Vector geographic data scale identification method, device and equipment and storage medium

Abstract

The application provides a vector geographic data scale recognition method, device and equipment and a storage medium, and relates to the technical field of scale recognition. The method comprises the following steps: performing semantic classification on an attribute field text sequence of vector geographic data based on a BERT model to obtain a layer semantic label; performing adaptive division on a hotspot area of the vector geographic data by kernel density estimation, and rasterizing the divided hotspot area into a rasterized image set; performing scale recognition by fusing the layer semantic label and visual features of the rasterized image set based on a ResNeXt visual feature extraction network, and integrating the recognition results of each subgraph by using a weighted voting mechanism to obtain a scale recognition result of the vector geographic data. The application solves the problem that the scale recognition accuracy is seriously insufficient when the existing method based on a single feature faces differences in geographic area distribution and diversity in layer semantics.

Description

Technical Field

[0001] This application relates to the field of scale recognition technology, and in particular to a method, apparatus, device and storage medium for vector geographic data scale recognition. Background Technology

[0002] Vector geographic data security assessment is a crucial process for determining the confidentiality level of data in accordance with relevant national confidentiality regulations. It is a necessary prerequisite for ensuring the compliant circulation and secure application of vector geographic data. According to the "Regulations on the Scope of State Secrets in Surveying and Mapping Geographic Information Management" (Natural Resources Development

[2020] No. 95), vector geographic data security assessment typically involves multiple indicators such as area, planar accuracy, and scale. Among these, the scale reflects the degree of abstraction and accuracy of representation of spatial entities in the data and is a core indicator for measuring the sensitivity and security level of the data.

[0003] However, in actual scale assessment work, due to inconsistent data collection and processing standards, some vector geographic data lacks scale metadata, requiring manual interpretation and estimation based on expert experience, which is inefficient and highly subjective. Furthermore, the complex spatial scenes and diverse semantic elements of vector geographic data further increase the difficulty of automatic scale identification. Therefore, how to achieve accurate and automated scale identification of vector geographic data in complex scenarios has become a critical bottleneck that urgently needs to be overcome in the field of geographic information data security assessment.

[0004] The key to automated scale identification lies in uncovering the robust mapping relationship between vector geographic data features and scale. Based on different features, existing automated vector geographic data scale identification algorithms can be divided into three categories: methods based on spatial relationships, geometric statistics, and visual features.

[0005] Algorithms based on spatial relationships typically rely on expert empirical formulas, constructing a mapping relationship between spatial relationships and scale through features such as spatial proximity, topological relationships between elements, or line-of-sight models. For example, the paper "Cheng X, Wu H, Ai T, Yang M. Detail Resolution: A New Model to Describe Level of Detail Information of Vector Line Data. In: Zhou, C., Su, F., Harvey, F., Xu, J.(eds) Spatial Data Handling in Big Data Era. Advances in Geographic Information Science. Springer, Singapore, 2017." discloses a rasterization method based on visual perception, which assists in scale discrimination by quantifying the visual fusion threshold of line elements at different scales. Although the method can reflect the relationship between spatial discernibility and scale to a certain extent, it lacks adaptive capability due to its reliance on manually set thresholds and empirical rules, and is prone to introducing subjective bias.

[0006] Algorithms based on geometric statistical features generally use statistical indicators such as point density, tortuosity, and spacing to construct threshold intervals, achieving a discrete mapping from data statistical feature values ​​to scale categories. For example, the literature "Liu Wanzeng, Chen Jun, Zhao Yong, et al. Basic problems and countermeasures for geographic information security assessment [J]. Journal of Spatiotemporal Information, 2023, 30(04): 474-481" utilizes the principles of natural laws and prior statistical knowledge, using a large amount of topographic vector geographic data as samples, to establish a mapping relationship between node density intervals and scale, and verifies the feasibility of the method through experiments with contour lines and water system elements. This research breaks through the limitations of traditional scale estimation methods and provides a new theoretical framework that can be calculated and verified for information security assessment. However, the same type of element may exhibit significant differences in different geographical regions. For example, the water network areas in the south and the water systems in the arid northwest have significantly different node densities and other characteristics. In this case, single geometric statistical quantities such as point density do not take into account the feature shift caused by regional differences, resulting in insufficient accuracy in cross-regional data scale identification.

[0007] Deep learning possesses powerful automatic feature learning capabilities, bringing new breakthroughs to scale recognition tasks. Visual feature-based methods typically rasterize vector geographic data and rely on deep convolutional neural networks (CNNs) to extract visual features such as geometric shapes of spatial elements, automatically learning the nonlinear mapping relationship between samples and scales. This allows for automatic scale recognition of unknown data during the inference stage. For example, the paper "Ren Jiaxin, Liu Wanzeng, Chen Jun, et al. Knowledge-guided intelligent scale recognition of fragmented raster topographic maps [J]. Acta Geodaetica et Cartographica Sinica, 2024, 53(01):146-157" constructs an expert knowledge image pyramid dataset, statistically determines the optimal recognition size for topographic maps of different sizes, and performs sub-map scale prediction based on the improved image classification model EffienctNet after data segmentation, integrating the results to achieve data scale recognition. This type of method innovatively introduces deep learning technology, overcoming the subjective errors of manual feature design and recognizing the scale differences in different geographic regions. It breaks through the limitations of traditional single-feature recognition methods and can be well applied to cross-regional vector geographic data, providing a scalable technical path for the intelligent processing of geographic data. However, different semantic elements, such as water systems and highways, may have different visual feature representations at the same scale. Therefore, a single visual algorithm lacks semantic discrimination ability and has insufficient accuracy in cross-semantic data scale recognition.

[0008] In summary, existing spatial relationship-based algorithms rely on manual feature extraction, which is prone to introducing subjective errors; algorithms based on geometric and statistical features do not consider geographical distribution and have limited accuracy in cross-regional data; algorithms based on visual features employ deep learning, which not only reduces human bias but is also applicable to cross-regional data. However, existing visual feature-based recognition methods have not yet considered visual feature shifts caused by semantic differences in elements, and therefore cannot be applied to cross-semantic vector geographic data. Summary of the Invention

[0009] This application provides a method, apparatus, device, and storage medium for vector geographic data scale recognition. By introducing the semantic classification module of the BERT model, the model is endowed with semantic classification capabilities. On this basis, the deep vision network ResNeXt is introduced into the vector geographic data scale recognition task, and the region is revealed by combining KDE segmentation features to establish a stable mapping from cross-regional vector geographic data to scale. Finally, a weighted voting mechanism is integrated to complete the data scale recognition, further improving the reliability of scale recognition.

[0010] Firstly, this application provides a method for identifying the scale of vector geographic data, including: Based on the BERT model, semantic classification is performed on the text sequences of attribute fields in vector geographic data to obtain layer semantic labels; Kernel density estimation is used to adaptively divide vector geographic data into hotspot areas, and the divided hotspot areas are rasterized into a set of rasterized images; Based on the ResNeXt visual feature extraction network, the visual features of the layer semantic labels and the rasterized image set are fused to perform scale recognition, and the recognition results of each sub-image are integrated using a weighted voting mechanism to obtain the scale recognition result of the vector geographic data; wherein, the sub-image is an image in the rasterized image set.

[0011] In one possible design, semantic classification of the text sequences of attribute fields in vector geographic data is performed based on the BERT model to obtain layer semantic labels, including: The attribute fields of the vector geographic data layer are parsed, the text information is extracted and segmented into words to form a word sequence. The word sequence is then input into the embedding layer of the BERT model. Through the summation of word embedding, segment embedding and location embedding, a high-dimensional vector representation of the text is obtained. The high-dimensional vector representation of the text is input into a multi-layer Transformer encoder, and contextual information is extracted through a self-attention mechanism to obtain the context representation sequence vector of all words; the context representation sequence vector of all words is then used as the basis for this process. The labeled vector serves as the global semantics. The global semantics are mapped to the semantic category space through a fully connected layer. An activation function is then added to calculate the probability of each text category label. The text classification label probability is taken as the semantic prediction result of the attribute field. For a vector geographic data layer containing multiple attribute fields, after obtaining the semantic prediction probability of each field, the semantic information of all fields is integrated by weighted averaging to obtain the comprehensive probability of the layer semantic classification. The semantic category with the highest comprehensive probability is selected as the layer semantic label of the vector geographic data.

[0012] In one possible design, the high-dimensional vector representation of the text is calculated using the following formula; In the formula, A high-dimensional vector representation of the text. E token For word embedding features, E segment For segment embedding features, E position Location embedding features; The high-dimensional vector representation of the text is input into a multi-layer Transformer encoder, and contextual information is extracted through a self-attention mechanism to obtain a sequence vector of contextual representations for all words; the sequence vector of contextual representations for all words is then used to extract contextual information. CLSThe labeled vectors serve as global semantics. A fully connected layer maps the global semantics to the semantic category space. An activation function is then added. The methods for calculating the probability of each text category label include: A multi-layer Transformer encoder is used to extract contextual information, wherein the multi-layer Transformer encoder calculates the inter-word relation function value using a self-attention mechanism: In the formula, , and It is the learned parameter matrix. , and These are query, key, and value matrices, respectively. It is a scaling factor. It is a normalized weight. head For attention output; The attention outputs are concatenated through a linear transformation layer to output a sequence vector of context representations for all words. ; Pick middle tagged vector This represents global semantics. A fully connected layer maps the global semantics to a semantic category space, and an activation function is added to... The classifier calculates the probability of each text category label. Output : In the formula, It is the output transformation matrix, used to further transform the concatenated multi-head attention output to the required output dimension; It is the bias vector.

[0013] In one possible design, kernel density estimation is used to adaptively segment vector geographic data into hotspot regions, and the segmented hotspot regions are rasterized into a set of rasterized images, including: Determine the two-dimensional kernel function and bandwidth; The spatial extent of the vector geographic data is uniformly divided into grids; for each grid, the kernel density estimate of all data points within the grid is calculated using the two-dimensional kernel function to generate a continuous density field. Based on the continuous density field, the information content of the grid is calculated; Based on the information content of each grid, grids with information content higher than a preset threshold are selected as hotspot areas. The vector geographic data within the hotspot areas are then rasterized to generate a set of rasterized images.

[0014] In one possible design, the Gaussian kernel function is used as the two-dimensional kernel function, and its calculation formula is as follows: In the formula, K ( u , v ) indicates with Gaussian kernel function as a variable This represents the coordinate offset of a data point relative to the kernel center. Represents an exponential function; The bandwidth is determined by the following formula: In the formula, The total number of data points. The standard deviation of the data coordinates. This is an adjustment factor used to balance the amount of data with the degree of smoothing. These are empirical parameters; A continuous density field is generated using the following formula: In the formula, Indicates in Density estimate at [location] Indicates the total number of grid data points. This represents the square of the bandwidth. Represents the Gaussian kernel function; x and y Represents the density estimation point. x i and y i This represents the i-th point within the grid. Based on the continuous density field, the information content of the grid is calculated using the following formula: In the formula, This represents the total number of coordinate points within the grid. i It is the index of the coordinate point within the grid.

[0015] In one possible design, based on the ResNeXt visual feature extraction network, scale bar recognition is performed by fusing the semantic labels of the layers with the visual features of the rasterized image set, including: The rasterized image set is input into a pre-trained deep convolutional neural network to extract visual feature vectors; The semantic labels of the layers are converted into semantic feature vectors through category embedding. The visual feature vectors, semantic feature vectors and geographic feature vectors are then fused to obtain the final fused feature vector. Based on the final fused feature vector, it is mapped to a category score through a fully connected layer and then... The function generates a probability distribution, selects the category with the highest probability as the sub-map scale regression result, and after processing the rasterized image set, obtains the sub-map scale discrimination set and the confidence set of each sub-map as the corresponding scale.

[0016] In one possible design, a weighted voting mechanism is used to integrate the identification results of each sub-map to obtain the scale identification results of the vector geographic data, including: Calculate the weighted total probability of voting at different scales using the following formula: In the formula, This represents the total number of subgraphs. This indicates the total number of scale categories. The indicator function represents the total probability of weighted voting at different scales. Indicates when subgraph Predicted scale category For the first i Scale label The value is 1 if the condition is met, and 0 otherwise. Subgraph The confidence level of the discrimination; The scale label with the highest weighted voting probability is taken as the final scale identification result of the vector geographic data.

[0017] Secondly, this application provides a vector geographic data scale identification device, the device comprising: The semantic classification module is configured to perform semantic classification on the text sequences of attribute fields of vector geographic data based on the BERT model to obtain layer semantic labels. The hotspot region segmentation module is configured to adaptively segment vector geographic data into hotspot regions using kernel density estimation, and then rasterize the segmented hotspot regions into a set of rasterized images. The scale recognition module is configured to perform scale recognition based on the ResNeXt visual feature extraction network, which integrates the semantic labels of the layers with the visual features of the rasterized image set, and uses a weighted voting mechanism to integrate the recognition results of each sub-map to obtain the scale recognition result of the vector geographic data; wherein, the sub-map is an image in the rasterized image set.

[0018] Thirdly, embodiments of this application provide an electronic device, including: at least one processor and a memory; the memory stores computer execution instructions; the at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to perform the vector geographic data scale identification method as described in the first aspect and various possible designs of the first aspect.

[0019] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions. When a processor executes the computer-executable instructions, it implements the vector geographic data scale identification method described in the first aspect and various possible designs of the first aspect.

[0020] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the vector geographic data scale identification method as described in the first aspect and various possible designs of the first aspect.

[0021] The vector geographic data scale identification method, apparatus, equipment, and storage medium provided in this application have at least the following beneficial effects: Addressing key challenges in intelligent scale identification of vector geographic data, and focusing on the insufficient accuracy of scale discrimination due to differences in geographic regional distribution and the semantic diversity of layers, this application introduces the BERT model to achieve semantic classification of vector geographic data. Based on ResNeXt's visual feature extraction and geographic region recognition, a precise mapping relationship is established from vector geographic data to scale categories. Experimental results show that the proposed method exhibits excellent accuracy and robustness across cross-regional and cross-semantic type data. Ablation experiments verify that the semantic classification module significantly improves the model's recognition ability, and the hybrid discrimination strategy based on KDE and voting mechanisms plays a positive role in further improving the algorithm's accuracy. Overall, this application provides an innovative and usable solution for intelligent scale identification of vector geographic data, offering theoretical support and a new technical path for the security assessment of vector geographic data. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0023] Figure 1 A comparison map of plain and mountain water system data provided in the embodiments of this application; Figure 2 A scale comparison map of highways and railways in a typical plain area provided for embodiments of this application; Figure 3A flowchart illustrating a vector geographic data scale identification method provided in this application embodiment; Figure 4 A flowchart illustrating a vector geographic data scale identification method provided in this application embodiment; Figure 5 A semantic classification flowchart provided for embodiments of this application; Figure 6 A flowchart illustrating the hotspot region division provided in this application embodiment; Figure 7 A flowchart for scale identification provided in the embodiments of this application; Figure 8 The following are scale detection result maps of different regions provided in the embodiments of this application; wherein, (a) AL.A algorithm; (b) AL.B algorithm; (c) the method proposed in this application; Figure 9 The scale detection results under different semantic features provided in the embodiments of this application; wherein, (a) AL.A algorithm; (b) AL.B algorithm; (c) the method proposed in this application; Figure 10 This is a schematic diagram of a portion of the map after rotational disturbance, provided in an embodiment of this application; wherein, (a) the original map; and (b) rotated 90°. Figure 11 The figure shows the experimental results of rotational disturbance provided in the embodiments of this application; Figure 12 The image shows the experimental results of attribute deletion perturbation provided in the embodiments of this application; Figure 13 A structural diagram of the vector geographic data scale recognition device provided in the embodiments of this application.

[0024] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0025] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0026] The collection, storage, use, processing, transmission, provision, and disclosure of financial data or user data involved in the technical solution of this application all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0027] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0028] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0029] The core of vector geographic data scale identification lies in extracting representative information from complex features and establishing a stable and accurate mapping relationship between them and the scale. However, at the same scale, vector geographic data from different geographic regions and with different semantic elements often exhibit significant feature differences. For example, Figure 1 This presentation showcases a comparison of vector geographic data of water systems at different scales in typical plain and mountainous areas. Figure 2 It presents a comparison of highway and railway data at different scales in a typical plain area.

[0030] Depend on Figure 1 It is evident that as the scale decreases, the simplification of the data continuously increases, and the spatial density of features gradually decreases. However, at the same scale, the density of water system data features in typical mountainous areas is far sparser than that in typical plain areas. This difference caused by geographical regional factors presents a challenge to the identification of vector geographic data scales.

[0031] Depend on Figure 2 It is evident that, at the same scale, the density of vector geographic data elements for highways in typical plain areas is significantly higher than that for railways. Railway data elements exhibit sparse characteristics across different scales, with relatively small differences between scales. Overcoming the scale-related differences caused by semantic factors is the second challenge in achieving accurate scale identification.

[0032] Based on the above challenges, this application provides a method for identifying the scale of vector geographic data. First, semantic feature invariants are designed using the attribute fields of vector geographic data to achieve semantic classification of elements, serving as an upstream task for scale identification and thus eliminating interference caused by semantic differences. Second, vector rasterization is used to represent the geographic regional characteristics of the data, serving as a midstream task for scale identification and thus eliminating the impact of regional differences on scale identification. Finally, based on a visual deep learning network, accurate scale identification is achieved. Figure 3 The overall process of the proposed method is shown. This method consists of three steps: semantic classification, hotspot region segmentation, and scale identification. First, the input vector data undergoes attribute parsing, leading to step ①, semantic classification. The BERT model is used to complete text input, feature extraction, and semantic inference, ultimately outputting semantic results such as water systems. Next, step ②, Hotspot Region Segmentation, involves grid division and KDE-based density analysis to identify hotspot regions. These regions are then rasterized, and geographic features are extracted to generate rasterized images. Finally, step ③, Scale Discrimination, inputs the rasterized images into the ResNeXt model for feature extraction and scale inference. Label integration and a voting mechanism are then used to output the final scale result.

[0033] Specifically, such as Figure 4 As shown, this vector geographic data scale identification method can be implemented through the following steps S10-S30.

[0034] S10: Based on the BERT model, perform semantic classification on the text sequences of attribute fields of vector geographic data to obtain layer semantic labels.

[0035] At the same map scale, different semantic vector layers often exhibit different characteristics, making it difficult for scale recognition models based on single features to be applicable to cross-semantic data environments. Therefore, semantic classification of vector layers is a crucial upstream task for achieving high-accuracy scale recognition, effectively eliminating the interference of semantic factors. However, due to the non-alignment characteristics of textual attributes in vector geographic data across different production environments and application scenarios, simple rule matching or field mapping struggles to achieve semantic consistency matching and accurate understanding, resulting in limited recognition performance in complex semantic scenarios.

[0036] BERT (Bidirectional Encoder Representations from Transformers) is a bidirectional deep semantic encoding model that effectively captures the semantic relationships of complex texts through self-supervised pre-training and bidirectional contextual modeling, and demonstrates outstanding performance in various text classification tasks.

[0037] BERT consists of multiple Transformer encoders, each layer containing a self-attention mechanism and a feedforward network. Through multi-head attention, it enhances the model's ability to capture different semantic patterns. For input vector geographic data text sequences... BERT encodes it as a high-dimensional semantic representation vector. Then, semantic classification is achieved through a fully connected layer and a softmax function.

[0038] Through the modeling process described above, BERT can extract deep semantic features from the attribute fields of vector geographic data, avoiding the semantic ambiguity caused by relying solely on keyword matching. It also establishes an accurate mapping from attribute field text sequences to layer semantics, achieving semantic classification of the data. Using the semantic classification results as one of the inputs to the deep learning network can eliminate the influence of semantic factors on the accuracy of scale determination, providing a foundation for building an accurate scale recognition model with cross-semantic and cross-regional difference perception capabilities.

[0039] S20: Adaptive hotspot region segmentation is performed on vector geographic data using kernel density estimation, and the segmented hotspot regions are rasterized into a set of rasterized images.

[0040] Vector geographic data often exhibits spatial non-uniformity and complex local topology. Directly extracting global features can easily weaken local salient features or enhance noise interference, thus affecting the accuracy of scale determination. Therefore, it is necessary to extract areas with significant feature distribution from the data.

[0041] Kernel density estimation (KDE) is a nonparametric probability density estimation method that generates a continuous density surface by estimating the probability density function of sample points, thereby intuitively reflecting the spatial distribution pattern and central tendency of data points (Dai et al., 2020). Its basic idea is to superimpose a kernel function at each sample point to smoothly estimate the overall density distribution, and its mathematical expression is shown in Equation (1).

[0042] (1) in, It is at point The estimated density at that location, It is the total number of data points. It's bandwidth, the width of the control core, which affects the smoothness. It's a kernel function. It refers to each data point in the dataset.

[0043] In scale discrimination tasks, KDE is used to identify high-density distribution areas in vector geographic data, enhancing local geometric and spatial distribution features. Adaptive segmentation of hotspot areas provides more representative input for subsequent extraction of geographic and visual features, thereby improving the algorithm's recognition accuracy.

[0044] S30: Based on the ResNeXt visual feature extraction network, the visual features of layer semantic labels and rasterized image sets are fused to perform scale recognition, and a weighted voting mechanism is used to integrate the recognition results of each sub-map to obtain the scale recognition results of vector geographic data; where the sub-map is an image in the rasterized image set.

[0045] Geographic data at different scales exhibit significant differences in line thickness, feature morphology, and element density. After rasterization, the visual distribution of vector geographic data effectively reflects these scale differences; therefore, visual features become an important basis for establishing stable mapping relationships across scales. Furthermore, influenced by factors such as population density, geographical environment, and climate conditions, vector geographic data from different regions at the same scale often exhibit different spatial distribution characteristics; these differences can also be characterized through visual features.

[0046] However, methods relying on manual judgment or based on manually designed features struggle to effectively capture the visual features of complex layers and have limited automation, making them unsuitable for large-scale applications. Deep convolutional neural networks (CNNs) offer automated feature extraction, boasting significantly stronger feature extraction capabilities compared to classic computer vision methods. ResNeXt is an improved residual network (ResNet) employing a "split-transform-merge" structure, combining grouped convolutions with residual connections to enhance the model's diversity and robustness in feature representation. Due to its high performance, ResNeXt has been widely applied to various computer vision tasks, particularly image classification, object detection, and medical image analysis.

[0047] Related experiments show that ResNeXt, due to the structured sparse connection characteristics of grouped convolution, decomposes the input space into complementary subspaces through a multi-branch feature fusion mechanism, breaking through the limitations of traditional neural networks' single-dimensional stacking, providing multi-granularity decoupling representation capabilities for complex image recognition, and effectively extracting the geometric nonlinear features of two-dimensional images.

[0048] Step S10 is a BERT-based semantic classification step. Semantic classification of vector geographic data layers is an upstream task for scale identification, aiming to eliminate the interference of semantic differences on scale identification. In some embodiments, such as... Figure 5 As shown, step S10 can be implemented through the following steps S101-S103 to achieve semantic classification.

[0049] S101: Analyze the attribute fields of the vector geographic data layer, extract text information and perform word segmentation to form a word sequence. Input the word sequence into the embedding layer of the BERT model. Through the sum of word embedding, segment embedding and location embedding, a high-dimensional vector representation of the text is obtained.

[0050] Specifically, text information is extracted from the attribute fields of vector geographic data layers, forming a text field sequence. For each attribute field, word segmentation is performed to convert it into a sequence of tokens. The BERT embedding layer then extracts vectors. First, token embedding maps each token to a fixed-dimensional vector space. Next, segment embedding distinguishes different input segments. Finally, word order information is introduced, and position embedding emphasizes positional relationships. Combining these three embedding methods yields a high-dimensional vector representation of the text. , as input to the Transformer: (2) In the formula, E token For word embedding features, E segment For segment embedding features, E position This is a location embedding feature.

[0051] S102: Input the high-dimensional vector representation of the text into a multi-layer Transformer encoder, extract contextual information through a self-attention mechanism, and obtain the contextual representation sequence vector of all words; use the contextual representation sequence vector of all words [ CLS The vectors marked with [] are used as global semantics. The global semantics are mapped to the semantic category space through a fully connected layer. An activation function is added to calculate the probability of each text category label.

[0052] Specifically, as shown in equations (3) and (4), a multi-layer Transformer encoder is used to extract contextual information. The Transformer calculates the inter-word relationship function value using a self-attention mechanism.

[0053] (3) (4) In the formula, , and It is the learned parameter matrix. , and These are query, key, and value matrices, respectively. It is a scaling factor used to ensure gradient stability. It is a normalized weight.

[0054] The attention output is fused through a linear transformation layer. Concatenate the sequences and output the context representation sequence vectors of all words. .

[0055] (5) In the formula, This indicates a splicing operation. head 1. head 2 and head n These are the first, second, and nth attention outputs, respectively.

[0056] Pick middle tagged vector This represents global semantics. A fully connected layer maps the global semantics to a semantic category space, and an activation function is added to... The classifier calculates the probability of each text category label. Output : (6) In the formula, It is the output transformation matrix, used to further transform the concatenated multi-head attention output to the required output dimension. It is the bias vector, which the model adjusts the output to make the network more adaptable to different data distributions.

[0057] S103: Take the text classification label probability as the semantic prediction result of the attribute field. For a vector geographic data layer containing multiple attribute fields, after obtaining the semantic prediction probability of each field, select the semantic category with the highest frequency as the layer semantic label of the vector geographic data.

[0058] Specifically, the category with the highest probability is taken as the semantic prediction result for the field, and the category is determined. for: (7) Set the total number of text messages to Once the semantic reasoning of all text information fields has been completed, the sequence is obtained. ,in, , and Let represent the predicted categories of the 1st, 2nd, and tth text messages, respectively. Finally, the semantic category with the highest frequency of occurrence is used as the layer semantic label for the vector geographic data. This semantic tag As one of the inputs for scale recognition.

[0059] Vector geographic data exhibits spatial non-uniformity; therefore, segmenting it into high-information quantum maps is an effective means to improve the accuracy of scale identification. Simultaneously, sub-map segmentation increases the sample size for scale discrimination, which, based on a voting mechanism, can improve the error tolerance of scale identification. Therefore, in some embodiments, a KDE-based adaptive segmentation method for hotspot areas of vector geographic data is proposed, specifically as follows: Figure 6 As shown, step S20 can be specifically implemented through the following steps S201-S204 to divide the hotspot area.

[0060] S201: Determine the two-dimensional kernel function and bandwidth.

[0061] Specifically, KDE calculation requires defining a two-dimensional kernel function and bandwidth. Because the Gaussian kernel function is sensitive to and smooths local density changes, it is suitable for the continuous distribution characteristics of geographic features. Therefore, the Gaussian kernel function is used as the two-dimensional kernel function, and its calculation formula is as follows: (8) in, This represents the coordinate offset of a data point relative to the kernel center. This represents an exponential function.

[0062] Bandwidth determines the smoothness of the kernel function. Insufficient bandwidth leads to noise, while excessive bandwidth results in over-smoothing and loss of local data features. The algorithm selects the optimal bandwidth through cross-validation, determining the optimal bandwidth by minimizing the estimation error on the validation set. The empirical formula for bandwidth is as follows: (9) In the formula, The total number of data points. The standard deviation of the data coordinates. This is an adjustment factor used to balance the amount of data with the degree of smoothing. These are empirical parameters.

[0063] S202: Divide the spatial extent of vector geographic data evenly into... k A grid is formed; for each grid, a two-dimensional kernel function is used to calculate the kernel density estimate of all data points in the grid, generating a continuous density field.

[0064] Specifically, grid partitioning and kernel density estimation. The vector geographic data space is uniformly divided into grids, and the KDE is calculated for each data point in the grid, as shown in Equation (10), to generate a continuous density field.

[0065] (10) In the formula, Indicates in Density estimate at [location] Indicates the total number of grid data points. This represents the square of the bandwidth. Represents the Gaussian kernel function; x and y Represents the density estimation point. x i and y i This represents the i-th point within the grid.

[0066] S203: Calculate the information content of the grid based on the continuous density field.

[0067] Specifically, for each grid cell, the information content of the grid cell is calculated based on the KDE results (i.e., the generated continuous density field): (11) In the formula, This represents the total number of coordinate points within the grid. i It is the index of the coordinate points within the grid. Grid information content. The larger the value, the more feature information is contained within the grid.

[0068] S204: Based on the information content of each grid, select grids with information content higher than a preset threshold as hotspot areas, and rasterize the vector geographic data within the hotspot areas to generate a set of rasterized images.

[0069] Specifically, setting a threshold The data is sorted according to the amount of information in each grid, and the data of the first part with more than the threshold is segmented. The vector geographic data is rasterized to generate image patches and normalized to meet the input requirements of the deep learning model, and the output is a set of rasterized images.

[0070] In some embodiments, step S30 uses an adaptive set of segmented sub-maps (i.e., a set of rasterized images) as two-dimensional data input, semantic classification results as input features, performs scale recognition of the considered region based on ResNeXt, and employs a voting mechanism to reduce the impact of false detections of individual sub-maps on the overall accuracy. Figure 7 As shown, step S30 can be specifically implemented through the following steps S301-S304 to identify the scale.

[0071] S301: Input the rasterized image set into a pre-trained deep convolutional neural network to extract visual feature vectors.

[0072] Specifically, the rasterized image set The input is a pre-trained deep convolutional neural network, which uses the ResNeXt-50 module for feature encoding. A multi-path convolution (Cardinal Group Convolution) structure captures visual features, including low-level local features such as edges and textures, as well as more complex and abstract features. The feature map size is reduced in the residual network module (ResidualBlock), effectively extracting the multi-level spatial distribution structure of the image to obtain the visual feature vector of the data. .

[0073] S302: Convert the layer semantic labels into semantic feature vectors through category embedding, and fuse the visual feature vectors, semantic feature vectors and geographic feature vectors to obtain the final fused feature vector.

[0074] Specifically, the semantic classification labels are converted into embedding vectors using the category embedding method. : (12) In the formula, Indicates the semantic label of the layer.

[0075] Concatenate multi-dimensional features: (13) In the formula, , and These are the weights of the semantic feature vector, visual feature vector, and geographic feature vector, respectively. and These are the extracted visual feature vector and geographic feature vector, respectively. It is a bias term. This is the final fused feature vector.

[0076] S303: Based on the final fused feature vector, it is mapped to a category score through a fully connected layer and then... The function generates a probability distribution, selects the category with the highest probability as the sub-map scale regression result, and after processing the rasterized image set, obtains the sub-map scale discrimination set and the confidence set of each sub-map as the corresponding scale.

[0077] Specifically, after unifying the dimensions of the fused feature vectors, spatial resolution is expanded using learnable convolutional kernels. Output mapped features: (14) In the formula, This indicates the transpose convolution operation. For hybrid activation functions, For spatial saliency maps, For the final fused feature vector, Conv 1×1 This is a unit convolution operation.

[0078] S304: Calculate the weighted total probability of voting for different scales: Use the scale label with the highest weighted total probability of voting as the final scale identification result of the vector geographic data.

[0079] Based on the final fused feature vector After being mapped to category scores through a fully connected layer, The function generates a probability distribution such that the sum of the probabilities for all scale categories is 1. (15) The category with the highest probability is selected as the scale regression result for this sub-image. After processing the rasterized image set, a scale discrimination set is obtained. and the set of confidence levels for each submap to be classified as that scale. ,in It represents the number of raster images.

[0080] To improve the algorithm's fault tolerance and thus enhance overall detection accuracy, a weighted voting mechanism is used to integrate subgraph prediction results. Assuming the number of scale labels is... That is, a unique set of tags As shown in equation (16).

[0081] (16) In the formula, This represents the total number of subgraphs. This indicates the total number of scale categories. The indicator function represents the total probability of weighted voting at different scales. Indicates when subgraph Predicted scale category For the first i Scale label The value is 1 if the condition is met, and 0 otherwise. Subgraph The discrimination confidence level; obtaining the regression probability distribution of the layer at different scales. Take the maximum probability. Corresponding scale This represents the final scale identification result for the data. The voting mechanism-based approach ensures that the prediction results are influenced not only by the frequency of occurrence but also by the confidence level of the subgraph prediction, resulting in greater overall accuracy and robustness.

[0082] The feasibility and progressiveness of the method proposed in this application will be explained in detail below through experiments and analysis.

[0083] Considering national geographic data confidentiality requirements, this embodiment selects publicly available vector geographic data as the experimental dataset. The dataset includes three scales: 1:250,000, 1:1,000,000, and 1:4,000,000, as shown in Table 1. To verify the algorithm's scale recognition capability in cross-regional scenarios, the experimental data includes four provinces: A, B, C, and D, covering typical geographic types such as plains, mountains, and hills. Simultaneously, to verify the performance of the scale recognition method in cross-semantic scenarios, the data content includes five semantic elements: water system lines, water system surfaces, highway lines, railway lines, and residential ground. 100 sub-regions were extracted from each province, with 500 maps at each scale, totaling 1500 experimental maps.

[0084] The pre-training and validation sets were divided in a 7:3 ratio, and the test set did not overlap with the training or validation sets. The experimental hardware used Windows 11 operating system, a 13th Gen Intel® Core™ i7-13700 2.10 GHz CPU, Python as the programming language, 64GB of memory, and an NVIDIA GeForce RTX 3090 GPU, with GPU acceleration implemented during the experiment.

[0085] Table 1 Experimental Dataset Information

[0086] Scale identification is a mapping process from vector geographic data to discrete scale categories, and the correct mapping scale fundamentally reflects the method's identification capability. However, a single accuracy metric can be misleading in class imbalance or fine-grained scale identification tasks. Therefore, this embodiment not only uses accuracy as the core quantitative evaluation metric, but also introduces multi-dimensional metrics such as precision, recall, F1 score, and Kappa coefficient to comprehensively measure the algorithm's true performance on class imbalance or fine-grained scales.

[0087] (1) Accuracy represents the proportion of map sheets with correctly identified scales out of the total number of map sheets. The calculation formula is as follows: (17) In the formula, This is the total number of map sheets. The number of map sheets correctly identified by the scale is indicated. The closer the accuracy value is to 100%, the lower the false detection rate of the algorithm in scale identification, and the higher the reliability of the classification results.

[0088] (2) Precision represents the proportion of samples predicted as positive that are actually positive. The formula is as follows: (18) In the formula, To correctly predict the number of samples that are classified as positive, This represents the number of samples that were incorrectly predicted as positive.

[0089] (3) Recall represents the proportion of samples that are actually positive that are correctly predicted. The formula is as follows: (19) In the formula, This represents the number of samples that were incorrectly predicted as negative.

[0090] (4) The F1 score is the harmonic mean of precision and recall, balancing the weights of the two. The calculation formula is as follows: (20) (5) The Kappa coefficient is used to measure the consistency between the classification result and the random classification. The calculation formula is as follows: (twenty one) In the formula, This refers to the observational consistency rate (the proportion of the actual classification that matches the prediction). This refers to the expected consistency rate (the proportion of consistency under random classification).

[0091] This embodiment designs multiple types of experiments to verify the performance of the proposed scale recognition method in terms of accuracy and robustness. The experiments use a node density-based algorithm (referring to "Liu Wanzeng, Wang Xinpeng, Zhao Tingting, et al. Scale Evaluation Method for Fragmented Topographic Vector Geographic Data [J]. Acta Geodaetica et Cartographica Sinica, 2024, 53(06):1013-1024") and a single visual feature-based algorithm (referring to "Ren Jiaxin, Liu Wanzeng, Chen Jun, et al. Knowledge-Guided Intelligent Scale Recognition of Fragmented Raster Topographic Maps [J]. Acta Geodaetica et Cartographica Sinica, 2024, 53(01):146-157") as comparison algorithms, abbreviated as AL.A and AL.B. AL.B uses the same multi-scale training strategy as the original paper, combining the pyramid dataset with subgraph integration. The single-scale training results are inferior to the multi-scale training results, therefore this will not be discussed further.

[0092] Accuracy Experiment: Accuracy is a crucial indicator for evaluating the precision and reliability of the mapping relationship between vector geographic data and scale. A highly accurate scale identification algorithm must maintain stable identification capabilities across geographical and semantic environments. Therefore, the algorithm's scale identification accuracy is evaluated at the provincial level, as shown in Table 2. Figure 8 As shown in Table 3, the accuracy of scale recognition is evaluated by semantic type. Figure 9 As shown.

[0093] Table 2. Results of Data Scale Identification in Different Regions

[0094] From Table 2 and Figure 8 It is evident that the AL.A algorithm has limited performance in cross-regional scenarios. Its accuracy in scale recognition exceeds 87% in provinces A and B; however, in province C, due to its mountainous and hilly terrain, fragmented road network, and scattered streams, the accuracy drops to 79.62% for 1:250,000 scale data. In northwestern province D, with its sparse roads and rivers, the accuracy for all three scale types is below 85%, reaching a minimum of 78.33%, which is insufficient for practical applications. The AL.B algorithm exhibits some adaptability to geographical differences, maintaining stable scale recognition accuracy across the four provinces, with a slight decrease in region D, but generally remaining above 88% and below 95%.

[0095] In contrast, the proposed method uses ResNeXt, a network with region recognition capabilities, as its main network, and combines it with a hotspot region segmentation strategy to effectively enhance the ability to extract features from the detected regions. It demonstrates superior cross-regional recognition performance on the dataset, achieving near 100% scale recognition accuracy in provinces A, B, and C, and maintaining over 95% accuracy in province D. In summary, the proposed method significantly outperforms the comparative algorithms in cross-regional dataset recognition, consistent with theoretical and experimental results.

[0096] Table 3. Recognition results of different semantic data scales

[0097] From Table 3 and Figure 9 As can be seen, the scale recognition accuracy of AL.A is mainly concentrated between 60% and 90%. However, the scale recognition accuracy for railway data shows a significant decrease, dropping to 60.5%. This is because the spatial density of railway data is significantly lower than that of highway data. If the algorithm uses fixed point density for discrimination, it is difficult to achieve effective data differentiation and scale mapping. AL.B's overall recognition accuracy is around 90%. Similarly, scale recognition accuracy fluctuates for railway and residential data, dropping to 72.25% and 79% respectively, demonstrating weak usability and instability in cross-semantic data recognition, making it difficult to adapt to real-world scenarios.

[0098] In contrast, the proposed method incorporates BERT semantic classification as a prerequisite task, effectively enhancing the model's ability to understand and discriminate across semantic features. On vector geographic data of five semantic elements, the algorithm demonstrates stable and high-precision scale recognition performance: the recognition accuracy is close to 100% for water system lines, water surface areas, and highway lines, and remains above 97% for railway lines and residential ground data. In summary, the proposed method exhibits high reliability and stability in cross-semantic scale recognition tasks, consistent with theoretical and experimental results.

[0099] This embodiment also conducted multi-dimensional index experiments. Tables 4 and 5 show the evaluation results of the proposed method on multi-dimensional indexes, comprehensively reflecting the overall performance of the proposed method in the scale recognition task. The experimental data show that the proposed method exhibits excellent performance in terms of accuracy, precision, recall, F1 score, and Kappa coefficient.

[0100] Table 4 Results of experimental indicators in different regions

[0101] Table 5 Results of Experimental Indicators for Different Semantic Data

[0102] In cross-regional experiments (Table 4), the method of this application achieved an accuracy of over 95.80%, with a maximum of 100.00%, in test data from four provinces: A, B, C, and D. The precision remained stable between 98.53% and 99.46%, indicating extremely high reliability in predicting positive samples. The recall ranged from 89.78% to 99.75%, demonstrating the algorithm's ability to effectively identify most samples at true scales. The F1 score, the harmonic mean of precision and recall, remained at a high level of 93.93% to 99.39%, indicating a good balance between precision and recall. The Kappa coefficient ranged from 0.857 to 0.897, showing significant consistency between the algorithm's classification results and random classification, indicating high reliability of the classification results. Of particular note is that in region D, despite the complex geographical features, the proposed method still maintains an accuracy of 95.80% and an F1 score of 94.64%, which fully demonstrates the robustness of the proposed method in complex geographical environments.

[0103] In cross-semantic experiments (Table 5), the proposed method performed excellently on five semantic elements: water system lines, water system surfaces, highway lines, railway lines, and residential ground. Accuracy ranged from 97.00% to 99.50%, with water system surfaces reaching 99.50% at a scale of 1:4,000,000. Precision remained stable between 98.22% and 99.55%, indicating extremely reliable predictions for positive samples. Recall ranged from 88.25% to 98.34%, demonstrating the algorithm's ability to effectively identify the true scale of various semantic elements. F1 scores remained high, ranging from 93.32% to 98.37%, with railway lines reaching 98.37% at a scale of 1:4,000,000. Kappa coefficients ranged from 0.842 to 0.881, indicating high consistency in classification results across different semantic data. In particular, for railway line data, despite its low spatial density, the method of this application still maintains an accuracy of 97.00% and an F1 score of 95.39%, demonstrating the excellent performance of the method of this application in processing low-density vector data.

[0104] In summary, the proposed method demonstrates excellent performance across multiple metrics, including accuracy, precision, recall, F1 score, and Kappa coefficient. It maintains high accuracy across various regional and semantic datasets while achieving a good balance between precision and recall. This fully demonstrates that the proposed method not only boasts high accuracy in scale identification tasks but also exhibits significant advantages in the comprehensiveness of its classification performance, thus meeting the high-precision requirements for scale identification in practical geographic information systems.

[0105] In practical applications, vector geographic data may undergo operations such as rotation, translation, and field editing. These processes can alter the spatial shape or attribute information of the data, but they do not change its scale definition. Therefore, the ability to maintain scale recognition accuracy under disturbances in data geometric and attribute features—that is, to possess strong robustness—is an important indicator of the usability of a scale recognition algorithm. To verify the robustness of the algorithm, this embodiment uses the dataset of Province A as the test sample and designs two types of disturbance experiments: rotation and attribute deletion.

[0106] Rotational perturbation experiment.

[0107] Vector geographic data may be rotated due to adjustments in coordinate systems or map projections, altering scale identification features and affecting accuracy. To verify the robustness of the scale identification algorithm against rotational perturbations, the data was rotated from 30° to 330° in 30° increments. The rotated portion of the data is shown below. Figure 10 As shown in Table 6, to reduce the impact of regional distribution and semantic differences on the results, this experiment selected three scales of data—riverline and river surface—as input, totaling 600 test objects. The scale recognition results after rotational perturbation are shown in Table 6. Figure 11 As shown.

[0108] Table 6. Identification results under rotational disturbance

[0109] From Table 6 and Figure 11 It can be seen that the recognition accuracies of AL.A, AL.B, and the proposed method are concentrated at 92.16%, 95.50%, and 100%, respectively. The scale discrimination accuracy of the algorithms remains stable under rotational perturbation, with small overall fluctuations. Although the rotation operation changes the spatial position and orientation distribution of vector features, the feature point density, layer visual structure, and semantic features remain unchanged, thus not significantly interfering with feature extraction and scale discrimination. Therefore, the proposed method exhibits strong robustness and scale recognition stability under rotational perturbation.

[0110] Attribute deletion perturbation experiment.

[0111] The deletion of attribute fields may alter the semantic features of vector geographic data, thus affecting the accuracy of scale identification. To comprehensively evaluate the robustness of the algorithm in the case of missing attributes, this experiment designed more detailed perturbation scenarios, including not only the deletion of 20%-80% of random attribute fields, but also specifically testing the impact of deleting key semantic fields and a "completely attribute-free table" scenario. Specifically, the fields deleted in the experiment included semantically key fields such as NAME, TYPE, and CLASS, as well as auxiliary fields such as ID (identifier), all of which were removed from the attribute table randomly. To ensure the objectivity of the experiment, all field deletion operations were based on a random sampling strategy, without selective deletion of specific fields. The experimental results are shown in Table 7 and... Figure 12 As shown.

[0112] Table 7 Results of attribute removal perturbation

[0113] From Table 7 and Figure 12 As can be seen, with the increase of attribute deletion perturbation intensity, the scale recognition results of AL.A and AL.B remain at a low level without significant change. In contrast, although the recognition accuracy of the proposed method decreases slightly, it still maintains a recognition accuracy of 98.87% when the attribute field deletion intensity reaches 80%, which is significantly higher than the 89.47% and 91.20% of the comparative algorithms. It is particularly noteworthy that in the extreme case of "completely no attribute table", the proposed method can still maintain a recognition accuracy of 95.45%, which fully demonstrates that the proposed method can still perform effective scale recognition based on visual features even in the absence of any semantic attribute information.

[0114] This is mainly due to the fact that the proposed method can achieve high-precision semantic understanding of incomplete attribute information based on sub-word-level semantic features and their contextual associations. At the same time, by fusing visual features, the method can still maintain high recognition ability even when attributes are missing. Especially in scenarios with no attribute table at all, the recognition accuracy of the proposed method remains above 95%, indicating that the visual feature extraction capability of the proposed method is excellent enough to independently complete the scale recognition task without relying on attribute information.

[0115] This application also provides a vector geographic data scale identification device for implementing the methods described in any of the above embodiments, such as... Figure 13 As shown, the vector geographic data scale identification device includes: The semantic classification module 1301 is configured to perform semantic classification on the text sequence of attribute fields of vector geographic data based on the BERT model to obtain layer semantic labels. The hotspot region segmentation module 1302 is configured to adaptively segment vector geographic data into hotspot regions using kernel density estimation, and rasterize the segmented hotspot regions into a set of rasterized images. The scale recognition module 1303 is configured to perform scale recognition based on the ResNeXt visual feature extraction network, which integrates the semantic labels of the layers and the visual features of the rasterized image set, and uses a weighted voting mechanism to integrate the recognition results of each sub-map to obtain the scale recognition result of the vector geographic data; wherein, the sub-map is an image in the rasterized image set.

[0116] This application provides an electronic device. The electronic device may include a processor and a memory, wherein the processor and the memory can communicate; exemplarily, the processor and the memory communicate via a communication bus.

[0117] The processor executes computer execution instructions stored in memory, causing the processor to perform the scheme in the above embodiments. The processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0118] The communication bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The system bus can be divided into address bus, data bus, control bus, etc. Transceivers are used to enable communication between database access devices and other computers (e.g., clients, read-write libraries, and read-only libraries). Memory may include random access memory (RAM) and may also include non-volatile memory.

[0119] The electronic device provided in this application embodiment can be the terminal device described in the above embodiments.

[0120] This application also provides a computer-readable storage medium storing computer instructions. When the computer instructions are executed on a computer, the computer performs the technical solution of the vector geographic data scale identification method described in the above embodiments.

[0121] This application also provides a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor can read the computer program from the computer-readable storage medium. When the at least one processor executes the computer program, it can implement the technical solution of the vector geographic data scale identification method in the above embodiments.

[0122] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.

[0123] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.

[0124] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.

[0125] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0126] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0127] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.

[0128] Buses can be Industry Standard Architecture (ISA) buses, Peripheral Component Interconnect (PCI) buses, or Extended Industry Standard Architecture (EISA) buses, etc. Buses can be categorized into address buses, data buses, control buses, etc.

[0129] The aforementioned storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0130] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic control unit or main control device.

[0131] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for scale identification of vector geographic data, characterized in that, The method includes: Based on the BERT model, semantic classification is performed on the text sequences of attribute fields in vector geographic data to obtain layer semantic labels; Kernel density estimation is used to adaptively divide vector geographic data into hotspot areas, and the divided hotspot areas are rasterized into a set of rasterized images; Based on the ResNeXt visual feature extraction network, the visual features of the layer semantic labels and the rasterized image set are fused to perform scale recognition, and the recognition results of each sub-image are integrated using a weighted voting mechanism to obtain the scale recognition result of the vector geographic data; wherein, the sub-image is an image in the rasterized image set.

2. The vector geographic data scale identification method according to claim 1, characterized in that, Based on the BERT model, semantic classification is performed on the text sequences of attribute fields in vector geographic data to obtain layer semantic labels, including: The attribute fields of the vector geographic data layer are parsed, the text information is extracted and segmented into words to form a word sequence. The word sequence is then input into the embedding layer of the BERT model. Through the summation of word embedding, segment embedding and location embedding, a high-dimensional vector representation of the text is obtained. The high-dimensional vector representation of the text is input into a multi-layer Transformer encoder, and contextual information is extracted through a self-attention mechanism to obtain the context representation sequence vector of all words; the context representation sequence vector of all words is then used as the basis for this process. The labeled vector serves as the global semantics. The global semantics are mapped to the semantic category space through a fully connected layer. An activation function is then added to calculate the probability of each text category label. The text classification label probability is taken as the semantic prediction result of the attribute field. For a vector geographic data layer containing multiple attribute fields, after obtaining the semantic prediction probability of each field, the semantic information of all fields is integrated by weighted averaging to obtain the comprehensive probability of the layer semantic classification. The semantic category with the highest comprehensive probability is selected as the layer semantic label of the vector geographic data.

3. The vector geographic data scale identification method according to claim 2, characterized in that, The high-dimensional vector representation of the text is calculated using the following formula; In the formula, A high-dimensional vector representation of the text. E token For word embedding features, E segment For segment embedding features, E position Location embedding features; The high-dimensional vector representation of the text is input into a multi-layer Transformer encoder, and contextual information is extracted through a self-attention mechanism to obtain a sequence vector of contextual representations for all words; the sequence vector of contextual representations for all words is then used to extract contextual information. CLS The labeled vectors serve as global semantics. A fully connected layer maps the global semantics to the semantic category space. An activation function is then added. The methods for calculating the probability of each text category label include: A multi-layer Transformer encoder is used to extract contextual information, wherein the multi-layer Transformer encoder calculates the inter-word relation function value using a self-attention mechanism: In the formula, , and It is the learned parameter matrix. , and These are query, key, and value matrices, respectively. It is a scaling factor. It is a normalized weight. head For attention output; The attention outputs are concatenated through a linear transformation layer to output a sequence vector of context representations for all words. ; Pick middle tagged vector This represents global semantics. A fully connected layer maps the global semantics to a semantic category space, and an activation function is added to... The classifier calculates the probability of each text category label. Output : In the formula, It is the output transformation matrix, used to further transform the concatenated multi-head attention output to the required output dimension; It is the bias vector.

4. The vector geographic data scale identification method according to claim 1, characterized in that, Kernel density estimation is used to adaptively segment vector geographic data into hotspot regions. The segmented hotspot regions are then rasterized into a set of rasterized images, including: Determine the two-dimensional kernel function and bandwidth; The spatial extent of the vector geographic data is uniformly divided into grids; for each grid, the kernel density estimate of all data points within the grid is calculated using the two-dimensional kernel function to generate a continuous density field. Based on the continuous density field, the information content of the grid is calculated; Based on the information content of each grid, grids with information content higher than a preset threshold are selected as hotspot areas. The vector geographic data within the hotspot areas are then rasterized to generate a set of rasterized images.

5. The vector geographic data scale identification method according to claim 4, characterized in that, The Gaussian kernel function is used as the two-dimensional kernel function, and its calculation formula is as follows: In the formula, K ( u , v ) indicates with Gaussian kernel function as a variable This represents the coordinate offset of a data point relative to the kernel center. Represents an exponential function; The bandwidth is determined by the following formula: In the formula, The total number of data points. The standard deviation of the data coordinates. This is an adjustment factor used to balance the amount of data with the degree of smoothing. These are empirical parameters; A continuous density field is generated using the following formula: In the formula, Indicates in Density estimate at [location] Indicates the total number of grid data points. This represents the square of the bandwidth. Represents the Gaussian kernel function; x and y Represents the density estimation point. x i and y i This represents the i-th point within the grid. Based on the continuous density field, the information content of the grid is calculated using the following formula: In the formula, This represents the total number of coordinate points within the grid. i It is the index of the coordinate point within the grid.

6. The vector geographic data scale identification method according to claim 1, characterized in that, Based on the ResNeXt visual feature extraction network, scale bar recognition is performed by fusing the semantic labels of the layers with the visual features of the rasterized image set, including: The rasterized image set is input into a pre-trained deep convolutional neural network to extract visual feature vectors; The semantic labels of the layers are converted into semantic feature vectors through category embedding. The visual feature vectors, semantic feature vectors and geographic feature vectors are then fused to obtain the final fused feature vector. Based on the final fused feature vector, it is mapped to a category score through a fully connected layer and then... The function generates a probability distribution, selects the category with the highest probability as the sub-map scale regression result, and after processing the rasterized image set, obtains the sub-map scale discrimination set and the confidence set of each sub-map as the corresponding scale.

7. The vector geographic data scale identification method according to claim 6, characterized in that, A weighted voting mechanism is used to integrate the identification results of each submap to obtain the scale identification results of the vector geographic data, including: Calculate the weighted total probability of voting at different scales using the following formula: In the formula, This represents the total number of subgraphs. This indicates the total number of scale categories. The indicator function represents the total probability of weighted voting at different scales. Indicates when subgraph Predicted scale category For the first i Scale label The value is 1 if the condition is met, and 0 otherwise. Subgraph The confidence level of the discrimination; The scale label with the highest weighted voting probability is taken as the final scale identification result of the vector geographic data.

8. A vector geographic data scale recognition device, characterized in that, The device includes: The semantic classification module is configured to perform semantic classification on the text sequences of attribute fields of vector geographic data based on the BERT model to obtain layer semantic labels. The hotspot region segmentation module is configured to adaptively segment vector geographic data into hotspot regions using kernel density estimation, and then rasterize the segmented hotspot regions into a set of rasterized images. The scale recognition module is configured to perform scale recognition based on the ResNeXt visual feature extraction network, which integrates the semantic labels of the layers with the visual features of the rasterized image set, and uses a weighted voting mechanism to integrate the recognition results of each sub-map to obtain the scale recognition result of the vector geographic data; wherein, the sub-map is an image in the rasterized image set.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the vector geographic data scale identification method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the vector geographic data scale identification method as described in any one of claims 1-7.

Images