Geographic road network topology consistency self-checking and repairing method based on multi-modal large language model

Abstract

The invention discloses a geographic road network topology consistency self-checking and repairing method based on a multi-mode large language model. The method comprises the following steps: obtaining geographic road network vector data, a high-resolution satellite remote sensing image, a vehicle GPS track sequence and a street view image of a key position of an area to be detected; extracting each modal feature by using a multi-modal encoder, and projecting the modal features to a unified language marking space through a space projection operator to realize semantic alignment; the aligned multi-modal features are input into a multi-modal large language model for self-check reasoning, and a diagnosis report containing road network topology errors is output; converting the diagnosis report into a solution diagram by using the model, and calling a code generation module to automatically write a repair code script; and executing the repair script to modify the road network data, feeding back the modified data to the model for closed-loop verification of anti-factual reasoning and virtual traversal semantic verification, and updating the production library after verification is passed. According to the invention, the accuracy of road network topology automatic repair is improved.

Description

Technical Field

[0001] This invention relates to the fields of geographic information system data processing and artificial intelligence application technology, and in particular to a method for self-checking and repairing geographic road network topology consistency based on a multimodal large language model. Background Technology

[0002] In the production and maintenance of geographic information data, high-precision road network topology is the foundation for navigation and positioning, traffic planning and autonomous driving. Existing road network data quality inspection and repair work usually relies on automated detection methods based on geometric rules and manual comparison operations.

[0003] With the development of remote sensing Earth observation technology and vehicle networking, although it is possible to acquire massive amounts of satellite remote sensing images, street view images, and GPS trajectory data reflecting vehicle driving behavior, these multi-source heterogeneous data are often in different feature spaces and data formats. Traditional processing methods usually require the establishment of independent analysis algorithms for different modal data or rely on operators to manually switch between different data views and visually interpret them to confirm the actual connectivity of the road.

[0004] When faced with complex urban road network environments, relying solely on geometric topology rules is insufficient to fully capture the physical traffic logic of roads. For example, there may be roads that appear to have continuous textures but are physically blocked, or areas where the road network is not geometrically connected but where vehicles frequently pass through. This presents challenges in data fusion and logical reasoning for establishing deep semantic relationships between multimodal data and achieving automated logical verification and closed-loop repair. Summary of the Invention

[0005] The purpose of this invention is to provide a method for self-checking and repairing the topology consistency of a geographic road network based on a multimodal large language model, so as to solve the problems pointed out in the background art.

[0006] This invention provides a method for self-checking and repairing the topological consistency of a geographic road network based on a multimodal large language model, comprising:

[0007] Acquire multimodal data of the area to be inspected, including geographic road network vector data, high-resolution satellite remote sensing images, vehicle GPS trajectory sequences, and street view images of key locations;

[0008] A multimodal encoder is used to extract the modal features of the multimodal data, and a spatial projection operator is used to project the modal features onto a unified language tag space to achieve semantic alignment between visual features and geographic descriptions.

[0009] The semantically aligned multimodal features are input into a multimodal large language model for self-inspection and reasoning. The road network topology map is compared with visual connectivity and behavioral patterns, and a diagnostic report containing road network topology errors is output.

[0010] The diagnostic report is transformed into a solution diagram using the multimodal large language model, and the code generation module is called to automatically write repair code scripts based on the solution diagram;

[0011] The script for fixing the code modifies the geographic road network vector data, and the modified road network data is fed back to the multimodal large language model for closed-loop verification of counterfactual reasoning. If the verification passes, the production database is updated.

[0012] Optionally, the step of extracting modal features from the multimodal data using a multimodal encoder and projecting the modal features onto a unified language tag space using a spatial projection operator includes:

[0013] The global contextual features of the high-resolution satellite remote sensing image and the fine semantic features of the street view image are extracted using a visual encoder.

[0014] The behavioral pattern features of the vehicle's GPS trajectory sequence are extracted using a trajectory encoder;

[0015] The spatial projection operator, which serves as a multilayer perceptron neural network structure, is used to project and fuse the geographic road network vector data, the global context features, the fine semantic features, and the behavioral pattern features, and these are then used as inputs to the multimodal large language model.

[0016] Optionally, the step of extracting behavioral pattern features of the vehicle GPS trajectory sequence using a trajectory encoder includes:

[0017] The original trajectory coordinates of the vehicle's GPS trajectory sequence are obtained as input;

[0018] The original trajectory coordinates are extracted using the trajectory encoder to obtain the behavioral pattern features, which are then used by the multimodal large language model to perform road network logic verification.

[0019] Optionally, the step of inputting the semantically aligned multimodal features into a multimodal large language model for self-checking inference includes:

[0020] By comparing the road network topology map with the high-resolution satellite remote sensing image, if the image shows continuous roads but there are no edge connections in the road network topology map, it is determined to be a false break error;

[0021] Logical verification is performed based on the vehicle's GPS trajectory sequence. If the trajectory data shows frequent turning behavior at intersections but the road network topology map does not establish turning restrictions or connection nodes, it is determined to be an error of missing intersection constraints.

[0022] By integrating the semantics of the street view image, if a physical obstacle or traffic sign is identified but the road network topology map shows it as a passable road segment, it is determined that the dead end is missing an end sign.

[0023] Optionally, the step of converting the diagnostic report into a solution diagram using the multimodal large language model includes:

[0024] Based on the divide-and-conquer strategy, the repair tasks in the diagnostic report are broken down into multiple atomic operations, including adding nodes, deleting edges, and modifying attributes.

[0025] Based on the execution logic order of the atomic operations, a directed acyclic graph is constructed as the solution graph.

[0026] Optionally, the step of automatically writing a repair code script based on the solution diagram by calling the code generation module includes:

[0027] For each atomic operation in the directed acyclic graph, the multimodal large language model is driven to generate a corresponding independent function or code snippet.

[0028] The individual functions or code snippets are arranged into the repair code script, which is configured to directly perform add, delete, and modify operations on the geographic information system database.

[0029] Optionally, the step of feeding the modified road network data back to the multimodal large language model for closed-loop verification of counterfactual reasoning includes:

[0030] Based on the multimodal large language model with the ability to judge and score the rationality of roads, the first rationality score of the road network structure before modification and the second rationality score of the road network structure after modification are obtained respectively.

[0031] Compare the first reasonableness score with the second reasonableness score;

[0032] If the difference between the second reasonableness score and the first reasonableness score exceeds a preset threshold, the counterfactual reasoning verification is deemed successful, confirming that the modification is effective and a positive optimization.

[0033] Optionally, before updating the production database, the following steps are also included:

[0034] Calculate the global topology matching ratio and local neighbor node accuracy of the modified road network data;

[0035] When the counterfactual reasoning verification passes, and both the global topology matching ratio and the accuracy of the local neighbor nodes meet the preset quality assessment threshold, the operation of updating the production database is executed.

[0036] Optionally, after determining that the counterfactual reasoning verification has passed, and before updating the production database, the method further includes performing a virtual traversal semantic verification step:

[0037] Based on the modified road network data, the target road segment that has undergone topology change and its multi-hop adjacent road segment are extracted to construct a local road network subgraph;

[0038] In the local road network sub-map, a virtual navigation start point and a virtual navigation end point that must be crossed the target road segment are set, and the high-resolution satellite remote sensing image and the street view image corresponding to the spatial location of the target road segment are retrieved;

[0039] Using the multimodal large language model to act as a virtual driver, the local road network sub-map, the virtual navigation start point, the virtual navigation end point, and visual obstacle features from the high-resolution satellite remote sensing image and the street view image are input into the multimodal large language model;

[0040] The multimodal large language model is driven to attempt to generate a continuous set of navigation instructions from the virtual navigation start point to the virtual navigation end point;

[0041] If the multimodal large language model is unable to generate the continuous navigation instruction set due to visual impairment features or topological logic conflicts, then the virtual traversal semantic verification step is determined to fail.

[0042] Optionally, after determining that the virtual traversal semantic verification step fails, the method further includes:

[0043] Capture the description of the blocking reason output by the multimodal large language model when it fails to generate the continuous navigation instruction set;

[0044] The description of the blocking cause is semantically parsed to generate corresponding negative topological constraints, which include prohibited node pairs or prohibited direction vectors.

[0045] The negative topological constraints are fed back to the construction process of the solution graph, and the multimodal large language model is triggered to re-execute the transformation of the diagnostic report and the writing of the repair code script based on the negative topological constraints, until the virtual traversal semantic verification step is passed.

[0046] The present invention has achieved the following beneficial effects:

[0047] This invention introduces a multimodal large language model and spatial projection technology to map geographic road network vector data, satellite remote sensing imagery, vehicle trajectory sequences, and street view images onto a unified language tagging space. This enables the model to comprehensively utilize visual features and behavioral pattern features to perform in-depth logical verification of the road network topology. This method not only automates the process from error detection and diagnostic report generation to code-level repair, but also establishes a dual verification system through counterfactual reasoning and virtual traversal semantic verification mechanisms. It can correct the geometric structure and traffic logic of road network data based on a multi-dimensional chain of evidence without relying on extensive manual intervention, thereby improving the topological consistency and physical authenticity of geographic road network data in complex scenarios. It also enhances the logical interpretability and reliability of the results during the data repair process.

[0048] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0049] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0051] Figure 1 This is a flowchart illustrating a method for self-checking and repairing geographic road network topology consistency based on a multimodal large language model, as described in an embodiment of the present invention.

[0052] Figure 2 This is a schematic diagram of the composition structure of a geographic road network topology consistency self-checking and repair system based on a multimodal large language model in an embodiment of the present invention. Detailed Implementation

[0053] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0054] like Figure 1 As shown in the figure, this invention discloses a method for self-checking and repairing the topological consistency of a geographic road network based on a multimodal large language model, including:

[0055] Step S11: Obtain multimodal data of the area to be inspected. The multimodal data includes geographic road network vector data, high-resolution satellite remote sensing images, vehicle GPS trajectory sequences, and street view images of key locations.

[0056] In this embodiment, it should be noted that the method of the present invention is typically deployed in a high-performance computing cluster or cloud server environment, interacting with the upstream data production platform through a standardized data interface. During actual execution, to ensure accurate alignment of heterogeneous data in time and space, the system first defines a geofence and an effective time window based on the requirements of the task to be inspected.

[0057] Specifically, geographic road network vector data, which serves as the primary object for self-inspection and repair, typically originates from professional map production databases (such as PostGIS databases) or Open Street Map (OSM) sources. This data is an abstract mapping of the physical road network in digital space. The acquisition process involves extracting all topological elements within a specified area from the database, including nodes and road segments. Nodes represent intersections or road endpoints and contain latitude and longitude coordinates; road segments represent road entities, containing not only coordinate sequences describing their geometry but also rich attribute information tables. These attribute information tables record detailed road classifications (e.g., highways, urban arterial roads, rural roads), traffic direction rules (one-way, two-way, reversible), number of lanes, design speed limits, and complex intersection turning restriction tables. After data acquisition, the system performs topological preprocessing on the vector data, constructing a directed graph data structure indexed by road segment IDs, and removing minor geometric noise caused by data storage precision (e.g., duplicate nodes, self-intersecting line segments) to ensure that the input model is a geometrically compliant base map.

[0058] Specifically, high-resolution satellite remote sensing imagery provides macroscopic-scale ground physical truth values ​​for verifying the geometric existence of road network data. This embodiment constrains the acquisition of image data to preferably use optical remote sensing images with a spatial resolution better than 0.5 meters (i.e., one pixel represents a 0.5m × 0.5m area on the ground) to ensure that computer vision algorithms can clearly identify fine features such as lane dividers (dotted and solid lines), road edges, physical barriers (such as green belts and guardrails), and intersection guide lines. After acquiring the original imagery, the system automatically performs a series of photogrammetric processing steps, including orthorectification and georegistration. The system utilizes a high-precision digital elevation model (DEM) to eliminate projection distortion caused by terrain undulations and satellite viewing angle tilt, ensuring that the pixel coordinate system of the imagery strictly coincides with the geographic coordinate system of the vector data (such as WGS-84 or GCJ-02), with registration errors controlled at the sub-pixel level. In addition, to enhance the ability to identify road surface materials (such as distinguishing between cement roads, asphalt roads and unpaved dirt roads), multispectral data can be acquired simultaneously, and the Normalized Difference Vegetation Index (NDVI) can be used to help determine whether the road is obscured by tree canopies, thereby avoiding misjudgments caused by visual obstruction.

[0059] Specifically, regarding vehicle GPS trajectory sequences, the system pulls massive amounts of floating car trajectory data from the traffic big data platform or vehicle-to-everything (V2X) cloud interface within a historical time window (e.g., the past 30 days) for the area to be inspected. The raw trajectory data often contains a large amount of noise, such as drift points caused by multipath effects due to obstruction from tall buildings, jump points caused by equipment malfunctions, and invalid dwell points caused by roadside parking. Therefore, this embodiment incorporates a trajectory cleaning and reconstruction module. The system uses a Kalman filter algorithm to smooth the raw positioning points and removes outliers based on speed thresholds and heading angle change rates. The cleaned data is reconstructed into continuous trajectory sequences. Each sequence not only contains latitude and longitude coordinates but also carries precise timestamps, instantaneous velocity vectors, and heading angle information. This data objectively depicts the actual traffic behavior of vehicles in the road network, revealing implicit logical conflicts such as roads appearing closed on the map but actually having vehicles passing through, or roads appearing clear on the map but never having been used by any vehicle.

[0060] Specifically, street view images of key locations provide first-person semantic information from a microscopic perspective, supplementing blind spots in satellite imagery. Due to the high cost of acquiring and processing street view data across the entire area, this invention employs an on-demand retrieval strategy. The system first automatically identifies potential complex nodes based on the topological characteristics of geographic road network vector data, such as multi-way intersections, interchange ramps, tunnel entrances and exits, and the terminal nodes of suspected dead-end roads, defining these locations as key locations. Subsequently, the system indexes the street view database based on the coordinates of these locations to obtain 360-degree panoramic images or vehicle-mounted front view images. Street view images can capture facade features that are not observable from a satellite perspective, such as traffic signs (no left turn, one-way streets, height and weight restrictions), road barriers (stone blocks, rising barriers), and traffic light status.

[0061] Step S12: Extract the modal features of the multimodal data using a multimodal encoder, and project the modal features onto a unified language tag space using a spatial projection operator to achieve semantic alignment between visual features and geographic descriptions.

[0062] Traditional GIS algorithms often process visual data and vector data separately, failing to achieve deep semantic interaction. This invention introduces feature encoding and spatial projection techniques to uniformly map data from different modalities into a high-dimensional vector space that a Large Language Model (LLM) can understand, enabling the model to uniformly process the semantic relationships between visual features, trajectory features, and text features.

[0063] Specifically, a multimodal encoder is used to extract modal features from the multimodal data, and a spatial projection operator is used to project these modal features onto a unified language tag space, including:

[0064] The first step involves using a visual encoder to extract global contextual features from high-resolution satellite remote sensing images and fine semantic features from street view images.

[0065] For high-resolution satellite remote sensing imagery, this embodiment employs a large-scale visual model based on the Transformer architecture (such as Vision Transformer, ViT) as the backbone network. Unlike traditional convolutional neural networks (CNNs) that focus on local texture, ViT effectively captures the global topological structure and long-range dependencies of roads in the imagery by dividing the large image into a series of fixed-size patches and introducing a self-attention mechanism. For example, it can identify a road that, although obscured by an overpass at one point, still maintains a trend of connectivity globally. The system inputs satellite imagery into the ViT encoder, and the output high-dimensional vector is the global context feature. This feature encodes the skeletal structure of the road network, connectivity trends, and the macroscopic semantics of the surrounding geographical environment (such as whether it is a mountainous area or a densely populated urban area).

[0066] For street view images, which contain fine-grained information such as traffic signs, road sign text, and road markings, this embodiment employs a visual encoder (such as a feature extraction network based on an improved SwinTransformer) with optical character recognition (OCR) and small object detection capabilities. This encoder focuses on traffic facility areas in the image, extracting feature vectors containing specific traffic rule semantics—that is, fine-grained semantic features. For example, when a "No Left Turn" sign appears in the street view, the feature vector generated by the encoder will be closely associated with concepts such as "prohibition" and "restraint" in the semantic space.

[0067] The second step involves using a trajectory encoder to extract behavioral pattern features from the vehicle's GPS trajectory sequence.

[0068] The behavior pattern features of vehicle GPS trajectory sequences are extracted using a trajectory encoder. Specifically, this includes: obtaining the original trajectory coordinates of the vehicle GPS trajectory sequence as input; extracting pattern features from the original trajectory coordinates using the trajectory encoder to obtain behavior pattern features, which are then used by a multimodal large language model to perform road network logic verification.

[0069] In practice, to transform discrete GPS trajectory data into tensors that can be computed by neural networks, the system first resamples the vehicle GPS trajectory sequence into a time step sequence with fixed time intervals. For each time step in the sequence... Construct a 5-dimensional feature vector Its mathematical expression is as follows: , In the formula, Representing time steps The longitude and latitude coordinates of the vehicle at that time; This represents the minimum-maximum normalization function based on the geographic bounding box of the region to be inspected, used to map coordinate values ​​to... The interval is calculated using the following formula: In the formula, Represents the coordinate components of the current time step (i.e.) or ); This represents the minimum value of the geographic bounding box of the area to be inspected in the corresponding coordinate dimension; This represents the maximum value of the geographic bounding box of the area to be inspected in the corresponding coordinate dimension. This represents the normalized instantaneous speed scalar, which is the current speed divided by the maximum speed limit designed for the road network (e.g., 120 km / h). This represents the change in the vehicle's heading angle at the current time step relative to the previous time step (in radians). These represent the sine and cosine trigonometric functions, respectively, used to map changes in angles into continuous numerical characteristics to avoid numerical discontinuities caused by the periodicity of angles.

[0070] Subsequently, the feature vector It is input into the trajectory encoder.

[0071] It is understandable that trajectory data is essentially sequential data containing spatiotemporal attributes. The trajectory encoder designed in this embodiment employs a bidirectional long short-term memory network (Bi-LSTM) or a Transformer-based time-series coding architecture. The input data is not only a latitude and longitude sequence but also incorporates dynamic features such as the rate of change of velocity and the rate of change of heading angle. By learning the flow patterns of massive trajectories, the encoder aggregates discrete point sets into behavioral pattern features that represent traffic logic. For example, if at an intersection, the flow direction features of a large number of trajectories all show straight or right turns, while the feature response value of the left-turn dimension is zero, this constitutes a strong implicit semantic, that is, left turns are physically or legally prohibited at this intersection. This behavioral pattern feature can compensate for the shortcomings of visual data in judging traffic rules.

[0072] The third step involves using the spatial projection operator, which serves as the structure of a multilayer perceptron neural network, to project and fuse geographic road network vector data, global contextual features, fine semantic features, and behavioral pattern features, which are then used as input to a multimodal large language model.

[0073] After extracting the original features of each modality, they cannot be directly processed by the large language model because they reside in different vector spaces with varying dimensions (e.g., visual features are 2048-dimensional, trajectory features are 512-dimensional). Therefore, this embodiment designs a learnable spatial projection operator. This operator is structured as a multilayer perceptron (MLP) or Q-Former structure, and its function is to map the feature vectors of each modality to the same dimension (e.g., 4096-dimensional) and semantic space as the word embeddings of the large language model.

[0074] During the pre-training phase, the system uses large-scale paired data such as remote sensing images-road network descriptions and street view images-traffic rule descriptions to conduct comparative learning training on the projection operator, enabling it to have alignment capabilities.

[0075] The contrastive learning training specifically employs the InfoNCE (Noise Contrastive Estimation) loss function for optimization. Assume a training batch contains... For a matched road network-image sample pair, denoted as For the first batch of this batch For each sample, its loss function Defined as: , In the formula, For the first The contrastive loss value for each sample pair; For the first The road network vector features of each sample after projection; For the first Visual features of the projected image corresponding to each sample (positive sample); The first in the batch Visual features of images of each sample; The cosine similarity function is used to calculate the cosine of the angle between two feature vectors in the semantic space. The formula is: In the formula, and These represent any two input feature vectors to be compared (in this embodiment, they specifically refer to the road network vector features). With the image visual features or ), This represents the L2 norm of a vector. The temperature coefficient hyperparameter is set to 0.07 in this embodiment; For the natural constant An exponential function with base 0; This indicates that all items within the batch... The image features are summed, where when When it is a positive sample pair, when When it is a negative sample pair.

[0076] This formula drives the spatial projection operator to achieve cross-modal semantic alignment by maximizing the similarity of positive sample pairs and minimizing the similarity of negative sample pairs.

[0077] After processing by the projection operator, visual and trajectory features are transformed into soft cues that the large language model can understand. For the road network vector data itself, the system uses a specialized geographic segmenter to serialize it into a structured text token sequence (such as a GeoJSON string), and then converts it into embeddings. Finally, the system concatenates the projected visual tokens, trajectory tokens, and vector data tokens to form a context sequence containing rich multimodal information, which is then input into the multimodal large language model.

[0078] Step S13: Input the semantically aligned multimodal features into the multimodal large language model for self-checking inference, compare the road network topology map with visual connectivity and behavioral patterns, and output a diagnostic report containing road network topology errors.

[0079] In this embodiment, the multimodal large language model (such as a GIS industry-specific model fine-tuned based on DeepSeek, Llama 3 or GPT-4 architecture) performs logical deduction and conflict detection on the input multimodal information based on the generalized world knowledge (such as traffic regulations and road design specifications) learned in its pre-training and the GIS professional knowledge learned in the fine-tuning.

[0080] Specifically, the semantically aligned multimodal features are input into a multimodal large language model for self-testing inference, including the following typical inference scenarios:

[0081] Scenario 1: Compare the road network topology map with high-resolution satellite remote sensing imagery. If the imagery shows continuous roads but the road network topology map shows no edge connections, it is judged as a false break error.

[0082] In this logical branch, the model first parses the input geographic road network vector data and identifies suspended nodes (i.e., dead-end points) with a degree of 1 using a graph traversal algorithm. Then, the model focuses its attention mechanism on the global contextual features of the satellite imagery corresponding to the node's coordinates. If the projected visual features strongly indicate a continuous, dark asphalt pavement texture at the breakpoint, and the pavement width and color are consistent with the road height before the breakpoint, while no obvious physical obstructions such as rivers or walls are detected, the model recognizes visual connectivity. At this point, the model further searches for trajectory behavior features passing through the area. If a continuous trajectory flow crosses the breakpoint, the model further confirms physical connectivity. Faced with multiple contradictions between vector data breaks, visual physical connectivity, and trajectory logical connectivity, the model uses logical reasoning to determine that the road network data has missing connectivity errors, i.e., false breakage errors. The model will describe this in detail in the diagnostic report: "The road network vector is broken at coordinates (x,y), but the satellite imagery shows that the pavement texture extends continuously to coordinates (x',y'), and there is historical trajectory coverage, indicating missing data. Confidence: High."

[0083] Scenario 2: Logical verification is performed by combining vehicle GPS trajectory sequences. If the trajectory data shows frequent turning behavior at intersections but the road network topology map does not establish turning restrictions or connection nodes, it is determined to be an error of missing intersection constraints.

[0084] In this logical branch, the model focuses on using behavioral pattern features from trajectory data for reverse reasoning. Assume that road network vector data shows a certain intersection node is omnidirectionally connected, and the attribute table does not define any turning restrictions (i.e., allowing straight, left, and right turns). However, the input behavioral pattern features show that in the past observation period, among tens of thousands of vehicles passing through this intersection, not a single vehicle performed a left turn, and all vehicles attempting to go to the left chose to go straight and then make a U-turn or turn right to detour. The large language model uses its logical inductive ability to deduce the cause from the statistical phenomenon of zero traffic flow: it is highly likely that there is a physical barrier at the intersection (which may not be visually obvious) or strict left-turn restrictions (such as time-limited restrictions). The model identifies a conflict between the permissible road network attributes and the prohibited behavior, determining it as a missing intersection constraint error. Conversely, if the road network shows two roads are not connected (such as upper and lower levels of an overpass), but trajectory data shows a large number of vehicles transferring between levels at this location (such as the existence of hidden ramps), the model will also determine it as a missing connection.

[0085] Scenario 3: By integrating the semantics of street view images, if a physical obstacle or traffic sign is identified but the road network topology map shows it as a passable road segment, it is determined to be a dead end with an incorrect termination sign.

[0086] In this logical branch, the model utilizes the fine semantic features of the street view for verification. Suppose the road network vectors show that a certain road is open and leads to the next intersection. However, the street view features received by the model clearly contain semantic embeddings of signs indicating construction ahead or dead ends, or identify features of physical obstacles such as walls or bollards. The large language model performs an adversarial comparison between these semantic representations of obstruction and the connectivity attributes of the road network vector data, thereby determining that the timeliness of the road network data lags behind changes in the physical world, and that the road segment is actually a dead end, characterized as a dead end with an incorrect termination sign.

[0087] Finally, the multimodal large language model outputs a structured diagnostic report. This report is a text containing natural language descriptions and structured key-value pairs, detailing the type of each error, its location coordinates, the IDs of the road network elements involved, the confidence score of the judgment, and the chain of evidence for the inference (e.g., judgment criteria: continuous satellite image texture and uninterrupted trajectory heatmap). This interpretable output provides precise instructions for subsequent automated repair.

[0088] Step S14: Use a multimodal large language model to convert the diagnostic report into a solution diagram, and call the code generation module to automatically write the repair code script based on the solution diagram.

[0089] Because topological editing of geographic road networks is highly interconnected and complex (for example, deleting a node may result in multiple dangling edges, or changing the direction of a road may render the entire area unreachable), directly generating a single repair instruction often leads to data structure corruption. Therefore, this invention introduces a solution graph as an intermediate state to ensure the rigor of the repair logic.

[0090] Specifically, the diagnostic report is transformed into a solution diagram using a multimodal large language model, including:

[0091] The first step: Based on the divide-and-conquer strategy, the repair tasks in the diagnostic report are broken down into multiple atomic operations, including adding nodes, deleting edges, and modifying attributes.

[0092] The large language model first performs semantic parsing on the diagnostic report, leveraging its built-in planning capabilities to decompose complex repair intentions into the most basic, indivisible atomic operations in GIS editing. For example, for a repair task involving a false fracture, the model breaks it down into a series of micro-steps: Operation A extracts the geometric coordinates of fracture 1; Operation B extracts the geometric coordinates of fracture 2; Operation C constructs a new geometric line segment connecting fracture 1 and fracture 2; Operation D inserts the new line segment into the road network layer; Operation E inherits the road grade attributes of the upstream road segment. This decomposition transforms ambiguous natural language instructions into deterministic, executable mathematical and logical operations.

[0093] The second step: Based on the execution logical order of atomic operations, construct a directed acyclic graph (DAG) as the solution graph.

[0094] After decomposing the atomic operations, the model needs to plan their execution sequence. In geographic information editing, the order of operations is crucial. Clearly, operation D (inserting a line segment) depends on the completion of operation C (building geometry), while operation C depends on operations A and B (obtaining coordinates). To clearly express this dependency, the model constructs a directed acyclic graph (DAG). Nodes in the graph represent atomic operations, and directed edges represent dependencies between operations. For example, if node A points to node C, it means that A must be completed before C can be executed. The process of constructing the DAG effectively avoids circular dependencies and logical deadlocks, ensuring the logical completeness of the repair solution.

[0095] Internally, the solution graph is instantiated as a list of objects, each atomic operation node containing four core fields: node_id (unique node identifier), operation (operation type, such as CREATE_LINK), params (operation parameters, such as start and end point IDs), and dependencies (a list of preceding dependent nodes). For example, an operation node for creating a connection is defined as:

[0096] {"node_id": "Op_C", "operation": "CREATE_LINK", "params": {"from": "Node_A", "to": "Node_B"}, "dependencies": ["Op_Check_A", "Op_Check_B"]}.

[0097] The orchestrator reads the list and uses a topological sorting algorithm to parse the execution order, ensuring that all prerequisite operations defined in its dependencies list have been executed before the current operation is executed.

[0098] The third step: Based on the solution diagram, call the code generation module to automatically write the repair code script.

[0099] Specifically, the code generation module is invoked based on the solution diagram to automatically write repair code scripts, including: for each atomic operation in the directed acyclic graph, driving the multimodal large language model to generate corresponding independent functions or code snippets; arranging each independent function or code snippet into a repair code script, and configuring the repair code script to directly perform add, delete, and modify operations on the geographic information system database.

[0100] In this step, the system utilizes a dedicated large model component with code generation capabilities. This component pre-loads commonly used GIS development libraries (such as ArcPy, PyQGIS, Shapely, and PostGISSQL). For each node in the DAG, the model generates specific Python functions or SQL transaction code. For example, for an attribute modification operation, the model generates code with the following logic:

[0101] Query the road chain with ID X, construct an Update statement, and set the OneWay field to True.

[0102] Subsequently, the system uses an orchestrator to assemble these scattered code fragments into a complete repair script according to the topological sorting result of the DAG. This script includes the necessary database connection, transaction initiation, exception handling, and transaction commit logic.

[0103] Step S15: Execute the repair code script to modify the geographic road network vector data, and feed the modified road network data back to the multimodal large language model for closed-loop verification of counterfactual reasoning. If the verification passes, update the production database.

[0104] In this embodiment, the system first executes a repair code script in an isolated memory sandbox environment. The sandbox environment creates a temporary copy of the current road network data using copy-on-write technology, ensuring that any unverified modifications do not contaminate the main production database. After the script execution is complete, the system generates a version of the road network data to be committed.

[0105] Secondly, perform bidirectional counterfactual reasoning based on semantic scoring.

[0106] The system maps the road network data of the version to be submitted back into high-dimensional feature vectors using the aforementioned spatial projection operator and inputs it into the multimodal large language model. At this point, the system sends an evaluation instruction to the model: "Assuming the current road network structure is as shown in the input, please evaluate its logical consistency with satellite imagery and historical trajectory data, and provide a reasonableness confidence score (0-100)." The model uses its internal attention mechanism to calculate the cross-attention between the road network vector token, visual token, and trajectory token. If the repaired road network (e.g., connecting two breakpoints) is highly aligned with the trajectory flow direction and covers the road surface texture in the imagery, the cross-attention will be significantly improved, thus generating a higher second reasonableness score. This reasonableness score... The specific calculation formula is as follows: Extract the cross-attention weight matrix of the last layer of the multimodal large language model Transformer. Let the input road network vector token set be... The input visual context token set is .

[0107] , In the formula, The calculated score for the rationality of the road network structure; This represents the total number of elements in the road network vector token set; Represents the first in the set of road network vectors One Token; Represents the first in the visual context set One Token; Represents Token For Token The normalized attention weight values, whose range is: ; This indicates the current network token. Find the visual token with the highest attention value among all visual tokens.

[0108] This formula is used to quantify whether each microstructure of the road network has strong visual evidence to support it in satellite imagery or street view.

[0109] At the same time, the system performs the same evaluation process on the original version of the road network data to obtain the first reasonableness score (denoted as ). The system calculates the second reasonableness score (denoted as...). The difference between the first reasonableness score and the second reasonableness score .

[0110] In this embodiment, a preset threshold is set. Used to filter random noise in model inference and ensure the effectiveness of the repair. This preset threshold... The value is typically set between 10 and 20 points (in a percentage-based scoring system), with 15 points being the preferred value. Only when the difference... Only then does the system determine that the repaired road network structure has a substantial improvement in semantic consistency.

[0111] The technical effect of setting this threshold is that the output of a multimodal large language model has a small variance characteristic. Only compared A difference of 1-2 points may be due to random fluctuations in the model's attention weights, which is insufficient to prove that the topological error has been accurately corrected. However, when the difference exceeds this threshold, it indicates that the model has found strong visual evidence on the new topology (significant concentration of attention weights), thus ruling out the possibility of illusion correction.

[0112] In addition, the system can also combine absolute score judgment, that is, require Only when the score exceeds the passing mark (e.g., >85 points) is the counterfactual reasoning deemed semantically valid. This logically confirms that the modified state is superior to the original state.

[0113] Furthermore, to mitigate potential cognitive biases in AI models, this invention introduces a quality gating mechanism based on deterministic statistics. Before updating the production database, the system must calculate and verify the following two hard metrics:

[0114] 1. Global Topology Matching Ratio: This metric measures the degree of similarity between the repaired road network geometry and the actual physical traffic trajectory skeleton, and is a key constraint to prevent geometric drift. The specific calculation process is as follows:

[0115] The system first calls an algorithm based on kernel density estimation (KDE) to aggregate the massive vehicle GPS trajectory sequences obtained in step S11 into a rasterized trajectory heatmap. Subsequently, an image skeleton extraction algorithm (such as a distance transform-based thinning algorithm or the Zhang-Suen fast parallel thinning algorithm) is used to extract the trajectory centerline skeleton with a single pixel width from the heatmap.

[0116] Next, the system projects the repaired geographic road network vector data onto the same pixel coordinate system as the trajectory centerline and performs a bidirectional Hausdorff distance calculation. The system establishes a buffer zone of a preset width (e.g., 2 meters, corresponding to half the width of a standard lane) on both sides of the vector road segment. The system then calculates the length of the trajectory centerline skeleton falling within this buffer zone. Total length of the trajectory centerline skeleton The ratio is the global topology matching ratio. If the repair operation (such as adding a new connection line) deviates significantly from the highlighted area of ​​the trajectory heatmap in spatial location (for example, the model incorrectly connects the road to the adjacent green belt), causing the ratio to fall below the preset first quality assessment threshold (such as 95%), the system will determine that the repair solution is semantically compliant but geometrically incorrect, and forcibly block the update.

[0117] 2. Local Neighbor Node Accuracy: This metric focuses on the graph theory rationality of the micro-topological structure surrounding the repair point. As a constrained planar graph, the degree distribution of nodes in a geographic road network exhibits significant engineering statistical patterns. The specific calculation process is as follows:

[0118] The system extracts a road network subgraph within the K-hop (K-Hop, e.g., K=2) neighborhood of the geometric objects involved in the repair operation. The system calculates the degree distribution entropy of all nodes in this subgraph and the variance of road segment lengths. Simultaneously, the system incorporates a graph theory-based rule library for detecting abnormal structures. For example, in non-roundabout areas, if the repair operation results in a supernode with a degree greater than 6, a debris segment less than 0.5 meters long, or causes an existing connected component to split into two isolated subgraphs, the system normalizes the count of these abnormal structures into a local neighbor node accuracy score. Only when this score exceeds a preset second quality assessment threshold (e.g., 99%) is the repair operation considered not to have violated the graph theory properties of the road network.

[0119] In a preferred embodiment, the system will generate a database transaction commit instruction and execute the operation of updating the production database only when the semantic scoring verification of counterfactual reasoning passes and both of the above two statistical hard indicators meet the requirements.

[0120] Step S16: After the counterfactual reasoning verification is passed and before the production database is updated, perform the virtual traversal semantic verification step.

[0121] In practical geographic information engineering, a highly deceptive high-order error exists: a road network may be geometrically complete and connected (connected and smooth), and appear reasonable on satellite imagery (with road surface texture), but in the physical world, it may be impassable due to roadblocks, bollards, temporary barriers, or traffic control. Traditional geometric detection and static image detection struggle to detect such logical dead ends. Therefore, this invention introduces a virtual traversal semantic verification mechanism.

[0122] The specific implementation process is as follows:

[0123] The first step: Construction of local road network subgraphs and scene instantiation.

[0124] The system first locates the area where the topology change has occurred based on the modified road network data. To reduce computational complexity, the system uses a breadth-first search algorithm to extract the target road segment with the topology change and its multi-hop adjacent road segments (e.g., extending 500 meters upstream and downstream), constructing a local road network subgraph. This subgraph retains complete road attributes (such as speed limits and number of lanes).

[0125] Next, the system sets virtual navigation start and end points in the local road network subgraph that must traverse the target road segment. The selection of these two points is adversarial, designed to force the navigation algorithm to traverse the newly repaired road segment. For example, if the edge between node A and node B is repaired, the start point is set at node S upstream of A, and the end point is set at node E downstream of B. Simultaneously, the system uses spatial indexing technology to retrieve high-resolution satellite remote sensing images and street view images that precisely correspond to the spatial location of the path, and attaches this visual data to the corresponding edge attributes of the subgraph, completing the instantiation of the simulation scene.

[0126] The second step: the role-playing and perceptual input of the multimodal large language model.

[0127] The system uses a multimodal large language model to act as a virtual driver. To activate this capability, the system constructs an input sequence containing system-level instructions, such as: "You are an experienced driver who strictly adheres to traffic rules. Now you need to determine whether the road ahead is actually passable based on visual perception and map data. Please observe whether there are any physical obstacles or no-entry signs on the road."

[0128] The system inputs local road network subgraphs (converted into adjacency matrices or structured text descriptions), virtual navigation start points, virtual navigation end points, and visual obstacle features from high-resolution satellite remote sensing images and street view images into a multimodal large language model. It is important to note that these visual obstacle features are high-dimensional semantic vectors extracted from the images by the aforementioned visual encoder. They explicitly encode obstacle information such as bollards, walls, pits, and no-entry signs, and are given extremely high weights in the model's attention layer.

[0129] The third step: attempting to generate and detect conflicts in a continuous navigation instruction set.

[0130] The system drives a multimodal large language model to attempt to generate a continuous set of navigation instructions from the virtual navigation start point to the virtual navigation end point. This is a sequence decision and generation task. The model needs to output instructions such as go straight for 200 meters or turn left at an intersection, just like real navigation software. However, before generating each instruction, the model must verify the current visual environment.

[0131] For example, when the model plans a path through a newly repaired target road segment, its attention mechanism focuses on the street view image features corresponding to that segment. If the street view shows the road is unobstructed, the model successfully generates the following: continue along the current road, passing through the newly connected road segment. If the street view shows a locked gate or a private road sign at the intersection, the logic gates within the model will trigger a conflict alarm: although the map topology shows connectivity, visual perception detects a hard physical blockage, preventing the execution of the straight-through command.

[0132] Fourth step: Determining the verification result.

[0133] If the multimodal large language model can successfully generate a complete and logically consistent instruction set that reaches the destination, the verification is considered successful. Conversely, if the multimodal large language model cannot generate a continuous navigation instruction set due to visual impairment features or topological logic conflicts, and stops or reports an error midway, the virtual traversal semantic verification step is considered unsuccessful.

[0134] Step S17: Negative topology constraint feedback and iterative repair.

[0135] When virtual traversal validation fails, simply reporting an error is insufficient. This invention constructs a negative topological constraint feedback mechanism.

[0136] The specific implementation process is as follows:

[0137] The first step: capturing and analyzing the description of the cause of the blockage.

[0138] The system utilizes a log capture module to capture the blocking reason description output by the multimodal large language model when it fails to generate a continuous navigation command set. This description is usually in natural language form, such as: Navigation failed, unable to pass through road segment with ID 1024 because a permanent concrete barrier was identified in the street view image and the road width is less than 1.5 meters.

[0139] Subsequently, the system invokes a dedicated semantic parser to perform structured parsing of the blockage reason description. The parser extracts the core entities: [Road Segment ID: 1024], [Obstacle Type: Concrete Wall], and [Location: Between Node A and Node B].

[0140] The second step: generating negative topological constraints.

[0141] Based on the analysis results, the system generates corresponding negative topological constraints. Traditional GIS restoration typically focuses only on positive constraints (i.e., where connectivity should exist). This invention introduces negative constraints, explicitly specifying where connectivity is absolutely prohibited. Negative topological constraints include prohibited node pairs (e.g., {Node_A, Node_B, Connection=False}) or prohibited direction vectors (e.g., {Edge_ID, Direction=OneWay_Reverse}). These constraints are encapsulated in a strong rule format.

[0142] The third step: Reconstruction of the solution graph and recursive execution.

[0143] The system feeds back the negative topological constraint to the solution graph construction process in step S14. At this point, the multimodal large language model is triggered again. It receives the newly generated negative constraint along with the original diagnostic report. Based on the objective of correcting the error and the constraint that A and B cannot be connected, the model performs a new logical planning.

[0144] The specific feedback mechanism employs dynamic prompt injection technology. The system transforms the negative topological constraints generated in step S17 into a prohibition list described in natural language, such as "Constraint: DO NOT connect Node X and Node Y due to physical barrier identified in StreetView," and appends it to the System Prompt context of the code generation module in step S14. At this point, when the multimodal large language model replans the directed acyclic graph, it uses the thought chain mechanism to retrieve this prohibition list, actively pruning repair paths containing prohibited connections, thereby generating a compliant new solution.

[0145] For example, the model might infer that since there's a wall between A and B preventing a direct connection, based on trajectory data, the vehicle detoured via the adjacent node C. Therefore, the correct solution isn't to connect AB, but rather to connect AC and CB. Based on this new reasoning, the model regenerates the atomic operation sequence, constructs a new DAG graph, and triggers the step of writing the repair code script.

[0146] The system will re-execute code generation, counterfactual reasoning verification, and virtual traversal verification. This process may be repeated multiple times until the final solution resolves the topological error and passes all physical and logical verifications, i.e., until the virtual traversal semantic verification step is deemed successful. To support the above complex calculation process, embodiments of the present invention also provide an apparatus (system) and electronic device for executing the method.

[0147] like Figure 2 As shown, the geographical road network topology consistency self-checking and repair system based on a multimodal large language model includes the following core functional modules in its logical architecture:

[0148] Multimodal Data Sensing and Preprocessing Module: This module is configured as a data access layer, responsible for high-concurrency interaction with the underlying PostGIS spatial database, object storage service (OSS), and real-time traffic flow data stream. This module includes a built-in coordinate system transformation unit to unify the spatial reference systems of data from different sources to the WGS-84 coordinate system. In addition, this module also includes an image pyramid construction unit for layering and tiling high-resolution satellite remote sensing imagery, and a trajectory cleaning unit for denoising, drift point removal, and map matching preprocessing of vehicle GPS trajectory sequences.

[0149] The Unified Semantic Space Projection Module integrates a visual encoder and a sequence encoder that run in parallel. The visual encoder uses a ViT-Large-based backbone network to transform image and street view pixel data into high-dimensional feature vectors; the sequence encoder uses a Transformer-XL architecture to capture long-term temporal dependency features of the trajectory flow. The core component of this module is a trainable cross-modal projection network that aligns heterogeneous feature vectors to the token embedding space of a large language model through multi-layer nonlinear mapping and layer normalization.

[0150] The core module for large-scale model inference and self-checking deploys a multimodal large language model fine-tuned with knowledge from the geographic information domain. This module includes a prompt word engineering engine for dynamically assembling input sequences containing task instructions, road network context descriptions, and multimodal feature vectors. It also includes a logic conflict detector to parse the model's output stream, identify specific types of topological errors such as false breaks and missing intersection constraints, and generate standardized JSON-formatted diagnostic reports.

[0151] Automated Repair and Code Generation Module: This module integrates a dedicated code generation model and graph construction algorithm engine. It receives diagnostic reports, transforms them into a directed acyclic graph (DAG) consisting of atomic operation nodes, and generates executable Python or SQL repair scripts for each node. The module includes a built-in syntax checker and sandbox environment to ensure that the generated code is syntactically correct and does not violate database integrity constraints during execution.

[0152] Closed-loop verification and feedback iteration module: This module includes a counterfactual reasoning engine, used to compare the road network rationality scores before and after repair; a virtual traversal simulation engine, used to perform path search on local subgraphs based on the A* algorithm or Dijkstra algorithm, and call the large model for visual consistency verification; and a negative constraint manager, used to extract the cause of the blockage and generate feedback signals when verification fails, driving the reasoning module to replan.

[0153] This invention also provides an electronic device intended to serve as a hardware carrier for performing the above-described methods. Considering that this invention relates to large-scale model inference and the processing of massive spatial data, the electronic device is preferably a high-performance server or server cluster.

[0154] Specifically, electronic devices include:

[0155] Memory is used to store computer programs and massive amounts of multimodal data. This memory adopts a multi-level storage architecture, including high-speed random access memory (DRAM) for loading model parameters and runtime intermediate variables, and non-volatile memory (such as NVMe SSD arrays) for persistent storage of geographic road network vector data, satellite image tile libraries, and trajectory index libraries.

[0156] The processor, coupled with memory, not only includes a general-purpose central processing unit (CPU, such as Intel Xeon or AMD EPYC series) for performing logic control, data scheduling, and operating system kernel tasks, but more importantly, it also includes dedicated AI accelerator cards, such as NVIDIA A100 / H100 GPUs or self-developed NPU chips. These accelerator cards are connected via high-speed interconnect buses (such as NVLink) and are specifically designed to perform intensive matrix multiplication and tensor convolution operations in multimodal large language models.

[0157] The processor is configured to execute a computer program to implement the steps of the self-check and repair method for geographic road network topology consistency based on a multimodal large language model in any of the foregoing embodiments.

[0158] Furthermore, embodiments of the present invention also provide a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it can implement the aforementioned self-testing and repair methods. This storage medium includes, but is not limited to, cloud object storage buckets, enterprise-grade disk arrays, optical discs, or flash drives.

[0159] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for self-checking and repairing the topological consistency of a geographic road network based on a multimodal large language model, characterized in that, include: Acquire multimodal data of the area to be inspected, including geographic road network vector data, high-resolution satellite remote sensing images, vehicle GPS trajectory sequences, and street view images of key locations; A multimodal encoder is used to extract the modal features of the multimodal data, and a spatial projection operator is used to project the modal features onto a unified language tag space to achieve semantic alignment between visual features and geographic descriptions. The semantically aligned multimodal features are input into a multimodal large language model for self-inspection and reasoning. The road network topology map is compared with visual connectivity and behavioral patterns, and a diagnostic report containing road network topology errors is output. The diagnostic report is transformed into a solution diagram using the multimodal large language model, and the code generation module is called to automatically write repair code scripts based on the solution diagram; The script for fixing the code modifies the geographic road network vector data, and the modified road network data is fed back to the multimodal large language model for closed-loop verification of counterfactual reasoning. If the verification passes, the production database is updated.

2. The method for self-checking and repairing the topology consistency of a geographic road network based on a multimodal large language model according to claim 1, characterized in that, The step of extracting modal features from the multimodal data using a multimodal encoder and projecting these features onto a unified language tagging space using a spatial projection operator includes: The global contextual features of the high-resolution satellite remote sensing image and the fine semantic features of the street view image are extracted using a visual encoder. The behavioral pattern features of the vehicle's GPS trajectory sequence are extracted using a trajectory encoder; The spatial projection operator, which serves as a multilayer perceptron neural network structure, is used to project and fuse the geographic road network vector data, the global context features, the fine semantic features, and the behavioral pattern features, and these are then used as inputs to the multimodal large language model.

3. The method for self-checking and repairing the topological consistency of a geographic road network based on a multimodal large language model according to claim 2, characterized in that, The step of extracting behavioral pattern features from the vehicle's GPS trajectory sequence using a trajectory encoder includes: The original trajectory coordinates of the vehicle's GPS trajectory sequence are obtained as input; The original trajectory coordinates are extracted using the trajectory encoder to obtain the behavioral pattern features, which are then used by the multimodal large language model to perform road network logic verification.

4. The method for self-checking and repairing the topology consistency of a geographic road network based on a multimodal large language model according to claim 1, characterized in that, The step of inputting semantically aligned multimodal features into a multimodal large language model for self-checking inference includes: By comparing the road network topology map with the high-resolution satellite remote sensing image, if the image shows continuous roads but there are no edge connections in the road network topology map, it is determined to be a false break error; Logical verification is performed based on the vehicle's GPS trajectory sequence. If the trajectory data shows frequent turning behavior at intersections but the road network topology map does not establish turning restrictions or connection nodes, it is determined to be an error of missing intersection constraints. By integrating the semantics of the street view image, if a physical obstacle or traffic sign is identified but the road network topology map shows it as a passable road segment, it is determined that the dead end is missing an end sign.

5. The method for self-checking and repairing the topological consistency of a geographic road network based on a multimodal large language model according to claim 1, characterized in that, The process of converting the diagnostic report into a solution diagram using the multimodal large language model includes: Based on the divide-and-conquer strategy, the repair tasks in the diagnostic report are broken down into multiple atomic operations, including adding nodes, deleting edges, and modifying attributes. Based on the execution logic order of the atomic operations, a directed acyclic graph is constructed as the solution graph.

6. The method for self-checking and repairing the topology consistency of a geographic road network based on a multimodal large language model according to claim 5, characterized in that, The step of automatically writing a repair code script based on the code generation module according to the solution diagram includes: For each atomic operation in the directed acyclic graph, the multimodal large language model is driven to generate a corresponding independent function or code snippet. The individual functions or code snippets are arranged into the repair code script, which is configured to directly perform add, delete, and modify operations on the geographic information system database.

7. The method for self-checking and repairing the topological consistency of a geographic road network based on a multimodal large language model according to claim 1, characterized in that, The step of feeding the modified road network data back to the multimodal large language model for closed-loop verification of counterfactual reasoning includes: Based on the multimodal large language model with the ability to judge and score the rationality of roads, the first rationality score of the road network structure before modification and the second rationality score of the road network structure after modification are obtained respectively. Compare the first reasonableness score with the second reasonableness score; If the difference between the second reasonableness score and the first reasonableness score exceeds a preset threshold, the counterfactual reasoning verification is deemed successful, confirming that the modification is effective and a positive optimization.

8. The method for self-checking and repairing the topological consistency of a geographic road network based on a multimodal large language model according to claim 7, characterized in that, Before updating the production database, the following is also included: Calculate the global topology matching ratio and local neighbor node accuracy of the modified road network data; When the counterfactual reasoning verification passes, and both the global topology matching ratio and the local neighbor node accuracy meet the preset quality assessment threshold, the operation of updating the production database is executed.

9. The method for self-checking and repairing the topological consistency of a geographic road network based on a multimodal large language model according to claim 7, characterized in that, After the counterfactual reasoning verification is passed and before the production database is updated, the process further includes performing a virtual traversal semantic verification step: Based on the modified road network data, the target road segment that has undergone topology change and its multi-hop adjacent road segment are extracted to construct a local road network subgraph; In the local road network sub-map, a virtual navigation start point and a virtual navigation end point that must be crossed the target road segment are set, and the high-resolution satellite remote sensing image and the street view image corresponding to the spatial location of the target road segment are retrieved; Using the multimodal large language model to act as a virtual driver, the local road network sub-map, the virtual navigation start point, the virtual navigation end point, and visual obstacle features from the high-resolution satellite remote sensing image and the street view image are input into the multimodal large language model; The multimodal large language model is driven to attempt to generate a continuous set of navigation instructions from the virtual navigation start point to the virtual navigation end point; If the multimodal large language model is unable to generate the continuous navigation instruction set due to visual impairment features or topological logic conflicts, then the virtual traversal semantic verification step is determined to be unsuccessful.

10. A method for self-checking and repairing geographic road network topology consistency based on a multimodal large language model according to claim 9, characterized in that, After determining that the virtual traversal semantic verification step fails, the method further includes: Capture the description of the blocking reason output by the multimodal large language model when it fails to generate the continuous navigation instruction set; The description of the blocking cause is semantically parsed to generate corresponding negative topological constraints, which include prohibited node pairs or prohibited direction vectors. The negative topological constraints are fed back to the construction process of the solution graph, and the multimodal large language model is triggered to re-execute the transformation of the diagnostic report and the writing of the repair code script based on the negative topological constraints, until the virtual traversal semantic verification step is passed.

Images