Historical place name spatio-temporal reconstruction method and device based on geographical context

Abstract

The application discloses a history place name space-time reconstruction method and device based on geographical context, relates to the field of geographic information management, and comprises the following steps: S1, constructing and training an optimized multi-modal language model; S2, acquiring history place name data; S3, processing the history place name data to form a global history place name annotation data set; S4, fusing the global history place name annotation data set; S5, extracting geographical context features; S6, constructing a history place name space-time state model and reconstructing to obtain a reconstruction result; and S7, visually displaying the reconstruction result. The space-time dislocation problem of history place names is solved through geographical context understanding and multi-source data fusion, and a multi-modal language model is introduced, so that the place name reconstruction makes up for the shortcomings of traditional methods. Through the modeling of geographical context, the reconstruction of history place names ensures the consistency of space and time, and in combination with the space-time background of historical changes, the timeliness and accuracy of data are enhanced, and the problems of missing history place name information and space-time dislocation are effectively solved.

Description

Method and apparatus for spatiotemporal reconstruction of historical place names based on geographical context Technical Field

[0001] This invention relates to the field of geographic information management, and in particular to a method and apparatus for spatiotemporal reconstruction of historical place names based on geographic context. Background Technology

[0002] In the new era, surveying and mapping geographic information plays a crucial role as a strategic data resource and a new factor of production. Place name information, as a core component, is not only a unique identifier for geographic entities but also carries profound historical and cultural significance. As a precise tool for spatial positioning, it plays an indispensable role in the economy, society, scientific research, national defense, and people's daily lives. In the context of the big data era, the importance of place name information is even more prominent. It has become a key "bridge" connecting multi-source big data and transcending domain boundaries, not only achieving precise positioning but also promoting the deep integration, exchange, and sharing of information, laying a solid foundation for the construction of a smart society. Particularly noteworthy is that historical place name data contains rich cultural connotations and practical value, demonstrating broad application potential. However, the lack of historical place name information for the corresponding year leads to "spatiotemporal misalignment" when users utilize historical remote sensing imagery data, affecting user experience and the application value of the data. Summary of the Invention

[0003] The purpose of this invention is to design a method and apparatus for spatiotemporal reconstruction of historical place names based on geographical context in order to solve the above-mentioned problems.

[0004] The present invention achieves the above objectives through the following technical solutions:

[0005] Methods for spatiotemporal reconstruction of historical place names based on geographical context include:

[0006] S1. Construct an initial multimodal language model and obtain a training dataset for training and optimization to obtain an optimized multimodal language model. The multimodal language model is used for the prediction of missing place names.

[0007] S2. Obtain historical place name data to be reconstructed through multiple channels; historical place name data includes geographic national conditions data, administrative boundary data, remote sensing and imagery data, and open Internet data. Geographic national conditions data includes land space monitoring data, land use data, and topographic elevation data. Administrative boundary data includes annual administrative division change data from the civil affairs department and administrative boundary data from land change surveys. Remote sensing and imagery data includes multi-temporal remote sensing images, historical satellite images, and orthophoto maps. Open Internet data includes geographic information public service platforms, social media place name annotations, and digitized historical maps.

[0008] S3. Process the historical place name data to form a comprehensive historical place name annotation dataset;

[0009] S4. Perform multi-source data fusion on the entire historical place name annotation dataset;

[0010] S5. Extract the geographic context features of the fused data;

[0011] S6. Construct a spatiotemporal state model of historical place names, and use the spatiotemporal state model of historical place names and a multimodal language model to complete the spatiotemporal reconstruction of historical place names and obtain the reconstruction results.

[0012] S7. Visualize the reconstruction results.

[0013] A spatiotemporal reconstruction device for historical place names based on geographical context includes:

[0014] Storage; storage is used to store computer programs;

[0015] An executor; the executor is used to execute a computer program in the storage, and when the computer program is executed, it implements the spatiotemporal reconstruction method of historical place names based on geographic context as described above.

[0016] The beneficial effects of this invention are as follows: it addresses the spatiotemporal misalignment problem of historical place names by using geographic context understanding and multi-source data fusion. Furthermore, it introduces a multimodal language model, ensuring that place name reconstruction not only relies on traditional image recognition and data matching techniques but also incorporates multiple perspectives from linguistics and geography, thus overcoming the shortcomings of traditional methods. Compared to traditional techniques, this method, through geographic context modeling, ensures that the reconstruction of historical place names does not depend on a single data source but achieves spatiotemporal consistency through multi-dimensional contextual analysis. This avoids common mismatches in traditional methods and enhances the timeliness and accuracy of data by incorporating the spatiotemporal background of historical changes, effectively solving the problems of missing historical place name information and "spatiotemporal misalignment." Attached Figure Description

[0017] Figure 1 is a schematic diagram of the processing of historical place name data;

[0018] Figure 2 is a schematic diagram of converting all place name data into a unified coordinate system;

[0019] Figure 3 shows schematic diagrams before and after geometric correction;

[0020] Figure 4 is a schematic diagram of multimodal feature fusion;

[0021] Figure 5 is a directory diagram of the organized map tile file data;

[0022] In Figure 3, the image on the left is the image before geometric correction, and the image on the right is the image after geometric correction. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0024] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0025] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0026] In the description of this invention, it should be understood that the terms "upper," "lower," "inner," "outer," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0027] Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0028] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, terms such as "set" and "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0029] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0030] Methods for spatiotemporal reconstruction of historical place names based on geographical context include:

[0031] S1. Construct an initial multimodal language model and obtain a training dataset for training and optimization to obtain an optimized multimodal language model. The multimodal language model is used for the prediction of missing place names.

[0032] The multimodal language model is trained using fused text, administrative boundaries, remote sensing imagery, and spatiotemporal sequences (place name change records). Combining Natural Language Processing (NLP) and Geospatial Analysis (GIS), a Transformer-based multimodal model (CLIP for multimodal image and text processing, and MViT for multimodal image processing) is employed for joint training, incorporating geographic context features and place name text semantics. The steps are as follows:

[0033] Training data preparation: Text data: place name descriptions, historical documents, or map annotations. Image data: map images, remote sensing images, or scanned copies of historical maps. Spatiotemporal attributes: latitude and longitude and timestamps corresponding to place names. Data partitioning is as follows: Training set (70%): used for model training. Validation set (15%): used for hyperparameter tuning and overfitting detection. Test set (15%): used for final performance evaluation.

[0034] Masked prediction training: Using a masked language model (MLM), input a randomly masked portion of place name text, predict the masked place names, and calculate the cross-entropy loss function. , is represented as: ;in, represents the real word at the i-th masked position; x represents the input sequence, including the mask markers; This represents the model parameters; N represents the total number of masked positions in a batch of data.

[0035] Spatiotemporal interpolation training: Using the STI training method, inputting partially missing spatiotemporal attributes (e.g., only longitude is provided, latitude is missing, place name time is missing, etc.), predicting the missing attribute values. The mean squared error loss function MSE is calculated, expressed as: Where n is the number of samples, This represents the model's prediction for the i-th sample. These are the actual sample values. If there are many outliers in the results, then the Huber loss function needs to be introduced.

[0036] Instruction fine-tuning: Optimizing the performance of multimodal language models on specific tasks. Fine-tuning is performed using labeled data (e.g., spatiotemporal attributes of known place names). A joint loss (weighted sum) is combined with masked language modeling (MLM) and sentence type identification (STI). Reinforcement learning optimizes the multimodal language model through human feedback (e.g., reasonableness ratings for place name predictions). The Proximal Policy Optimization (PPO) algorithm is used to maximize the reward function (e.g., prediction accuracy).

[0037] S2. Obtain historical place name data to be reconstructed through multiple channels; historical place name data includes geographic national conditions data, administrative boundary data, remote sensing and imagery data, and open Internet data. Geographic national conditions data includes land space monitoring data, land use data, and topographic elevation data. Administrative boundary data includes annual administrative division change data from the civil affairs department and administrative boundary data from land change surveys. Remote sensing and imagery data includes multi-temporal remote sensing images, historical satellite images, and orthophoto maps. Open Internet data includes geographic information public service platforms, social media place name annotations, and digitized historical maps.

[0038] S3. Process historical place name data to form a comprehensive historical place name annotation dataset. The processing of historical place name data will aim to build a multi-source historical place name database, strictly following six major processes: preprocessing, fusion, matching analysis, verification, quality inspection, and database construction. Specifically, this includes:

[0039] Benchmark selection: Based on the obtained data, the historical place name data of the most recent year was used as the benchmark.

[0040] Data Supplementation and Detection: Utilizing historical place name data from the most recent year, a 500m radius is used as the influence range based on the place name's location. This is compared with other data from that year using buffer analysis to detect the existence of duplicate place names. The buffer is represented as follows: Where A represents target geographic feature data from other years, r is the buffer distance (currently 500m), d is the distance function (usually Euclidean distance), and x is the current place name and location information. The distance function is as follows: , It is the starting latitude. It is the latitude of the endpoint. It is the longitude of the starting point. It is the endpoint longitude; if there is a place name with the same name under the buffer rule, the baseline data is used directly; if there is no place name with the same name, other collected data is used for supplementary detection.

[0041] Data verification: Data verification is conducted through internet queries (such as Baidu, Gaode geocoding, and reverse geocoding services) and satellite imagery data.

[0042] Baseline Dataset: Coordinate System 1: All data is converted to a unified coordinate system, using the CGCS2000 National Geodetic Coordinate System, with latitude and longitude (geodetic coordinates) as the coordinate format. Data processing techniques such as ETL or professional GIS software are used for batch coordinate transformation. The flowchart is shown in Figure 2. Datasource and datasource2 are the operation data source and coordinate system reference data source, respectively. Time Base Alignment: Date information should be expressed in Gregorian calendar, and time information should be expressed in Beijing time. Vector data, raster data, and text data are converted to GeoJSON, GeoTIFF (raster data storage format), and CSV (comma-separated document) standard formats. Geometric Correction: For data with geometric position errors or deformations, control point registration can be used. Feature points are extracted using the SIFT scale-invariant feature transform algorithm, and RANSAC (random sampling consensus algorithm) can be combined to eliminate incorrect matches. The SIFT scale-invariant feature transform algorithm can detect key points in the image that are scale-invariant and rotation-invariant. The descriptor of each key point is a 128-dimensional vector composed of 8 directional histograms in a 4×4 grid. ,in It is a Gaussian weighted function. Figure 3 shows a schematic diagram of the pixel gradient direction before and after geometric correction.

[0043] S4. Perform multi-source data fusion on the entire region's historical place name annotation dataset; perform multi-source data fusion on historical place name data with administrative division data, annual administrative division change data from the civil affairs department, and administrative boundary data from land change surveys, specifically including:

[0044] Spatial data fusion: Using GIS software (such as ArcGIS) to merge multi-source historical place name data into complete geographic entities according to administrative divisions, with townships as the smallest geographic entity.

[0045] Attribute fusion: Through the "Data Source" field, multi-source historical place name data are dynamically spliced ​​together, and attribute information is fused based on the smallest geographic entity.

[0046] Sensitive information processing: Classified fields (military place names) need to be matched separately using a keyword database and encrypted (using the national cryptographic algorithm SM3+ encryption).

[0047] S5. Extract the geographic context features of the fused data; historical place name geographic entity data has multimodal information. Context feature extraction is performed on the spatial, attribute, and spatiotemporal features of historical place name entities, specifically including:

[0048] Spatial Feature Extraction: This example uses VGG-16 for spatial information encoding. VGG-16 is a convolutional neural network model that includes a feature module and a classification module, possessing powerful feature extraction capabilities. In this example, the last fully connected layer and the Softmax normalized exponential function layer of the VGG-16 are removed to obtain pre-trained visual features, resulting in a 4096-dimensional embedding of entity spatial information. ;

[0049] Attribute Feature Extraction: Numerical information in historical place name entities exists in the form of key-value pairs (k,v), where k represents the attribute key and v represents the value. For example, the place name "Government Service Hall" has attribute information (Type: Government Agency). This example uses a BERT pre-trained language model to obtain the 768-dimensional embedding of the attribute key k. Then, a 768-dimensional fully connected layer is used to obtain the embedding of the value v. Finally, the key embedding and value embedding are concatenated, and a fully connected layer is used to obtain the 768-dimensional embedding of the entity's numerical information. ;

[0050] Spatiotemporal Feature Extraction: The spatiotemporal information of historical place names exists in multiple dimensions, including attributes and space. This information can typically be converted into spatiotemporal textual descriptions to record the spatiotemporal features of historical place names. ERT is considered effective for extracting textual features. In this example, a pre-trained BERT-based improved model, Sentence-BERT, is used as the textual feature extractor, which can obtain good sentence-level semantic representations. This paper uses the results of the last hidden layer of the BERT model to obtain a 768-dimensional embedding of the entity textual information. ;

[0051] Multimodal feature fusion: Each element of the fusion vector corresponds to the feature interaction between different modal features. This method introduces supernode multimodal fusion based on bilinear fusion. The multimodal fusion method can be extended to any number of modalities; extracted feature values , , Figure 4 shows a schematic diagram of multimodal feature fusion, where the supernode representation is formed through tensor fusion. , Represented as: Where W is the weight matrix and b is the bias. It is a tensor outer product The resulting high-dimensional tensor , m represents the m-th mode; Z is represented as: ;in, , and These are spatiotemporal features, spatial features, and attribute features, respectively. This method simultaneously captures the entity's unimodal, bimodal, and trimodal interaction information. To reduce complexity, a low-rank tensor decomposition w is used to replace the weight vector W, expressed as: The smallest k that makes the decomposition effective is called the rank of the tensor; vector set It is called the decomposition factor of the original weight tensor, where m represents the m-th mode and i represents the i-th decomposition factor.

[0052] S6. Construct a spatiotemporal state model of historical place names, and use the spatiotemporal state model of historical place names and a multimodal language model to complete the spatiotemporal reconstruction of historical place names and obtain the reconstruction results. The spatiotemporal state model includes three parts: spatiotemporal state set of place name regions, spatiotemporal state set of place name sub-regions, and regional grid coding.

[0053] Spatiotemporal Status Set of Place Name Regions: For each place name region, maintain a time-series database or use a version control system to track its changes over time. Record each modification to the region's status, including the reason for the change (such as administrative division adjustments), the date, and the specific changes.

[0054] Spatiotemporal state set of place name sub-regions: For place name sub-regions (such as blocks and villages), a similar data structure is used to construct a parent-child relationship graph, clarifying the subordinate relationship between each sub-region and its position relative to the main region.

[0055] Regional Grid Coding: This example designs a two-level regional grid coding system: Level 1 coding: Generates a coding prefix based on the main attributes of the place name (such as administrative level, type, etc.). For example, "provincial level" can be coded as "P", "city level" as "C", etc. Level 2 coding: Assigns location codes based on geographical location. Existing geographic grid systems (UTM coordinate systems) can be used to map geographical locations to numerical strings. The two coding parts are concatenated to form a complete grid coding string. For example, "PC1234567890", where "PC" represents the city level and "1234567890" is the specific location code.

[0056] Historical place name reconstruction in time and space:

[0057] Initial place name identification: Extracting known place names and their spatiotemporal attributes from historical documents and remote sensing images.

[0058] Missing place name prediction: Use a multimodal language model to predict place names for missing years and verify their rationality by combining geographical context (such as topography and water system).

[0059] Spatiotemporal continuity modeling: Continuous spatiotemporal distribution of place names is generated using spatiotemporal interpolation algorithms (such as IDW inverse proportional weight interpolation and ST-ResNet spatiotemporal network model). This includes interpolation point place name address location information. ,in, Let i be the position of the i-th neighboring point; The distance between the i-th neighboring point and the center point; p is a power parameter constant with a value of 2.

[0060] Cross-validation: Compare with land change survey data and administrative division data from the civil affairs department to correct the prediction results.

[0061] S7. Visualize the reconstruction results. Based on the publicly available electronic map tile data production standards, use relevant GIS platform software to create map tiles with level 7-17 place name annotations covering the entire area for each year. This includes determining tile standards, tile format, tile range, and tile data parameters. Map tiles are image units formed by dividing a map with a defined geographical coverage area according to certain grid division rules. This is achieved through the map caching function of the GIS software platform, based on the tile data production standards.

[0062] Map tile specifications: The starting point for map tile division is 180 degrees west longitude and 90 degrees north latitude, increasing in rows and columns towards the east and south; the size of each map tile is 256 pixels * 256 pixels; the map tile data format is PNG; the map scale and resolution comply with the "Electronic Map Data Specification for Geographic Information Public Service Platform" (CH / Z 9011-2011);

[0063] Map Tile Pyramid Rule: To allow various map services from distributed nodes to be overlaid, a unified pyramid layering rule must be adopted, with each layer having a fixed display scale (i.e., the ground resolution of the tile). The display scale is calculated as follows: , Level indicates the scale level, with a minimum of 0. The screen resolution is 96 dpi. The Earth's major radius is 6,378,137 meters, which is the parameter specified by the 2000 National Geodetic Coordinate System.

[0064] Map tile file data organization: The "Map Tile Dataset" is the root directory of the map tile file data, as shown in Figure 5. The directories under it are map tile hierarchy (directory name naming convention: "L + level", L1, L2, L3, ...). The map tile hierarchy directory is followed by the row of the map tile matrix at that level (directory name naming convention: "R + row number", RO, R1, R2, ...). The row directory contains the specific map tile files (file name naming convention: "C + column number" c0.png (or C0jpg), C1.png (or C1.jpg), C2.png (or C2.jpg), ...).

[0065] Create a new service in a Geographic Information System (GIS) professional service engine platform software (such as GeoServer or SuperMap iServer), select the WMTS1.0.0 Web map tile service protocol, and set the coordinate system to the National Geographic Coordinate System CGCS2000 (EPSG:4490).

[0066] Use APIs such as Leaflet and OpenLayers (web map development framework) to display place name map services and complete the visualization of historical place name spatiotemporal data.

[0067] This method effectively reconstructs historical place names by understanding the geographical context and constructing a multimodal language model. At the same time, it cross-validates historical place names by integrating data such as national geographic conditions and land space monitoring data, administrative boundary data from land change surveys, annual administrative division and change data from civil affairs departments, data from geographic information public service platforms, and publicly available internet data. This allows for a more complete and accurate reconstruction of historical place name data from different years and periods, thereby solving the problem of "spatiotemporal misalignment" and improving the completeness and accuracy of the data.

[0068] A spatiotemporal reconstruction device for historical place names based on geographical context includes:

[0069] Storage; storage is used to store computer programs;

[0070] An executor; the executor is used to execute a computer program in the storage, and when the computer program is executed, it implements the spatiotemporal reconstruction method of historical place names based on geographic context as described above.

[0071] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. A method for spatiotemporal reconstruction of historical place names based on geographical context, characterized in that, include: S1. Construct an initial multimodal language model and obtain a training dataset for training and optimization to obtain an optimized multimodal language model. The multimodal language model is used for the prediction of missing place names. The initial training optimization of the multimodal language model specifically includes: (1) dividing the training dataset into a training set, a validation set, and a test set; (2) using a masked language model (MLM) to randomly mask the training set, using the multimodal language model to predict the masked place names, and calculating the cross-entropy loss function. (3) Using the STI training method, construct partially missing spatiotemporal attributes based on the training set, use a multimodal language model to predict the missing attribute values, and calculate the mean squared error loss function (MSE); (4) Combine the cross-entropy loss function (5) Optimize the multimodal language model using the mean square error loss function (MSE); (6) Fine-tune the multimodal language model using human feedback to obtain the optimized multimodal language model; S2. Obtain historical place name data to be reconstructed through multiple channels; Historical place name data includes geographic national conditions data, administrative boundary data, remote sensing and image data and open Internet data. Geographic national conditions data includes land space monitoring data, land use data and topographic elevation data. Administrative boundary data includes annual administrative division change data of the civil affairs department and administrative boundary data of land change survey. Remote sensing and image data includes multi-temporal remote sensing images, historical satellite images and orthophoto maps. Open Internet data includes geographic information public service platform, social media place name labeling and digitized results of historical maps; S3. Process the historical place name data to form S4. Converge the historical place name annotation dataset across the entire region; S5. Extract the geographic context features of the fused data; S6. Construct a spatiotemporal state model of historical place names, and use the spatiotemporal state model of historical place names and a multimodal language model to complete the spatiotemporal reconstruction of historical place names and obtain the reconstruction results; The spatiotemporal reconstruction of historical place names using the spatiotemporal state model of historical place names and a multimodal language model specifically includes: 1) Initial place name identification: Extract known place names and their spatiotemporal attributes from historical documents and remote sensing images; 2) Missing place name prediction: Predict place names for missing years using the optimized multimodal language model, and verify the rationality by combining geographic context features; 3) Spatiotemporal continuity modeling: Generate a continuous spatiotemporal distribution of place names through a spatiotemporal interpolation algorithm, and interpolate the location information of the generated place names. Represented as: ； Let i be the position of the i-th neighboring point; The distance between the i-th neighboring point and the center point; p is a power parameter constant; 4) Cross-validation: compare with the land change survey data and the administrative division data of the civil affairs department to correct the prediction results; S7, visualize the reconstruction results.

2. The method for spatiotemporal reconstruction of historical place names based on geographical context according to claim 1, characterized in that, S3 specifically includes: S31, selecting historical place name data from the most recent year as a baseline; S32, using the selected baseline, and based on the location of the place name, using a radius of 500m as the influence range, comparing it with other data from that year, using buffer analysis P, to detect the existence of identical place names. If a place name with the same name exists within the buffer, the selected baseline is directly adopted; otherwise, other collected place name data is used for supplementary detection; S33, verifying place name data through internet queries or satellite imagery data; S34, converting all place name data into a unified coordinate system and aligning all place name data by time; S35, checking all place name data for geometric position errors and / or geometric deformations. If so, the control point registration method is used, the SIFT scale-invariant feature transformation algorithm is used to extract feature points, and the RANSAC random sampling consensus algorithm is combined to eliminate erroneous matches, achieving geometric correction and obtaining a comprehensive historical place name annotation dataset.

3. The method for spatiotemporal reconstruction of historical place names based on geographical context according to claim 1, characterized in that, In S4, multi-source data fusion specifically includes: S41, Spatial data fusion: using GIS software to merge multi-source historical place name data into complete geographic entities according to administrative divisions, with townships as the smallest geographic entity; S42, Attribute fusion: dynamically splicing multi-source historical place name data through the "Data Source" field, and fusion of attribute information based on the smallest geographic entity; S43, Sensitive information processing: matching classified fields separately through a keyword database and encrypting them.

4. The method for spatiotemporal reconstruction of historical place names based on geographical context according to claim 1, characterized in that, In S5, extracting the geographic context features of the fused data specifically includes: S51, Spatial Feature Extraction: using the modified VGG-16 model to encode spatial information and obtain the spatial features of entity spatial information. The modified VGG-16 model removes the last fully connected layer and the Softmax layer. S52, Attribute Feature Extraction: Numerical information in historical place name entities exists in the form of key-value pairs (k,v), where k represents the attribute key and v represents the value. A 768-dimensional embedding of the attribute key k is obtained using a BERT pre-trained language model; the embedding of the value v is obtained using a 768-dimensional fully connected layer. The key embedding and value embedding are connected, and a fully connected layer is used to obtain the attribute features of the entity's numerical information. S53. Spatiotemporal Feature Extraction: The spatiotemporal information attributes and multi-dimensional spatial information in historical place name entities are converted into spatiotemporal text information descriptions, and the Sentence-BERT text feature extractor is used to extract the spatiotemporal features of historical place names. S54, Multimodal Feature Fusion: Supernodes formed through tensor fusion As a geographical context feature, it is represented as: Where W is the weight matrix and b is the bias. It is a high-dimensional tensor formed by the tensor outer product ⊗. , m is the m-th mode; Z is defined as: ;in, 、 and These are the spatiotemporal features, spatial features, and attribute features, respectively. A low-rank tensor decomposition w is used to replace the weight vector W, expressed as: The smallest k that makes the decomposition effective is called the rank of the tensor, and the vector set. The factor is called the decomposition factor of the original weight tensor, where i represents the i-th factor.

5. The method for spatiotemporal reconstruction of historical place names based on geographical context according to claim 4, characterized in that, In S6, the historical place name spatiotemporal state model includes a place name regional spatiotemporal state set, a place name sub-region spatiotemporal state set, and regional grid coding. The place name regional spatiotemporal state set maintains a time-series database or uses a version control system to track its changes over time, recording each modification to the regional state, including the reason for the change, the date, and the specific changes. The place name sub-region spatiotemporal state set uses a similar data structure to construct a parent-child relationship graph for each place name sub-region, clarifying the hierarchical relationships between sub-regions and their positions relative to the main region. The regional grid coding is a two-level regional grid coding: the first-level coding generates a coding prefix based on the main attributes of the place name. The second level of coding assigns location codes based on geographical location.

6. The method for spatiotemporal reconstruction of historical place names based on geographical context according to claim 1, characterized in that, Specifically, S7 includes: ① Determining tile specifications and dividing the map with a defined geographic coverage area into several image units based on the tile specifications; ② Adopting a unified pyramid layering rule to fix the display ratio of each layer; ③ Organizing and constructing map tile file data; ④ Creating a new service in the Geographic Information System (GIS) professional service engine platform software, selecting the WMTS1.0.0 Web map tile service protocol, and setting the coordinate system to the EPSG:4490 national geographic coordinate system CGCS2000; ⑤ Using the place name map service to complete the visualization of historical place name spatiotemporal data.

7. A device for spatiotemporal reconstruction of historical place names based on geographical context, characterized in that, include: Storage; The storage device is used to store a computer program; the executor is used to execute the computer program in the storage device, and when the computer program is executed, it implements the spatiotemporal reconstruction method of historical place names based on geographical context as described in any one of claims 1-6.

Images