Urban land function identification method and system based on ai

Abstract

The invention discloses an ai-based urban land function identification method and system, and belongs to the technical field of artificial intelligence data processing, and the method comprises the steps: obtaining a high-resolution remote sensing image, a dynamic population thermodynamic diagram and POI data of a target region, and constructing a multi-modal original data set; performing space-time alignment and characterization processing on the data set to generate standardized multi-modal data; weighting the remote sensing image based on the standardized data in combination with population thermodynamic and POI data to generate a plot function feature vector; classifying the feature vectors and outputting a preliminary land use function classification result; generating a credibility score based on the classification result; when the credibility is lower than a threshold value, analyzing the feature vector to generate land use function mixing proportion data; and integrating the classification result, the credibility score and the mixing proportion data to generate a land use function identification report. According to the method, multi-modal data fusion is adopted, and a function mixing degree dynamic analysis strategy is combined, so that high-precision identification of urban functional areas and quantitative analysis of mixed functional plots can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence data processing, in particular to an urban land function recognition method and system based on AI. BACKGROUND

[0002] Urban land function recognition refers to the process of classifying and distinguishing the social and economic functions carried by different plots in the city, such as residence, commerce, industry, and public service, through technical means. It is a basic work for urban planning, resource allocation, dynamic monitoring, and scientific management. With the development of information technology, using artificial intelligence algorithms to process multi-source urban big data has become the mainstream direction in this field. By building computer models to automatically analyze and understand complex urban spatial structures, large-scale and efficient function recognition is achieved.

[0003] Existing data-based urban land function recognition methods usually use high-resolution remote sensing images as the main data source, and use computer vision technology to extract physical form features such as buildings, roads, and green spaces for classification. Some methods further integrate point of interest (POI) data or mobile signaling data, combining these social and economic activity data features with image features for simple splicing or combination, and then inputting them into machine learning classifiers such as support vector machines or deep neural networks to improve recognition accuracy through more dimensions of information.

[0004] However, the existing technical solutions have deficiencies in data processing and analysis logic. First, the fusion method between multi-source heterogeneous data is relatively simple, and simple feature splicing cannot fully exploit the deep associations between different modal data. For example, how do the spatio-temporal rhythms of human activities affect the interpretation of physical space forms? The lack of this interactive relationship limits the recognition accuracy of the model. Second, most existing methods are limited to single-label classification frameworks and cannot effectively represent mixed-function land that is prevalent in cities. Their output results often oversimplify complex realities. In addition, most methods only provide a single classification result, lacking a quantitative evaluation mechanism for their own prediction uncertainty, and users cannot judge the reliability of the recognition results, which greatly limits their credibility in key decision-making applications. SUMMARY

[0005] To solve the above problems, the present application provides an urban land function recognition method and system based on AI, which uses multi-modal data fusion and combines a function mixing degree dynamic analysis strategy to achieve high-precision recognition of urban function areas and quantitative analysis of mixed-function plots.

[0006] The above objectives can be achieved through the following solutions: An AI-based urban land use function recognition method and system, comprising: acquiring high-resolution remote sensing images, dynamic population heat maps, and POI data of a target area to construct a multi-modal original data set; performing spatio-temporal alignment and feature processing on the data set to generate standardized multi-modal data; generating a land parcel function feature vector based on the standardized data combined with population heat and POI data; classifying the feature vector to output a preliminary land use function classification result; generating a credibility score based on the classification result; when the credibility is lower than a threshold, analyzing the feature vector to generate land use function mixed proportion data; integrating the classification result, the credibility score, and the mixed proportion data to generate a land use function recognition report.

[0007] Optionally, the spatio-temporal alignment and feature processing of the multi-modal original data set to generate standardized multi-modal data comprises: extracting building contours and ground textures from the high-resolution remote sensing image data to generate an image feature map; aggregating population densities of the dynamic population heat map data at different time periods to generate a population dynamic curve; counting the distribution densities of different formats in the POI data to generate a POI density distribution map; and grid aligning the image feature map, the population dynamic curve, and the POI density distribution map to generate standardized multi-modal data.

[0008] Optionally, the generation of the land parcel function feature vector comprises: calculating the correlation weight of the population dynamic curve and the POI density distribution map at each spatial position on the image feature map to generate an attention map; multiplying the attention map and the image feature map to obtain an attention weighted feature map; and performing cross-modal fusion on the attention weighted feature map, the population dynamic curve, and the POI density distribution map to generate a land parcel function feature vector.

[0009] Optionally, the classification processing of the land parcel function feature vector to output a preliminary land use function classification result comprises: obtaining a multi-modal fusion feature vector using the land parcel function feature vector; and performing classification decision on the multi-modal fusion feature vector to output a preliminary land use function classification result.

[0010] Optionally, the multiplication of the attention map and the image feature map to obtain an attention weighted feature map comprises: performing spatial position alignment processing on the attention map and the image feature map to obtain an attention map and image feature map pair; and performing element-wise multiplication operation on the attention map and image feature map pair to obtain an attention weighted feature map.

[0011] Optionally, the function mixing degree analysis on the plot function feature vector to generate land use function mixing proportion data comprises: decomposing the plot function feature vector into a morphology feature sub-vector, a people flow feature sub-vector and an industry format feature sub-vector; calculating the preliminary land use function classification result based on the morphology feature sub-vector, the people flow feature sub-vector and the industry format feature sub-vector to obtain a contribution weight; and quantifying the contribution proportion of the preliminary land use function classification result in the plot based on the contribution weight to generate land use function mixing proportion data.

[0012] Optionally, the integration of the preliminary land use function classification result, the reliability score and the land use function mixing proportion data to generate a land use function recognition report comprises: fusing the preliminary land use function classification result and the land use function mixing proportion data to generate a refined classification label; associating the reliability score with the refined classification label to form a classification result with a reliability label; and visualizing the classification result with a reliability label to generate a land use function recognition report.

[0013] Optionally, the visualization of the classification result with a reliability label to generate a land use function recognition report comprises: spatial rasterization rendering of the classification result with a reliability label to obtain a land use function heat distribution map; and reliability superposition of the land use function heat distribution map to generate a land use function recognition report.

[0014] Optionally, the generation of a reliability score based on the preliminary land use function classification result comprises: statistical feature extraction of the preliminary land use function classification result to obtain land use function classification feature distribution data; and evaluation of data consistency of the land use function classification feature distribution data to generate a reliability score.

[0015] Based on the same inventive concept, the application also provides an AI-based urban land function recognition system, which comprises: a data acquisition module for acquiring high-resolution remote sensing image data, dynamic population heat map data and point of interest (POI) data of a target area, and generating a multi-modal original data set; a data preprocessing module for performing spatio-temporal alignment and feature processing on the multi-modal original data set, and generating standardized multi-modal data; a feature fusion module for weighting the high-resolution remote sensing image data based on the standardized multi-modal data, combining the dynamic population heat map data and the POI data, and generating a land parcel function feature vector; a classification decision module for performing classification processing on the land parcel function feature vector, and outputting a preliminary land use function classification result; a credibility evaluation module for generating a credibility score based on the preliminary land use function classification result; a function mixing degree analysis module for performing function mixing degree analysis on the land parcel function feature vector if the credibility score is lower than a preset threshold, and generating land use function mixing ratio data; and a report generation module for integrating the preliminary land use function classification result, the credibility score and the land use function mixing ratio data, and generating a land use function recognition report.

[0016] Compared with the prior art, the application has the following advantages: The application realizes significant technical improvement at the data processing level by introducing a multi-modal data deep fusion and adaptive analysis mechanism. Specifically, the method uses population dynamics and industry distribution and other social and economic information to intelligently guide the model to focus on and weight key physical features in remote sensing images through an attention mechanism. This deep interaction and fusion of cross-modal information enables the generated feature vector to more deeply reflect the inherent nature of land use function, thereby significantly improving the recognition accuracy of the classification model, especially in distinguishing complex urban land parcels with similar physical forms but significantly different actual functions.

[0017] The application proposes an adaptive analysis strategy based on classification result uncertainty, enhancing the intelligence and real-world applicability of the method. Instead of providing a single classification label, the method first calculates a credibility score to assess the reliability of the classification result. For areas with low credibility and ambiguous functions, the system automatically switches to a more detailed function mixing degree analysis mode to quantitatively reveal the composition ratio of multiple functions within the land parcel. This dual-channel processing mechanism enables the method to flexibly cope with the functional complexity and mixing of urban space, providing richer and more accurate analysis results than traditional single classification.

[0018] The application greatly improves the practicability and interpretability of the final result by integrating multi-level analysis information and effectively visualizing. It systematically integrates the information of three dimensions of main function classification, function mixing ratio and result reliability into a unified report framework. By superimposing and rendering the reliability and function distribution diagram, the user can intuitively grasp the function layout and identify the reliability distribution of the result, which makes the complex analysis data into easy-to-understand decision intelligence, and provides more comprehensive, reliable and deep insight data support for urban planning and management.

[0019] Other features and advantages of the present application will be set forth in the following description, and in part will be apparent from the description, or can be learned by practice of the present application. The objects and other advantages of the present application will be realized and attained by the structure particularly pointed out in the written description and claims thereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0021] Figure 1 is a flowchart of an ai-based urban land function recognition method according to an embodiment of the present application.

[0022] Figure 2 is a dynamic population heat map according to an embodiment of the present application.

[0023] Figure 3 is a POI density distribution map according to an embodiment of the present application.

[0024] Figure 4 is an attention weight map according to an embodiment of the present application.

[0025] Figure 5 is a structural diagram of an ai-based urban land function recognition system according to an embodiment of the present application. DETAILED DESCRIPTION

[0026] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0027] Referring Figure 1 One embodiment of the present application proposes an AI-based urban land use function recognition method, which adopts multi-modal data fusion and combines with a function mixing degree dynamic analysis strategy to achieve high-precision recognition of urban functional areas and quantitative analysis of mixed-function blocks.

[0028] The method of the embodiment specifically includes: Obtaining high-resolution remote sensing image data, dynamic population heat map data and point of interest (POI) data of a target area to generate a multi-modal original data set; Performing spatio-temporal alignment and feature processing on the multi-modal original data set to generate standardized multi-modal data; Based on the standardized multi-modal data, combining the dynamic population heat map data and the POI data, weighting the high-resolution remote sensing image data to generate a block function feature vector; Performing classification processing on the block function feature vector to output a preliminary land use function classification result; Based on the preliminary land use function classification result, generating a credibility score; If the credibility score is lower than a preset threshold, performing function mixing degree analysis on the block function feature vector to generate land use function mixing proportion data; Integrating the preliminary land use function classification result, the credibility score and the land use function mixing proportion data to generate a land use function recognition report.

[0029] Specifically, first, the high-resolution remote sensing image, the dynamic population heat map and the POI data are unified into standardized analysis units through spatio-temporal alignment, and the dynamic population heat map is as shown in FIG. Figure 2 The key is that it is not simply superimposed data, but innovatively uses population dynamics and POI industry distribution as semantic guidance to intelligently weight the remote sensing image features, thereby generating a deep fusion feature vector that can represent physical form, human activity rhythm and social and economic function. On this basis, the method establishes a double-channel analysis mechanism, that is, first performing preliminary classification, and then generating a credibility score according to the classification result. This score serves as a decision switch, if the credibility is high, the preliminary result is accepted, if the credibility is low, a deeper function mixing degree analysis is automatically triggered to quantitatively decompose the composite function of the block. Finally, by integrating the information of the main function classification, the credibility evaluation and the mixing proportion in three dimensions, a comprehensive recognition report is formed.

[0030] Optionally, the spatio-temporal alignment and feature processing on the multi-modal original data set to generate standardized multi-modal data includes: Building outlines and surface textures are extracted from the high-resolution remote sensing image data to generate an image feature map; By aggregating the population density of the dynamic population heatmap data at different time periods, a dynamic population curve is generated; The distribution density of different business types in the Points of Interest (POI) data is statistically analyzed to generate a POI density distribution map; The image feature map, population dynamic curve, and POI density distribution map are aligned into a grid to generate standardized multimodal data.

[0031] Specifically, firstly, for high-resolution remote sensing image data, image segmentation or edge detection algorithms are used to extract the contour information of buildings and land parcels. Simultaneously, texture analysis techniques such as gray-level co-occurrence matrix (GLCM) or wavelet transform are employed to quantify the visual characteristics of land cover, such as the texture complexity and directionality of vegetation, water bodies, and exposed surfaces. These two parts of information together constitute an image feature map containing both physical morphology and surface material information. Secondly, for dynamic population heatmap data, which typically consists of population spatial distribution maps across multiple time slices, this method aggregates the population density values ​​of specific spatial units, such as pre-defined grid units, across all time slices to form a time-series vector. This vector is the population dynamic curve, accurately depicting the rhythm of human activity intensity changes within that unit over a day or a cycle. Thirdly, for Points of Interest (POI) data, spatial density analysis is performed separately for different business types, such as commercial, office, residential, and public services. A commonly used method is kernel density estimation, generating a continuous density surface for each business type. These surfaces are superimposed to form a multi-channel POI density distribution map, as shown in the image. Figure 3 As shown, each channel represents the spatial agglomeration degree of a specific business type. Finally, to achieve multimodal data fusion, grid alignment is required. This involves establishing a unified georeferenced grid system covering the entire target area, projecting the previously generated image feature map, population dynamics curve, and POI density distribution map onto this grid system. For each grid cell, the pixel values ​​of its corresponding image feature map, the complete population dynamics curve, and the density values ​​of the POI density distribution map for each channel are extracted and combined to form a high-dimensional feature vector. This vector is then normalized to eliminate dimensional differences between different features, ultimately producing standardized multimodal data. The data for each grid cell can be represented as: , in, Represents grid cell Standardized multimodal data; It is a feature vector extracted from the image feature map that describes the physical morphology and surface texture of the unit; is a population dynamic curve vector representing the rhythm of population activity of the unit; is a density feature vector representing the functional format configuration of the unit extracted from the POI density distribution map.

[0032] Optionally, the generating the land parcel functional feature vector comprises: calculating the correlation weight of the population dynamic curve and the POI density distribution map at each spatial position on the image feature map to generate an attention map; multiplying the attention map and the image feature map to obtain an attention weighted feature map; cross-modal fusion of the attention weighted feature map, the population dynamic curve and the POI density distribution map to generate a land parcel functional feature vector.

[0033] Specifically, first, the correlation weight is calculated to generate an attention map. The core of this step is to establish a model to evaluate the indicative strength of the socio-economic activity information represented by the population dynamic curve and the POI density distribution map to the physical environment information represented by the image feature map at each spatial position. In specific implementation, for each grid unit, the population dynamic curve vector and the feature vector of the POI density distribution map are spliced to form a comprehensive socio-economic feature vector. The vector is input into a pre-trained or jointly trained neural network module, which is usually composed of several fully connected layers and an activation function such as Sigmoid. The function is to learn the nonlinear mapping from the socio-economic features to the attention weight. The output of the network is a scalar value between 0 and 1, which is the correlation weight of the spatial position. The attention weight is as shown in Figure 4 . Repeat this operation for all grid units in the target area to generate a two-dimensional weight matrix corresponding to the image feature map in the spatial dimension, which is the attention map. The calculation process can be represented as: , wherein, is the correlation weight at the grid unit ; represents the Sigmoid activation function for restricting the output value between 0 and 1; and are the learnable parameters of the neural network module, i.e. the weight matrix and the bias vector; and are the population dynamic curve vector and the POI density feature vector of the unit, respectively; and The concatenation operation of the vector is represented. Secondly, the generated attention map is multiplied with the image feature map to obtain the attention weighted feature map. This step applies the correlation weight obtained in the previous step to the original image feature. The specific operation is to align the attention map and the image feature map in the exact spatial position, and then perform element-wise multiplication operation on each corresponding grid cell. That is, the correlation weight value of the cell is multiplied with each element in the image feature vector of the cell respectively. After this operation, the corresponding image features of those areas identified as more important by the socio-economic activity information will be enhanced or highlighted in the new feature map, while the image features of the areas with lower relevance will be suppressed. Finally, the attention weighted feature map is cross-modal fused with the population dynamic curve and the POI density distribution map to generate the final plot function feature vector. This is the information summary step. For each grid cell, its image feature vector after attention weighting, the original population dynamic curve vector and the POI density distribution map feature vector are concatenated in a predetermined order to form a single vector with higher dimension and more comprehensive information. The final generated long vector is the plot function feature vector, which not only contains the physical appearance of the plot, but also integrates the dynamic human activity rhythm and the distribution of social function industry, and the physical features have been intelligently adjusted according to the importance of the latter two.

[0034] Optionally, the classification processing of the plot function feature vector includes: The multi-modal fusion feature vector is obtained by using the plot function feature vector; The multi-modal fusion feature vector is classified and decided to output the preliminary land use function classification result.

[0035] Specifically, first, a multi-modal fusion feature vector is obtained using the plot function feature vector. Here, the plot function feature vector obtained from the previous step is already a highly fused feature representation. In order to further enhance the discriminative ability of the features and adapt to the needs of the classification task, it is usually input into one or more fully connected neural network layers. The role of these network layers is to perform nonlinear transformation and deep abstraction on the input features, and to integrate the features that have been preliminarily fused between different modalities at a higher level, so as to learn the combined patterns that are more critical to classification decision-making. The output of this process is a multi-modal fusion feature vector that has been refined and optimized again, and its dimension and representation are more suitable for the final classification decision. Second, a preliminary land use function classification result is output by making a classification decision on the multi-modal fusion feature vector. The multi-modal fusion feature vector obtained in the previous step is input into the last layer of the classification model, i.e., the classification decision layer. This layer is usually a fully connected layer with a Softmax activation function. The function of the Softmax function is to convert the input vector, i.e., the confidence score of the model for each class, into a probability distribution. The value of each element in the distribution represents the probability that the plot belongs to the corresponding functional class, and the sum of all elements is 1. Finally, the class with the highest probability value is selected as the preliminary land use function classification result of the plot. The entire decision-making process can be summarized by the following formula: , wherein, represents the probability that the given original plot function feature vector is classified as class ; is the multi-modal fusion feature vector obtained after intermediate layer processing; and are the weight matrix and bias vector of the classification decision layer, respectively, which are parameters learned through model training; represents a specific land use function class, such as residential area, commercial area, or industrial area, etc. The final preliminary land use function classification result is the class with the maximum probability value.

[0036] Optionally, the multiplying the attention map and the image feature map to obtain an attention weighted feature map comprises: aligning the attention map and the image feature map in spatial position to obtain an attention map and image feature map pair; performing element-wise multiplication operation on the attention map and image feature map pair to obtain an attention weighted feature map.

[0037] Specifically, first, spatial position alignment processing is performed to obtain an attention map and an image feature map pair. This step aims to ensure that each weight value in the attention map accurately corresponds to a spatial cell in the image feature map. In the pre-processing, both the image feature map and the attention map generated from social and economic data are mapped to the same unified geographic reference grid system. Therefore, the image feature map can be regarded as a three-dimensional tensor, with the first two dimensions representing spatial positions (rows and columns of the grid), and the third dimension being the image feature vector of the position; while the attention map is a two-dimensional matrix, with the same dimensions as the first two dimensions of the image feature map, and each element in the matrix is the associated weight of the corresponding spatial position. The spatial position alignment processing is to confirm and utilize this inherent spatial co-site relationship to pair each grid cell associated weight with its corresponding image feature vector to form a computable attention map and image feature map pair. Then, element-wise multiplication is performed on the attention map and image feature map pair to generate an attention-weighted feature map. This operation is a direct application of the attention mechanism. For each grid cell , the scalar associated weight in the attention map is multiplied by the corresponding image feature vector in the image feature map pair. This multiplication is not a matrix multiplication, but a scalar and vector multiplication, i.e., the scalar weight is multiplied by each component of the image feature vector . This process is repeated for each grid cell in the entire target area. This operation can be represented by the following formula: , where, represents the attention-weighted feature vector generated at grid cell ; is the scalar associated weight of the corresponding position obtained from the attention map, with a value between 0 and 1, reflecting the importance of social and economic activity information at that position for function identification; is the original image feature vector of the position, containing physical information such as building outlines and surface textures. After this operation, the output attention-weighted feature map has the same dimensions as the original image feature map, but its internal feature values have been rescaled according to the importance of social and economic information.

[0038] Optionally, the land use function feature vector is decomposed into a morphological feature sub-vector, a people flow feature sub-vector, and a business format feature sub-vector. ​Contribution degree weights are obtained by calculating the preliminary land use function classification result based on the morphological feature sub-vector, the people flow feature sub-vector, and the industry format feature sub-vector. Based on the contribution degree weights, the proportion of the preliminary land use function classification result in the plot is quantified to generate land use function mixing proportion data.

[0039] Specifically, first, the plot function feature vector is decomposed. The feature vector itself is a high-dimensional vector formed by splicing features of different modal data sources. The decomposition operation is to reverse split the long vector back to its original constituent part according to the preset feature structure. Specifically, it is divided into three sub-vectors with clear physical and socio-economic significance, namely the morphological feature sub-vector, which is derived from the attention-weighted high-resolution remote sensing image features; the people flow feature sub-vector, which is derived from the population dynamic curve extracted from the dynamic population heat map data; and the industry format feature sub-vector, which is derived from the industry density distribution of the point of interest POI data statistics. Secondly, based on the three decomposed feature sub-vectors, the preliminary land use function classification result is calculated to obtain the contribution degree weights. This step aims to evaluate the contribution degree of each data modality to determining that the plot has a certain specific function. For each preset possible land use function category, such as residential, commercial, industrial, etc., the model internally stores a set of "function prototype" features learned through training. When calculating, the morphological feature sub-vector, the people flow feature sub-vector, and the industry format feature sub-vector of the plot are respectively calculated with the prototype features corresponding to each function category for similarity or matching degree. For example, by calculating the similarity of the people flow feature sub-vector of the plot and the standard "commercial function" prototype people flow curve, the contribution of the people flow dimension to the commercial function of the plot is obtained. After performing this operation for all modalities and all function categories, a set of quantified contribution degree weights is obtained. The weights can be represented as: , wherein, represents the contribution degree weight of the i-th modality, such as morphology, people flow, or industry format, to the k-th land use function category; is a similarity calculation function, such as cosine similarity; is the i-th modality feature sub-vector of the plot; , is the i-th function category stored in the model; , is the k-th land use function category stored in the model. ​prototype feature vector on one modality, which is learned in the model training phase. Finally, based on the calculated contribution weights, the composition proportion of the preliminary land use function classification result in the plot is quantified, and the land use function mixed proportion data is generated. This step integrates the contribution information of all modalities and converts it into a direct proportion distribution. For each candidate function category First, the contribution weights from the shape, flow, and format dimensions are fused to obtain a comprehensive score of the function category. Then, a normalization function such as Softmax is used to convert the comprehensive scores of all function categories into a probability distribution, and the sum of the distribution is 1. Each value in the distribution represents the composition proportion of the corresponding function in the plot. The final proportion vector is the land use function mixed proportion data. Its calculation can be represented as: , where, is the final mixed proportion of function category ; is the comprehensive score of function category ; and the summation symbol represents the traversal summation of the scores of all candidate function categories .

[0040] Optionally, the integration of the preliminary land use function classification result, the reliability score, and the land use function mixed proportion data to generate a land use function recognition report includes: Fusing the preliminary land use function classification result and the land use function mixed proportion data to generate a refined classification label; Associating the reliability score with the refined classification label to form a classification result with reliability annotation; Visualizing the classification result with reliability annotation to generate a land use function recognition report.

[0041] ​Specifically, firstly, the preliminary land use function classification result and the land use function mixed proportion data are fused to generate a refined classification label. The preliminary land use function classification result is the dominant function label of a plot, and the land use function mixed proportion data reveals the quantitative composition of each function component in the plot. The purpose of this step is to combine the two and create a more descriptive classification label. For example, if the preliminary classification result of a plot is "commercial district", and its land use function mixed proportion data shows "seventy percent commercial and thirty percent residential", then the refined classification label can be generated as "commercial-dominated commercial-residential mixed district". This fusion operation can be performed based on a set of predefined rules that determine the final mixed function description according to the proportion threshold of the dominant function and the secondary function. This step converts a single, discrete classification result into a continuous, multi-dimensional function description that is closer to the complexity of reality. Secondly, the confidence score is associated with the newly generated refined classification label to form a classification result with confidence annotation. The confidence score is an indicator of the reliability of the preliminary classification result. This step is to directly attach this quantitative indicator to the refined classification label of each plot. For example, the identification result of the aforementioned plot can be recorded as "commercial-dominated commercial-residential mixed district, confidence 0.85". This annotation allows the user of the report to not only understand the qualitative function of the plot, but also immediately know the confidence level of the judgment. In particular, for areas with low confidence, the user will be reminded that the function of the area may be more complex or lack data support, and further attention is needed. Finally, the classification result with confidence annotation is visualized to generate the final land use function identification report. This step is to convert abstract data into intuitive geographic information maps and statistical charts. The core visualization result is usually a land use function map, where each plot or grid cell on the map is assigned a different color or symbol according to its refined classification label, forming a colorful urban function mosaic map. At the same time, the confidence score can also be visualized on the map, for example, by the depth, transparency or superimposed shading of the color, so that low-confidence areas are immediately apparent on the map. The report can also include area proportion statistics charts of various functions, function mixing degree heat maps and other auxiliary charts to provide multi-angle analysis views. This series of maps, tables and necessary text descriptions together form a detailed and diverse land use function identification report.

[0042] Optionally, the visualizing the classification result with confidence annotation to generate a land use function identification report comprises: spatially rasterizing the classification result with confidence annotation to obtain a land use function heat distribution map; superimposing the confidence on the land use function heat distribution map to generate a land use function identification report.

[0043] Specifically, first, the classification results with credibility labels are spatially rasterized to obtain a land use function heat map. The "heat map" here does not refer to temperature, but borrows its concept to intuitively show the distribution intensity and aggregation of different functions in space. This operation first establishes a geographic coordinate system corresponding to the analysis grid. Then, a unique and highly recognizable color is set for each refined classification label, such as pure residential area, commercial-dominated mixed residential area, etc. Next, each analysis grid cell is traversed, its refined classification label is read, and the corresponding preset color is used to fill the location of the cell on the map. After rendering all cells, a thematic map is formed, which is a land use function heat map that separates different functional areas by color. This map intuitively shows the spatial layout, range boundary, and aggregation pattern of various functional areas in the city. Secondly, the land use function heat map generated in the previous step is superimposed with credibility to generate the final land use function identification report. This step aims to integrate the certainty information of the classification results into spatial visualization. Various visual encoding techniques can be used in implementation. One commonly used method is to use transparency or color saturation. The credibility score of each grid cell is linearly mapped to the Alpha channel value or saturation of the rendered color. For example, a cell with a credibility score of 1 has a fully opaque and highest saturation functional color; while a cell with a lower credibility score, its color becomes more transparent or tends to gray. After processing, users can not only identify the function type by color, but also intuitively perceive the reliability of the identification result by the brightness or dullness of the color. Those areas with color blur and high transparency are areas with high model uncertainty, which need to be paid special attention to by the user. Finally, this visualization map that integrates both function type and credibility information, together with necessary legends, titles, and statistical analysis charts, constitutes a complete and detailed land use function identification report.

[0044] Optionally, the generating a credibility score based on the preliminary land use function classification result comprises: performing statistical feature extraction on the preliminary land use function classification result to obtain land use function classification feature distribution data; evaluating data consistency of the land use function classification feature distribution data to generate a credibility score.

[0045] Specifically, firstly, statistical features are extracted from the preliminary land use function classification results to obtain land use function classification feature distribution data. Here, the preliminary land use function classification results do not simply refer to the final selected function label with the highest probability, but rather to the complete probability distribution vector output by the classification model's decision layer, such as the Softmax layer. This vector contains the probability values ​​of a plot belonging to each predefined function category, with a sum of 1. This probability vector itself constitutes the core land use function classification feature distribution data. From this distribution data, various statistical features can be extracted, such as the peak value of the probability distribution (i.e., the highest probability value) or the entropy of the probability distribution. Secondly, the data consistency of the land use function classification feature distribution data is evaluated to generate a confidence score. Data consistency here refers to the certainty or explicitness of the classification probability distribution. A highly consistent distribution is characterized by one category having a probability much higher than all other categories, indicating that the model has strong confidence in the classification result. Conversely, a low-consistency distribution is characterized by multiple categories having relatively similar probability values, resulting in a flat distribution and indicating that the model hesitates among multiple possibilities, leading to high uncertainty in the outcome. One of the most common and effective methods for assessing this consistency is to calculate the information entropy of the probability distribution. Information entropy is a standard measure of uncertainty; the higher the entropy value, the greater the uncertainty and the worse the data consistency. Therefore, the credibility score can be defined as a function inversely proportional to the normalized information entropy. Its calculation formula can be expressed as: , in, The final credibility score has a value range of 0 to 1; In the land use function classification characteristic distribution data, the land parcel belongs to the first... The probability of each functional category, which is directly output by the classifier; It represents the total number of predefined land use categories; summation symbol. Indicates all Iterate through and sum the values ​​of each category; It is a logarithm with base 2. The denominator is... It is the maximum possible entropy value of the probability distribution, used to normalize the entropy value to the interval between 0 and 1.

[0046] Based on the same inventive concept, such as Figure 5 As shown, the present invention also provides an AI-based urban land use function identification system, the system comprising: The data acquisition module is used to acquire high-resolution remote sensing image data, dynamic population heat map data, and point of interest (POI) data of the target area, and generate a multimodal raw dataset. The data preprocessing module is used to perform spatiotemporal alignment and feature processing on the original multimodal dataset to generate standardized multimodal data; a feature fusion module configured to generate a land parcel functional feature vector by weighting the high-resolution remote sensing image data based on the standardized multi-modal data, in combination with the dynamic population heat map data and point of interest (POI) data; a classification decision module configured to perform classification processing on the land parcel functional feature vector and output a preliminary land use functional classification result; a credibility evaluation module configured to generate a credibility score based on the preliminary land use functional classification result; a functional mixing degree analysis module configured to perform functional mixing degree analysis on the land parcel functional feature vector and generate land use functional mixing proportion data if the credibility score is lower than a preset threshold; a report generation module configured to generate a land use functional identification report by integrating the preliminary land use functional classification result, the credibility score, and the land use functional mixing proportion data.

[0047] To verify the feasibility of the present application in implementation, the present application is applied to a certain bureau. The bureau aims to conduct a fine land use functional status survey and update of a rapidly developing high-tech park in the city. Traditional manual survey methods are time-consuming and costly, and it is difficult to accurately identify and quantify emerging and highly mixed land parcels, resulting in lagging planning map updates. The bureau hopes to use the method and system of the present application to achieve efficient, accurate, and dynamic identification of urban land use functions.

[0048] In this embodiment, the bureau uses the method and system proposed in the present application to identify and analyze the land use functions of the core area of the high-tech park. The system first obtains high-resolution remote sensing image data of the third quarter of 2024, continuous one-week hourly dynamic population heat map data provided by communication operators, and point of interest (POI) data provided by mainstream map service providers, and generates a multi-modal raw data set covering the entire park area.

[0049] In the data preprocessing stage, the system performs spatio-temporal alignment and feature processing on the multi-modal raw data set. First, building outlines and surface texture information are extracted from the remote sensing image to generate an image feature map. Second, population density data at different time periods are aggregated to generate a population dynamic curve reflecting the rhythm of human flow activities for each land parcel. At the same time, the distribution density of different formats in the park, such as technology companies, restaurants, apartments, convenience stores, etc., is counted to generate a multi-channel POI density distribution map. Finally, the image feature map, population dynamic curve, and POI density distribution map are projected onto a 50m x 50m geographic reference grid system and normalized to generate standardized multi-modal data, laying a foundation for subsequent analysis.

[0050] In the core analysis stage, the system generates plot function feature vectors based on standardized multi-modal data. Taking plots A and B in the park as examples, the POI data of plot A is dominated by technology companies, and its population dynamic curve presents the typical "9-to-5" weekday characteristics. Plot B is mixed with a large number of restaurants, cafes, and some apartment buildings, and its population dynamic curve shows peaks in the afternoon and evening on weekdays. The system generates an attention map accordingly, with a high concentration of attention weights on plot A and a more dispersed weight on plot B. This attention map is then multiplied element by element with the image feature map to obtain an attention-weighted feature map, i.e., visual features more closely related to socioeconomic activities are enhanced. Finally, the system performs cross-modal fusion of the weighted feature map with the population dynamic curve and the POI density distribution map to generate information-rich plot function feature vectors for plots A and B.

[0051] Subsequently, the system classifies the plot function feature vectors. Through a pre-trained classification model, plot A is initially classified as an "industrial research and development zone", with a highly concentrated probability distribution output by the classification decision layer, resulting in a confidence score of 0.95. Plot B is initially classified as a "commercial service area", but the probability distribution shows that the model is hesitant between "commercial service area" and "residential area", resulting in a confidence score of only 0.68.

[0052] Since the confidence score of plot B is 0.68, which is lower than the system's preset threshold of 0.80, the system automatically initiates function mix degree analysis for it. The system decomposes the function feature vector of plot B into three feature sub-vectors: morphology, people flow, and business format. By calculating the similarity of each sub-vector to the "commercial service" and "residential" function prototypes, it finds that the business format feature contributes significantly to the commercial function, while the people flow feature also contributes significantly to the residential function. Finally, the system quantifies the functional composition ratio of plot B as "65% commercial service function and 35% residential function".

[0053] Finally, the system integrates all analysis results to generate a land use function recognition report. The label of plot A is determined as "industrial research and development zone" with a confidence score of 0.95. The preliminary classification result of plot B, "commercial service area", is fused with the mixed proportion data to generate a refined classification label "commercial service-oriented commercial-residential mixed area" with an associated confidence score. In the final visual report, the land use functions of the entire park are presented in the form of a color map, with plot A rendered in high saturation blue representing industrial research and development, and plot B rendered in orange representing commercial-residential mixed use with slightly higher color transparency, intuitively indicating the complexity of the region's function and the uncertainty of the recognition.

[0054] By comparing with the results of artificial field investigation conducted at the same time by the bureau, the recognition accuracy of the method reaches 94%, and especially in recognizing mixed-function plots, it has a significant advantage compared with the traditional method with an accuracy of 78%. The entire recognition process only takes 2 days, which is much faster than the 3 weeks required by artificial investigation. The data show that the application can efficiently and accurately identify the function of urban land, especially can reveal the functional complexity of complex areas, and provide strong technical support for fine management and scientific planning of cities.

[0055] Table 1 Target area plot function preliminary classification and credibility evaluation data table Table 2 Typical plot function mixing degree analysis result data table Table 3 Comparison data table of recognition effect of new and old methods From the data of the above table 1 to table 3, it can be seen that the application has excellent performance in the application of urban land function identification. Table 1 shows that the system can quickly classify different plots, and automatically generate credibility scores based on classification probability distribution, successfully screening out functionally complex B plots and D plots for in-depth analysis. Table 2 specifically shows the fine analysis capability of the system for complex plots. By quantifying the contribution of multi-modal features, the realistic function mixing ratio and refined label are obtained. The comparison data in table 3 clearly proves that the application has an overwhelming advantage in recognition accuracy and work efficiency compared with the traditional investigation method, especially in identifying an increasing number of mixed-function plots, the accuracy is significantly improved, which can provide more real and timely decision basis for urban planning.

[0056] It should be noted that the electrical connection between the above-mentioned units does not necessarily represent the direct connection of the line, and the indirect connection mode can also be applied to the embodiments of the application as long as the purpose of the application is achieved. The above described is only an exemplary embodiment of the application, and cannot limit the scope of the application.

[0057] That is, any equivalent changes and modifications made according to the teachings of the application are still within the scope of the application. Other embodiments of the application will be readily apparent to those skilled in the art upon considering the specification and practice of the true disclosure. The present application is intended to cover any variations, uses, or adaptive changes to the application that follow the general principles of the application and include common knowledge or conventional technical means in the art not disclosed by the application.

Claims

1. An AI-based method for identifying urban land use functions, characterized in that, The method includes: Acquire high-resolution remote sensing image data, dynamic population heat map data, and point of interest (POI) data for the target area to generate a multimodal raw dataset; The original multimodal dataset is subjected to spatiotemporal alignment and feature generation to produce standardized multimodal data; Based on the standardized multimodal data, combined with the dynamic population heat map data and point of interest (POI) data, the high-resolution remote sensing image data is weighted to generate a land parcel functional feature vector. The functional feature vectors of the land parcels are classified, and preliminary land use function classification results are output. Based on the preliminary land use function classification results, a confidence score is generated; If the confidence score is lower than a preset threshold, the functional mix degree of the land parcel is analyzed to generate land use function mix ratio data. By integrating the preliminary land use function classification results, the confidence score, and the land use function mixing ratio data, a land use function identification report is generated.

2. The AI-based urban land use function identification method according to claim 1, characterized in that, The process of performing spatiotemporal alignment and feature generation on the original multimodal dataset to produce standardized multimodal data includes: Building outlines and surface textures are extracted from the high-resolution remote sensing image data to generate an image feature map; By aggregating the population density of the dynamic population heatmap data at different time periods, a dynamic population curve is generated; The distribution density of different business types in the Points of Interest (POI) data is statistically analyzed to generate a POI density distribution map; The image feature map, population dynamic curve, and POI density distribution map are aligned into a grid to generate standardized multimodal data.

3. The AI-based urban land use function identification method according to claim 2, characterized in that, The generated land parcel functional feature vector includes: The correlation weights of the population dynamic curve and POI density distribution map at each spatial location on the image feature map are calculated to generate an attention map. The attention map is multiplied by the image feature map to obtain an attention-weighted feature map; The attention-weighted feature map is fused with the population dynamic curve and POI density distribution map across modes to generate a land parcel functional feature vector.

4. The AI-based urban land use function identification method according to claim 1, characterized in that, The classification process of the functional feature vector of the land parcel, and the output of the preliminary land use function classification result, includes: Using the functional feature vectors of the land parcels, a multimodal fusion feature vector is obtained; The multimodal fused feature vector is used to make a classification decision, and a preliminary land use function classification result is output.

5. The AI-based urban land use function identification method according to claim 3, characterized in that, The step of multiplying the attention map with the image feature map to obtain the attention-weighted feature map includes: The attention map and the image feature map are spatially aligned to obtain an attention map and image feature map pair. Element-wise multiplication is performed on the attention map and the image feature map to obtain the attention-weighted feature map.

6. The AI-based urban land use function identification method according to claim 1, characterized in that, The step of performing functional mixing degree analysis on the functional feature vector of the land parcel to generate land use function mixing ratio data includes: The functional feature vector of the land parcel is decomposed into morphological feature sub-vectors, pedestrian flow feature sub-vectors, and business type feature sub-vectors; Based on the morphological feature sub-vector, pedestrian flow feature sub-vector, and business type feature sub-vector, the contribution weight is calculated on the preliminary land use function classification results. Based on the contribution weight, the proportion of the preliminary land use function classification results within the land parcel is quantified to generate mixed land use function ratio data.

7. The AI-based urban land use function identification method according to claim 1, characterized in that, The process of integrating the preliminary land use function classification results, the confidence score, and the land use function mixing ratio data to generate a land use function identification report includes: By integrating the preliminary land use function classification results with the land use function mixing ratio data, a detailed classification label is generated; The confidence score is associated with the refined classification label to form a classification result with confidence label; The classification results with confidence level labels are visualized to generate a land use function identification report.

8. The AI-based urban land use function identification method according to claim 7, characterized in that, The step of visualizing the classification results with confidence labels to generate a land use function identification report includes: The classification results with confidence labels are rendered in a spatial raster format to obtain a land use function heat map. The credibility of the land use function heat map is overlaid to generate a land use function identification report.

9. The AI-based urban land use function identification method according to claim 1, characterized in that, The generation of a credibility score based on the preliminary land use function classification results includes: Statistical features are extracted from the preliminary land use function classification results to obtain land use function classification feature distribution data; The data consistency of the land use function classification feature distribution data is evaluated, and a confidence score is generated.

10. An AI-based urban land use function identification system, applied to the AI-based urban land use function identification method as described in any one of claims 1-9, characterized in that, The system includes: The data acquisition module is used to acquire high-resolution remote sensing image data, dynamic population heat map data, and point of interest (POI) data of the target area, and generate a multimodal raw dataset. The data preprocessing module is used to perform spatiotemporal alignment and feature processing on the original multimodal dataset to generate standardized multimodal data; The feature fusion module is used to weight the high-resolution remote sensing image data based on the standardized multimodal data, combined with the dynamic population heat map data and point of interest (POI) data, to generate a land parcel functional feature vector. The classification decision module is used to classify the functional feature vectors of the land parcels and output preliminary land use function classification results. The credibility assessment module is used to generate a credibility score based on the preliminary land use function classification results; The functional mixing degree analysis module is used to analyze the functional mixing degree of the land parcel's functional feature vector and generate land use function mixing ratio data if the confidence score is lower than a preset threshold. The report generation module is used to integrate the preliminary land use function classification results, the confidence score, and the land use function mixing ratio data to generate a land use function identification report.

Images