A feature recognition method and system for geographic objects facing three-dimensional GIS

Abstract

The present application relates to a kind of element identification method and system of geographic object for three-dimensional GIS, wherein the method includes: intercepting the visual screenshot of target geographic object and selecting multiple grid points, obtaining the geographic coordinates corresponding to each grid point pixel coordinate, form sparse control point pair and be serialized as sparse text;Obtain elevation value and elevation value statistics and be serialized as elevation text;Visual feature description text is generated based on visual description text;Splicing sparse text, elevation text and visual feature description text get splicing prior vector;Based on splicing prior vector and visual screenshot output normalized pixel coordinates;Visual screenshot and normalized pixel coordinates are input into separated two-stage visual segmentation model to extract pixel coordinate sequence, obtain pixel coordinate sequence;It is pasted back to visual screenshot, obtain full map coordinate sequence;It is converted into vector feature under geographic coordinates, obtain the identification result of target geographic object.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional geographic information system technology, and mainly to a method and system for feature identification of geographic objects in three-dimensional GIS. Background Technology

[0002] 3D Geographic Information Systems (3D GIS) have become a core foundational platform for spatial planning, emergency command, resource inspection, and terrain analysis. Traditional 3D GIS relies on engines like ArcGIS and Cesium to achieve basic functions such as 3D scene rendering, viewpoint navigation, and single-point geographic picking, but it still has significant shortcomings in intelligent interaction. Meanwhile, artificial intelligence technologies such as large language models, multimodal visual models, and general image segmentation models are developing rapidly, providing a technological foundation for the transformation of GIS from traditional manual operation to natural language-driven, visually intelligent interaction. Currently, the industry has attempted to combine various AI models with 3D GIS, but existing technical solutions suffer from architectural flaws, insufficient scene adaptation, low performance, and fragmented links, failing to meet the actual business needs of open-word querying, pixel-level target segmentation, and automated vector feature import in real-world 3D scenes.

[0003] Chinese invention patent application CN120339399A discloses a method and apparatus for enhancing the spatial perception capability of a multimodal large model. The technical solution includes extracting feature description information of at least one object in a target image and / or target video using a multimodal large model, and generating initial structured data of the at least one object based on the feature description information; performing gridding processing on the target image and / or target video, and adding visual cues containing location information to multiple grids in the target image and / or target video, to combine the visual cues and the initial structured data to generate structured data containing coordinate information and description information of the at least one object; based on the structured data, locating the target region corresponding to the at least one object in the target image and / or target video, and optimizing the coordinate information through the target region to obtain the actual object coordinates of the at least one object; mapping the actual object coordinates back to the system coordinates of the target image and / or target video to obtain the actual spatial positioning result of the at least one object; however, the above... The aforementioned technical solutions only perform meshing on ordinary images / videos. The location information carried by the mesh is merely the image's own pixel index and template virtual coordinates, which cannot establish a mapping between screen pixels and real geographic coordinates. It is only applicable to ordinary two-dimensional planar images and cannot output real geographic coordinates, only obtaining virtual coordinates of the image system, which cannot be directly used for GIS spatial business. At the same time, the above technical solutions only coarsely locate the target area and optimize the basic coordinates. They do not introduce a two-stage segmentation model with encoder-decoder separation such as SAM / SAM2, and cannot extract the target pixel-level contours. They can only obtain rough coordinates of rectangular boxes and cannot generate polygonal vector features, lacking the ability to extract fine-grained feature boundaries. In addition, the above technical solutions only complete the coarse location coordinate optimization. They do not have a complete process of automatically converting the normalized boxes output by the multimodal model into segmentation model point / box prompts, adaptive cropping, and coordinate pasting. Manual intervention is required to generate segmentation prompts, and it is impossible to achieve a fully automated flow from natural language to detection to fine segmentation.

[0004] Therefore, there is an urgent need for an intelligent visual interaction method that can adapt to 3D GIS scenarios, accurately complete georeferencing and fine feature segmentation, and automatically generate analyzable vector elements. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a method and system for feature identification of geographic objects in 3D GIS.

[0006] The technical solution of the present invention is as follows: On the one hand, this invention proposes a feature identification method for geographic objects in 3D GIS, the method comprising: S1. Capture a visual screenshot of the target geographic object; select multiple grid points on the visual screenshot and obtain the geographic coordinates corresponding to the pixel coordinates of each selected grid point; construct sparse control point pairs based on the pixel coordinates and geographic coordinates, and serialize them into sparse text. S2. Obtain the elevation value of the target geographic object, calculate the elevation value statistics based on the elevation value, and serialize it into elevation text; input the visual description text of the target geographic object, and input the preset large language model to generate visual feature description text; S3. Concatenate the sparse text, elevation text, and visual feature description text to obtain a concatenated prior vector; input the concatenated prior vector and visual screenshot into a multimodal visual model for recognition, and output the normalized pixel coordinates of the target geographic object; input the visual screenshot and normalized pixel coordinates into a two-stage split visual segmentation model to extract the pixel coordinate sequence of the target geographic object and normalize it to obtain a normalized pixel coordinate sequence. S4. Paste the normalized pixel coordinate sequence back to the visual screenshot to obtain the normalized coordinate sequence of the whole image; convert the normalized coordinate sequence of the whole image into vector features corresponding to the target geographic object under geographic coordinates to obtain the recognition result of the target geographic object.

[0007] Preferably, steps S1-S4 are all executed under the same viewport state at the same time, wherein the viewport state is uniquely determined by the viewport range and the viewing angle direction.

[0008] Preferably, the concatenated prior vectors and visual screenshots are input into the multimodal vision model. The specific steps are as follows: The visual screenshot is standardized to obtain a standardized screenshot; The prior vector and the preset coordinate reference anchoring instruction are concatenated into a prompt text. The coordinate reference anchoring instruction is used to indicate that the geometric result output by the multimodal vision model is represented in normalized coordinate form. The prompt text and standardized screenshots are combined into a multimodal input payload according to the interface specifications of the multimodal vision model; The multimodal input payload is submitted to the inference engine of the multimodal vision model to generate the original normalized coordinate results; Clamping is performed on each coordinate value of the original normalized coordinate result to constrain each coordinate value to the interval [0,1], thereby obtaining the normalized pixel result of the target geographic object.

[0009] Preferably, each coordinate value is constrained to the interval [0,1], where 0 corresponds to the left or lower boundary of the standardized image, and 1 corresponds to the right or upper boundary of the standardized image, specifically: If the current coordinate value is less than 0, then the current coordinate value will be forcibly assigned to 0; If the current coordinate value is greater than 1, then the current coordinate value will be forcibly assigned the value 1; If the current coordinate value is between 0 and 1, then the current coordinate value remains unchanged.

[0010] Preferably, the split two-stage visual segmentation model includes a preloaded image encoder and a lazy-loaded cue decoder, wherein: The image encoder is used to receive input visual screenshots and output dense feature embedding tensors; The cue decoder is used to receive the input dense feature embedding tensor and coordinate cue, and generate a probability mask map based on the dense feature embedding tensor and coordinate cue.

[0011] Preferably, the specific steps for segmenting the coordinate results based on the split two-stage visual segmentation model are as follows: The normalized bounding box in the coordinate results is converted into coordinate cues for the split two-stage visual segmentation model. The coordinate cues include the center point, top left corner point, and bottom right corner point of the normalized bounding box. Based on the center point of the normalized bounding box, a preset cropping window is driven to crop a visual sub-image from the visual screenshot that matches the input resolution of the image encoder. Input the visual subgraph into the image encoder to obtain the dense feature embedding tensor corresponding to the cropped subgraph; The coordinate cue and dense features are embedded into the tensor input cue decoder to obtain the probability mask of the clipped sub-image, and then binarized to obtain the binary mask. A contour tracing algorithm is performed on the connected regions with a value of 1 in the binary mask image to extract the pixel coordinate sequence of the target geographic object in the cropped sub-image, and then normalized to obtain the normalized pixel coordinate sequence.

[0012] Preferably, the specific steps for storing vector features in the vector business layer of the 3D GIS are as follows: Determine whether a persistent segmented layer exists in a 3D GIS; If it does not exist, create a persistent segmentation layer; If they exist, the vector features will be automatically added as new features to the persistent segmentation layer.

[0013] On the other hand, the present invention also provides a feature identification system for geographic objects in 3D GIS, the system comprising: The image acquisition module captures visual screenshots of the target geographic object; The text generation module selects multiple grid points at fixed intervals on the visual screenshot, obtains the geographic coordinates corresponding to the pixel coordinates of each selected grid point, constructs sparse control point pairs based on the pixel coordinates and geographic coordinates, and serializes them into sparse text; inputs the visual description text of the target geographic object, and inputs a preset large language model to generate visual feature description text. The coordinate segmentation module concatenates sparse text, elevation text, and visual feature description text to obtain a concatenated prior vector. The concatenated prior vector and visual screenshot are input into a multimodal visual model for recognition, and the normalized pixel coordinates of the target geographic object are output. The visual screenshot and normalized pixel coordinates are input into a two-stage split visual segmentation model to extract the pixel coordinate sequence of the target geographic object and normalize it to obtain a normalized pixel coordinate sequence. The coordinate transformation module pastes the normalized pixel coordinate sequence back onto the visual screenshot to obtain the normalized coordinate sequence of the entire image; it then transforms the normalized coordinate sequence of the entire image into vector features corresponding to the target geographic object under geographic coordinates to obtain the recognition result of the target geographic object.

[0014] In another aspect, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in the present invention.

[0015] In another aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method described in the present invention.

[0016] The present invention has the following beneficial effects: 1. This invention utilizes the ray casting principle of a 3D GIS engine to establish sparse control point pairs between pixels and real geographic coordinates in real time. It injects these pairs into a multimodal model through text serialization soft registration, eliminating the need for training and fine-tuning of the multimodal visual model. This enables accurate geographic positioning of targets in any 3D oblique view or terrain undulation scene, improves the mapping accuracy of pixel-to-geographic coordinate conversion, and enhances the model compatibility and deployment convenience of the solution. 2. This invention utilizes the operating principle of preloading the encoder, lazy loading the decoder, and feature embedding caching of the split two-stage visual segmentation model to automatically convert the normalized detection boxes output by the multimodal model into segmentation point boxes and combine them with adaptive clipping and coordinate backfit logic to achieve pixel-level fine contour extraction of ground features, improve the completeness and precision of target boundary extraction, and enhance the human-computer interaction response speed in 3D scenes. 3. This invention converts segmented pixel coordinates into standard vector features in batches and automatically persists them to the GIS vector business layer, realizing the asset management of visual recognition results, improving the implementation efficiency of professional businesses such as 3D GIS spatial query, area statistics, buffer analysis, and topology calculation, and opening up a complete business loop from natural language map viewing to spatial analysis. Attached Figure Description

[0017] Figure 1 This is a detailed flowchart of an embodiment of the present invention; Figure 2 This is a schematic diagram of the first identification result according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the second identification result according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the third identification result in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be understood that the step numbers used in the text are for ease of description only and are not intended to limit the order in which the steps are performed.

[0020] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0021] The terms “comprising” and “including” indicate the presence of the described feature, whole, step, operation, element and / or component, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or collections thereof.

[0022] The term “and / or” refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes these combinations.

[0023] Example 1: See Figure 1 This invention provides a method for feature identification of geographic objects in 3D GIS, the method comprising: S1. Capture a visual screenshot of the target geographic object in the current viewport of the 3D GIS engine; When the scene vision of the 3D GIS engine is at any angle, the front end generates a screenshot of the current viewport at the same time. S2. Select multiple grid points on the visual screenshot and call the ray projection function of the 3D GIS engine to obtain the geographic coordinates corresponding to the pixel coordinates of each selected grid point. Uniform sampling along the pixel width and height of the visual screenshot 100 pixels, of which This indicates the width of the visual screenshot. Indicates the height of the visual screenshot; For example, take a screenshot with a resolution of 1920×1080: The number of sampling points in the width direction is 1920 / 2+1=961; The number of sampling points in the height direction is 1080 / 2+1=541; Total number of control points: 961 × 541 ≈ 520,000; The ray casting function is specifically SceneView.toMap(screenPoint) of the 3D GIS engine or other similar ray casting APIs; S3. Based on pixel coordinates and geographic coordinates, sparse control point pairs are constructed and serialized into sparse text; For example: pixel coordinates (120, 80) correspond to geographic coordinates (116.391, 39.912), pixel coordinates (240, 80) correspond to geographic coordinates (116.395, 39.911), and so on; The sparse control point pairs are dynamically established by the 3D GIS engine during runtime through ray projection, rather than by applying a predefined grid template or four-point homography transformation registration to the 2D image, nor are they derived from the raster GeoTransform of the registered orthophoto. S4. Call the digital elevation model query interface to obtain the elevation value of the target geographic object, calculate the elevation value statistics based on the elevation value, and serialize it into elevation text; S5. Input the visual description text of the target geographic object, and input the preset large language model to generate the visual feature description text; The visual descriptive text includes phrases such as "roof solar panels" and "pool on the right." The visual feature description text is, for example, “The photovoltaic panels are regular rectangular arrays ranging from dark blue to blue-black, arranged neatly, with smooth surfaces. They often appear in patches on rooftops, ground, or hillsides in the image base map, with clear and regular geometric boundaries.” S6. Concatenate the sparse text, elevation text, and visual feature description text to obtain the concatenated prior vector. The aforementioned splicing of prior vectors refers to a technique that guides the multimodal vision model to output normalized coordinates after referencing the grid during the inference phase, without modifying the network weights of the multimodal vision model, without additional training or fine-tuning, simply by serializing the coordinate reference grid into text and injecting it into the prompt of the multimodal vision model; the "soft" meaning is that the registration information exists in the inference context in the form of discrete text tokens, rather than being fixed into model parameters or an additional registration network. S7. Input the stitched prior vector and visual screenshot into the multimodal visual model, and output the normalized pixel coordinates of the target geographic object. The specific steps are as follows: Obtain the original width and original height of the visual screenshot; normalize the visual screenshot according to the input size required by the visual encoder of the multimodal visual model (e.g., scaling) to obtain a normalized screenshot; The prior vector and the preset coordinate reference anchoring instruction are concatenated into a prompt text. The coordinate reference anchoring instruction is used to indicate that the geometric result output by the multimodal vision model is represented in normalized coordinate form. The prompt text and standardized screenshots are combined into a multimodal input payload according to the interface specifications of the multimodal vision model; The multimodal input payload is submitted to the inference engine of the multimodal vision model. The sampling temperature parameter of the inference engine is set to be lower than a preset threshold, so that the multimodal vision model generates the original normalized coordinate result under the context constraint of the spliced ​​prior vector. Clamping is performed on each coordinate value of the original normalized coordinate result to constrain each coordinate value to the [0,1] interval, thereby obtaining the normalized pixel result of the target geographic object, including at least one of the normalized bounding box quadruple, normalized point coordinates, or normalized polygon vertex sequence. Each coordinate value is constrained to the interval [0,1], where 0 corresponds to the left or lower boundary of the standardized image, and 1 corresponds to the right or upper boundary of the standardized image; the value range of each component of the original normalized geometric result is [0,1], specifically: If the current coordinate value is less than 0, then the current coordinate value will be forcibly assigned to 0; If the current coordinate value is greater than 1, then the current coordinate value will be forcibly assigned the value 1; If the current coordinate value is between 0 and 1, then keep the current coordinate value unchanged; In this embodiment, the mesh sampling, viewport screenshotting, multimodal visual model inference, and geometric inverse calculation are all based on the same viewport state at the same time, so that the geometric inverse calculation under the three-dimensional oblique dynamic camera does not produce temporal offset; S8. Input the visual screenshot and normalized pixel coordinates into the two-stage visual segmentation model, extract the pixel coordinate sequence of the target geographic object and normalize it to obtain the normalized pixel coordinate sequence. While multimodal large models (such as GPT-4V and LLaVA) can "understand" images and point out where the target is, they can usually only output rough rectangular bounding boxes. In GIS operations, the rectangular bounding boxes will include a lot of background, trees or shadows, making them unsuitable for area statistics or accurate mapping. Using visual segmentation models (such as SAM), the rough rectangular bounding boxes provided by the large model serve as cues, driving the segmentation model to extract the precise outer contours (jagged, irregular polygons) of the target at the pixel level. This step is a key step in improving accuracy by transforming "visual perception results" into "geometric data that can be calculated by GIS." The segmentation process targets a single target geographic object specified in a natural language query (e.g., "that solar panel" or "that reservoir"). The segmentation model generates a binary mask for this object, in which pixels belonging to the target have a value of 1, and the background has a value of 0. This mask is accurate to every pixel in the screenshot, so the boundary can fit the edge of a roof or the shoreline of a water body. The split-type two-stage visual segmentation model includes SAM, SAM2, and SAM3. HQ, MobileSAM, FastSAM, EfficientSAM, etc.; S81, the split two-stage visual segmentation model includes a preloaded image encoder and a lazy-loaded cue decoder, wherein: The image encoder is used to receive input visual screenshots and output dense feature embedding tensors; The cue decoder is used to receive the input dense feature embedding tensor and coordinate cue, and generate a probability mask map based on the dense feature embedding tensor and coordinate cue; S82. Convert the normalized bounding box in the coordinate result into coordinate hints for the split two-stage visual segmentation model. The coordinate hints include the center point, upper left corner point, and lower right corner point of the normalized bounding box. Based on the center point of the normalized bounding box, a preset cropping window is driven to crop a visual sub-image from the visual screenshot that matches the input resolution of the image encoder. Input the visual subgraph into the image encoder to obtain the dense feature embedding tensor corresponding to the cropped subgraph; Coordinate cues and dense features are embedded into a tensor input cue decoder to obtain a probability mask of the cropped sub-image, which is then binarized to obtain a binary mask. , among which, when hour, , Indicates the first Line 1 The pixel values ​​of the binary mask image. Indicates the first Line 1 The pixel values ​​of the column probability mask; when hour, , This represents the preset binarization threshold. ; A contour tracing algorithm is performed on connected regions with a value of 1 in the binary mask image to extract the pixel coordinate sequence of the target geographic object in the cropped sub-image, and then normalized to obtain a normalized pixel coordinate sequence. , Indicates the width of the cropped sub-image. Indicates the height of the cropped sub-image. Indicates the first pixel coordinates; S9. Paste the normalized pixel coordinate sequence back to the visual screenshot to obtain the normalized coordinate sequence of the entire image. ,in , , Indicates the first A full-map normalized coordinate system, This indicates the starting pixel coordinates of the cropped sub-image within the visual screenshot. Indicates the width of the cropped sub-image. Indicates the height of the cropped sub-image. This represents the normalized starting coordinates of the top-left corner of the cropped sub-image within the visual screenshot. , ; S10. Use a preset conversion method to convert the segmented coordinate results into vector features corresponding to the target geographic object under geographic coordinates. The vector element is a graphic object with clearly defined geographic coordinates (latitude and longitude) and attribute information; The vector features include the type, geometry (e.g., polygon, line segment, or point) and attribute information (e.g., target geographic object name, source model, generation time, etc.) of the target geographic object. The conversion method includes linear inverse mapping or bilinear interpolation; S11. Store vector features in the vector business layer of the 3D GIS for subsequent spatial query and analysis operations; Determine whether a persistent segmented layer exists in a 3D GIS; If it does not exist, create a persistent segmentation layer; If they exist, the vector features will be automatically added as new features to the persistent segmentation layer; In one specific embodiment, please refer to Figure 2 Input the visual description text "circle the photovoltaic panels on the current map". The photovoltaic panels are identified and their range is segmented by recognizing the features of the photovoltaic panels on the remote sensing image. The geometric range of the photovoltaic panels is converted into corresponding vector elements to obtain the first recognition result. The vector elements are surface elements (polygons) and are accompanied by attribute information such as "category", for example { "category": "photovoltaic panel", "source model": SAM2, "generation time": "YYYY-MM-DD",}, which can be used for subsequent statistical analysis. In another specific embodiment, please refer to Figure 3 The system takes a visual description text as input: "Identify and obtain the area and water surface elevation of the reservoir on the map." It identifies and segments the reservoir's range using image features of the water body, while simultaneously querying and constructing an elevation grid. The reservoir's geometric range is converted into corresponding geographic vector polygon elements, yielding a second identification result. These vector elements are polygonal polygons with attached attribute information such as "category," "area," and "elevation," for example: {"category": "reservoir", "area": ​​"1.25", "area unit": "square kilometers", "water surface elevation": "512.3", "elevation unit": "meter", "source model": SAM2, "generation time": "YYYY-MM-DD",}, which can be used for subsequent statistical analysis. In another specific embodiment, please refer to Figure 4 The system takes a visual description text, "What are the terrain features of the current map?", as input. It identifies landforms and feature categories through image features, and simultaneously constructs an elevation grid. The geometric centers of features are converted into corresponding vector elements, yielding a third identification result. These vector elements are point elements (geometric center points of features) with accompanying "category" attribute information, such as {"Category": "Mountain Top", "Elevation": "863.0", "Elevation Unit": "Meters", "Terrain Category": "Mountainous Area", "Source Model": SAM2, "Generation Time": "YYYY-MM-DD",}. The elevation grid is used to analyze terrain undulations and indicate the terrain category of the large model (mountainous area, plains, hills, etc.).

[0024] Example 2: This embodiment provides a feature identification system for geographic objects in 3D GIS, the system comprising: The image acquisition module captures visual screenshots of the target geographic object; The text generation module selects multiple grid points at fixed intervals on the visual screenshot and obtains each selected grid; it constructs sparse control point pairs based on pixel coordinates and geographic coordinates and serializes them into sparse text; it inputs the visual description text of the target geographic object and inputs a preset large language model to generate visual feature description text. The coordinate segmentation module concatenates sparse text and visual feature description text to obtain a concatenated prior vector; it inputs the concatenated prior vector and visual screenshot into a multimodal visual model to output the normalized pixel coordinates of the target geographic object; and it segments the coordinate result based on a separate two-stage visual segmentation model to obtain the segmented coordinate result. The coordinate transformation module converts the segmented coordinate results into vector features corresponding to the target geographic object under geographic coordinates; and stores the vector features in the vector business layer of the 3D GIS for subsequent spatial query and analysis operations.

[0025] Example 3: This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a feature identification method for geographic objects in 3D GIS as described in any one of Embodiment 1.

[0026] Example 4: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a feature identification method for geographic objects in a 3D GIS as described in any one of Embodiment 1.

[0027] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, A and B simultaneously, or B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, at least one of a, b, and c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0028] Those skilled in the art will recognize that the units and algorithm steps described in the embodiments disclosed herein can be implemented using electronic hardware, computer software, or a combination of electronic hardware and software. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0029] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0030] In the several embodiments provided in this application, any function, if implemented as a software functional unit and sold or used as an independent product, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0031] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for feature identification of geographic objects in 3D GIS, characterized in that, The method includes: S1. Capture a visual screenshot of the target geographic object; select multiple grid points on the visual screenshot and obtain the geographic coordinates corresponding to the pixel coordinates of each selected grid point; construct sparse control point pairs based on the pixel coordinates and geographic coordinates, and serialize them into sparse text. S2. Obtain the elevation value of the target geographic object, calculate the elevation value statistics based on the elevation value, and serialize it into elevation text; input the visual description text of the target geographic object, and input the preset large language model to generate visual feature description text; S3. Concatenate the sparse text, elevation text, and visual feature description text to obtain a concatenated prior vector; input the concatenated prior vector and visual screenshot into a multimodal visual model for recognition, and output the normalized pixel coordinates of the target geographic object; input the visual screenshot and normalized pixel coordinates into a two-stage split visual segmentation model to extract the pixel coordinate sequence of the target geographic object and normalize it to obtain a normalized pixel coordinate sequence. S4. Paste the normalized pixel coordinate sequence back to the visual screenshot to obtain the normalized coordinate sequence of the whole image; convert the normalized coordinate sequence of the whole image into vector features corresponding to the target geographic object under geographic coordinates to obtain the recognition result of the target geographic object.

2. The feature identification method for geographic objects in 3D GIS according to claim 1, characterized in that, Steps S1-S4 are all executed under the same viewport state at the same time, where the viewport state is uniquely determined by the viewport range and the viewing angle direction.

3. The feature identification method for geographic objects in 3D GIS according to claim 1, characterized in that, The concatenated prior vectors and visual screenshots are input into the multimodal vision model. The specific steps are as follows: The visual screenshot is standardized to obtain a standardized screenshot; The prior vector and the preset coordinate reference anchoring instruction are concatenated into a prompt text. The coordinate reference anchoring instruction is used to indicate that the geometric result output by the multimodal vision model is represented in normalized coordinate form. The prompt text and standardized screenshots are combined into a multimodal input payload according to the interface specifications of the multimodal vision model; The multimodal input payload is submitted to the inference engine of the multimodal vision model to generate the original normalized coordinate results; Clamping is performed on each coordinate value of the original normalized coordinate result to constrain each coordinate value to the interval [0,1], thereby obtaining the normalized pixel result of the target geographic object.

4. The feature identification method for geographic objects in 3D GIS according to claim 3, characterized in that, Each coordinate value is constrained to the interval [0,1], where 0 corresponds to the left or lower boundary of the standardized image, and 1 corresponds to the right or upper boundary of the standardized image, specifically: If the current coordinate value is less than 0, then the current coordinate value will be forcibly assigned to 0; If the current coordinate value is greater than 1, then the current coordinate value will be forcibly assigned the value 1; If the current coordinate value is between 0 and 1, then the current coordinate value remains unchanged.

5. The feature identification method for geographic objects in 3D GIS according to claim 1, characterized in that, The split two-stage visual segmentation model includes a preloaded image encoder and a lazy-loaded cue decoder, wherein: The image encoder is used to receive input visual screenshots and output dense feature embedding tensors; The cue decoder is used to receive the input dense feature embedding tensor and coordinate cue, and generate a probability mask map based on the dense feature embedding tensor and coordinate cue.

6. The feature identification method for geographic objects in 3D GIS according to claim 5, characterized in that, The specific steps for segmenting the coordinate results based on the split two-stage visual segmentation model are as follows: The normalized bounding box in the coordinate results is converted into coordinate cues for the split two-stage visual segmentation model. The coordinate cues include the center point, top left corner point, and bottom right corner point of the normalized bounding box. Based on the center point of the normalized bounding box, a preset cropping window is driven to crop a visual sub-image from the visual screenshot that matches the input resolution of the image encoder. Input the visual subgraph into the image encoder to obtain the dense feature embedding tensor corresponding to the cropped subgraph; The coordinate cue and dense features are embedded into the tensor input cue decoder to obtain the probability mask of the clipped sub-image, and then binarized to obtain the binary mask. A contour tracing algorithm is performed on the connected regions with a value of 1 in the binary mask image to extract the pixel coordinate sequence of the target geographic object in the cropped sub-image, and then normalized to obtain the normalized pixel coordinate sequence.

7. The feature identification method for geographic objects in 3D GIS according to claim 1, characterized in that, The specific steps for storing vector features into the vector business layer of a 3D GIS are as follows: Determine whether a persistent segmented layer exists in a 3D GIS; If it does not exist, create a persistent segmentation layer; If they exist, the vector features will be automatically added as new features to the persistent segmentation layer.

8. A feature identification system for geographic objects in 3D GIS, characterized in that, The system includes: The image acquisition module captures visual screenshots of the target geographic object; The text generation module selects multiple grid points at fixed intervals on the visual screenshot, obtains the geographic coordinates corresponding to the pixel coordinates of each selected grid point, constructs sparse control point pairs based on the pixel coordinates and geographic coordinates, and serializes them into sparse text; inputs the visual description text of the target geographic object, and inputs a preset large language model to generate visual feature description text. The coordinate segmentation module concatenates sparse text, elevation text, and visual feature description text to obtain a concatenated prior vector. The concatenated prior vector and visual screenshot are input into a multimodal visual model for recognition, and the normalized pixel coordinates of the target geographic object are output. The visual screenshot and normalized pixel coordinates are input into a two-stage split visual segmentation model to extract the pixel coordinate sequence of the target geographic object and normalize it to obtain a normalized pixel coordinate sequence. The coordinate transformation module pastes the normalized pixel coordinate sequence back onto the visual screenshot to obtain the normalized coordinate sequence of the entire image; it then transforms the normalized coordinate sequence of the entire image into vector features corresponding to the target geographic object under geographic coordinates to obtain the recognition result of the target geographic object.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images