameters of the camera to determine the basic visible data range, acquires historical browsing trajectory data to determine the predicted line-of-sight range, and obtains a set of preloaded candidate tiles based on the basic visible data range and the predicted line-of-sight range. It then performs priority sorting and preloading on the data tiles in the preloaded candidate tile set to build an intelligent dynamic scheduling mechanism.

[0031] Specifically, the first step is to obtain the real-time state parameters of the current camera (including position, attitude angle, field of view angle, near and far clipping plane distance, etc.) based on the real-time state parameters. Then, a hexahedral view frustum consisting of the near plane, far plane, left plane, right plane, top plane, and bottom plane is constructed based on these parameters. The standard plane equation for each plane is then calculated. The calculation of the standard plane equation is a well-known existing technique in computer graphics. Its core is based on the camera's view matrix and projection matrix. By extracting each row of the projection matrix and performing normalization, the standard plane equation Ax + By + Cz + D = 0 for the six clipping planes of the view frustum is directly derived. This will not be described in detail here. The next step involves traversing the axis-aligned bounding boxes of all data tiles to be loaded in the multi-scale data pyramid. For each axis-aligned bounding box, the eight vertices are substituted into six standard plane equations for sign determination: if all vertices are located outside any plane, the data tile to be loaded is determined to be completely invisible and excluded; if all vertices are located inside all planes, the data tile to be loaded is determined to be completely visible; if some vertices are located inside the plane and some are located outside, the data tile to be loaded is determined to be partially visible. Further, the Euclidean distance from the center point of the bounding box of each fully visible and partially visible data tile to the camera position is calculated as the view distance value of the data tile to be loaded. The view distance value is compared with the maximum visible distance threshold of the corresponding level in the five-dimensional spatial hierarchy model. Data tiles to be loaded with view distance values ​​greater than the maximum visible distance threshold are removed, and the set of the remaining fully visible and partially visible data tiles to be loaded is determined as the basic visible data range.

[0032] In the process of determining the predicted field of view, historical browsing trajectory data of users is collected and preprocessed (including noise reduction, normalization, trajectory segmentation, etc.). The preprocessed historical browsing trajectory data is then used to train a Long Short-Term Memory (LSTM)-based network. This paper describes a gaze prediction model based on LSTM (Laser-Sensitive Timing). The model takes a sequence of camera states across M consecutive frames and interaction contexts (such as zooming, click events, and dwell time) as input, outputting the predicted camera state and corresponding predicted gaze range within a future T-millisecond time window. It calculates the difference between the predicted gaze range and the basic visible data range to obtain a preloaded candidate tile set, filtering out data tiles that are predicted to be seen but are not currently visible. For each data tile in the preloaded candidate tile set, a comprehensive preload priority score is calculated. The score is based on the distance of the data tile from the predicted gaze center, semantic importance weight, and data update frequency. This score is weighted, accumulated, and normalized to obtain the comprehensive preload priority score for each data tile. The data tiles are then sorted according to their comprehensive preload priority scores from highest to lowest to form a preload queue, thus constructing an intelligent dynamic scheduling mechanism. Simultaneously, combining GPU instantiation and asynchronous loading techniques, tiles within the basic visible data range are loaded synchronously, while tiles in the preload queue are loaded asynchronously in the background, thereby eliminating rendering stutters and improving browsing smoothness.

[0033] The data tile to be rendered is denoted as the data tile to be rendered. The geometric complexity metric and camera distance parameter of the data tile to be rendered are obtained. The comprehensive LOD switching threshold of the data tile to be rendered is determined. Based on the comprehensive LOD switching threshold of all data tiles to be rendered, an adaptive LOD transition mechanism is constructed.

[0034] Specifically, the method for determining the overall LOD switching threshold of the data tile to be rendered is as follows: Obtain the geometric information of the data tile to be rendered, calculate the geometric complexity metric and camera distance parameter, where the geometric complexity metric is the ratio of the vertex density of the data tile to the maximum allowed vertex density of the corresponding layer; the camera distance parameter is the ratio of the Euclidean distance from the center point of the axis-aligned bounding box of the data tile to the current camera position to the maximum visible distance threshold of the corresponding layer of the current five-dimensional spatial hierarchy model; obtain the semantic importance weight of the data tile to be rendered, and preset corresponding weight coefficients (e.g., ...) for the semantic importance weight, geometric complexity metric, and camera distance parameter. , , And satisfy The three factors are multiplied by their respective weight coefficients and then summed to obtain the overall LOD switching threshold of the data tile to be rendered.

[0035] In constructing the adaptive LOD transition mechanism, the range of the comprehensive LOD switching threshold is divided into N intervals, each interval corresponding to a LOD rendering level. The smaller the interval value, the higher the precision of the corresponding LOD rendering level. The comprehensive LOD switching threshold of each data tile to be rendered is calculated in real time. When the comprehensive LOD switching threshold of the data tile to be rendered crosses from one interval to an adjacent interval, the LOD level switching is triggered. When the LOD level switching is triggered, the transition method is selected according to the switching direction: when switching from a low-precision level to a high-precision level, geometric morphing is used to achieve smooth deformation of the geometric shape through vertex interpolation; when switching from a high-precision level to a low-precision level, texture gradient transition is used to achieve visual softening through progressive adjustment of texture resolution. Furthermore, the transition duration is dynamically set for each transition method, and the transition duration is inversely proportional to the current camera movement speed—when the camera movement speed is stationary, the transition duration is set to the preset maximum value; when the camera moves at high speed, the transition duration is shortened to ensure responsiveness, thereby avoiding visual jumps during level switching and ensuring smooth interaction.

[0036] Under the preset coordinate system, based on the adaptive LOD transition mechanism and the intelligent dynamic scheduling mechanism, the spatiotemporal fusion of physical entities and observation data is performed to create a unified 2D and 3D rendering model.

[0037] Specifically, firstly, a rendering environment with a unified preset coordinate system (such as the WGS84 coordinate system) is initialized, and a complete transformation pipeline from geographic coordinates to screen pixel coordinates is established. Multi-source heterogeneous spatial data and real-time monitoring data are uniformly transformed into the preset coordinate system for spatiotemporal alignment. Real-time detection of the camera's current viewpoint: When a 2D overhead view is detected, the terrain elevation is set to zero and 2D vector tile data is loaded for planar map rendering to meet the planar accuracy requirements of professional surveyors; when a 3D tilted viewpoint is detected, a nonlinear mapping relationship is constructed to dynamically calculate the terrain rise height and perform terrain fitting processing, while loading the corresponding data tiles based on an intelligent dynamic scheduling mechanism; the nonlinear mapping relationship is a nonlinear mapping relationship between the terrain rise height, the viewpoint tilt, and the camera height, expressed as:

[0038] in, To raise the altitude of the terrain, The preset maximum terrain elevation is determined by statistically analyzing the actual maximum terrain elevation difference in the target area using existing terrain elevation statistical analysis methods, combined with the system rendering requirements. The viewing angle sensitivity coefficient is determined using existing user visual experience testing and rendering effect evaluation methods, combined with the terrain display smoothness requirements under different viewing angle tilt angles. For the angle of view tilt, For camera height, The decay function decreases with camera height. Based on the multi-scale spatial rendering theory in existing technology, it is obtained by fitting the correspondence between terrain display effect and user visual comfort at different camera heights, which will not be described in detail here. When the camera height is greater than the preset macro threshold (set autonomously based on experience), low-resolution terrain tiles are loaded to display macro mountain shapes. When the camera height is less than the preset micro threshold, high-level high-resolution terrain tiles are automatically requested to display micro terrains such as gullies and cliffs. At the same time, two-dimensional vector elements (road network, water system, land parcels) are dynamically attached to the three-dimensional terrain surface through vector slicing to achieve true two-dimensional and three-dimensional integration. Then, terrain fitting processing is performed. The processing method is as follows: In the three-dimensional tilt state, the terrain elevation at the center point of the bottom surface of the UAV tilt photography model is sampled in real time and recorded as the bottom surface terrain elevation. The actual terrain elevation at the current position in the elevation tile pyramid is sampled and recorded as the current terrain elevation. The vertical distance difference between the bottom surface terrain elevation and the current terrain elevation is calculated. The vertical distance difference is written as the vertical correction amount into the vertical offset component of the transformation matrix of the UAV tilt photography model to complete the ground fitting processing. Finally, based on the intelligent dynamic scheduling mechanism, the corresponding data tiles are loaded, and a continuous mapping relationship from a two-dimensional plane coordinate system to a three-dimensional spatial coordinate system is established through terrain lift calculation. This ensures that two-dimensional vector elements (road network, water system, land parcels) maintain spatial consistency during the two-dimensional to three-dimensional switching process, avoiding issues such as element drift and misalignment. However, during the reconstruction process of the UAV oblique photogrammetry model, due to camera calibration errors, GPS positioning errors, insufficient point cloud matching accuracy in complex terrain areas (such as cliffs and ravines), and discrepancies between the assumed ground reference surface during model reconstruction and the actual terrain, errors will inevitably occur. The problem of the bottom being suspended or sunken into the ground is addressed by terrain fitting processing. This process calculates the elevation difference between the bottom of the model and the actual terrain, dynamically adjusts the vertical position of the model, and ensures that the model is accurately positioned on the ground, completely eliminating the clipping phenomenon. Terrain fitting processing can track changes in terrain elevation in real time and dynamically adjust the model position to ensure that the model and the terrain always remain in a fitted state. In addition, terrain fitting processing is not only applicable to UAV oblique photography models, but also to all 3D solid models such as architectural white models and infrastructure models. It achieves unified adaptation of 3D models from different sources and in different formats to the terrain, greatly improving the system's compatibility and scalability.

[0039] Based on an adaptive LOD transition mechanism, various data tiles in the multi-scale data pyramid are rendered in layers to obtain a multi-scale integrated 2D and 3D rendering scene. During the rendering process, five scale thresholds and JSON style configuration files are preset. The engine automatically switches the data source and symbolization style according to the current map scale: the global-regional-local level uses a semi-transparent filling of region boundaries, blue dynamic water system patches, and a labeling strategy that only displays key place names and monitoring station names; the object level uses solid line outlines to fill plots, displays plot codes or area parameters, and enables a click-highlighting labeling strategy; the JSON style configuration file supports runtime hot updates, and the system can adjust the map style or business rules without recompiling the code.

[0040] Further, a spatiotemporal correlation mapping table between physical entities and monitoring data is established, and spatiotemporal completion is performed on the real-time monitoring data to obtain completed monitoring data. The completed monitoring data is then mapped to the surface or surrounding area of ​​the corresponding physical entity according to the spatiotemporal correlation mapping table, and the corresponding rendering method is selected based on the data type to realize the spatiotemporal fusion rendering modeling of physical entities and dynamic monitoring data.

[0041] Specifically, physical entities include UAV oblique photography models, land parcels, buildings, infrastructure, and monitoring stations; a spatiotemporal mapping table records the unique identifier, spatial range, associated sensor ID, data update cycle, and corresponding hierarchical level of each physical entity; a spatiotemporal index structure for monitoring data is established, recording the collection timestamp, spatial coordinates, sensor type, data confidence level, and unique identifier of the associated physical entity for each monitoring data point; responding to the user's preset timeline selection command, the system determines the user-specified target time and retrieves the monitoring data for the target time based on the spatiotemporal index structure; for monitoring data missing at the target time, a Kriging interpolation algorithm is used for spatiotemporal completion, and the completed monitoring data is rendered on the surface of the 3D model of the corresponding associated physical entity's hierarchical level in the form of a particle system, dynamic heatmap, or time series label, with the rendering transparency and color intensity adjusted according to the data confidence level, achieving a leap from "seeing" to "managing," truly reflecting the virtual-real fusion value of digital twins.

[0042] In processing the oblique photogrammetry model, in response to the model loading trigger command, the 3DTiles dataset of the UAV oblique photogrammetry model, which has undergone vector individualization processing, is obtained from the model pyramid. Accurate spatial registration in the WGS84 coordinate system is completed based on the 3D bounding box and ground control point coordinates in the model metadata, without the need for secondary translation. The transformation matrix of the root node of the spatially registered UAV oblique photogrammetry model is read to extract the bottom elevation of the model, denoted as the bottom terrain elevation. The actual terrain elevation at the current location is obtained from the sampling elevation tile pyramid, denoted as the current terrain elevation. The vertical distance difference between the bottom terrain elevation and the current terrain elevation is calculated. Simultaneously, based on terrain slope and roughness index, suspended or abnormal terrain is identified: when the terrain slope exceeds a preset terrain slope threshold or the roughness index exceeds a preset roughness index threshold, it indicates the existence of complex terrain areas such as cliffs, gullies, or steep slopes; otherwise, it is determined to be a flat terrain area. For flat terrain areas, the single-point sampling results of the center point are directly used for correction, that is, the vertical distance difference is written as the vertical correction amount into the vertical offset component of the UAV oblique photogrammetry model transformation matrix to adjust the model. Vertical offset can achieve good fit with low computational cost and high efficiency. For complex terrain areas, multi-point sampling correction is performed. The correction algorithm is as follows: select R support points (R≥4) evenly distributed on the bottom surface of the model, sample the actual terrain elevation corresponding to each support point, and use the least squares method to fit the best fitting plane that matches the bottom surface of the model. Calculate the average elevation difference and normal vector angle between the bottom surface of the model and the best fitting plane (or the single-point elevation plane of the flat terrain area) to obtain a comprehensive correction amount including translation and rotation. Dynamically adjust the root node transformation matrix of the model according to the comprehensive correction amount to make the bottom surface of the model parallel and closely fit the best fitting plane, completely eliminating the anomaly of the bottom of the model being suspended or sunken underground. Load the plot feature vector data and fit it to the terrain surface in the form of vector surface or line. Set the rendering order so that it is drawn above the fitted UAV oblique photography model to avoid being occluded by the model. Respond to user interaction commands to query plot attribute information and trigger the highlighting of associated models. At the same time, overlay real-time or historical observation data to the object space level to realize the spatiotemporal fusion rendering of physical entities and dynamic monitoring data.

[0043] In one specific embodiment, the complete technical implementation path from data organization to front-end interaction is described, such as... Figure 3 As shown; I. Organization and Construction of Five-Dimensional Spatial Data 1. Define the five-dimensional spatial hierarchy Based on business needs, five levels of spatial data range and accuracy standards are defined: Global-Regional-Local Space (Macro Situation): Spatial scope: Covers the entire administrative region of the base; Scale: 1:500,000 to 1:200,000; Content: Only the main water system, backbone road network, administrative boundaries and macro-ecological pattern are retained; Regional spatial (administrative unit) scope: natural regions and administrative units; scale: 1:100,000 to 1:50,000; content: displays detailed provincial, municipal, and county boundaries; Local spatial (refined management) scope: functional zones and township boundaries; scale: 1:25000 to 1:10000; content: presenting contour lines, functional zone boundaries, small watershed ditches, and patches of nature reserves.

[0044] Object space (entity interaction) scope: single plot of land, building, monitoring station; scale: 1:5000; content: clearly identify the outline of land features, their area and internal structure, and support attribute query; Observational data level (data-real fusion) scope: superimposed on the object level; content: real-time or historical monitoring data (atmosphere, water, vegetation, soil, rocks), dynamically presented in the form of point elements, charts, and labels.

[0045] 2. Multi-scale data pyramid modeling For different data types, appropriate storage and processing technologies are adopted: Vector data is organized using a vector pyramid, and road networks, water systems, and land parcels are simplified, aggregated, and symbolized using deep learning-based artificial intelligence extraction methods. Example: The global-regional-local space only displays the main roads, while the object space loads complete geometric details (various types of map features). The image data is organized using a remote sensing image pyramid to construct the COG (Cloud Optimized GeoTIFF) format, with each level of resolution matched to the scale to ensure fast loading; Terrain data is organized using elevation tiles and stored hierarchically using the Terrarium format, supporting real-time terrain rendering and mountain shadow analysis. The OSGB format oblique photogrammetry model is converted to 3DTiles to organize the UAV model data. The OSGB is converted to 3DTiles format, and the model point data is encapsulated in a JSON file for direct rendering by the front end.

[0046] II. Viewpoint-dependent dynamic scheduling and rendering 1. Viewpoint-driven dynamic scheduling mechanism The system schedules data based on view frustum clipping and viewing distance judgment: it filters and loads only tiles within the current field of view that meet the screen error threshold; prefetching and unloading: when the user zooms, it triggers the prefetching of data at different levels; it uses GPU instantiation and asynchronous loading technology to eliminate stuttering; LOD transition: it uses geometric morphing or texture gradients to avoid visual jumps when switching levels.

[0047] 2. Scale linkage and rendering rules The engine automatically switches between data sources and styles (defined via a JSON style file) based on a preset five-level scale threshold, as shown in Table 1. Table 1 Global - Regional - Local The area boundaries are filled with semi-transparent material; waterways are represented by a dynamic blue membrane; labels display the province, city, county, and the name of the data observation station. object level Drone oblique photography model; ground-hugging; 2D vector image overlaid on top. Ordinary model: ground-hugging plots with electronic fence outlines; click to display electronic fence; annotations display plot codes or area parameters. III. Realization of Integrated 2D and 3D Scenes 1. Multi-scale terrain loading and seamless switching Data foundation: Construct a multi-resolution terrain pyramid covering the entire region (e.g., based on a 12.5-meter ALOS DEM converted to Terrarium / Mapbox format).

[0048] Viewpoint-driven scheduling: Macro view (greater than 5 kilometers): Load low-resolution terrain (level 8-10) to display macro-mountain shapes; Micro view (less than 2 kilometers): Automatically request high-level tiles (level 14-16) to present micro-terrain such as ravines and cliffs.

[0049] Seamless switching between 2D and 3D: Coordinate system: WGS84 is used uniformly; Logic: Terrain elevation is set to zero in 2D state; As the viewpoint tilts or the height decreases, the terrain rises; 2D vector elements (road network, water system) are dynamically fitted to the 3D terrain surface through vector tiles.

[0050] 2. Loading and fitting of oblique photogrammetry models ("Important Facilities" scene) When a user clicks the "Important Facilities" button on the interface, the following steps are performed: Step 1: The button triggers a request; the system sends a request to the server to obtain the 3DTiles dataset of all drone models. The models have been pre-processed into vector individualization and include LOD levels.

[0051] Step 2: Model spatial registration; using the 3D bounding box and positioning point coordinates in the model metadata, the model is directly placed in the corresponding WGS84 geographic coordinates without secondary translation.

[0052] Step 3: Terrain Fitting Processing (Horizon Adjustment); For cases where the model bottom is suspended or sunk into the terrain due to reconstruction errors, the system automatically reads the transformation matrix of the model root node, extracts the bottom elevation value, and calculates the correction amount based on the terrain elevation of the current location (by sampling terrain tiles), dynamically adjusting the model position so that it "sits" on the ground.

[0053] Step 4: Integration with land parcel features; Overlay: Load land parcel features in SHP format (farmland boundaries, protected area boundaries); Rendering: Land parcel features are rendered as vector surfaces / lines and are aligned with the terrain surface. The rendering order is set so that they are always drawn on top of the model to avoid being occluded by the model; Interaction: Users can click on a land parcel to query attribute information, while the model remains highlighted, achieving integrated interaction between facilities and land types.

[0054] In one specific embodiment, the specific implementation of the gaze prediction model and the adaptive LOD transition mechanism is described in detail: During the training phase of the gaze prediction model, browsing trajectory data from 1000 users on the platform over the past 30 days were collected. This included camera position sequences (longitude, latitude, altitude), attitude angle sequences (pitch, yaw, roll), zoom level sequences, dwell time, and click / interaction records. The data underwent preprocessing: abnormal jump points were removed, trajectory smoothing was performed, and continuous trajectories were segmented into 5-second windows. An LSTM network was trained using the preprocessed data. The network input layer had a dimension of M×6 (M=10 frames, each frame containing 3D position + 2D attitude angle + 1D zoom level), the hidden layer contained 128 LSTM units, and the output layer had a dimension of 6 (predicting the camera state T=500 milliseconds in the future). After training, the model achieved a gaze prediction accuracy of over 85% on the test set.

[0055] In the actual operation of adaptive LOD transition, taking an important building (semantic importance weight 0.9) and surrounding ordinary fences (semantic importance weight 0.2) as an example; when the camera is 500 meters away from the building, the comprehensive LOD switching threshold of the building tiles is calculated as: T = 0.5 × 0.3 + 0.9 × 0.5 + 0.2 × 0.2 = 0.64 (assuming geometric complexity of 0.3, camera distance normalization of 0.2, and weight coefficients). , , When the camera moves to a distance of 5 kilometers, the T-value increases and falls into the low-precision range, triggering a texture gradient transition. The building model gradually switches to a simplified vector outline, and the transition time is dynamically adjusted according to the camera's movement speed (if the camera is stationary, the transition time is set to 800 milliseconds; if the camera moves at high speed, it is shortened to 200 milliseconds). Users can hardly perceive the visual jump. Ordinary fences, due to their low semantic weight, have already switched to minimalist symbols when the camera is 2 kilometers away, effectively saving rendering resources.

[0056] In one embodiment, a spatiotemporal AI digital base integrated 2D / 3D modeling system is disclosed, such as... Figure 2 As shown, the system includes: The spatial data organization module is used to acquire multi-source heterogeneous spatial data, divide the multi-source heterogeneous spatial data into five spatial levels from macro to micro: global, regional, local, object, and observation data level, and configure scale thresholds, data accuracy standards and semantic labeling systems for each level to construct a five-dimensional spatial hierarchy model; and construct multi-scale data pyramids for different types of multi-source heterogeneous spatial data. The scheduling mechanism construction module is used to obtain the current real-time status parameters of the camera, determine the basic visible data range, obtain historical browsing trajectory data to determine the predicted line of sight range, obtain a set of preloaded candidate tiles based on the basic visible data range and the predicted line of sight range, perform priority sorting and preloading on the data tiles in the preloaded candidate tile set, and build an intelligent dynamic scheduling mechanism. The transition mechanism construction module is used to record the data tiles to be rendered as data tiles to be rendered, obtain the geometric complexity metric and camera distance parameters of the data tiles to be rendered, determine the comprehensive LOD switching threshold of the data tiles to be rendered, and construct an adaptive LOD transition mechanism based on the comprehensive LOD switching threshold of all data tiles to be rendered. The rendering and modeling module is used to perform spatiotemporal fusion two-dimensional and three-dimensional integrated rendering and modeling of physical entities and observation data in a preset coordinate system, based on an adaptive LOD transition mechanism and an intelligent dynamic scheduling mechanism.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0058] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0059] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A spatiotemporal AI digital base two- or three-dimensional integrated modeling method, characterized in that, The methods include: We acquire multi-source heterogeneous spatial data and divide it into five spatial levels from macro to micro: global, regional, local, object, and observation data. We also configure scale thresholds, data accuracy standards, and semantic labeling systems for each level to construct a five-dimensional spatial hierarchy model. Furthermore, we construct multi-scale data pyramids for different types of multi-source heterogeneous spatial data. The system obtains the current real-time status parameters of the camera, determines the basic visible data range, obtains historical browsing trajectory data to determine the predicted line of sight range, and obtains a set of preloaded candidate tiles based on the basic visible data range and the predicted line of sight range. The system then performs priority sorting and preloading on the data tiles in the preloaded candidate tile set to build an intelligent dynamic scheduling mechanism. The data tile to be rendered is denoted as the data tile to be rendered. The geometric complexity metric and camera distance parameter of the data tile to be rendered are obtained. The comprehensive LOD switching threshold of the data tile to be rendered is determined. Based on the comprehensive LOD switching threshold of all data tiles to be rendered, an adaptive LOD transition mechanism is constructed. Under the preset coordinate system, based on the adaptive LOD transition mechanism and the intelligent dynamic scheduling mechanism, the spatiotemporal fusion of physical entities and observation data is performed to create a unified 2D and 3D rendering model.

2. The space-time AI digital foundation two-three-dimensional integrated modeling method according to claim 1, characterized in that, Multi-source heterogeneous spatial data includes five types of data: vector data, image data, terrain data, and UAV oblique photogrammetry model data. Multi-scale data pyramids include vector pyramids, remote sensing image pyramids, elevation tile pyramids, and model pyramids. Based on vector data, convolutional neural networks are used to perform semantic segmentation and classification labeling on each vector element of the vector data to obtain land cover type labels. Basic importance scores are assigned to different land cover type labels according to preset business rules. The spatial distribution density, topological connectivity and user historical access frequency of vector elements are obtained. Combined with the basic importance scores, the semantic importance weight of each land cover type label is calculated. Based on the semantic importance weight, each vector element is differentiated, simplified, aggregated and symbolized to construct a vector pyramid. Based on image data, a remote sensing image pyramid containing multiple resolution levels was constructed using cloud-optimized GeoTIFF format; Based on the terrain data, the terrain elevation data is extracted from the terrain data, and the terrain elevation data is encoded into RGB image tiles using the Terrarium format to construct a multi-level elevation tile pyramid. Based on UAV oblique photogrammetry model data, the original OSGB format data is converted and lightweighted to generate a 3DTiles format model pyramid containing multiple levels of detail.

3. The space-time AI digital foundation two-three-dimensional integrated modeling method according to claim 1, characterized in that, The method for determining the basic visible data range is as follows: Obtain the real-time state parameters of the current camera; construct a hexahedral view frustum composed of the near plane, far plane, left plane, right plane, top plane, and bottom plane based on the real-time state parameters, and calculate the standard plane equation for each plane; Traverse the axis-aligned bounding boxes of all data tiles to be loaded in the multi-scale data pyramid. For each axis-aligned bounding box, substitute the eight vertices into the six standard plane equations and perform sign judgment. If all vertices are located outside any plane, the data tile to be loaded is determined to be completely invisible and excluded; if all vertices are located inside all planes, the data tile to be loaded is determined to be completely visible; if some vertices are located inside the plane and some are located outside, the data tile to be loaded is determined to be partially visible. Calculate the Euclidean distance from the center point of the bounding box of each fully visible and partially visible data tile to be loaded to the camera position, and use it as the view distance value of the data tile to be loaded. Compare the view distance value with the maximum visible distance threshold of the corresponding level of the five-dimensional spatial hierarchy model, and remove the data tiles to be loaded whose view distance value is greater than the maximum visible distance threshold. The set of the remaining fully visible and partially visible data tiles to be loaded is determined as the basic visible data range.

4. The space-time AI digital foundation two-three-dimensional integrated modeling method according to claim 3, characterized in that, The method for constructing an intelligent dynamic scheduling mechanism is as follows: Collect and preprocess user historical browsing trajectory data, and train a gaze prediction model based on a long short-term memory network using the preprocessed historical browsing trajectory data. The gaze prediction model takes the camera state sequence of M consecutive frames and the interaction context as input, and outputs the predicted camera state and the corresponding predicted gaze range within the next T millisecond time window. Calculate the difference between the predicted gaze range and the basic visible data range to obtain a preloaded candidate tile set. For each data tile in the preload candidate tile set, a comprehensive preload priority score is calculated. The data tiles are then sorted in descending order of the comprehensive preload priority score to obtain the preload queue, thereby constructing an intelligent dynamic scheduling mechanism.

5. The space-time AI digital foundation two-three-dimensional integrated modeling method according to claim 1, characterized in that, The method for constructing an adaptive LOD transition mechanism is as follows: The range of values ​​for the overall LOD switching threshold is divided into N intervals, each interval corresponding to a LOD rendering level. The smaller the interval value, the higher the precision of the corresponding LOD rendering level. The overall LOD switching threshold for each data tile to be rendered is calculated in real time; when the overall LOD switching threshold of the data tile to be rendered crosses from one interval to an adjacent interval, the LOD level switching is triggered. When LOD level switching is triggered, the transition method is selected according to the switching direction. When switching from a low-precision level to a high-precision level, geometric morphing is used for transition, and texture gradient transition is used when switching from a high-precision level to a low-precision level. The transition duration is dynamically set for each transition method, and the transition duration is inversely proportional to the current camera movement speed. When the current camera movement speed is stationary, the transition duration is set to the preset maximum value; in this way, an adaptive LOD transition mechanism is constructed.

6. The space-time AI digital foundation two-three-dimensional integrated modeling method according to claim 5, characterized in that, The method for determining the overall LOD switching threshold of the data tiles to be rendered is as follows: Obtain the geometric information of the data tile to be rendered, calculate the geometric complexity metric and camera distance parameter. The geometric complexity metric is the ratio of the vertex density of the data tile to be rendered to the maximum allowed vertex density of the corresponding data tile level. The camera distance parameter is the ratio of the Euclidean distance from the center point of the axis-aligned bounding box of the data tile to the current camera position to the maximum visible distance threshold of the corresponding level of the current five-dimensional spatial hierarchy model. Obtain the semantic importance weight of the data tile to be rendered, preset corresponding weight coefficients for the semantic importance weight, geometric complexity measure and camera distance parameter respectively, multiply the semantic importance weight, geometric complexity measure and camera distance parameter by their respective weight coefficients and sum them to obtain the comprehensive LOD switching threshold of the data tile to be rendered.

7. The space-time AI digital foundation two-three-dimensional integrated modeling method according to claim 1, characterized in that, The method for performing spatiotemporal fusion of physical entities and observation data into a unified 2D / 3D rendering model is as follows: Initialize a rendering environment with a unified preset coordinate system, establish a complete transformation pipeline from geographic coordinates to screen pixel coordinates, and uniformly transform multi-source heterogeneous spatial data and real-time monitoring observation data to the preset coordinate system for spatiotemporal alignment. The system detects the current viewpoint of the camera in real time. When the viewpoint is detected to be a two-dimensional top view, the terrain elevation is set to zero and two-dimensional vector tile data is loaded for planar map rendering. When the viewpoint is detected to be a three-dimensional tilt view, a non-linear mapping relationship is constructed to dynamically calculate the terrain elevation and perform terrain fitting processing. At the same time, the corresponding data tiles are loaded based on an intelligent dynamic scheduling mechanism. Based on the adaptive LOD transition mechanism, various data tiles in the multi-scale data pyramid are rendered in layers to obtain a multi-scale integrated 2D and 3D rendering scene. A spatiotemporal mapping table between physical entities and observation data is established, and spatiotemporal completion is performed on the real-time monitored observation data to obtain completed observation data. The completed observation data is then mapped to the surface or surrounding area of ​​the corresponding physical entity according to the spatiotemporal mapping table, and the corresponding rendering method is selected based on the data type to realize the spatiotemporal fusion rendering modeling of physical entities and observation data.

8. The spatiotemporal AI digital base two-dimensional and three-dimensional integrated modeling method according to claim 7, characterized in that, The method for dynamically calculating terrain lift by constructing a nonlinear mapping relationship is as follows: The nonlinear mapping relationship is the nonlinear mapping relationship between terrain lift height, viewing angle tilt, and camera height, expressed as: in, To raise the altitude of the terrain, The preset maximum terrain elevation. The visual sensitivity coefficient, For the angle of view tilt, For camera height, The decay function decreases with camera height; The terrain fitting process is as follows: In a three-dimensional tilted state, the terrain elevation at the center point of the bottom surface of the UAV tilted photography model is sampled in real time and recorded as the bottom terrain elevation. The actual terrain elevation at the current position in the elevation tile pyramid is sampled and recorded as the current terrain elevation. The vertical distance difference between the bottom terrain elevation and the current terrain elevation is calculated. The vertical distance difference is written as a vertical correction amount into the vertical offset component of the transformation matrix of the UAV tilted photography model to complete the ground fitting process.

9. The spatiotemporal AI digital base two-dimensional and three-dimensional integrated modeling method according to claim 7, characterized in that, The method for obtaining complete observation data is as follows: Physical entities include UAV oblique photography models, land parcels, buildings, infrastructure, and monitoring stations. The spatiotemporal correlation mapping table records the unique identifier, spatial range, associated sensor ID, and the corresponding level of each physical entity's data update cycle. Establish a spatiotemporal index structure for the observation data, recording the acquisition timestamp, spatial coordinates, sensor type, data confidence level, and unique identifier of the associated physical entity for each observation data point; In response to the user's preset timeline selection command, the system determines the target time specified by the user, retrieves the observation data for the target time based on the spatiotemporal index structure, performs spatiotemporal completion of the missing observation data for the target time using the Kriging interpolation algorithm, and renders the completed observation data as a particle system, dynamic heat map, or time series label on the surface of the 3D model of the corresponding associated physical entity level, and adjusts the rendering transparency and color intensity according to the data confidence level.

10. A spatiotemporal AI digital base 2D / 3D integrated modeling system, implementing the spatiotemporal AI digital base 2D / 3D integrated modeling method according to any one of claims 1-9, characterized in that the system include: The spatial data organization module is used to acquire multi-source heterogeneous spatial data, divide the multi-source heterogeneous spatial data into five spatial levels from macro to micro: global, regional, local, object, and observation data level, and configure scale thresholds, data accuracy standards and semantic labeling systems for each level to construct a five-dimensional spatial hierarchy model; and construct multi-scale data pyramids for different types of multi-source heterogeneous spatial data. The scheduling mechanism construction module is used to obtain the current real-time status parameters of the camera, determine the basic visible data range, obtain historical browsing trajectory data to determine the predicted line of sight range, obtain a set of preloaded candidate tiles based on the basic visible data range and the predicted line of sight range, perform priority sorting and preloading on the data tiles in the preloaded candidate tile set, and build an intelligent dynamic scheduling mechanism. The transition mechanism construction module is used to record the data tiles to be rendered as data tiles to be rendered, obtain the geometric complexity metric and camera distance parameters of the data tiles to be rendered, determine the comprehensive LOD switching threshold of the data tiles to be rendered, and construct an adaptive LOD transition mechanism based on the comprehensive LOD switching threshold of all data tiles to be rendered. The rendering and modeling module is used to perform spatiotemporal fusion two-dimensional and three-dimensional integrated rendering and modeling of physical entities and observation data in a preset coordinate system, based on an adaptive LOD transition mechanism and an intelligent dynamic scheduling mechanism.

Images