Method and System for Generating a Terrestrial Geographic Mapping Model Based on Artificial Intelligence

Abstract

The present invention provides a method and system for generating a terrestrial geographic mapping model based on artificial intelligence. First, a set of geographic mapping data including satellite remote sensing images, surface elevation, etc. of the target area is obtained. Then, terrain feature interpretation is carried out on it to generate a set of terrain semantics, surface coverage, and mapping task-related features. Next, a geospatial encoder is called to perform multi-scale spatial modeling on the relevant feature set to obtain a set of geospatial representation vectors. Then, this set of geospatial representation vectors is dynamically fused with the preset dynamic features of the geographic entity topological map to generate a set of geographic entity-related features. Finally, a geographic mapping analysis model is trained based on the set of geographic entity-related features and the set of mapping task-related features to generate a geographic mapping optimization strategy, which is fed back to the mapping terminal to trigger the dynamic adjustment operation of mapping parameters, so as to realize the generation of a more optimized terrestrial geographic mapping model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology. Specifically, it relates to a method and system for generating a terrestrial geographic mapping model based on artificial intelligence. Background Art

[0002] In the field of terrestrial geographic mapping, traditional geographic mapping methods mainly rely on manual interpretation and analysis of single data sources, and their efficiency and accuracy are often restricted by many factors. On the one hand, when dealing with complex and changeable geographic environments, traditional methods often have difficulty in comprehensively and accurately capturing geographic features and their interrelationships due to the lack of effective integration and utilization of multi-source heterogeneous data. For example, although satellite remote sensing image data provides rich surface information, it lacks the support of key data such as surface elevation and geological structure data, which limits the accuracy and depth of terrain interpretation; while surface elevation data and geological structure feature data can reflect the three-dimensional structure of the terrain and geological background, but lack of fusion analysis with remote sensing image data, it is difficult to form a comprehensive understanding of the geographic environment.

[0003] On the other hand, traditional mapping methods often adopt fixed modeling methods and scales, which are difficult to meet the requirements of different geographic environments and mapping tasks, resulting in weak model generalization ability and unable to accurately reflect the complex spatial relationships and attribute associations between geographic entities. In addition, traditional methods also lack effective means for mining and utilizing historical mapping task log data, and cannot extract valuable mapping experience and knowledge from it to guide subsequent mapping work. Summary of the Invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, embodiments of the present invention provide a method for generating a terrestrial geographic mapping model based on artificial intelligence, and the method includes:

[0005] Obtain a set of geographic mapping data for the target area, where the set of geographic mapping data includes satellite remote sensing image data, surface elevation data, geological structure feature data, and historical mapping task log data;

[0006] Perform terrain feature interpretation processing on the set of geographic mapping data to generate a set of terrain semantic features, a set of surface coverage features, and a set of mapping task association features;

[0007] Call a geospatial encoder to perform multi-scale spatial modeling processing on the set of terrain semantic features and the set of surface coverage features to generate a set of geospatial representation vectors for the target area;

[0008] Performing dynamic feature fusion processing based on the set of geospatial representation vectors and a preset geospatial entity topology map to generate a set of geospatial entity association features, where the geospatial entity topology map includes spatial topology relationships and attribute association relationships between entity nodes;

[0009] Training a geospatial mapping analysis model according to the set of geospatial entity association features and the set of mapping task association features to generate an optimized geospatial mapping strategy for the target area, and feeding back the optimized geospatial mapping strategy to the mapping terminal to trigger dynamic adjustment operations of mapping parameters.

[0010] On the other hand, an embodiment of the present invention further provides a system for generating a land geospatial mapping model based on artificial intelligence, including a processor and a machine-readable storage medium, where the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0011] Based on the above aspects, after the embodiment of the present application forms a set of mapping data that comprehensively reflects the geographical features of the target area by fusing satellite remote sensing image data, surface elevation data, geological structure feature data, and historical mapping task log data, through terrain feature interpretation processing, a set of terrain semantic features, a set of surface coverage features, and a set of mapping task association features are generated. Further, a geospatial encoder performs multi-scale spatial modeling processing on the set of terrain semantic features and the set of surface coverage features to generate a set of geospatial representation vectors, realizing an abstract and structured expression of geospatial information, and effectively capturing the spatial dependence relationship and hierarchical structure between geospatial entities. Through dynamic feature fusion processing with a preset geospatial entity topology map, the generated set of geospatial entity association features not only fuses the spatial topology relationship and attribute association relationship of geospatial entities, but also dynamically reflects the interaction mechanism between geospatial entities, providing a more interpretable and predictive feature representation for the training of the geospatial mapping analysis model. Finally, the geospatial mapping analysis model trained based on the set of geospatial entity association features and the set of mapping task association features can comprehensively consider the spatial distribution, attribute features of geospatial entities, and the execution requirements of mapping tasks, and generate a highly targeted optimized geospatial mapping strategy. By feeding back this strategy to the mapping terminal to trigger dynamic adjustment operations of mapping parameters, the intelligence and adaptive optimization of the mapping process are realized, effectively improving the flexibility and response speed of mapping operations, while reducing the dependence on human intervention, and significantly improving the overall efficiency and result quality of geospatial mapping. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic flowchart of the execution process of the method for generating a land geospatial mapping model based on artificial intelligence provided by an embodiment of the present invention.

[0013] Figure 2 It is a schematic diagram of exemplary hardware and software components of a land geographic mapping model generation system based on artificial intelligence provided by an embodiment of the present invention. Detailed implementation manners

[0014] The present invention will be specifically described below with reference to the accompanying drawings of the specification. Figure 1 It is a flowchart of a method for generating a land geographic mapping model based on artificial intelligence provided by an embodiment of the present invention. The method for generating a land geographic mapping model based on artificial intelligence will be introduced in detail below.

[0015] Step S110: Obtain a set of geographic mapping data for the target area, where the set of geographic mapping data includes satellite remote sensing image data, surface elevation data, geological structure feature data, and historical mapping task log data.

[0016] For example, in this geographic mapping task, it can be set to conduct mapping work on a vast and geologically complex area, which contains various topographies and landforms, such as mountains, plains, rivers, and areas with different degrees of vegetation, as well as some artificial building facilities, and there have been multiple mapping tasks carried out here in the past.

[0017] First, obtain the set of geographic mapping data. For example, the satellite remote sensing image data is obtained through a professional satellite observation system. The satellite is equipped with an optical sensor that can capture image information of the target area at different times and different spectral bands. For example, on a sunny day in spring, the satellite passes over the target area along a set orbit and takes a series of color images covering the visible light band and part of the near-infrared band. The above satellite remote sensing image data records rich surface information of the target area, including the color and distribution of vegetation, the reflection characteristics of water bodies, and the outlines of artificial buildings.

[0018] The surface elevation data is obtained by means of a high-precision topographic survey device, such as an airborne lidar (LiDAR) system, which is carried on an aircraft and flies over the target area along a predetermined route. By emitting laser beams and receiving reflected signals, the distance from the ground to the aircraft is measured, thereby generating detailed surface elevation data. Thus, digital elevation model (DEM) data with centimeter-level accuracy for the target area is obtained, which shows the undulations of the terrain, such as the altitude of mountains and the depth of valleys.

[0019] Geological structure feature data are collected through various geological exploration means. For example, ground-penetrating radar is used to detect the geological structure within a certain depth range below the ground surface. By utilizing the differences in the propagation characteristics of electromagnetic waves in different geological media, it draws a cross-sectional view of the underground geological structure, identifying information such as the distribution of rock layers and the location of faults. At the same time, combined with on-site geological surveys, rock samples are collected for composition analysis and age determination to further enrich the geological structure feature data. For example, through ground-penetrating radar detection, an underground fault is discovered. Its trend is roughly east-west, and the depth is approximately between 50 meters and 100 meters underground. Analysis of the rock samples collected on-site shows that the rocks in this area are mainly sedimentary rocks, and their formation age is about millions of years ago.

[0020] The historical mapping task log data is sourced from the records of past mapping projects carried out in the target area. It details information such as the time of each mapping task, the mapping equipment used, the mapping path planning parameters, the sensor configuration parameters, and the quality assessment indicators for the collected data. For example, in a mapping task three years ago, a multispectral camera model XYZ-100 was used, and 5 different spectral bands were set for data collection. The mapping path covered most of the main terrains in the target area. The quality assessment indicators of the data collected at that time showed that the clarity of the image data reached over 85% in some areas, but there were certain degrees of occlusion and blurring in some densely vegetated areas.

[0021] Step S120: Perform terrain feature interpretation processing on the geographical mapping data set to generate a terrain semantic feature set, a surface coverage feature set, and a mapping task association feature set.

[0022] In this embodiment, first, multi-scale image segmentation processing is performed on the satellite remote sensing image data. Using an advanced image segmentation algorithm, the satellite remote sensing image is divided according to different scales to generate multiple image segmentation units. For example, with 100×100 pixels as a basic unit for preliminary segmentation, and then gradually expanding the scale to perform segmentation at different scales such as 500×500 pixels and 1000×1000 pixels. Each image segmentation unit contains pixel-level spectral features and spatial position encoding. Taking an image segmentation unit of 500×500 pixels as an example, the pixel-level spectral features of this unit record the reflection values of each pixel in different spectral bands. For example, the reflection value in the visible light red band is 120, the green band is 100, and the blue band is 80, etc.; its spatial position encoding clarifies the position of this unit in the entire image. For example, the abscissa is 2000 and the ordinate is 3000, indicating its specific coordinate position in the satellite remote sensing image.

[0023] Next, call the pre-trained semantic segmentation model to perform land cover type annotation processing on each image segmentation unit. This semantic segmentation model has been trained with a large amount of image data and can accurately identify different land cover types. After processing, a surface cover feature set is generated. For example, in a large image area, the semantic segmentation model identifies a part of the area as vegetation cover type, specifically forest vegetation, with a vegetation coverage rate of over 70%; another part of the area is labeled as a water body distribution area, which is a river with a width of about 50 meters; and there are also some areas determined to be artificial building areas, including a large factory and some auxiliary buildings, with a floor area of about 50,000 square meters.

[0024] Then, perform terrain gradient calculation processing on the surface elevation data. By calculating the ratio of the height difference between adjacent elevation points to the horizontal distance, a terrain undulation feature set is generated. For example, in a certain mountainous area, after calculation, it is found that the terrain gradient is relatively large in some areas, indicating a steep terrain, with a terrain undulation of 20 meters per 100 meters, while in the plain area, the terrain gradient is relatively small, and the terrain undulation is only 1 meter per 100 meters. At the same time, perform fault line identification processing on the geological structure feature data. Using geological radar data and on-site investigation information, a geological structure boundary feature set is identified through a set algorithm. For example, in a certain part of the target area, an obvious fault line is successfully identified, and the geological features on both sides of the fault line are significantly different. One side is relatively hard granite, and the other side is relatively soft shale.

[0025] Perform feature fusion processing on the terrain undulation feature set and the geological structure boundary feature set to generate a terrain semantic feature set. For example, by comprehensively analyzing the areas with a large terrain undulation and close to the fault line, it is found that the above areas may have the characteristics of frequent geological activities and rapid terrain changes, thus forming corresponding feature descriptions in the terrain semantic feature set.

[0026] Perform task parsing processing on the historical surveying and mapping task log data. Carefully extract the historical surveying and mapping path planning parameters, sensor configuration parameters, and data quality evaluation indicators to generate a surveying and mapping task correlation feature set. For example, from the log of a past surveying and mapping task, it is extracted that the surveying and mapping path planning adopted a spiral scanning method, the flight altitude was maintained at 500 meters, and the flight speed was 50 meters per second; in terms of sensor configuration, an optical sensor of model ABC-200 was used, and 3 bands were set for data acquisition; the data quality evaluation indicators show that the accuracy of the collected data is about 90%, but there are data missing situations in some areas.

[0027] Step S130, call the geospatial encoder to perform multi-scale spatial modeling processing on the terrain semantic feature set and the surface cover feature set to generate a geospatial representation vector set of the target area.

[0028] In this embodiment, a multi-scale processing architecture of a geospatial encoder can be constructed, which includes a local feature extraction module, a regional context modeling module, and a global space fusion module. The surface coverage features of a single image segmentation unit are subjected to sliding window processing by the local feature extraction module. For example, for an image segmentation unit containing vegetation coverage, a 3×3 pixel sliding window is used to move pixel by pixel within the unit to extract local texture features and edge features. During the window movement process, the color changes and texture details of the pixels in the window are analyzed, and it is found that the texture of the vegetation area presents irregular filamentous and blocky features, and the pixel color changes at the edge are more obvious, thereby extracting the above-mentioned local features.

[0029] The regional context modeling module calculates the spatial association weights of the terrain semantic features of adjacent image segmentation units. For example, two adjacent image segmentation units, one for mountainous terrain and the other for plain terrain, are used to calculate the spatial association weights. The module analyzes the feature similarity and spatial distance differences between them. For example, by calculating the differences in terrain relief, geological structure, and spatial distance, it was found that the terrain relief of the mountain unit is 15 meters per 100 meters, while that of the plain unit is 2 meters per 100 meters. The geological structures of the two units also differ significantly, and the spatial distance is 500 meters. Based on the above factors, the regional context association weight is calculated, and the features of the adjacent units are weighted and aggregated based on this regional context association weight to obtain the regional aggregated features.

[0030] The global spatial fusion module analyzes cross-scale dependencies of multi-scale features. It comprehensively analyzes local texture features and regional aggregation features at different scales to generate a global spatial dependency map. For example, it was found that small-scale local texture features (such as the microtexture of vegetation) and large-scale regional terrain features (such as the direction of mountain ranges) have certain correlations, which are reflected in the global spatial dependency map.

[0031] Finally, the local texture features, regional aggregate features, and global dependency graph are concatenated to generate a set of geospatial representation vectors. For example, the locally extracted vegetation texture feature vectors, the regionally aggregated terrain feature vectors, and the feature vectors of the global dependency graph are concatenated in a certain order to form a complete geospatial representation vector, thereby generating a set of geospatial representation vectors for the target area.

[0032] Step S140 , performing dynamic feature fusion processing based on the geographic space representation vector set and a preset geographic entity topology map to generate a geographic entity association feature set, wherein the geographic entity topology map includes spatial topological relationships and attribute association relationships between entity nodes.

[0033] In this embodiment, an entity node set of the target area and a topological edge set between the entity node sets can be extracted from a preset geographical entity topological map. The entity nodes contain multiple attributes, such as spatial location attributes, geological attributes, and historical survey record attributes. For example, in the geographical entity topological map, there is an entity node representing a mountain range. Its spatial location attribute defines the longitude and latitude range of the mountain range in the target area. The geological attribute indicates that it is mainly composed of granite. The historical survey record attribute records the time of past surveys of the mountain range, the survey methods, and the relevant data characteristics collected.

[0034] Calculate the multi-dimensional similarity set between each geospatial representation vector in the geospatial representation vector set and each entity node in the entity node set. Taking a geospatial representation vector and an entity node representing a mountain range as an example, calculate their spatial distance similarity. By calculating the spatial location distance between the area represented by the vector and the mountain range entity node, the spatial distance similarity is obtained as 0.8; calculate the geological attribute matching degree and find that there is a certain similarity between the geological characteristics of the area represented by the vector and the granite geological attribute of the mountain range, and the geological attribute matching degree is 0.7; calculate the historical survey record correlation degree. According to past survey records, it is found that there is a certain correlation between them in terms of survey time and data characteristics, and the historical survey record correlation degree is 0.6. Thus, a multi-dimensional similarity set is obtained.

[0035] Based on the multi-dimensional similarity set, construct a dynamic feature selection gating mechanism to generate a set of dynamic feature weight coefficients corresponding to each entity node. For example, for the entity node representing a mountain range, according to the multi-dimensional similarity set, determine its feature enhancement weight as 0.7 and its feature suppression weight as 0.3. This set of weight coefficients is used to perform weighted processing on the entity node attributes.

[0036] Perform weighted processing on the entity node attributes of the entity node set according to the set of dynamic feature weight coefficients. For the entity node representing a mountain range, multiply its spatial location attribute, geological attribute, and historical survey record attribute by the corresponding weight coefficients respectively. For example, after multiplying the spatial location attribute by the feature enhancement weight of 0.7, its importance in the fusion process is enhanced; after multiplying the geological attribute by the feature suppression weight of 0.3, its influence in the fusion process is relatively weakened. Thus, a set of weighted entity node attributes is generated.

[0037] Perform cross-modal feature concatenation processing on the weighted set of entity node attributes and the set of geospatial representation vectors to generate a fused set of geospatial entity association features. For example, concatenate the weighted mountain entity node attribute vector and the geospatial representation vector according to certain rules to form a fused feature vector. Finally, perform feature dimensionality reduction processing on the fused set of geospatial entity association features, and convert the high-dimensional feature vector into a low-dimensional and representative vector through a specific algorithm to generate a set of geospatial entity association features.

[0038] Step S150: Train a geospatial mapping analysis model based on the set of geospatial entity association features and the set of mapping task association features to generate an optimized geospatial mapping strategy for the target area, and feedback the optimized geospatial mapping strategy to the mapping terminal to trigger dynamic adjustment operations of mapping parameters.

[0039] In this embodiment, a dual-stream input architecture of the geospatial mapping analysis model can be constructed. The first input stream processes the set of geospatial entity association features, and the second input stream processes the set of mapping task association features. Perform spatial topological reasoning processing on the set of geospatial entity association features through the first input stream. Construct an adjacency matrix of the graph attention network based on the entity node topological edge set in the set of geospatial entity association features. For example, for entity nodes representing different geospatial entities (such as mountains, rivers, cities, etc.), dynamically calculate the weights of the adjacency matrix according to the spatial distance similarity and geological attribute matching degree between them. If the spatial distance between a mountain and a river is relatively close and there is a certain correlation in geological attributes, the weight of the adjacency matrix between them is 0.6; if the spatial distance between a mountain and a city is relatively far and there are significant differences in geological attributes, the weight is 0.2. Perform sparsification processing on the adjacency matrix, remove the edges with weights lower than a preset threshold (such as 0.3), and generate a pruned topological edge subset. Perform feature propagation processing on the pruned topological edge subset through a multi-layer graph attention network, iteratively aggregate the attribute features of adjacent nodes, and generate a spatial topological reasoning result containing global spatial dependency relationships.

[0040] Perform task context modeling processing on the mapping task - associated feature set through the second input stream. Perform spatio - temporal position encoding processing on the historical mapping path planning parameters. For example, encode the time information (such as mapping records at different time points) and spatial coordinate information in the historical mapping path planning to generate a path planning encoding vector containing timestamps and spatial coordinates. Perform temporal sequence modeling processing on the path planning encoding vector through a bidirectional gated recurrent unit to capture the forward and backward dependencies during the historical task execution process and generate a task temporal feature vector. Perform multi - modal embedding processing on the sensor configuration parameters, mapping the discrete sensor configuration parameters (such as sensor models, band settings, etc.) into continuous sensor feature vectors with the same dimension as the task temporal feature vector. Calculate the dynamic association weights between the task temporal feature vector and the sensor feature vector through a cross - attention layer to generate a task context feature that fuses task context and sensor configuration.

[0041] Input the updated geographical entity - associated feature set and the task context feature into the cross - stream fusion module and perform alignment processing through the multi - head attention mechanism. Perform linear transformation processing on the updated geographical entity - associated feature set to generate a set of spatial topology query vectors; perform linear transformation processing on the task context feature to generate a set of task key vectors and a set of task value vectors. Calculate the similarity scores between the set of spatial topology query vectors and the set of task key vectors to generate a cross - modal attention weight distribution. Perform weighted aggregation processing on the set of task value vectors according to this weight distribution to generate a set of single - head attention features. Concatenate multiple sets of single - head attention features and perform non - linear transformation processing to generate a jointly optimized feature set after cross - stream fusion.

[0042] Perform multi - task prediction processing based on the jointly optimized feature set. Through the path planning prediction branch, perform spatial rasterization processing on the jointly optimized feature set, divide the target area into small spatial grids, calculate the path feasibility probability of each grid, and generate a path feasibility probability map. Search for the optimal path sequence in this map based on the dynamic programming algorithm to generate a set of mapping path planning parameters. For example, plan a new mapping path with the flight altitude varying between 400 meters and 600 meters, the flight speed maintained at 45 meters per second, and the heading angle adjusted in real - time according to the terrain and the target area.

[0043] Through the sensor configuration branch, perform band importance ranking processing on the jointly optimized feature set, calculate the spectral resolution contribution degree and terrain matching degree of each band, and generate a set of band selection priority parameters. For example, determine that in the next mapping task, the near - infrared band and the green band are preferentially selected for data acquisition.

[0044] Perform distribution consistency analysis on the jointly optimized feature set through the quality control branch, generate dynamic confidence interval thresholds based on historical data quality evaluation indicators, and generate a set of data acquisition quality control thresholds. For example, set the signal-to-noise ratio threshold for collecting image data to 30 dB and the geometric distortion rate threshold to 5%.

[0045] Balance the gradient backpropagation ratio of the multi-task prediction branch through the adaptive loss weight allocation mechanism, and iteratively optimize the prediction accuracy of each branch.

[0046] Finally, feedback the geodetic survey optimization strategy to the surveying and mapping terminal to trigger the dynamic adjustment operation of surveying and mapping parameters. Convert the set of surveying and mapping path planning parameters into a set of flight control instructions. For example, convert the planned flight altitude sequence, airspeed sequence, and heading angle sequence into specific instructions. The flight altitude instruction is to gradually rise from 400 meters to 500 meters within a certain period of time, the airspeed instruction is to maintain 45 m / s, and the heading angle instruction is to rotate 30 degrees clockwise, etc. Input the set of flight control instructions into the navigation control system of the surveying and mapping terminal to generate a real-time waypoint coordinate sequence, and perform collision detection processing on the real-time waypoint coordinate sequence and the preset surveying and mapping area boundary. If it is found that a certain waypoint coordinate is close to the preset boundary, generate a set of waypoint correction instructions to adjust the waypoint position. Merge the set of waypoint correction instructions with the set of sensor band selection parameters into a set of device control signals to trigger the filter switching operation and exposure parameter adjustment operation of the multi-spectral camera of the surveying and mapping terminal. For example, switch the filter to the near-infrared band and adjust the exposure time to 0.01 seconds. Perform signal-to-noise ratio analysis and geometric distortion detection processing on the real-time image data collected by the surveying and mapping terminal based on the set of data acquisition quality control thresholds. If it is found that the signal-to-noise ratio of a certain part of the image data is lower than 30 dB, generate a set of quality anomaly marked data. Input the set of quality anomaly marked data into the geodetic survey analysis model for resampling path prediction processing, generate a set of supplementary surveying and mapping path parameters, and update the set of surveying and mapping path planning parameters to ensure the quality and integrity of subsequent surveying and mapping data.

[0047] Based on the above steps, in the embodiment of the present application, after integrating satellite remote sensing image data, surface elevation data, geological structure feature data, and historical survey task log data to form a surveying and mapping data set that comprehensively reflects the geographical features of the target area, through terrain feature interpretation processing, a terrain semantic feature set, a surface coverage feature set, and a surveying and mapping task correlation feature set are generated. Further, the geospatial encoder performs multi-scale spatial modeling processing on the terrain semantic feature set and the surface coverage feature set to generate a geospatial representation vector set, realizing the abstract and structured expression of geospatial information, effectively capturing the spatial dependence relationship and hierarchical structure among geographical entities. Through dynamic feature fusion processing with a preset geographical entity topology map, the generated geographical entity correlation feature set not only integrates the spatial topology relationship and attribute association relationship of geographical entities, but also dynamically reflects the interaction mechanism among geographical entities, providing a more interpretable and predictive feature representation for the training of the geodetic survey analysis model. Finally, the geodetic survey analysis model trained based on the geographical entity correlation feature set and the surveying and mapping task correlation feature set can comprehensively consider the spatial distribution, attribute characteristics of geographical entities, and the execution requirements of surveying and mapping tasks to generate a highly targeted geodetic survey optimization strategy. This strategy triggers dynamic adjustment operations of surveying and mapping parameters by feeding back to the surveying and mapping terminal, realizing the intelligent and adaptive optimization of the surveying and mapping process, effectively improving the flexibility and response speed of surveying and mapping operations, while reducing the dependence on human intervention, and significantly improving the overall efficiency and result quality of geodetic surveying and mapping.

[0048] In a possible implementation manner, step S120 includes:

[0049] Step S121, performing multi-scale image segmentation processing on the satellite remote sensing image data to generate a plurality of image segmentation units, and each image segmentation unit includes pixel-level spectral features and spatial position encoding.

[0050] Taking the smallest scale as an example, each image segmentation unit is set to 50×50 pixels. At this smallest scale, the entire image is divided block by block. For example, in a vegetation area, an image segmentation unit of 50×50 pixels is divided. Each pixel in this unit has its specific spectral characteristics. In the red, green, and blue bands of visible light, the reflection values of some pixels in this unit are 130 in the red band, 150 in the green band, and 110 in the blue band respectively. The above pixel-level spectral characteristics record the color information of this area. At the same time, through the coordinate system of the image, a spatial position code is assigned to this unit. Assuming that the upper left corner coordinates of this unit in the image are (1000, 2000), its position in the entire image is thus determined. Then, segmentation is performed at a larger scale of 200×200 pixels. In an area containing some water bodies and vegetation, the newly divided image segmentation unit integrates the information of more pixels. Its pixel-level spectral characteristics are the comprehensive manifestation of the reflection values of many pixels in this area, and the spatial position code also corresponds to the position of this larger unit in the image. Through such multi-scale image segmentation processing, multiple image segmentation units of different scales are generated, and each unit contains unique pixel-level spectral characteristics and spatial position codes.

[0051] Step S122: Call the pre-trained semantic segmentation model to perform ground object category annotation processing on each image segmentation unit, and generate a surface coverage feature set. Among them, the ground object categories include vegetation coverage types, water body distribution areas, and artificial building areas.

[0052] In an area containing multiple ground objects, for an image segmentation unit of 200×200 pixels, the semantic segmentation model makes a judgment based on a large number of image data features it has learned. For example, the spectral characteristics of most pixels in this unit have a high degree of matching with the characteristics of vegetation in the model. For example, the reflection value in the near-infrared band shows the unique high-reflection characteristic of vegetation, and the texture characteristics also conform to the irregular distribution pattern of vegetation. Therefore, it can be labeled as a vegetation coverage type. In another image segmentation unit near a river, by analyzing the spectral characteristics of the pixels, the reflection value is high in the blue band, and the shape presents a long and narrow strip, which conforms to the characteristics of the water body, so it is labeled as a water body distribution area. For an area with buildings, the pixels in the image segmentation unit present regular geometric shapes and are significantly different from the surrounding natural ground objects in spectral characteristics, and are labeled as an artificial building area. By processing all image segmentation units, a surface coverage feature set is generated, which details the distribution of different ground objects in the image.

[0053] Step S123: Perform terrain gradient calculation processing on the surface elevation data to generate a terrain undulation feature set, and perform fault line identification processing on the geological structure feature data to generate a geological structure boundary feature set.

[0054] In the mountainous area part, two adjacent elevation measurement points are selected. The elevation of point A is 800 meters, and the elevation of point B is 900 meters. The measured horizontal distance between the two points is 200 meters. According to the calculation method of terrain gradient, the terrain gradient is equal to the elevation difference divided by the horizontal distance, that is, (900 - 800) ÷ 200 = 0.5. This value indicates that for every 1-meter horizontal distance advanced in this area, the elevation will rise by 0.5 meters. By performing such calculations on multiple adjacent elevation points in the entire area, a terrain undulation feature set is generated, clearly showing the undulation changes of the terrain. At the same time, fault line identification processing is performed on the geological structure feature data. Through the data of geological radar and on-site investigation information, obvious differences in the reflection signals of different strata are found in a certain area. After detailed analysis, it is determined that there is a sudden change interface of a stratum at a depth of 60 meters underground. Combining the analysis results of rock samples, it is judged that this interface is a fault line, and thus a geological structure demarcation feature set is generated, clarifying the demarcation line of different geological structures.

[0055] Step S124, perform feature fusion processing on the terrain undulation feature set and the geological structure demarcation feature set to generate a terrain semantic feature set.

[0056] In a mountainous area near the fault line, the terrain undulation is large, and the terrain gradient reaches 0.8. Combining the geological structure demarcation features, the fault line in this area may have caused crustal movement, thereby resulting in drastic changes in the terrain. Integrate the relevant information in these two feature sets, such as combining the specific values and change trends of the terrain undulation with the position and trend of the fault line, etc., to generate a terrain semantic feature set, which more comprehensively describes the terrain and geological features of this area.

[0057] Step S125, perform task parsing processing on the historical surveying and mapping task log data, extract the historical surveying and mapping path planning parameters, sensor configuration parameters, and data quality evaluation indicators, and generate a surveying and mapping task correlation feature set.

[0058] In the log of a previous surveying and mapping task, the surveying and mapping path planning parameters were detailed. At that time, a grid-like surveying and mapping path planning was adopted. A surveying line was set every 500 meters horizontally and every 800 meters vertically to cover the entire target area. In terms of sensor configuration, a multispectral sensor model XYZ-500 was used, and 7 different spectral bands were set, namely the red, green, blue, cyan, and purple bands of visible light, as well as the near-infrared and short-wave infrared bands. In terms of data quality evaluation indicators, by analyzing the collected image data, it was found that the clarity of the images reached 88% in most areas, but in some areas with dense vegetation, due to vegetation occlusion, the clarity dropped to 75%. By parsing the data in the above historical surveying and mapping task log, the historical surveying and mapping path planning parameters, sensor configuration parameters, and data quality evaluation indicators were extracted to generate a set of associated features for the surveying and mapping task.

[0059] In a possible implementation manner, step S130 includes:

[0060] Step S131, constructing a multi-scale processing architecture for the geospatial encoder, where the multi-scale processing architecture includes a local feature extraction module, a regional context modeling module, and a global spatial fusion module.

[0061] Step S132, performing a sliding window process on the surface cover features of a single image segmentation unit through the local feature extraction module to extract local texture features and edge features.

[0062] Taking a vegetation-covered image segmentation unit as an example, a 5×5 pixel sliding window is used. Starting from the upper left corner of the unit, the window moves sequentially. When the window moves to a certain position, the texture features of the pixels within the window are analyzed, and it is found that the color changes between the pixels show irregular filaments, which is the manifestation of the texture features of the vegetation leaves; at the same time, observing the changes in the edge pixels of the window, it is found that the colors of the edge pixels are somewhat different from those of the internal pixels, thus constituting the edge features. Through the above sliding window process, the local texture features and edge features of this image segmentation unit are extracted.

[0063] Step S133, calculating the spatial association weights of the terrain semantic features of adjacent image segmentation units through the regional context modeling module to generate regional context association weights, and performing weighted aggregation processing on the adjacent unit features based on the regional context association weights to obtain regional aggregation features.

[0064] For example, select two adjacent image segmentation units. One unit is in a mountainous area with a terrain undulation degree of 0.6, and the other unit is in a plain with a terrain undulation degree of 0.1. First, analyze the feature similarity between them and find that there are certain differences in their geological structures, but there is a certain correlation in their altitudes. The mountainous unit has a higher altitude, and the plain unit has a lower altitude. Then calculate the spatial distance difference. The spatial distance between the two units in the image is measured to be 500 pixels. Considering the above factors, through a specific calculation method, assuming the weight of the feature similarity is 0.6 and the weight of the spatial distance difference is 0.4, calculate the regional context association weight. The specific calculation process is as follows: The feature similarity score is set to 0.4 by comparing features such as terrain undulation degree and geological structure; the spatial distance difference score is set to 0.6 according to the distance. Then the regional context association weight = 0.6×0.4 + 0.4×0.6 = 0.48. Based on this regional context association weight, perform weighted aggregation processing on the features of adjacent units. Assign a weight of 0.52 to the features of the mountainous unit and a weight of 0.48 to the features of the plain unit, and add the terrain semantic features of the two units after weighting to obtain the regional aggregation feature, which more comprehensively reflects the comprehensive features of adjacent regions.

[0065] Step S134, analyze the cross-scale dependence relationship of the multi-scale features through the global spatial fusion module to generate a global spatial dependence relationship map.

[0066] In the entire target area, there is a certain connection between the local texture features at a small scale, such as the microscopic texture of vegetation, and the terrain features at a large scale, such as the trend of mountains. For example, in mountainous areas, the growth direction of vegetation is often affected by the trend of mountains. By comprehensively analyzing features at different scales, a global spatial dependence relationship map is generated. In the map, the features at different scales and their associated relationships are presented graphically, clearly showing the mutual dependence of geographical features within the entire area.

[0067] Step S135, perform feature splicing processing on the local texture features, regional aggregation features, and global dependence relationship map to generate a set of geospatial representation vectors. Among them, the spatial association weight calculation is achieved by analyzing the feature similarity and spatial distance difference between adjacent units.

[0068] For example, the vegetation texture feature vectors extracted locally, the terrain feature vectors aggregated regionally, and the feature vectors corresponding to the global dependence relationship map can be spliced in a certain order. For instance, first place the local texture feature vectors in the front, then splicing the regional aggregation feature vectors, and finally splicing the feature vectors of the global dependence relationship map to form a complete geospatial representation vector. By processing all relevant features, a set of geospatial representation vectors for the target area is generated.

[0069] In a possible implementation manner, step S140 includes:

[0070] Step S141, extracting a set of entity nodes of the target area and a set of topological connection edges between the sets of entity nodes from the geographical entity topological map, where each entity node includes the spatial position attribute of the entity node, the geological attribute of the entity node, and the historical survey record attribute of the entity node.

[0071] In this embodiment, in the geographical entity topological map, the target area includes multiple entity nodes. For example, there is an entity node representing a large mountain range. Its spatial position attribute clarifies the specific geographical location of the mountain range in the target area, and the longitude and latitude range is accurate to a certain extent, such as between 30°N - 31°N and 110°E - 111°E; the geological attribute indicates that the mountain range is mainly composed of granite, and the characteristics of the granite are also recorded in detail, such as relevant parameters such as hardness and density; the historical survey record attribute records the detailed information of past surveys of the mountain range, including the time of each survey, such as different periods three years ago, five years ago, etc., as well as the survey methods used and the characteristics of the data collected. There is also an entity node representing a river. The spatial position attribute determines the direction and flowing range of the river. The geological attribute shows that the surrounding geology of the river is sedimentary rock. The historical survey record attribute includes the measurement records of data such as the river flow and depth and the corresponding survey means. The above entity nodes are interconnected through the set of topological connection edges. For example, there is a topological connection edge between the mountain range and a nearby river, and the weight of the connection edge represents a certain degree of association between them. This weight may be determined based on factors such as their spatial distance and geological influence.

[0072] Step S142, calculating a multi-dimensional similarity set between each geographical space representation vector in the geographical space representation vector set and each entity node in the set of entity nodes. The multi-dimensional similarity set includes the spatial distance similarity between the geographical space representation vector and the entity node, the geological attribute matching degree between the geographical space representation vector and the entity node, and the historical survey record association degree between the geographical space representation vector and the entity node.

[0073] For example, let's calculate the spatial distance similarity between a geospatial representation vector and a physical node representing a mountain range. First, determine the coordinates of the center of the region represented by the geospatial representation vector, assuming it's 30.5°N, 110.5°E, and compare them with the coordinates of the center of the physical node representing the mountain range. The difference in longitude and latitude is converted to actual distance. Assuming the actual distance corresponding to each degree of longitude and latitude change is known, the calculated spatial distance between the two is 50 kilometers. According to the set distance similarity calculation rule, closer distances have higher similarity. Assuming a similarity of 1 for distances between 0 and 10 kilometers and a linearly decreasing similarity between 10 and 100 kilometers, the calculated spatial distance similarity in this example is 0.8 (calculation process: (100 - 50) ÷ (100 - 10) × (1 - 0.5) + 0.5 = 0.8, meaning, first calculate the distance proportion, then calculate the similarity based on the set decreasing rule).

[0074] Then, we calculate the geological attribute matching degree. Analysis shows that the main geological component of the area represented by this geographic spatial representation vector is granite, which is the same as the geological attributes of the mountain entity node, so the geological attribute matching degree is 1 (if the geological components are partially the same, the matching degree is calculated based on the proportion of the same components).

[0075] Finally, the correlation degree of historical surveying records is calculated. By reviewing historical records, it is found that the area represented by this geospatial representation vector partially overlaps with the mountain range in past surveys, and that the surveying time and equipment used are the same. Based on these correlation factors and the established calculation rules, a comprehensive calculation is performed, assuming factors such as the percentage of overlapping area, the degree of overlap in surveying time, and the degree of equipment similarity. The correlation degree of historical surveying records is 0.7 (calculation process: assuming the percentage of overlapping area is 0.4, the degree of overlap in surveying time is 0.3, and the degree of equipment similarity is 0.3. Add these factors and divide by 3 to get 0.3 + 0.4 + 0.3 ÷ 3 = 0.7). This yields a multidimensional similarity set between this geospatial representation vector and the mountain entity node, including a spatial distance similarity of 0.8, a geological attribute match of 1, and a historical surveying record correlation of 0.7. This calculation is repeated for each geospatial representation vector in the set of geospatial representation vectors and each node in the set of entity nodes, resulting in a complete multidimensional similarity set.

[0076] Step S143: construct a dynamic feature selection gating mechanism based on the multidimensional similarity set, and the dynamic feature selection gating mechanism generates a dynamic feature weight coefficient set corresponding to each entity node, wherein the dynamic feature weight coefficient set includes the feature enhancement weight of the entity node and the feature suppression weight of the entity node.

[0077] Taking the mountain entity node as an example, the multi-dimensional similarity set between it and each geographical space representation vector is synthesized. Suppose the spatial distance similarity weight is set to 0.3, the geological attribute matching degree weight is 0.4, and the historical surveying and mapping record correlation weight is 0.3. First, calculate a comprehensive score. Taking a certain geographical space representation vector as an example, its comprehensive score = 0.8×0.3 + 1×0.4 + 0.7×0.3 = 0.85 (by multiplying the similarity of each dimension by the corresponding weight and then adding them up). Sort the comprehensive scores of all geographical space representation vectors and the mountain entity node, and generate a set of dynamic feature weight coefficients according to the sorting result and the set rules. For example, for the features associated with the geographical space representation vector with a higher comprehensive score, a higher feature enhancement weight is given, suppose it is 0.7; for the features associated with the vector with a lower comprehensive score, a lower feature suppression weight is given, suppose it is 0.3. In this way, a set of dynamic feature weight coefficients corresponding to the mountain entity node is generated, including the feature enhancement weight 0.7 and the feature suppression weight 0.3. Such processing is performed on each entity node to generate a set of dynamic feature weight coefficients corresponding to all entity nodes.

[0078] Step S144, perform weighted processing on the entity node attributes of the entity node set according to the set of dynamic feature weight coefficients, and generate a weighted entity node attribute set, where the weighted entity node attribute set includes the entity node attributes strengthened by the feature enhancement weight and the entity node attributes weakened by the feature suppression weight.

[0079] Taking the mountain entity node as an example, its spatial location attribute, geological attribute, and historical surveying and mapping record attribute will all be weighted. For the spatial location attribute, multiply by the feature enhancement weight 0.7, which means the importance of this attribute in the subsequent fusion process is enhanced; for the geological attribute, multiply by the feature suppression weight 0.3, and its influence in the fusion is relatively weakened; the historical surveying and mapping record attribute is also multiplied by the feature enhancement weight 0.7. After such weighted processing, a weighted entity node attribute set is generated, where the spatial location attribute is strengthened, the geological attribute is weakened, and the historical surveying and mapping record attribute is strengthened. Such weighted processing is performed on the attributes of all entity nodes to obtain a weighted entity node attribute set.

[0080] Step S145, perform cross-modal feature splicing processing on the weighted entity node attribute set and the geographical space representation vector set to generate a set of fused geographical entity association features.

[0081] In this embodiment, the weighted mountain entity node attribute vector and the geospatial representation vector can be concatenated in a certain order. For example, first place the weighted spatial position attribute vector of the mountain entity node in the front, then concatenate the geological attribute vector, then concatenate the historical survey record attribute vector, and finally concatenate the geospatial representation vector. In this way, the attribute sets of all entity nodes after weighting and the geospatial representation vector sets are concatenated to generate a fused geospatial entity association feature set.

[0082] Step S146: Perform feature dimensionality reduction processing on the fused geospatial entity association feature set to generate the geospatial entity association feature set.

[0083] Suppose the fused geospatial entity association feature set is a high-dimensional vector with a dimension of 100. The principal component analysis (PCA) method is used for dimensionality reduction processing. First, calculate the covariance matrix of this high-dimensional vector. Each element of the covariance matrix represents the covariance between two dimensions. Then, perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues and eigenvectors. Sort according to the magnitudes of the eigenvalues, and select the eigenvectors corresponding to the top k largest eigenvalues (assuming k = 30). Project the original high-dimensional vector into the low-dimensional space spanned by these k eigenvectors, thereby reducing the 100-dimensional vector to 30 dimensions. After such feature dimensionality reduction processing, the final geospatial entity association feature set is generated. While retaining important information, this geospatial entity association feature set reduces the data dimension.

[0084] In a possible implementation manner, step S150 includes:

[0085] Step S151: Construct a two-stream input architecture for the geospatial survey analysis model. The first input stream of the two-stream input architecture processes the geospatial entity association feature set, and the second input stream of the two-stream input architecture processes the survey task association feature set.

[0086] In this scenario, the geospatial entity association feature set contains the feature information after fusion processing of various geospatial entities such as mountains, rivers, and cities; the survey task association feature set covers relevant parameters such as path planning, sensor configuration, and data quality assessment in historical survey tasks.

[0087] Step S152: Perform spatial topology reasoning processing on the geospatial entity association feature set through the first input stream to generate a spatial topology reasoning result. The spatial topology reasoning processing includes constructing a graph attention network based on the entity node topology edge set in the geospatial entity association feature set, and iteratively updating the node features through node feature similarity calculation to generate an updated geospatial entity association feature set.

[0088] In a possible implementation, step S152 includes:

[0089] Step S1521, construct the adjacency matrix of the graph attention network based on the entity node topology edge set in the geographical entity association feature set, and the weight of the adjacency matrix is dynamically calculated and generated by the spatial distance similarity and geological attribute matching degree of the entity nodes.

[0090] In this embodiment, in the above region, there are multiple geographical entity nodes, such as entity nodes representing mountains, rivers, cities, etc. Taking the mountain node and the river node as examples, to determine their weights in the adjacency matrix, it is necessary to calculate the spatial distance similarity and geological attribute matching degree of the entity nodes.

[0091] When calculating the spatial distance similarity, assume that the spatial position coordinates of the mountain node are (1000, 2000), and the spatial position coordinates of the river node are (1500, 2200). First, calculate the differences in the horizontal and vertical coordinates. The horizontal coordinate difference is 1500 - 1000 = 500, and the vertical coordinate difference is 2200 - 2000 = 200. Then calculate the sum of the squares of the differences, that is, the square of 500 plus the square of 200. The square of 500 is 250000, and the square of 200 is 40000. The sum of the two is 290000. Then take the square root of it to get the spatial distance of about 538.5 (which may be fine-tuned according to specific distance measurement standards in actual calculations). Assume that the similarity is 1 when the spatial distance is between 0 - 300, and the similarity decreases linearly between 300 - 600. Then calculate the spatial distance similarity as (600 - 538.5) ÷ (600 - 300) × (1 - 0.5) + 0.5 = 0.6 (first calculate the distance ratio, and then calculate the similarity according to the decreasing rule).

[0092] For the geological attribute matching degree, assume that the mountain is mainly composed of granite, and the geology around the river is sedimentary rock. According to the pre-set geological attribute matching rules, the matching degree of different rock types is assumed to be 0.2 (the specific matching degree is determined according to detailed geological classification and association rules).

[0093] Integrate the spatial distance similarity and geological attribute matching degree to calculate the weight of the adjacency matrix. Assume that the weight of the spatial distance similarity is 0.6, and the weight of the geological attribute matching degree is 0.4. Then the weight of the adjacency matrix between the mountain node and the river node is 0.6 × 0.6 + 0.4 × 0.2 = 0.44 (multiply each part of the similarity by the corresponding weight and then add them). In this way, calculate the relationships between all entity nodes and construct the adjacency matrix of the graph attention network.

[0094] Step S1522, perform sparsification processing on the adjacency matrix, and remove the edges with weights lower than the preset threshold to generate a pruned topological edge subset.

[0095] Preset a threshold value, such as 0.3. Check each weight value in the adjacency matrix. For the edges with weights lower than 0.3, remove them. For example, there is a city node and a remote mountain area node. After calculation, the weight of the adjacency matrix between them is 0.2, which is lower than the preset threshold of 0.3. Then, remove this edge. Through such operations, a pruned subset of topological edges is generated, removing some weakly associated connections, making the graph structure more concise and highlighting the main relationships.

[0096] In step S1523, perform feature propagation processing on the pruned subset of topological edges through a multi-layer graph attention network, iteratively aggregating the attribute features of adjacent nodes to generate a spatial topological inference result containing global spatial dependency relationships.

[0097] Taking the first-layer graph attention network as an example, for each node, aggregate the attribute features of adjacent nodes according to its adjacent nodes and edge weights. Assume a node has its own attribute feature vector, and its adjacent nodes also have their own attribute feature vectors. Taking a mountain node as an example, one of its adjacent river nodes has attribute features such as flow rate and water quality, and another adjacent city node has attribute features such as population and building type. According to the weights in the adjacency matrix with the above adjacent nodes, perform weighted summation on the attribute features of the adjacent nodes. For example, the weight with the river node is 0.4, and the weight with the city node is 0.5. Multiply the attribute features of the river node by 0.4, multiply the attribute features of the city node by 0.5, and then add the product of its own attribute features multiplied by a weight (assume it is 0.1) to obtain the updated feature vector. This process is repeated in the multi-layer graph attention network, and each layer performs a new round of aggregation based on the updated node features of the previous layer. Through multiple iterations, the node can obtain information of nodes at farther distances, thereby generating a spatial topological inference result containing global spatial dependency relationships.

[0098] In step S153, perform task context modeling processing on the mapping task-associated feature set through the second input stream to generate task context features. The task context modeling processing includes performing temporal encoding on the historical mapping path planning parameters in the mapping task-associated feature set to generate a task temporal encoding vector, and performing cross-attention association processing on the task temporal encoding vector and the sensor configuration parameters.

[0099] In a possible implementation manner, step S153 includes:

[0100] In step S1531, perform spatio-temporal position encoding processing on the historical mapping path planning parameters to generate a path planning encoding vector containing timestamps and spatial coordinates.

[0101] In past surveying and mapping tasks, the time and path information of each surveying and mapping were recorded. For example, the first surveying and mapping task started on January 1, 2020, and the path started from the upper left corner of the area (coordinates (0, 0)) and advanced along a specific trajectory; the second surveying and mapping task was carried out on March 5, 2021, and the starting point of the path was (100, 100). The time information was digitized. Assuming the date was converted to the number of days from January 1, 2000, the distance for January 1, 2020 was 7305 days, and the distance for March 5, 2021 was 7670 days. The spatial coordinate information of the path was combined with the time information to form a path planning coding vector containing timestamps and spatial coordinates. For example, for the first surveying and mapping task, the coding vector could be represented as (7305, 0, 0) (this is only a simplified example, and actually may contain more dimensional information to accurately represent the path).

[0102] Step S1532: Perform a temporal modeling process on the path planning coding vector through a bidirectional gated recurrent unit to capture the forward and backward dependencies during the execution of historical tasks and generate a task temporal feature vector.

[0103] Taking the coding vectors of the first and second surveying and mapping tasks as examples, the influence of the path of the first surveying and mapping task on the path of the second surveying and mapping task and the change trend of the second surveying and mapping task relative to the first can be analyzed. During the processing, the bidirectional gated recurrent unit updates the hidden state at the current moment based on the input at the current moment and the hidden state at the previous moment. For example, when processing the coding vector of the second surveying and mapping task, it can refer to the hidden state of the coding vector of the first surveying and mapping task at the previous moment, combine the current input (i.e., the coding vector of the second surveying and mapping task), and determine how much information from the previous moment to retain and how much current information to receive through the gating mechanism. In this way, the forward and backward dependencies during the execution of historical tasks are captured, and a task temporal feature vector is generated. This task temporal feature vector can reflect the evolution and laws of the surveying and mapping path in the time dimension.

[0104] Step S1533: Perform multi-modal embedding processing on the sensor configuration parameters to map the discrete parameter types into continuous sensor feature vectors with the same dimension as the task temporal feature vector.

[0105] For example, a sensor of model ABC-900 has been used, and 6 bands are set for data acquisition. Then, the above discrete parameters are mapped into a continuous sensor feature vector that is consistent with the dimension of the task time series feature vector through a specific mapping function. Suppose the dimension of the task time series feature vector is 10. For the sensor model ABC-900, it is mapped into a 10-dimensional vector through a predefined embedding matrix, and the value of each dimension of the vector is determined according to the relevant features of the sensor model. For the band setting, similar mapping processing is also performed on the information of each band, and finally all sensor configuration parameters are mapped into a continuous sensor feature vector.

[0106] Step S1534, calculate the dynamic association weight between the task time series feature vector and the sensor feature vector through a cross-attention layer, and generate a task context feature that fuses the task context and the sensor configuration.

[0107] In this embodiment, the cross-attention layer will compare the eigenvalue of the task time series feature vector and the sensor feature vector in each dimension. For example, in a certain dimension, the value of the task time series feature vector is 0.6, and the value of the sensor feature vector is 0.8. By calculating the similarity between them, assuming the dot product method is used (in fact, a more complex calculation method may be used), the similarity value is 0.6×0.8 = 0.48. Such calculations are performed for all dimensions, and then through a softmax function (simply understood here as normalizing the above similarity value so that the sum is 1), an association weight distribution is obtained. According to the above weights, the task time series feature vector and the sensor feature vector are weighted and fused to generate a task context feature that fuses the task context and the sensor configuration. This task context feature synthesizes the relevant information of the historical mapping path and the sensor configuration at different times.

[0108] Step S154, input the updated geographical entity association feature set and the task context feature into a cross-flow fusion module, and perform alignment processing on the spatial topology feature and the task context feature through a multi-head attention mechanism to generate a jointly optimized feature set.

[0109] In a possible implementation manner, step S154 includes:

[0110] Step S1541, perform a linear transformation process on the updated geographical entity association feature set to generate a set of spatial topology query vectors.

[0111] In this embodiment, the updated geographical entity association feature set contains rich geographical entity information, such as the features of entities such as mountains, rivers, and cities after spatial topological reasoning processing. Taking the mountain entity as an example, its features may include multiple dimensions such as spatial location, geological attributes, and associations with other entities. Suppose the updated feature vector of the mountain entity is a 50-dimensional vector, and each dimension represents different feature information. Through a predefined linear transformation matrix, the size of which is 50×30 (the dimensions here are determined according to specific requirements and model design), the feature vector of the mountain entity is linearly transformed. The process of linear transformation is to multiply each dimension value of the feature vector by the element at the corresponding position in the linear transformation matrix, and then add the product results column by column to obtain a new 30-dimensional vector. For example, for the first dimension value 0.8 of the mountain feature vector, it is multiplied by the 30 elements in the first row of the linear transformation matrix respectively, and then the 30 products are added to obtain the first dimension value of the new vector; and so on, such operations are performed on all 50 dimensions, and finally a 30-dimensional spatial topological query vector is generated. Such linear transformation processing is performed on all entity feature vectors in the updated geographical entity association feature set to generate a set of spatial topological query vectors.

[0112] Step S1542: Perform linear transformation processing on the task context features to generate a set of task key vectors and a set of task value vectors.

[0113] The task context features incorporate information such as historical surveying and mapping paths and sensor configurations. Suppose the task context feature vector is a 40-dimensional vector, and it is also processed through two predefined linear transformation matrices. The first linear transformation matrix has a size of 40×20 and is used to generate a set of task key vectors. The task context feature vector is linearly transformed, and the calculation method is similar to the previous one, that is, each dimension value of the feature vector is multiplied by the corresponding element in the matrix and added column by column to obtain a 20-dimensional task key vector. This operation is performed on all task context feature vectors to generate a set of task key vectors. The second linear transformation matrix has a size of 40×20 and is used to generate a set of task value vectors. According to the same linear transformation method, the task context feature vector is transformed into a 20-dimensional task value vector, and then a set of task value vectors is generated.

[0114] Step S1543: Calculate the similarity scores between the set of spatial topological query vectors and the set of task key vectors to generate a cross-modal attention weight distribution.

[0115] Take a vector in the spatial topology query vector set and a vector in the task key vector set as an example. Suppose the spatial topology query vector is (0.6, 0.8, …, 0.4), and the task key vector is (0.5, 0.7, …, 0.3), both of which are 20-dimensional vectors. When calculating the similarity score, the dot product method is used. Multiply the values of the corresponding dimensions of the two vectors, and then add up all the products. That is, 0.6×0.5 + 0.8×0.7 + … + 0.4×0.3. Suppose the calculation result is 10.2 (more dimensions and complex data are involved in actual calculations). Perform such calculations for each vector in the spatial topology query vector set and each vector in the task key vector set to obtain a series of similarity scores. Then, process the above similarity scores through a softmax function. The role of the softmax function is to convert the above similarity scores into a probability distribution, making the sum of all scores equal to 1. This probability distribution is the cross-modal attention weight distribution. For example, after being processed by the softmax function, the probability corresponding to a certain similarity score is 0.2, indicating that the attention weight between the spatial topology query vector and the task key vector is 0.2.

[0116] Step S1544, perform weighted aggregation processing on the task value vector set according to the cross-modal attention weight distribution to generate a single-head attention feature set.

[0117] For each vector in the task value vector set, perform weighting according to the cross-modal attention weight distribution calculated previously. Suppose there are three vectors in the task value vector set, namely vector A (0.2, 0.3, …, 0.1), vector B (0.4, 0.5, …, 0.2), and vector C (0.1, 0.6, …, 0.3), and the corresponding attention weights are 0.3, 0.5, and 0.2 respectively. The process of weighted aggregation is to multiply each task value vector by the corresponding attention weight and then add the results. That is, (0.3×(0.2, 0.3, …, 0.1)) + (0.5×(0.4, 0.5, …, 0.2)) + (0.2×(0.1, 0.6, …, 0.3)). First, calculate the multiplications separately to get (0.06, 0.09, …, 0.03), (0.2, 0.25, …, 0.1), and (0.02, 0.12, …, 0.06). Then, add the three results corresponding to the dimensions to obtain a new vector (0.28, 0.46, …, 0.19), which is a vector in the single-head attention feature set. Perform such weighted aggregation processing on all combinations of spatial topology query vectors and task value vectors to generate a single-head attention feature set.

[0118] Step S1545, splice multiple single-head attention feature sets and perform non-linear transformation processing to generate a jointly optimized feature set after cross-stream fusion.

[0119] Suppose 5 single-head attention feature vectors are generated, namely vector 1, vector 2, vector 3, vector 4, and vector 5. These 5 vectors are concatenated in sequence to form a new high-dimensional vector. For example, vector 1 is 20-dimensional, vector 2 is 20-dimensional, and so on. The dimension of the concatenated vector is 20×5 = 100 dimensions. Then, a non-linear transformation is performed on this 100-dimensional vector using a non-linear function such as the ReLU function (Rectified Linear Unit). For each dimension value in the vector, if the value is greater than 0, it remains unchanged; if the value is less than 0, it is changed to 0. For example, a certain dimension value in the vector is -0.2, which becomes 0 after being processed by the ReLU function; a certain dimension value is 0.5, which remains unchanged. Through such non-linear transformation processing, a jointly optimized feature set after cross-stream fusion is generated, which fuses the spatial topological information of geographical entities and the context information of mapping tasks.

[0120] Step S155: Based on the jointly optimized feature set, perform multi-task prediction processing to generate a mapping path planning parameter set, a sensor band selection parameter set, and a data acquisition quality control threshold set.

[0121] In a possible implementation manner, step S155 includes:

[0122] Step S1551: Perform spatial rasterization processing on the jointly optimized feature set through a path planning prediction branch to generate a path feasibility probability map, and search for an optimal path sequence in the path feasibility probability map based on the dynamic programming algorithm to generate a mapping path planning parameter set.

[0123] For example, the target area can be divided into spatial grids with a size of 10×10 meters. For each grid, the path feasibility probability is calculated based on the geographical entity information and mapping task information in the jointly optimized feature set. For example, if a certain grid is located in a mountainous area and, according to historical mapping data and geographical entity features, the terrain in this area is complex and the passage is difficult, then combining the above information, through a pre-trained model or rule, the path feasibility probability of this grid is calculated to be 0.2. For a grid located on a plain and close to an existing mapping path, the path feasibility probability is calculated to be 0.8. In this way, a path feasibility probability map of the entire target area is generated. Based on the dynamic programming algorithm, the optimal path sequence is searched in this path feasibility probability map. Starting from the starting point, consider the path feasibility probability of each grid and the cost to the next grid (such as the cost determined by factors such as distance and terrain complexity). Assume the starting point is (0, 0), and there are several adjacent grids around it. Calculate the comprehensive score from the starting point to each adjacent grid, and the comprehensive score takes into account the path feasibility probability and the cost. For example, the path feasibility probability to an adjacent grid A is 0.6, and the cost is 2; the path feasibility probability to adjacent grid B is 0.7, and the cost is 3. Through a certain calculation method (such as dividing the probability by the cost), the comprehensive score of grid A is 0.3, and the comprehensive score of grid B is 0.23. Select the grid with the highest comprehensive score as the next point of the path, and then use this new point as the starting point to continue repeating the above process until reaching the end point or traversing all possible paths. In this way, a set of mapping path planning parameters is generated to determine parameters such as flight altitude, flight speed, and heading angle. For example, the flight altitude of a certain section of the planned path is 450 meters, the flight speed is 35 meters per second, and the heading angle is 45 degrees.

[0124] Step S1552, perform band importance ranking processing on the jointly optimized feature set through the sensor configuration branch, calculate the spectral resolution contribution degree and terrain matching degree of each band, and generate a set of band selection priority parameters.

[0125] The combined optimization feature set contains the topographic information of geographical entities and sensor-related information. For each band, its performance under different terrains is analyzed. Taking the near-infrared band as an example, in the vegetation-covered area, the near-infrared band can well reflect the health status and growth of vegetation, and the contribution degree of spectral resolution is assumed to be 0.8. The vegetation coverage area in this region is relatively large, and the matching degree with the terrain is relatively high, assumed to be 0.7. For the green light band in the visible light, in the water area, it can clearly reflect information such as the color and transparency of the water body. The contribution degree of spectral resolution is assumed to be 0.6, and the water body is widely distributed in this region, and the matching degree with the terrain is 0.6. By comprehensively considering the contribution degree of spectral resolution and the terrain matching degree of each band under different terrains, the comprehensive score of each band is calculated. For example, the comprehensive score of the near-infrared band = 0.8×0.7 = 0.56; the comprehensive score of the green light band = 0.6×0.6 = 0.36. According to the above comprehensive scores, the bands are sorted to generate a set of band selection priority parameters, and the selection order such as the near-infrared band being prior to the green light band is determined.

[0126] Step S1553, perform distribution consistency analysis on the combined optimization feature set through the quality control branch, generate a dynamic confidence interval threshold based on the historical data quality evaluation index, and generate a set of data acquisition quality control thresholds.

[0127] The historical data quality evaluation indexes include image clarity, data integrity, etc. Assume that the average value of image clarity in the historical data is 80%, and the standard deviation is 5%. According to statistical principles, the dynamic confidence interval threshold is calculated at a 95% confidence level. First, calculate the lower limit of the confidence interval. The lower limit = average value - 1.96×standard deviation (1.96 is the Z value corresponding to the 95% confidence level), that is, 80% - 1.96×5% = 70.2%; the upper limit = average value + 1.96×standard deviation, that is, 80% + 1.96×5% = 89.8%. In this way, a set of data acquisition quality control thresholds is generated, stipulating that during the subsequent data acquisition process, the image clarity should be maintained between 70.2% and 89.8%.

[0128] Step S1554, balance the gradient backpropagation ratio of the multi-task prediction branch through the adaptive loss weight allocation mechanism, and iteratively optimize the prediction accuracy of each branch.

[0129] During the multi-task prediction process, the path planning prediction branch, the sensor configuration branch, and the quality control branch will all generate prediction results and corresponding losses. The adaptive loss weight allocation mechanism will automatically adjust the weights according to the loss situation of each branch. For example, in a certain round of training, the loss of the path planning prediction branch is relatively large, while the losses of the sensor configuration branch and the quality control branch are relatively small. According to the adaptive mechanism, the proportion of gradient backpropagation of the path planning prediction branch will be appropriately increased, and the proportion of gradient backpropagation of the other two branches will be reduced. Suppose the original proportion of gradient backpropagation of the path planning prediction branch is 0.3, the sensor configuration branch is 0.3, and the quality control branch is 0.4. Since the loss of the path planning prediction branch is large, its proportion of gradient backpropagation is adjusted to 0.5, the sensor configuration branch is adjusted to 0.2, and the quality control branch is adjusted to 0.3. In this way, the gradient backpropagation of each branch is dynamically balanced during each training, and the prediction accuracy of each branch is iteratively optimized, so that the entire geodetic survey analysis model can more accurately generate the survey path planning parameter set, the sensor band selection parameter set, and the data acquisition quality control threshold set, providing a more optimized strategy for the actual survey work.

[0130] Step S156, perform multi-task prediction processing based on the jointly optimized feature set to generate a survey path planning parameter set, a sensor band selection parameter set, and a data acquisition quality control threshold set, construct a loss function according to the parameter differences between the survey path planning parameter set, the sensor band selection parameter set, and the data acquisition quality control threshold set and the historical survey task log data, and optimize the parameters of the geodetic survey analysis model through backpropagation to generate a geodetic survey optimization strategy.

[0131] For example, the flight altitude of a certain section of the path in the historical survey task log is 450 meters, while the currently planned flight altitude is 500 meters. Calculate the difference between the two as 50 meters; there are differences between the sensor band selection in the historical data and the currently generated priority parameter set, and calculate the difference value through a certain quantization method. Combine the above difference values, and according to the set loss function formula (such as the idea of mean square error, taking the sum of the squares of each parameter difference as the loss value), calculate the loss value. Through the backpropagation algorithm, the loss value is backpropagated to each parameter of the geodetic survey analysis model, and the parameters are adjusted to reduce the loss value. After multiple iterative trainings, the parameters of the geodetic survey analysis model are optimized, and finally a geodetic survey optimization strategy is generated to guide the subsequent survey work in the target area and improve the survey efficiency and data quality.

[0132] For example, in a possible implementation manner, step S150 further includes:

[0133] Step S157: Convert the set of mapping path planning parameters into a set of flight control instructions, where the set of flight control instructions includes a sequence of flight altitudes, a sequence of airspeeds, and a sequence of heading angles.

[0134] In this embodiment, the set of mapping path planning parameters includes detailed path planning information. For example, when mapping in this area, the planned path needs to cross mountains, rivers, and pass by some urban areas. The optimized set of mapping path planning parameters clarifies the flight altitude, airspeed, and heading angle at each stage. Taking a specific path planning as an example, during the process of traveling from point A to point B in the area, the sequence of flight altitudes is as follows: the initial flight altitude is 400 meters, and it gradually rises to 600 meters when approaching the mountains to avoid the higher terrain of the mountains; the sequence of airspeeds is: the initial airspeed is 40 m / s, it is appropriately increased to 45 m / s in relatively flat areas, and decreased to 35 m / s when approaching complex terrain to ensure the accuracy and safety of mapping; the sequence of heading angles is: the heading angle is 30 degrees when starting from point A, and it is adjusted to 45 degrees according to the river direction when passing through the river. The above parameters of flight altitude, airspeed, and heading angle are converted into a set of flight control instructions in a specific format, and each instruction in the set of flight control instructions clarifies the flight parameters that the mapping terminal should follow at a certain moment or stage.

[0135] Step S158: Input the set of flight control instructions into the navigation control system of the mapping terminal to generate a sequence of real-time waypoint coordinates, and perform collision detection processing on the sequence of real-time waypoint coordinates and the preset mapping area boundary to generate a set of waypoint correction instructions.

[0136] After the navigation control system of the surveying and mapping terminal receives the flight control instruction set, it calculates the real-time waypoint coordinate sequence according to the current position information and time information of the surveying and mapping terminal, combined with the parameters in the flight control instruction. For example, if the surveying and mapping terminal is currently located at the coordinates (1000, 2000), and according to the flight control instruction, it flies at a speed of 40 meters per second and a heading angle of 30 degrees. After 10 seconds, through trigonometric calculation (here is a detailed explanation: in the plane rectangular coordinate system, when flying along a 30-degree heading angle, the horizontal displacement is the speed multiplied by the time and then multiplied by cos30 degrees, that is, 40×10×cos30 degrees is approximately equal to 400×0.866 = 346.4; the vertical displacement is the speed multiplied by the time and then multiplied by sin30 degrees, that is, 40×10×sin30 degrees = 400×0.5 = 200. The original coordinates are (1000, 2000), then the new coordinates are (1000 + 346.4, 2000 + 200), that is, (1346.4, 2200)), and the next waypoint coordinate is obtained as (1346.4, 2200), and so on to generate the entire real-time waypoint coordinate sequence. The preset surveying and mapping area boundary is a pre-set range. For example, in the surveying of this area, the preset boundary is a rectangular area, and the coordinates of its four vertices are (500, 1500), (1500, 1500), (1500, 2500), and (500, 2500) respectively. Compare each coordinate in the real-time waypoint coordinate sequence with the preset surveying and mapping area boundary to check whether there is a waypoint exceeding the boundary. If there is a waypoint exceeding the boundary, for example, the calculated coordinate of a certain waypoint is (1600, 2300), which exceeds the maximum value of the x-axis of the preset boundary, which is 1500, then a waypoint correction instruction set is generated. The waypoint correction instruction set will recalculate the appropriate waypoint coordinates according to the situation of exceeding the boundary. For example, adjust this waypoint to (1500, 2300) to ensure that the surveying and mapping path is always within the preset area.

[0137] Step S159, merge the waypoint correction instruction set and the sensor band selection parameter set into a device control signal set, and trigger the filter switching operation and exposure parameter adjustment operation of the multispectral camera of the surveying and mapping terminal.

[0138] For the sensor band selection parameter set, the bands to be selected under different regions and mapping requirements have been determined according to the previous multi-task prediction processing. For example, in the vegetation-covered area, based on the analysis of the spectral resolution contribution degree and terrain matching degree of each band, it is determined to preferentially select the near-infrared band and the green band. After the waypoint correction instruction set and the sensor band selection parameter set are merged, an equipment control signal set is formed. The equipment control signal set transmits the above information to the mapping terminal, triggering the filter switching operation of the multi-spectral camera. For example, when the mapping terminal reaches the waypoint in the vegetation-covered area, the sensor band selection information in the equipment control signal set instructs the multi-spectral camera to switch to the filters corresponding to the near-infrared band and the green band. At the same time, according to the lighting conditions and mapping requirements of this area, the equipment control signal set will also adjust the exposure parameters. Suppose in this area, based on the analysis of historical data and the current environment, the calculated appropriate exposure time is 0.05 seconds and the aperture size is f / 8. The equipment control signal set transmits the above exposure parameters to the multi-spectral camera to adjust its exposure settings to ensure that the quality of the acquired image data meets the requirements.

[0139] Step S1510, perform signal-to-noise ratio analysis and geometric distortion detection processing on the real-time image data collected by the mapping terminal based on the data acquisition quality control threshold set, and generate a quality anomaly marker data set.

[0140] The set of data acquisition quality control thresholds stipulates the ranges of various quality indicators for image data. For example, in the surveying and mapping of this area, it is stipulated that the signal-to-noise ratio of the image data should not be lower than 30 dB, and the geometric distortion rate should not exceed 5%. For the real-time image data collected by the surveying and mapping terminal, signal-to-noise ratio analysis is carried out. Taking a certain frame of image data as an example, first calculate the total intensity of the signal in this frame of image. Assuming that by statistically summing the gray values of all pixels in the image, the total signal intensity is obtained as 10,000 (this is just an example to illustrate the calculation method, and the actual calculation may be more complex). Then calculate the intensity of the noise. By analyzing the distribution and characteristics of the random noise in the image, a specific algorithm is used to estimate the noise intensity. Assuming that the estimated noise intensity is 100. According to the definition of the signal-to-noise ratio (signal-to-noise ratio = total signal intensity / noise intensity), the signal-to-noise ratio of this frame of image is 10,000 / 100 = 100 dB (the actual calculation may involve more accurate measurement and calculation methods). For geometric distortion detection, by comparing with a pre-set standard image geometric model, analyze whether the shape, position, etc. of the objects in the image have deformed. For example, in the standard image, if a straight line appears significantly bent in the actual image, determine the geometric distortion rate by measuring and calculating the degree of bending. Assuming that in the image of a certain area, the edge of an object that should originally be a straight line has a certain degree of bending. After measurement and calculation, the geometric distortion rate is 3%. Compare the above analysis results with the set of data acquisition quality control thresholds. If the signal-to-noise ratio is lower than 30 dB or the geometric distortion rate exceeds 5%, then mark this part of the image data as quality abnormal and generate a set of quality abnormal marked data.

[0141] Step S1511, input the set of quality abnormal marked data into the geographical surveying and mapping analysis model for resampling path prediction processing, generate a set of supplementary surveying and mapping path parameters, and update the set of surveying and mapping path planning parameters.

[0142] In this embodiment, the quality anomaly marked data set contains relevant information of the collected image data that does not meet the quality requirements, such as the location where the quality anomaly occurs, the corresponding waypoint coordinates, etc. Input the above data into the geodetic survey and analysis model, and the geodetic survey and analysis model analyzes based on the above information and data such as the previous geographical entity association feature set and the survey task association feature set. For example, it is found that the quality anomaly of the image data in a certain area is caused by the unreasonable settings of the flight altitude and speed of the survey path in this area. Through the analysis of the geographical entity features (such as terrain undulation, geological structure, etc.) and historical survey data in this area, the geodetic survey and analysis model re-predicts a more suitable resampling path. Suppose the original survey path has a flight altitude of 500 meters and a speed of 40 meters per second in this area. After model analysis, the resampling path is planned to adjust the flight altitude to 550 meters and reduce the speed to 30 meters per second to ensure that better-quality image data can be collected in this area. After generating the supplementary survey path parameter set, merge and update it with the original survey path planning parameter set. For example, find the corresponding area and flight segment in the original survey path planning parameter set, and adjust parameters such as flight altitude and speed according to the supplementary survey path parameter set, so as to update the entire survey path planning parameter set, provide more accurate and optimized path guidance for subsequent survey work, and improve the quality and efficiency of survey data.

[0143] Figure 2 FIG. shows a schematic diagram of exemplary hardware and software components of an artificial intelligence-based terrestrial geodetic survey model generation system 100 that can implement the ideas of the present application. For example, the processor 120 can be used on the artificial intelligence-based terrestrial geodetic survey model generation system 100 and is used to execute the functions in the present application.

[0144] The artificial intelligence-based terrestrial geodetic survey model generation system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the artificial intelligence-based terrestrial geodetic survey model generation method of the present application. Although only one server is shown in the present application, for convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0145] For example, the artificial intelligence-based terrestrial geographic mapping model generation system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as disks, ROM, or RAM, or any combination thereof. Exemplarily, the artificial intelligence-based terrestrial geographic mapping model generation system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of the present application may be implemented according to these program instructions. The artificial intelligence-based terrestrial geographic mapping model generation system 100 further includes an I / O interface 150 between the computer and other input / output devices.

[0146] For ease of explanation, only one processor is described in the artificial intelligence-based terrestrial geographic mapping model generation system 100. However, it should be noted that the artificial intelligence-based terrestrial geographic mapping model generation system 100 in the present application may also include multiple processors. Therefore, the steps performed by one processor described in the present application may also be jointly performed or separately performed by multiple processors. For example, if the processor of the artificial intelligence-based terrestrial geographic mapping model generation system 100 performs step A and step B, it should be understood that step A and step B may also be jointly performed by two different processors or separately performed in one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0147] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the above artificial intelligence-based terrestrial geographic mapping model generation method is implemented.

[0148] It should be noted that, in order to simplify the description of the present invention disclosure and thus help the understanding of one or more embodiments of the invention, in the previous description of the embodiments of the present invention, sometimes multiple features are merged into one embodiment, drawing, or description thereof.

Claims

1. A method for generating a terrestrial geographic mapping model based on artificial intelligence, characterized in that, The method includes: Obtaining a set of geodetic survey data for a target area, where the set of geodetic survey data includes satellite remote sensing image data, surface elevation data, geological structure feature data, and historical survey task log data; Performing terrain feature interpretation processing on the set of geodetic survey data to generate a set of terrain semantic features, a set of surface coverage features, and a set of survey task association features; Invoking a geospatial encoder to perform multi-scale spatial modeling processing on the set of terrain semantic features and the set of surface coverage features to generate a set of geospatial representation vectors for the target area; Performing dynamic feature fusion processing based on the set of geospatial representation vectors and a preset geospatial entity topology map, where the geospatial entity topology map includes spatial topology relationships and attribute association relationships between entity nodes, to generate a set of geospatial entity association features; Training a geodetic survey analysis model according to the set of geospatial entity association features and the set of survey task association features to generate an optimized geodetic survey strategy for the target area, and feeding back the optimized geodetic survey strategy to a survey terminal to trigger dynamic adjustment operations of survey parameters; The training of the geodetic survey analysis model according to the set of geospatial entity association features and the set of survey task association features to generate an optimized geodetic survey strategy for the target area includes: Constructing a dual-stream input architecture for the geodetic survey analysis model, where the first input stream of the dual-stream input architecture processes the set of geospatial entity association features, and the second input stream of the dual-stream input architecture processes the set of survey task association features; Performing spatial topology reasoning processing on the set of geospatial entity association features through the first input stream to generate a spatial topology reasoning result. The spatial topology reasoning processing includes constructing a graph attention network based on a set of topological edges between entity node sets in the set of geospatial entity association features, and iteratively updating node features through node feature similarity calculation to generate an updated set of geospatial entity association features; Performing task context modeling processing on the set of survey task association features through the second input stream to generate task context features. The task context modeling processing includes performing temporal encoding on historical survey path planning parameters in the set of survey task association features to generate a task temporal encoding vector, and performing cross-attention association processing on the task temporal encoding vector and sensor configuration parameters; Inputting the updated set of geospatial entity association features and the task context features into a cross-stream fusion module, and performing alignment processing on spatial topology features and task context features through a multi-head attention mechanism to generate a set of joint optimization features; Performing multi-task prediction processing based on the set of joint optimization features to generate a set of survey path planning parameters, a set of sensor band selection parameters, and a set of data acquisition quality control threshold values; Constructing a loss function according to parameter differences between the set of survey path planning parameters, the set of sensor band selection parameters, the set of data acquisition quality control threshold values, and parameters in the historical survey task log data, and optimizing parameters of the geodetic survey analysis model through backpropagation to generate an optimized geodetic survey strategy.

2. The method for generating a terrestrial geographic mapping model based on artificial intelligence according to claim 1, wherein, Performing terrain feature interpretation processing on the geographical mapping data set to generate a terrain semantic feature set, a surface coverage feature set, and a mapping task correlation feature set, including: Performing multi-scale image segmentation processing on the satellite remote sensing image data to generate a plurality of image segmentation units, each image segmentation unit including pixel-level spectral features and spatial position encoding; Invoking a pre-trained semantic segmentation model to perform ground object category annotation processing on each image segmentation unit to generate a surface coverage feature set, wherein the ground object categories include vegetation coverage types, water body distribution areas, and artificial building areas; Performing terrain gradient calculation processing on the surface elevation data to generate a terrain undulation feature set, and performing fault line identification processing on the geological structure feature data to generate a geological structure boundary feature set; Performing feature fusion processing on the terrain undulation feature set and the geological structure boundary feature set to generate a terrain semantic feature set; Performing task parsing processing on the historical mapping task log data, extracting historical mapping path planning parameters, sensor configuration parameters, and data quality evaluation indicators, and generating a mapping task correlation feature set.

3. The method for generating a terrestrial geographic mapping model based on artificial intelligence according to claim 1, characterized in that, Invoking a geospatial encoder to perform multi-scale spatial modeling processing on the terrain semantic feature set and the surface coverage feature set to generate a geospatial representation vector set of the target area, including: Constructing a multi-scale processing architecture of the geospatial encoder, the multi-scale processing architecture including a local feature extraction module, a regional context modeling module, and a global spatial fusion module; Performing a sliding window process on the surface coverage features of a single image segmentation unit through the local feature extraction module to extract local texture features and edge features; Calculating spatial association weights for the terrain semantic features of adjacent image segmentation units through the regional context modeling module to generate regional context association weights, and performing weighted aggregation processing on the terrain semantic features of adjacent image segmentation units based on the regional context association weights to obtain regional aggregation features; Performing cross-scale dependence relationship analysis on the multi-scale features through the global spatial fusion module to generate a global spatial dependence relationship graph; Performing feature stitching processing on the local texture features, the regional aggregation features, and the global dependence relationship graph to generate a geospatial representation vector set; wherein the spatial association weight calculation is realized by analyzing the feature similarity and spatial distance difference between adjacent image segmentation units.

4. The method for generating a terrestrial geographic mapping model based on artificial intelligence according to claim 1, wherein Performing dynamic feature fusion processing based on the geospatial representation vector set and a preset geospatial entity topology map to generate a geospatial entity association feature set, including: Extracting an entity node set of the target area and a topological edge set between the entity node sets from the geospatial entity topology map, wherein each entity node includes the spatial position attribute of the entity node, the geological attribute of the entity node, and the historical mapping record attribute of the entity node; Calculate the multi-dimensional similarity set between each geospatial representation vector in the geospatial representation vector set and each entity node in the entity node set. The multi-dimensional similarity set includes the spatial distance similarity between the geospatial representation vector and the entity node, the geological attribute matching degree between the geospatial representation vector and the entity node, and the historical survey record association degree between the geospatial representation vector and the entity node; Construct a dynamic feature selection gating mechanism based on the multi-dimensional similarity set. The dynamic feature selection gating mechanism generates a set of dynamic feature weight coefficients corresponding to each entity node. Among them, the set of dynamic feature weight coefficients contains the feature enhancement weight of the entity node and the feature suppression weight of the entity node; Perform weighted processing on the entity node attributes of the entity node set according to the set of dynamic feature weight coefficients to generate a weighted set of entity node attributes. Among them, the weighted set of entity node attributes includes the entity node attributes strengthened by the feature enhancement weight and the entity node attributes weakened by the feature suppression weight; Perform cross-modal feature splicing processing on the weighted set of entity node attributes and the geospatial representation vector set to generate a fused set of geospatial entity association features; Perform feature dimensionality reduction processing on the fused set of geospatial entity association features to generate the set of geospatial entity association features.

5. The method for generating a terrestrial geographic mapping model based on artificial intelligence according to claim 1, wherein The spatial topology inference processing of the set of geospatial entity association features through the first input stream to generate a spatial topology inference result includes: Construct an adjacency matrix of the graph attention network based on the topology edge set between the entity node sets in the set of geospatial entity association features. The weight of the adjacency matrix is dynamically calculated and generated by the spatial distance similarity and geological attribute matching degree of the entity node; Perform sparsification processing on the adjacency matrix to remove the edges with weights lower than the preset threshold to generate a pruned subset of topology edges; Perform feature propagation processing on the pruned subset of topology edges through a multi-layer graph attention network, iteratively aggregate the attribute features of adjacent nodes, and generate a spatial topology inference result containing global spatial dependence relationships.

6. The method for generating a terrestrial geographic mapping model based on artificial intelligence according to claim 1, wherein The task context modeling processing of the survey task association feature set through the second input stream to generate task context features includes: Perform spatio-temporal position encoding processing on the historical survey path planning parameters to generate a path planning encoding vector containing timestamps and spatial coordinates; Perform temporal sequence modeling processing on the path planning encoding vector through a bidirectional gated recurrent unit to capture the forward and backward dependence relationships during the historical task execution process and generate a task temporal feature vector; Perform multi-modal embedding processing on the sensor configuration parameters to map the discrete parameter types into continuous sensor feature vectors with the same dimension as the task temporal feature vector; Calculate the dynamic association weight between the task temporal feature vector and the sensor feature vector through a cross-attention layer to generate task context features that fuse task context and sensor configuration.

7. The method for generating an artificial intelligence-based terrestrial geographic mapping model according to claim 1, wherein Input the updated geographical entity association feature set and the task context features into the cross-flow fusion module, and align the spatial topological features and the task context features through the multi-head attention mechanism to generate a jointly optimized feature set, including: Perform a linear transformation on the updated geographical entity association feature set to generate a set of spatial topological query vectors; Perform a linear transformation on the task context features to generate a set of task key vectors and a set of task value vectors; Calculate the similarity scores between the set of spatial topological query vectors and the set of task key vectors to generate a cross-modal attention weight distribution; Perform a weighted aggregation on the set of task value vectors according to the cross-modal attention weight distribution to generate a set of single-head attention features; Concatenate multiple sets of single-head attention features and perform a non-linear transformation to generate a jointly optimized feature set after cross-flow fusion.

8. The method for generating an artificial intelligence-based terrestrial geographic mapping model according to claim 1, wherein Perform multi-task prediction processing based on the jointly optimized feature set to generate a set of mapping path planning parameters, a set of sensor band selection parameters, and a set of data acquisition quality control threshold values, including: Perform a spatial rasterization process on the jointly optimized feature set through the path planning prediction branch to generate a path feasibility probability map, and search for an optimal path sequence in the path feasibility probability map based on the dynamic programming algorithm to generate a set of mapping path planning parameters; Perform a band importance ranking process on the jointly optimized feature set through the sensor configuration branch, calculate the spectral resolution contribution degree and terrain matching degree of each band, and generate a set of band selection priority parameters; Perform a distribution consistency analysis on the jointly optimized feature set through the quality control branch, and generate a dynamic confidence interval threshold based on the historical data quality evaluation index to generate a set of data acquisition quality control threshold values; Balance the gradient backpropagation ratio of the multi-task prediction branches through the adaptive loss weight allocation mechanism, and iteratively optimize the prediction accuracy of each branch.

9. An artificial intelligence-based terrestrial geographic mapping model generation system, characterized in that, The artificial intelligence-based terrestrial geographical mapping model generation system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the artificial intelligence-based terrestrial geographical mapping model generation method according to any one of claims 1-8 above.

Images