A method for probabilistic quantitative identification of mixed states of urban functional areas by fusing multi-source heterogeneous data

Abstract

This invention discloses a method for probabilistic quantification of mixed states in urban functional areas by integrating multi-source heterogeneous data, belonging to the fields of geographic information science, smart cities, and artificial intelligence. The method includes: S1, constructing a traffic analysis area; S2, acquiring multi-source data; S3, preprocessing the multi-source data acquired in S2; S4, generating true probability distribution labels; S5, extracting socio-semantic features; S6, extracting dynamic activity features; S7, extracting physical scene features; S8, feature fusion and distribution prediction; S9, model training; and S10, determining the mixed state and outputting the results. This invention achieves refined quantification of the mixed states of urban functional areas, outputting a continuous probability distribution vector rather than a single category label, significantly improving the information density and application value of the recognition results.

Description

Technical Field

[0001] This invention belongs to the fields of geographic information science, smart city and artificial intelligence technology, and specifically relates to a method for probabilistic quantitative identification of mixed states of urban functional areas by integrating multi-source heterogeneous data. Background Technology

[0002] Urban functional zones (UFZs) refer to areas within a city that perform specific economic and social functions, such as residential areas, commercial areas, and industrial areas. Accurately identifying the distribution and attributes of UFZs is fundamental to urban planning, resource allocation, and smart city management. Traditional urban functional zone identification primarily relies on land use or remote sensing imagery, focusing on physical attributes such as morphology and density, while neglecting the social significance behind the space and the social and cultural connotations inherent in its spatial form.

[0003] With the development of geographic information technology, the technical means in this field have undergone the following evolution:

[0004] (1) Shift from land cover to social function: Traditional land use and land cover (LULC) classification mainly relies on remote sensing satellite imagery, focusing on the physical characteristics of the land surface (such as vegetation, water bodies, and impermeable surfaces). However, modern urban studies pay more attention to socio-economic functions, that is, the types of social activities that humans carry out in a specific area. For example, high-rise buildings may be residential buildings or office buildings, which are difficult to distinguish based on physical appearance alone.

[0005] (2) Application of multi-source sensing data: In order to capture social functional attributes, existing technologies have begun to introduce physical sensing data and social sensing data, such as high-resolution remote sensing images and street view images. They have the advantages of wide coverage and uniform format, and can reflect the three-dimensional spatial structure of the city. Mobile phone signaling, taxi trajectory, and social media check-in data can reflect the dynamic activity patterns of people.

[0006] (3) Introduction of deep learning models: With the development of artificial intelligence, methods based on convolutional neural networks (CNN), long short-term memory networks (LSTM) and natural language processing (NLP) models (such as Word2Vec and BERT) are widely used to extract high-dimensional features of urban areas to replace traditional manual feature engineering.

[0007] (4) The need for label distribution learning: In reality, urban blocks are often mixed-use, that is, the same area may contain residential, commercial and public service facilities at the same time. Therefore, the research trend is shifting from single label classification (treating the area as a pure class) to multi-label distribution prediction (quantifying the proportion of each function).

[0008] Despite the progress made in identifying urban functional zones, existing technologies still have the following core shortcomings that restrict their application in refined urban management.

[0009] (1) Semantic gap and recognition bias caused by a single data source: When relying solely on remote sensing or street view, it is impossible to distinguish areas with similar visual features but different functions (the phenomenon of different spectrums of the same object). For example, it is difficult to distinguish between residential buildings and public service buildings with similar appearances. When relying solely on POI or social media data, there is a lack of spatial scale information of geographical entities, and the data itself is sparsity and noise, making it difficult to accurately define the physical boundaries of functional areas.

[0010] (2) Lack of quantitative description of mixed-use status: Most existing technologies treat UFZ identification as a single-label classification task, meaning that for a given area, the model can only output one category (e.g., commercial area). This ignores the fact that mixed land use is common in urban development (e.g., mixed commercial and residential areas with ground floor shops and upper floors). Existing methods cannot answer the question of how much of the area belongs to commercial and how much belongs to residential.

[0011] (3) The ground truth generation method lacks scientific rigor: When training deep learning models, real labels are needed as a benchmark. Existing label generation methods are usually simply based on the number of POIs or their area ratio. This method has a serious flaw: it ignores the scarcity and importance of functions. For example, a train station or a top-tier hospital may occupy less land than a large park, but its dominant role in urban functions is far greater. The simple area ratio method underestimates the weight of key facilities. At the same time, this method ignores functional activity; static POI data cannot reflect the actual frequency of facility use. An abandoned shopping mall and a bustling shopping center may have no difference in POI data, but their functional attributes are completely different.

[0012] (4) The model loss function is not suitable for mixed distribution prediction: When dealing with urban function classification, existing deep learning models generally use the cross-entropy loss function. This function aims to widen the gap between the correct class and other classes, tending to output hard classification results that are either one or the other, and is not suitable for fitting a fine probability distribution. Summary of the Invention

[0013] To address the problems existing in the prior art, this invention proposes a probabilistic quantitative identification method for the mixed state of urban functional areas that integrates multi-source heterogeneous data. This method achieves refined quantification of the mixed state of urban functional areas and outputs a continuous probability distribution vector instead of a single category label. This significantly improves the information density and application value of the identification results, providing richer and more instructive data support for refined urban management than traditional classification results.

[0014] Technical Solution: To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0015] A method for probabilistic quantitative identification of mixed states of urban functional areas by fusing multi-source heterogeneous data includes the following steps:

[0016] S1. Traffic Analysis Area Construction: Obtain basic geographic data of the study area, including administrative boundaries, road networks, water systems, and buildings;

[0017] S2. Multi-source data acquisition: Acquire multi-source data corresponding to the traffic analysis area;

[0018] S3. Perform coordinate unification, spatial clipping, outlier removal, duplicate record deletion, and semantic cleaning on the multi-source data obtained in S2.

[0019] S4. Generating labels based on true probability distribution;

[0020] S5. Social semantic feature extraction: Encode the traffic analysis area-level interest point set and social text set to obtain social semantic features;

[0021] S6. Dynamic Activity Feature Extraction: An LSTM network is used to extract dynamic activity features, and the hidden state at the last time step is taken as the traffic analysis area. Dynamic activity characteristics ;

[0022] S7. Physical Scene Feature Extraction: For multiple images or multiple scene data within the same traffic analysis area, the visual feature vector and semantic feature vector are concatenated or added to obtain a single-sample comprehensive feature vector. ;

[0023] S8. Feature Fusion and Distribution Prediction: The extracted features are fused, and the fused feature vector is input into the distribution predictor to obtain the prediction result;

[0024] S9. Model Training: Using the true probability distribution labels generated in S4 as supervision signals, the distribution predictor in S8 is trained. The optimal parameter combination of the distribution predictor is obtained by minimizing the total loss function.

[0025] S10, Mixed state determination and result output.

[0026] As a preferred option, the specific content of S1 is as follows:

[0027] Highways, arterial roads, secondary arterial roads, local roads, and connecting roads are extracted. Different buffer widths are set according to road level, and road vectors are buffered and superimposed with the boundary of the study area to form road constraint segmentation surfaces. The road constraint segmentation surfaces are rasterized and vectorized to obtain candidate spatial cells. Fragmented cells with areas below the threshold or abnormal shapes are deleted to obtain the traffic analysis area set. Specifically:

[0028] ;

[0029] in, This represents the i-th traffic analysis zone. This indicates the total number of traffic analysis zones.

[0030] Preferably, in S2, the multi-source data includes:

[0031] Interest point data, denoted as a set. ;gather This includes name, category, latitude and longitude, and address information;

[0032] Social text data, denoted as set ;gather This includes latitude and longitude, address information, check-in / review text, and posting time;

[0033] Crowd activity data, denoted as set ;gather This includes latitude and longitude data of taxi trajectories, latitude and longitude data of ride-hailing vehicle trajectories, mobile phone signaling data, location heatmap data, or mobile application trajectory data;

[0034] Physical scene data, denoted as set ;gather It includes a collection of street view images and a collection of remote sensing images.

[0035] As a preferred option, the specific implementation details in S4 are as follows:

[0036] First, based on traffic analysis area Belongs to category Number of points of interest Calculate the category frequency item;

[0037] Then, based on the number of traffic analysis zones Calculate the inverse document frequency term;

[0038] Based on domain unit set Calculate the space radiation term;

[0039] A dynamic activity level is constructed based on trajectory and check-in data;

[0040] Further, construct text category semantic support items based on the social text set;

[0041] Next, construct the traffic analysis area. In functional categories Overall weight ;

[0042] Finally, the comprehensive weights of each category are normalized to obtain the true probability distribution labels.

[0043] As a preferred option, the specific implementation details in S5 are as follows:

[0044] First, based on the traffic analysis area Construct a static functional vector based on the number of interest points in each category;

[0045] Then, the mapping is applied to the traffic analysis area. The social text collection is segmented, keywords are extracted, and semantic encoding is performed, and the interest point vector and the text semantic vector are concatenated;

[0046] Finally, after passing through a fully connected mapping layer, we obtain socio-semantic features with a unified dimension. .

[0047] As a preferred option, the specific implementation details in S6 are as follows:

[0048] First, the dynamic activity of each traffic analysis area at multiple time steps is constructed into an activity time series. ;

[0049] Then, after standardizing and normalizing the activity time series, it is input into a recurrent neural network, a temporal convolutional network, or a Transformer temporal modeling network.

[0050] Finally, an LSTM network is used to extract dynamic activity features.

[0051] As a preferred option, the specific implementation details in S7 are as follows:

[0052] For each traffic analysis zone Multiple street view images were collected, and a traffic analysis area was set up. Corresponding physical scene data set ;

[0053] Visual features of each physical scene sample are extracted using convolutional neural networks, visual Transformers, or basic visual models. Spatial redundancy is removed using global average pooling units, and the visual feature vector of a single image is output. ;

[0054] Extract semantic features from physical scene features, and enhance the descriptive text output by the text construction unit. A multi-head self-attention mechanism is used to extract global semantic features and output the corresponding semantic feature vector. This establishes cross-modal associations between the physical environment in the visual and semantic spaces;

[0055] For multiple images or multiple scene data within the same traffic analysis area, visual feature vectors will be used. With semantic feature vector By concatenating or adding the samples, a single-sample comprehensive feature vector can be obtained. .

[0056] As a preferred option, the specific implementation details in S8 are as follows:

[0057] Traffic analysis area socio-semantic features Dynamic activity characteristics and physical scene features The fusion is performed to obtain the fused feature vector. :

[0058] =[ ];

[0059] Subsequently, the feature vectors will be fused. Input distribution predictor; assuming the distribution predictor includes a fully connected layer and a Softmax output layer, then the traffic analysis area... The predicted probability distribution vector is:

[0060] ;

[0061] in, Traffic analysis area The predicted probability distribution vector for each functional category; This represents the output layer weight matrix. Indicates the bias term. Represents the fused feature vector; This represents the normalized exponential function.

[0062] As a preferred option, the specific implementation details in S9 are as follows:

[0063] The true probability distribution labels generated in S4 are used as supervision signals to train the distribution predictor in S8;

[0064] KL divergence is used as the loss function; for any traffic analysis area Its distribution fitting loss Defined as:

[0065] ;

[0066] in, Represents any traffic analysis area Distribution fitting loss; Traffic analysis area The true probability distribution label for the k-th function; This represents the traffic analysis area predicted by the model. Predicted probability for the k-th function; Indicates the total number of preset function categories; This represents a very small constant set to prevent the denominator from being zero or to prevent abnormal logarithmic calculations.

[0067] The overall loss function over the entire traffic analysis area for:

[0068] ;

[0069] in, Indicates the total number of traffic analysis zones; Represents any traffic analysis area Distribution fitting loss;

[0070] To enhance spatial continuity, a smoothing constraint term based on adjacency relationships is constructed. Let E be the set of adjacency edges in the traffic analysis area, then the smoothing loss function is... Defined as:

[0071] ;

[0072] in, Traffic analysis area The predicted probability distribution vector; Traffic analysis area Adjacent traffic analysis area The predicted probability distribution vector; Traffic analysis area Traffic Analysis Area There is a spatial adjacency relationship between them; This represents the squared L2 norm between the predicted probability distribution vectors of two adjacent traffic analysis zones;

[0073] The total loss function is obtained by synthesis. Specifically:

[0074] + ;

[0075] in, Indicates the smoothing constraint weight parameters; Represents the smoothing loss function; This represents the overall loss function.

[0076] As a preferred option, the specific implementation details in S10 are as follows:

[0077] For traffic analysis area Predicted probability distribution vector The category with the highest probability is determined as the dominant functional category. Then calculate the distribution entropy. The traffic analysis area is determined to be a single-function area, a dual-function mixed area, or a multi-function mixed area.

[0078] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0079] (1) This invention achieves refined quantification of the mixed state of urban functional areas, significantly improving the information density and application value of the identification results. This invention aims to output a continuous probability distribution vector, rather than a single category label. This quantification result not only answers what the area is, but also answers how much each type of function accounts for; providing richer and more instructive data support for refined urban management (such as mixed land use planning and on-demand allocation of public resources) than traditional classification results.

[0080] (2) This invention significantly improves the recognition accuracy of homogeneous heterogeneous complex regions and enhances the model's anti-interference capability. By introducing temporal features of crowd activity, it utilizes the unique patterns of different functional areas in the time dimension to assist in judgment. According to theoretical analysis of the model, compared with single-modal methods, the fusion model of this invention is expected to achieve an improvement of about 5% in the cosine similarity index and significantly reduce the error of Chebyshev distance. This proves that this method can effectively utilize the complementarity of multimodal data and can greatly improve the accuracy and confidence of recognition.

[0081] (3) This invention constructs a scientific labeling system that balances functional scarcity and dynamic activity, eliminating the bias of traditional training data. This invention innovatively combines the TF-IDF algorithm with dynamic population heatmaps to generate the training labels required by the model. Through the TF-IDF algorithm, it is expected to effectively increase the weights of scarce and iconic functions within a region, solving the bias of underestimating key functions in traditional label generation. By introducing dynamic population heatmaps, it is expected that the generated labels can more realistically reflect the actual functional intensity undertaken by the region in urban operations, ensuring the objectivity and reliability of the model training benchmark from the source.

[0082] (4) This invention optimizes the convergence objective of model training, which can improve the model's fitting efficiency to complex distributions. This invention uses KL divergence or JS divergence, which are specifically used to measure the similarity of probability distributions, as the loss function. This loss function can keenly capture the subtle differences in morphology between the predicted distribution and the true distribution, and is expected to effectively guide the model parameters to optimize towards fitting the proportional structure rather than class determination. According to theoretical derivation, this method is expected to achieve a low value of about 0.264 in the L1 distance index, which measures the distribution difference, indicating that the predicted results generated by the model are expected to achieve a high degree of consistency with the real situation in terms of numerical distribution.

[0083] (5) This invention achieves fully automated end-to-end processing, significantly reducing labor costs and improving operational efficiency. Based on a mature deep learning architecture, this invention is expected to enable automated, end-to-end processing of various heterogeneous data (images, time series, POIs). It can achieve rapid and automated identification and monitoring of large urban areas, and once the system is deployed, it can respond to newly input urban data in seconds. This will greatly save the cost and time of manual surveys and make dynamic monitoring of urban functional evolution (such as monthly or quarterly updates) possible, significantly improving the efficiency of urban planning and management. Attached Figure Description

[0084] Figure 1 This is the overall flowchart of the method of the present invention;

[0085] Figure 2 This is the core network structure diagram of the present invention;

[0086] Figure 3 This is a schematic diagram of the data overview of the research area of ​​this invention. Detailed Implementation

[0087] The present invention will be further illustrated below with reference to specific embodiments. These embodiments are implemented based on the technical solutions of the present invention, and it should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0088] This embodiment provides a method for probabilistic quantification identification of mixed states in urban functional zones by fusing multi-source heterogeneous data. It utilizes multimodal deep learning technology to integrate physical spatial perception (street view, POI) and social perception (pedestrian flow time series) data to perform refined prediction of the mixed functional probability distribution of urban functional zones (UFZ). This invention provides an end-to-end deep learning framework comprising four main modules: a data input module, a data preprocessing module, a probability distribution label generator, a multimodal feature extractor, and a distribution predictor.

[0089] In this embodiment, the traffic analysis area refers to the urban spatial analysis unit constructed under the constraints of road boundaries, which is used to carry multi-source data such as points of interest, social text, trajectories, and physical scenes.

[0090] Single functional area: refers to a region in which a certain traffic analysis area is significantly dominated by a certain category in the predicted probability distribution.

[0091] Dual-function mixed zone: refers to a traffic analysis area that is mainly composed of two types of functions.

[0092] Multifunctional mixed zone: refers to a traffic analysis area where the functional categories are relatively evenly distributed and it is difficult to summarize it by a single or two dominant functions.

[0093] True probability distribution labels: These are the functional category distribution labels used to supervise model training, reflecting the intensity of traffic analysis areas in each preset functional category.

[0094] Distribution predictor: refers to a neural network module used to output the probability distribution of traffic analysis areas across various functional categories.

[0095] This invention uses a traffic analysis zone as the basic spatial unit and achieves probabilistic quantitative identification of urban functional zones by integrating point-of-interest data, social text data, crowd activity data, and physical scene data. The overall process is as follows: Figure 1 As shown, it includes: traffic analysis area construction, multi-source data acquisition, multi-source data preprocessing, real probability distribution label generation, multi-branch feature extraction, feature fusion and distribution prediction, mixed state determination and result output.

[0096] Specifically, the following steps are included:

[0097] S1. Traffic Analysis Area Construction: Obtain basic geographic data of the study area, including administrative boundaries, road networks, water systems, and buildings;

[0098] Highways, arterial roads, secondary arterial roads, branch roads, and connecting roads are extracted. Different buffer widths are set according to the road level to buffer the road vectors and superimpose them with the boundary of the study area to form a road constraint segmentation surface.

[0099] The road constraint segmentation surface is rasterized and vectorized to obtain candidate spatial cells. Fragmented cells with areas below the threshold or abnormal shapes are deleted to obtain the traffic analysis area set. Specifically:

[0100]

[0101] in, Let N represent the i-th traffic analysis zone, and N represent the total number of traffic analysis zones.

[0102] To facilitate subsequent multi-source data fusion, the geometric center of each traffic analysis area is denoted as... The boundary polygon is The area is .

[0103] If the area of ​​a traffic analysis zone is smaller than a preset threshold Then, the traffic analysis zone is merged with its adjacent traffic analysis zone that has the longest boundary contact length, until the following condition is met: .

[0104] By using the above methods, the boundary of the traffic analysis area can be made to fit the road and natural dividing elements as closely as possible, thereby improving the consistency between the spatial analysis unit and the actual functional boundary.

[0105] In this embodiment, firstly, main roads, secondary roads, tertiary roads, arterial roads, expressways and various connecting roads are extracted, and road buffer zones of different widths are constructed according to road level; then, the road buffer zones are merged with the main urban area boundary of the study area to form a road constraint segmentation surface; then, through steps such as merging road network buffer zones, vector to raster conversion, raster processing and vectorization, the basic research unit TAZ is generated.

[0106] Compared to regular grids, TAZs constructed based on road boundaries more closely resemble the actual form of urban functional zones. Roads, waterways, and administrative boundaries often play a clear separating role in urban space, enabling the division of residential, commercial, industrial, and public service functions into spatial units that better align with actual operational logic. The method proposed in this application can establish a spatial carrier suitable for multi-source data overlay and functional probability identification under the constraints of urban road networks.

[0107] S2. Multi-source data acquisition: Acquire multi-source data corresponding to the traffic analysis area, such as... Figure 3 As shown, the multi-source data includes:

[0108] (1) Points of interest data, denoted as a set: The collection includes name, category, latitude and longitude, and address information.

[0109] (2) Social text data, denoted as a set: The collection includes latitude and longitude coordinates, address information, check-in / comment text, and publication time.

[0110] (3) Crowd activity data, denoted as set The dataset includes latitude and longitude data of taxi trajectories, latitude and longitude data of ride-hailing vehicle trajectories, mobile phone signaling data, location heatmap data, or mobile application trajectory data.

[0111] (4) Physical scene data, denoted as set The collection includes street view imagery and remote sensing imagery.

[0112] The data used in the study area are shown in Table 1 below.

[0113] Table 1 Data used in the study area

[0114]

[0115] like Figure 3 As shown, this figure illustrates the spatial distribution of point-of-interest data, social text data, crowd activity data, and physical scene data in the main urban area of ​​the study region. The physical scene data is represented by online remote sensing images in the figure, but it also includes street view images.

[0116] S3. Data Preprocessing and Spatial Association: Perform coordinate unification, spatial pruning, outlier removal, duplicate record deletion and semantic cleaning on the multi-source data obtained in S2;

[0117] Multi-source data are projected onto the same coordinate system reference system.

[0118] Delete records located outside the boundaries of the study area.

[0119] Remove data with abnormal latitude and longitude coordinates, and remove data with abnormal time fields or duplicate sampling for trajectory data.

[0120] Denoise the social text by removing URLs, hashtags, @mentions, emojis, and stop words.

[0121] Sort the trajectory data by time and delete records with abnormal speed and spatial jumps.

[0122] Remove severely occluded, pure black images, and images with excessively low resolution from the image data.

[0123] The cleaned data is mapped to the traffic analysis area according to its spatial location. For any point of interest... If its spatial location is within the traffic analysis area Boundary polygon Inside, define the mapping function. for:

[0124]

[0125] belong Similarly, for any text record, trajectory record, and physical scene data, spatial inclusion relationships or nearest neighbor relationships are used to map them to the corresponding traffic analysis area, resulting in multi-source samples at the traffic analysis area level. Specifically:

[0126]

[0127] in, These respectively represent mappings to the traffic analysis area. The set of points of interest, the set of social texts, the set of crowd activity data, and the set of physical scene data.

[0128] S4. Generating labels based on true probability distribution;

[0129] S401. First, calculate the category frequency term. Let the i-th traffic analysis zone be... The k-th type of function is Then the statistical traffic analysis area Belongs to category Number of points of interest This constitutes the category counting vector. Specifically:

[0130]

[0131] The specific formula for calculating the category frequency term is as follows: .

[0132] in, A very small constant set to prevent the denominator from being zero or to prevent abnormal logarithmic calculations. Traffic analysis area The category frequency term for the k-th function; Traffic analysis area Belongs to category Number of points of interest; Traffic analysis area Belongs to category Number of points of interest. This indicates the total number of preset function categories.

[0133] S402, Secondly, let the included category be... The number of traffic analysis zones is The inverse document frequency term is then defined as:

[0134]

[0135] in, This represents the inverse document frequency term for the k-th type of function. This indicates the total number of traffic analysis zones.

[0136] S403. Next, consider the spatial clustering of similar functions within the traffic analysis area. Let the traffic analysis area be... center With center at and radius at, The set of domain units is denoted as Then the space radiation term is defined as:

[0137]

[0138] in, Traffic analysis area Spatial radiation term for the k-th type of function; Indicates the neighborhood traffic analysis area Belongs to category Number of points of interest; Indicates the neighborhood traffic analysis area Belongs to category Number of points of interest; Indicates the total number of preset function categories; A very small constant set to prevent the denominator from being zero or to prevent abnormal logarithmic calculations.

[0139] S404. Further, construct a dynamic activity item based on trajectory and check-in data. Let the traffic analysis area be [data missing] in the t-th time step. The number of trajectory records is The number of check-in text records is Then, standardization processing is performed separately, specifically as follows:

[0140]

[0141]

[0142] in, Traffic analysis area The standardized value of the trajectory activity at the t-th time step; Traffic analysis area The standardized value of the sign-in text activity at the t-th time step; , Let represent the mean and standard deviation of the trajectories of all traffic analysis areas at time step t, respectively. , Let represent the mean and standard deviation of the check-in text for all traffic analysis areas at time step t, respectively. A very small constant set to prevent the denominator from being zero or to prevent abnormal logarithmic calculations.

[0143] The two types of dynamic activities are weighted and fused to obtain the activity level at each time step. Specifically:

[0144]

[0145] in, The value range is from 0 to 1. The weighting coefficient for the fusion of trajectory activity data and check-in text activity data, with a value range of 0 ≤ ≤1. When The larger the value, the higher the weight of trajectory activity data in the dynamic activity calculation; when The smaller the time, the higher the weight of the check-in text activity data.

[0146] The traffic analysis area is obtained by averaging and normalizing over all time steps. Dynamic activity item Specifically:

[0147]

[0148] in, Traffic analysis area Normalized dynamic activity at the t-th time step, where T represents the total number of observation time steps.

[0149] S405, in addition, traffic analysis area The semantic support of text categories is calculated for the associated social text set D.

[0150] social text collection The text records are segmented, denoised, and have keywords extracted or semantically encoded. Specifically, TF-IDF, word vector models, topic classification models, BERT, or other pre-trained language models can be used as text encoding functions, denoted as . Then the semantic representation of the r-th social text. for:

[0151]

[0152] in, This indicates mapping to the traffic analysis area. The rth social text record.

[0153] Traffic analysis area The semantic representations of all social texts within the region are subjected to average pooling or weighted pooling to obtain the semantic vectors of the traffic analysis area-level texts, specifically:

[0154]

[0155] in, Traffic analysis area The text semantic vector. If the traffic analysis area If there are no social text records, then... Set it to a zero vector, or supplement it with the text semantic vector of the adjacent traffic analysis area. This indicates mapping to the traffic analysis area. The total number of social text records; This represents the semantic representation of the r-th social text.

[0156] Furthermore, for the k-th functional category Construct category semantic prototype text based on the category name, keyword set, or typical descriptive text of the functional category. And the category semantic prototype vector is obtained through the same text encoding function, specifically:

[0157]

[0158] in, The semantic description text representing the k-th functional category. This represents the category semantic prototype vector of the k-th functional category. This represents the text encoding function.

[0159] Based on traffic analysis zone-level text semantic vectors With category semantic prototype vector Semantic similarity between them, calculating traffic analysis areas Functional categories The text category semantic support is as follows:

[0160]

[0161] in, Traffic analysis area social text collection For the kth functional category The text category semantic support is denoted by cos(·), where cos(·) represents the cosine similarity. Through the above normalization process, The value range is from 0 to 1. The larger the area, the greater the traffic analysis area. The more social text content within the area supports the functional category of the traffic analysis area, the better. .

[0162] In another implementation, keyword matching scores, topic classification probabilities, or category probabilities output by a pre-trained language model can also be used as the basis for classification. And normalize it to the range of 0 to 1.

[0163] S406, Constructing a Traffic Analysis Area In functional categories Overall weight Specifically:

[0164]

[0165] in, , , This is the adjustment coefficient. Traffic analysis area The category frequency term for the k-th function; This represents the inverse document frequency term for the k-th type of functionality; Traffic analysis area Spatial radiation term for the k-th type of function; Traffic analysis area The dynamic activity level item; Traffic analysis area social text collection For the kth functional category Text category semantic support.

[0166] S407. Finally, normalize the comprehensive weights of each category to obtain the true probability distribution labels. Specifically:

[0167]

[0168] in, Traffic analysis area In functional categories Overall weighting; Traffic analysis area Weights on the m-th class of functionality; Indicates the total number of preset function categories; A very small constant set to prevent the denominator from being zero or to prevent abnormal logarithmic calculations.

[0169] Thus, the traffic analysis area is obtained. True probability distribution label vector Specifically:

[0170]

[0171] Through the above method, the present invention converts traditional single labels into probability distribution labels, which can more accurately express the mixed functional status of the traffic analysis area.

[0172] S5. Socio-semantic feature extraction: Extraction of interest point sets at the traffic analysis zone level. Social text collection Encode the socio-semantic features to obtain them.

[0173] S501, First, based on the traffic analysis area Construct a static feature vector based on the number of interest points in each category:

[0174]

[0175] in, This indicates mapping to the traffic analysis area. A set of points of interest Traffic analysis area The number of interest points for the k-th type of function.

[0176] S502, Then, map to the traffic analysis area social text collection Word segmentation, keyword extraction, and semantic encoding are performed. The text representation is obtained using TF-IDF, word vector models, or pre-trained language models, denoted as... The interest point vector and the text semantic vector are then concatenated to obtain:

[0177]

[0178] in, This indicates mapping to the traffic analysis area. A set of points of interest Traffic analysis area The text semantic vector; Indicates traffic analysis area The social semantic feature vector before fusion is obtained by concatenating the interest point count vector and the text semantic vector.

[0179] S503, then through a fully connected mapping layer, yields socio-semantic features with a unified dimension. :

[0180]

[0181] in, Indicates traffic analysis area The pre-fusion socio-semantic feature vector is obtained by concatenating the interest point count vector and the text semantic vector. This represents a nonlinear mapping function in the sociosemantic branch, used to map the pre-fusion sociosemantic feature vectors to sociosemantic features of a unified dimension.

[0182] In this embodiment, it should be noted that the text semantic vector This serves as a text representation at the traffic analysis zone level, which can be used for subsequent socio-semantic feature extraction. Furthermore, its similarity to the semantic prototypes of each functional category can be used to calculate the text category semantic support in S4. .

[0183] S6. Dynamic activity feature extraction;

[0184] S601. First, the dynamic activity of each traffic analysis area at multiple time steps is constructed into an activity time series. Specifically:

[0185]

[0186] in, Traffic analysis area The dynamic activity at time step t.

[0187] S602. Then, after standardizing and normalizing the activity time series, it is input into a recurrent neural network, a temporal convolutional network, or a Transformer temporal modeling network.

[0188] S603. Finally, in the preferred embodiment, an LSTM network is used to extract dynamic activity features, the recursive form of which can be expressed as:

[0189]

[0190] in, Indicates the first The hidden state at each time step. Traffic analysis area The dynamic activity level at time step t; Indicates the first The hidden state at each time step.

[0191] Take the hidden state at the last time step as the traffic analysis area. Dynamic activity characteristics Specifically:

[0192]

[0193] in, This represents the hidden state at the last time step.

[0194] This step allows us to extract the temporal activity patterns of residential areas, commercial areas, industrial areas, and public service areas at different times.

[0195] In this embodiment, by statistically analyzing trajectory records and check-in records according to time steps, the activity intensity sequence of each TAZ in multiple time periods can be obtained; after standardization, normalization and time series modeling, dynamic activity features for function recognition can be formed.

[0196] The technical significance of dynamic activity characteristics lies in the fact that different urban functional areas typically exhibit different temporal usage patterns. For example, residential areas may show obvious commuting tidal characteristics during the morning and evening, commercial and public service areas may be more active during the day or specific consumption periods, while industrial and office areas may show strong regularity on weekdays. The method in this application can extract temporal differences in urban functions from population activity data, which can be used to help distinguish areas with similar visual forms but different social functions.

[0197] S7. Physical scene feature extraction;

[0198] S701, For each traffic analysis zone Collect multiple street view images, or extract corresponding remote sensing images, building outlines, building heights, street cross-sectional elements, or real-world photographs. Establish a traffic analysis area. Corresponding physical scene data set for:

[0199]

[0200] in, This represents the m-th physical scene sample (image).

[0201] This embodiment applies to each sample Simultaneously perform visual features Image semantic features Extraction.

[0202] S702. Visual features of each physical scene sample are extracted using a convolutional neural network, visual Transformer, or visual base model. Spatial redundancy is removed using a global average pooling unit, and the visual feature vector of a single image is output. Specifically:

[0203]

[0204] in, Represents the visual encoding function. This represents the Global Average Pooling (GAP) operator.

[0205] S703. When extracting semantic features of physical scene features, cross-modal associations of the physical environment are established by textualizing and then vectorizing the image content. This includes the following sub-steps:

[0206] Step 1: Use a pre-trained image description model (such as a visual transformer based on an Encoder-Decoder architecture) to process the street scene image. Automatic identification is performed to recognize physical elements (such as buildings, vegetation, traffic signs, road surface materials, etc.) and generate descriptive text. Specifically:

[0207]

[0208] in, This represents the Image Captioning Function. In the field of deep learning, it is typically a multimodal model (such as a visual transformer based on an Encoder-Decoder architecture) that converts unstructured pixel-level image data into structured natural language text.

[0209] Step 2: To correct potential biases in automated annotation, combine the image's built-in category metadata. (Semantic enhancement is performed on the POI category and administrative division attributes associated with the shooting location, and the final descriptive phrases are formed through text building units.) Specifically:

[0210]

[0211] in, Indicates text-level concatenation or text construction operations; This represents descriptive text; This represents the semantically enhanced image description text. Representation with street view images The corresponding category metadata includes the POI category associated with the shooting point, road name, administrative division attribute, or surrounding functional category information.

[0212] Step 3: Use a pre-trained BERT model as the encoder to process the enhanced descriptive text. Semantic encoding is performed. The hidden state of the [CLS] bits of the BERT output layer is taken as the semantic feature vector of the image. This vector can capture the deep semantic connections of image content within the urban functional context, specifically:

[0213]

[0214] in, The semantic encoding unit employs a pre-trained deep bidirectional Transformer encoder as its core architecture. It enhances the descriptive text output by the text construction unit. A multi-head self-attention mechanism is used to extract global semantic features and output the corresponding semantic feature vector. This establishes cross-modal associations between the physical environment in the visual and semantic spaces.

[0215] S704. For multiple images or multiple scene data within the same traffic analysis area, the visual feature vectors will be... With semantic feature vector By concatenating or adding the samples, a single-sample comprehensive feature vector can be obtained. .

[0216] Aggregation is performed using average pooling to obtain the physical scene features at the traffic analysis zone level. Specifically:

[0217]

[0218] in, This represents the total number of physical scene samples within the i-th traffic analysis zone; Traffic analysis area The comprehensive feature vector of the m-th physical scene sample within the range; This represents the physical scene characteristics of the i-th traffic analysis zone.

[0219] Attention pooling can also be used to obtain the physical scene perception features at the traffic analysis zone level. Specifically:

[0220]

[0221]

[0222] in, Let be the attention weight for the m-th scene sample. This represents the total number of physical scene samples within the i-th traffic analysis zone. This represents the comprehensive feature vector of the m-th physical scene sample within the traffic analysis area.

[0223] S8. Feature fusion and distribution prediction;

[0224] Traffic analysis area socio-semantic features Dynamic activity characteristics and physical scene features The fusion is performed to obtain the fused feature vector. :

[0225] =[ ]

[0226] Subsequently, the feature vectors will be fused. Input distribution predictor. Assuming the distribution predictor includes a fully connected layer and a Softmax output layer, then the traffic analysis region... The predicted probability distribution vector is:

[0227]

[0228] in, Traffic analysis area The predicted probability distribution vector for each functional category; This is the output layer weight matrix. For bias terms, To fuse feature vectors; This represents the normalization exponential function, used to convert the unnormalized scores of the output layer into probability distributions for each functional category.

[0229] Traffic Analysis Area The predicted probability distribution vector is represented as:

[0230] =[ , ]

[0231] in, Traffic analysis area The predicted probability distribution vector for each functional category; K represents the total number of preset functional categories. This represents the traffic analysis area predicted by the model. The predicted probability for the k-th function satisfies:

[0232]

[0233] This step allows the multimodal information of the traffic analysis area to be uniformly mapped to a functional category probability distribution.

[0234] S9, Model Training;

[0235] The true probability distribution labels generated in S4 are used as supervision signals to train the distribution predictor in S8. The true probability distribution label vector... Represented as:

[0236] =[ , ]

[0237] The predicted probability distribution vector is represented as:

[0238] =[ , ]

[0239] in, Traffic analysis area The predicted probability distribution vector for each functional category; Traffic analysis area The true probability distribution label for the k-th function; This represents the traffic analysis area predicted by the model. The predicted probability for the k-th function.

[0240] In a preferred embodiment, KL divergence is used as the loss function. For any traffic analysis region... Its distribution fitting loss Defined as:

[0241]

[0242] in, Traffic analysis area True probability distribution labels for k-class functions; This represents the traffic analysis area predicted by the model. Predicted probability for the k-th function; Indicates the total number of preset function categories; A very small constant set to prevent the denominator from being zero or to prevent abnormal logarithmic calculations.

[0243] The overall loss function over the entire traffic analysis area for:

[0244]

[0245] in, Indicates the total number of traffic analysis zones; Represents any traffic analysis area Distribution fitting loss .

[0246] Furthermore, to enhance spatial continuity, a smoothing constraint term based on adjacency relationships is constructed. Let E be the set of adjacency edges in the traffic analysis zone, then the smoothing loss function... Defined as:

[0247]

[0248] in, Traffic analysis area The predicted probability distribution vector; Traffic analysis area Adjacent traffic analysis area The predicted probability distribution vector; Traffic analysis area Traffic Analysis Area There is a spatial adjacency relationship between them; The L2 norm squared represents the difference between the predicted probability distribution vectors of two adjacent traffic analysis zones, and is used to measure the difference in functional probability distributions between adjacent traffic analysis zones.

[0249] The total loss function is obtained by synthesis. :

[0250] +

[0251] in, To smooth out the constraint weight parameters; For smoothing loss function; This is the overall loss function.

[0252] The optimal combination of parameters for the distribution predictor is obtained by minimizing the total loss function.

[0253] S10. Mixed state determination and result output;

[0254] For traffic analysis area Predicted probability distribution vector The category with the highest probability is determined as the dominant functional category. Specifically:

[0255]

[0256] in, This represents the traffic analysis area predicted by the model. Predicted probability for the k-th function; Function category index, with values ​​1, 2, ..., K.

[0257] To determine whether the traffic analysis area belongs to a single-function zone, a dual-function mixed zone, or a multi-function mixed zone, its distribution entropy is further calculated. Specifically:

[0258]

[0259] in, This represents the traffic analysis area predicted by the model. Predicted probability for the k-th function; Indicates the total number of preset function categories; A very small constant set to prevent the denominator from being zero or to prevent abnormal logarithmic calculations.

[0260] If the following conditions are met: and Then determine the traffic analysis area. It is a single-function area. , These represent the dominant function probability threshold, used to determine whether a certain function category has obvious dominance, and the distribution entropy threshold, used to determine whether the function distribution is sufficiently concentrated.

[0261] set up Let these represent the first and second largest probabilities in the predicted probability distribution, respectively, if they satisfy: and Then determine the traffic analysis area. It is a dual-function hybrid area. , The thresholds representing the sum of probabilities of the first two major functional categories are used to determine whether the traffic analysis area is mainly composed of two types of functions, and the threshold representing the probability of the second major functional category is used to determine whether the second type of function has a sufficient contribution.

[0262] In other cases, determine the traffic analysis area. It is a multi-functional hybrid zone.

[0263] The final output includes the probability distribution of functional categories, dominant functional categories, mixed functional labels, and their corresponding thematic maps for each traffic analysis area, including: urban functional area probability distribution map, dominant functional zoning map, and mixed functional intensity classification map.

[0264] like Figure 1 , Figure 2 As shown, firstly, basic geographic data of the administrative boundaries, road network, water system, and buildings of the study area are acquired, and a traffic analysis zone set (TAZ) is constructed according to S1. In this embodiment, fragmented units with areas below a threshold are merged to form a reasonably regular traffic analysis zone set.

[0265] In this embodiment, this application uses XX City as the research area. The main urban area of ​​XX City, as the most economically active region, possesses abundant POI data, taxi trajectory data, and Weibo check-in data, laying a solid foundation for multi-source data fusion analysis. Furthermore, its data density and quality are higher than those of economically underdeveloped non-urban built-up areas. Therefore, this application selects the main urban area of ​​XX City, including Districts XX1, XX2, XX3, XX4, XX5, and XX6, as the research object.

[0266] This example divides the urban area into unified spatial analysis units, namely Transportation Analysis Zones (TAZs). For each spatial analysis unit, the following three types of heterogeneous data are obtained:

[0267] AOI / POI data includes functional categories (such as commercial, residential, industrial, public services, etc.), geographic location information, and spatial distribution, which are used to construct a functional attribute matrix.

[0268] Population activity data: including mobile phone signaling data, LBS service heat maps and mobile APP trajectory data, used to statistically analyze the intensity of population activity in the study area over a continuous time period (e.g., 50 consecutive time steps, covering weekdays and weekends) to form a time series.

[0269] Street view image data: such as Baidu Street View or Gaode Street View, to obtain images / street views of the center point and its surroundings of each study area from multiple angles, as static feature inputs of the physical environment.

[0270] like Figure 1 As shown, the data undergoes preprocessing, including TAZ spatial unit partitioning, time series alignment, street view sampling and cleaning, and POI / AOI mapping. Subsequently, the preprocessed data is input into the true probability distribution label generation module. This module consists of a functional scarcity calculation unit, a dynamic activity calculation unit, a comprehensive weighted label generation unit, and a true probability distribution label output unit, used to generate true probability distribution label vectors. , as a soft label for supervised learning.

[0271] 1. Calculate Functional Scarcity: The functional scarcity calculation unit is used to calculate the category frequency item and the inverse document frequency item. For the i-th traffic analysis area... and k-class functions The category frequency term is defined as:

[0272]

[0273] in, Traffic analysis area The category frequency term for the k-th function; This indicates the category within the traffic analysis area. Number of points of interest; Traffic analysis area Belongs to category The number of points of interest; K represents the total number of preset function categories; A very small constant set to prevent the denominator from being zero or to prevent abnormal logarithmic calculations.

[0274] The inverse document frequency term is defined as:

[0275]

[0276] in, The inverse document frequency term represents the k-th function category; N represents the total number of traffic analysis zones; Indicates the inclusion of categories The number of traffic analysis zones.

[0277] 2. Calculate the traffic analysis zone-level dynamic activity level: The dynamic activity level calculation unit is used to calculate the traffic analysis zone-level dynamic activity level item. .

[0278]

[0279] in, Traffic analysis area The dynamic activity level item; Traffic analysis area Normalized dynamic activity at time step t; T represents the total number of observation time steps.

[0280] 3. Comprehensive Weighting: The comprehensive weighting label generation unit is used to generate a traffic analysis area by comprehensively considering category frequency, category scarcity, spatial radiation, dynamic activity, and text semantic support. In functional categories Overall weighting:

[0281]

[0282] in, Traffic analysis area In functional categories Overall weighting; Traffic analysis area Spatial radiation term for the k-th type of function; Traffic analysis area social text collection For the kth functional category Text category semantic support; , , This is the adjustment coefficient.

[0283] 4. Generate probability distribution vector: Finally, the true probability distribution label output unit normalizes the comprehensive weights of all functional categories to generate true probability distribution labels:

[0284]

[0285] in, Traffic analysis area In functional categories Overall weighting; Traffic analysis area Weights on the m-th class of functionality; Indicates the total number of preset function categories; A very small constant set to prevent the denominator from being zero or to prevent abnormal logarithmic calculations.

[0286] Thus, the traffic analysis area is obtained. True probability distribution label vector :

[0287]

[0288] in, Traffic analysis area The true probability distribution label for the k-th function. Through the above method, this invention transforms traditional single-function labels into probability distribution labels, enabling a quantitative expression of the mixed composition ratio of different functional categories within the traffic analysis area.

[0289] Furthermore, socio-semantic features are extracted according to S5. Extract dynamic activity features according to S6 Extract physical scene features according to S7 Based on S8, feature fusion and probability distribution prediction are performed to obtain the predicted probability distribution vector of the traffic analysis area. .

[0290] In this embodiment, the multimodal feature extractor adopts a dual-branch parallel structure, including a social perception branch and a physical perception branch, which respectively extract social perception features and physical perception features to ensure that features of different modalities are fully extracted. Specifically, as follows... Figure 2 As shown, it includes:

[0291] 1. Social perception branch:

[0292] Step 1: Static Function Vector Construction. Count the number of POIs of each category within the traffic analysis area and construct the feature vector of the points of interest.

[0293] Step 2, Text Semantic Encoding. Extract social media text associated with the region and extract text semantic vectors using a pre-trained language model (such as BERT).

[0294] Step 3: Feature Fusion. The result vectors generated in Steps 1 and 2 are concatenated and mapped through a fully connected layer to output a socio-semantic feature vector. .

[0295] Step 4: Extract social perception features (dynamic temporal branch) and output temporal features. This involves extracting temporal data on population activity in the region (e.g., 50...). 1D time series input to a stacked long short-term memory network (LSTM);

[0296] In this embodiment, a three-layer LSTM is configured, containing 128 hidden neurons.

[0297] LSTM is used to extract periodic and long-term dependent characteristics from time series data (such as weekday morning and evening rush hours and weekend pedestrian flow changes).

[0298] Step 5: Take the hidden state of the last LSTM layer at the last time step as the temporal feature vector. (For example dimension).

[0299] 2. Physical perception branch: Extract physical perception features (static spatial branch) and output physical perception features;

[0300] Step 1: Extract visual features and output visual feature vectors;

[0301] 1) Input multiple street view images of the input region into the ResNet-101 feature extraction unit;

[0302] A pre-trained deep residual network (ResNet-101) is used.

[0303] 2) Input the result of step 1 into the global average pooling unit, and output the visual feature vector through global average pooling;

[0304] Remove the fully connected layers at the end of the network and output the visual feature vector through global average pooling. (For example dimension).

[0305] Step 2: Extract semantic features and output semantic feature vectors;

[0306] 1) Based on the image description / keyword generation unit, a pre-trained image description model is used to automatically textify the street view images in the area and extract keywords or text segments that describe the physical environment;

[0307] 2) Based on text-based building blocks, combined with the image's built-in category metadata, descriptive text is formed;

[0308] 3) Based on the BERT semantic encoding unit, the BERT model is used as the encoder to perform semantic vectorization processing on the automatically generated descriptive text, and output semantic feature vectors. (For example (Dimension), thereby establishing cross-modal associations between vision and semantics.

[0309] Step 3: Extract visual feature vectors from all street view images within the area. and semantic feature vectors The input is averaged based on the region-level average pooling / aggregation unit, and average pooling is performed to form the final physical sensing features of the region.

[0310] This invention constructs a multimodal deep learning architecture that can effectively extract and fuse static visual features, text semantic features, and dynamic temporal features. It utilizes the complementarity between different modal data (for example, using the tidal pattern of pedestrian flow to help determine building attributes) to achieve high-precision functional recognition in complex urban scenarios.

[0311] like Figure 1 As shown, based on the feature fusion unit, the three channel features output by the multimodal feature extractor are fused.

[0312] 1. Multimodal Feature Fusion: A concatenation strategy is adopted to concatenate the feature vectors of the three channels, specifically as follows:

[0313]

[0314] in, This represents the fused vector. This represents a vector concatenation operation; This represents a dynamic temporal feature vector extracted from temporal data of crowd activities using LSTM or other temporal modeling networks. This represents a static visual feature vector extracted from street view images, remote sensing images, or other physical scene data through a convolutional neural network, a visual Transformer, or a basic visual model. It represents the text semantic feature vector extracted from POI / AOI functional category information, social text data, or image description text using TF-IDF, word vector model, BERT, or other semantic encoding models.

[0315] 2. Input the fused vector into the fully connected mapping unit; input the result data from the fully connected mapping unit into the Dropout unit; input the result data from the Dropout unit into the Softmax output layer;

[0316] fused vector (e.g., 2944) The 1-dimensional feature is input into a fully connected layer containing a Dropout layer, which enables deep interaction and dimensionality reduction of cross-modal features through nonlinear mapping.

[0317] 3. Input the fused features into the distribution predictor and output the probability distribution of the mixed state of urban functional areas;

[0318] The network ends with a Softmax activation function layer, which outputs a predicted probability distribution vector. The dimension of this vector is equal to the number of functional categories, K.

[0319] Subsequently, the model is trained using the S9 loss function, and the optimal parameters are obtained through iterative updates. For a given traffic analysis area, the model output may be:

[0320] =[0.52,0.28,0.03,0.10,0.05,0.02]

[0321] The six probability values ​​correspond to residential, commercial, industrial, public management and public services, science, education and culture, and leisure and entertainment, respectively. The highest probability is for the residential category, with a probability of 0.52. If the conditions for determining mixed-function zones are met simultaneously, it can be identified as a mixed residential-commercial area.

[0322] For example, for another traffic analysis area, the model output is:

[0323] =[0.83,0.06,0.01,0.05,0.03,0.02]

[0324] Since the highest probability is significantly higher than other categories and the distribution entropy is low, the traffic analysis area is determined to be a single residential functional area.

[0325] This embodiment demonstrates that the present invention can not only identify the dominant functional category of a traffic analysis area, but also quantify the proportion of its mixed functions, thereby meeting the needs of refined identification of urban functional areas, mixed land use assessment, and planning decision support.

[0326] like Figure 1 As shown, the predicted probability distribution vector True probability distribution label vector Physically perceived features are input into the distributed loss function;

[0327] A probability distribution similarity metric is used as the loss function. Specifically, the KL divergence metric is used to predict the probability distribution vector. Label vector with true probability distribution The differences between them. For any traffic analysis area Its KL divergence loss is defined as:

[0328]

[0329] in, Represents any traffic analysis area The distribution fitting loss; K represents the total number of preset function categories; Traffic analysis area The true probability distribution label for the k-th function; This represents the traffic analysis area predicted by the model. Predicted probability for the k-th function; A minimal constant is set to prevent the denominator from being zero or to prevent abnormal logarithmic calculations. The overall KL divergence loss over the entire N traffic analysis region is:

[0330]

[0331] in, This represents the overall loss function; N represents the total number of traffic analysis zones.

[0332] In this embodiment, JS divergence (Jensen-Shannon Divergence) or L1 distance can be selected as the loss function according to business needs, so as to guide the model parameters to converge towards fitting a refined proportional structure.

[0333] This invention selects a loss function mechanism suitable for label distribution learning, which can keenly perceive the subtle differences in morphology between the predicted distribution and the true distribution, and guide the model parameters to converge toward fitting a refined proportional structure, rather than simply pursuing classification accuracy.

[0334] Optimization objective: To minimize This allows for the optimization of model parameters, enabling the model to focus on the mixed proportion structure of urban functional areas rather than traditional classification tasks, thereby achieving accurate quantification of mixed functions.

[0335] Optimizer: The Adam optimizer is preferred, and an appropriate learning rate is set (e.g., ).

[0336] The method of this invention can be implemented in a distributed computing system or high-performance computing cluster that includes the following components:

[0337] Hardware environment: Equipped with a high-performance graphics processor (GPU, such as NVIDIA Tesla V100 / A100) and ample memory (RAM) to support the training and inference of deep learning models such as ResNet-101 and BERT.

[0338] Software environment: Based on the Python programming language, using mainstream deep learning frameworks (such as TensorFlow or PyTorch) for model building, training, and deployment. Geographic information processing can utilize tool libraries such as GeoPandas or PostGIS.

[0339] Proof of experimental accuracy and reliability of results: This embodiment provides statistical data for accuracy evaluation, as shown in Table 2 below.

[0340] Table 2 Accuracy Assessment Statistics

[0341]

[0342] Table 2 presents the accuracy evaluation statistics for the study area. The evaluation sample size was 200. The overall mean precision (OA) was 0.84, the mean precision (P) was 0.77, the mean recall (R) was 0.87, and the mean F-score was 0.82. These results indicate that within the selected sample range, the method proposed in this application can achieve good dominant function identification performance and has a high recall rate, demonstrating that the model has strong identification coverage ability for samples from different functional areas.

[0343] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for probabilistic quantitative identification of mixed states of urban functional areas by fusing multi-source heterogeneous data, characterized in that, Includes the following steps: S1. Traffic Analysis Area Construction: Obtain basic geographic data of the study area, including administrative boundaries, road networks, water systems, and buildings; S2. Multi-source data acquisition: Acquire multi-source data corresponding to the traffic analysis area; S3. Perform coordinate unification, spatial clipping, outlier removal, duplicate record deletion, and semantic cleaning on the multi-source data obtained in S2. S4. Generating labels based on true probability distribution; S5. Social semantic feature extraction: Encode the traffic analysis area-level interest point set and social text set to obtain social semantic features; S6. Dynamic Activity Feature Extraction: An LSTM network is used to extract dynamic activity features, and the hidden state at the last time step is taken as the traffic analysis area. Dynamic activity characteristics ; S7. Physical Scene Feature Extraction: For multiple images or multiple scene data within the same traffic analysis area, the visual feature vector and semantic feature vector are concatenated or added to obtain a single-sample comprehensive feature vector. ; S8. Feature Fusion and Distribution Prediction: The extracted features are fused, and the fused feature vector is input into the distribution predictor to obtain the prediction result; S9. Model Training: Using the true probability distribution labels generated in S4 as supervision signals, the distribution predictor in S8 is trained. The optimal parameter combination of the distribution predictor is obtained by minimizing the total loss function. S10, Mixed state determination and result output.

2. The method for probabilistic quantification identification of mixed states of urban functional areas by fusing multi-source heterogeneous data according to claim 1, characterized in that, In S1, the specific content is as follows: Highways, arterial roads, secondary arterial roads, local roads, and connecting roads are extracted. Different buffer widths are set according to road level, and road vectors are buffered and superimposed with the boundary of the study area to form road constraint segmentation surfaces. The road constraint segmentation surfaces are rasterized and vectorized to obtain candidate spatial cells. Fragmented cells with areas below the threshold or abnormal shapes are deleted to obtain the traffic analysis area set. Specifically: ； in, This represents the i-th traffic analysis zone. This indicates the total number of traffic analysis zones.

3. The method for probabilistic quantification identification of mixed states of urban functional areas by fusing multi-source heterogeneous data according to claim 1, characterized in that, In S2, multi-source data includes: Interest point data, denoted as a set. ;gather This includes name, category, latitude and longitude, and address information; Social text data, denoted as set ;gather This includes latitude and longitude, address information, check-in / review text, and posting time; Crowd activity data, denoted as set ;gather This includes latitude and longitude data of taxi trajectories, latitude and longitude data of ride-hailing vehicle trajectories, mobile phone signaling data, location heatmap data, or mobile application trajectory data; Physical scene data, denoted as set ;gather It includes a collection of street view images and a collection of remote sensing images.

4. The method for probabilistic quantification identification of mixed states of urban functional areas by fusing multi-source heterogeneous data according to claim 1, characterized in that, In S4, the specific implementation details are as follows: First, based on traffic analysis area Belongs to category Number of points of interest Calculate the category frequency item; Then, based on the number of traffic analysis zones Calculate the inverse document frequency term; Based on domain unit set Calculate the space radiation term; A dynamic activity level is constructed based on trajectory and check-in data; Further, construct text category semantic support items based on the social text set; Next, construct the traffic analysis area. In functional categories Overall weight ; Finally, the comprehensive weights of each category are normalized to obtain the true probability distribution labels.

5. The method for probabilistic quantification identification of mixed states of urban functional areas by fusing multi-source heterogeneous data according to claim 1, characterized in that, In S5, the specific implementation details are as follows: First, based on the traffic analysis area Construct a static functional vector based on the number of interest points in each category; Then, the mapping is applied to the traffic analysis area. The social text collection is segmented, keywords are extracted, and semantic encoding is performed, and the interest point vector and the text semantic vector are concatenated; Finally, after passing through a fully connected mapping layer, we obtain socio-semantic features with a unified dimension. .

6. The method for probabilistic quantification identification of mixed states of urban functional areas by fusing multi-source heterogeneous data according to claim 1, characterized in that, In S6, the specific implementation details are as follows: First, the dynamic activity of each traffic analysis area at multiple time steps is constructed into an activity time series. ; Then, after standardizing and normalizing the activity time series, it is input into a recurrent neural network, a temporal convolutional network, or a Transformer temporal modeling network. Finally, an LSTM network is used to extract dynamic activity features.

7. The method for probabilistic quantification identification of mixed states of urban functional areas by fusing multi-source heterogeneous data according to claim 1, characterized in that, In S7, the specific implementation details are as follows: For each traffic analysis zone Multiple street view images were collected, and a traffic analysis area was set up. Corresponding physical scene data set ; Visual features of each physical scene sample are extracted using convolutional neural networks, visual Transformers, or basic visual models. Spatial redundancy is removed using global average pooling units, and the visual feature vector of a single image is output. ; Extract semantic features from physical scene features, and enhance the descriptive text output by the text construction unit. A multi-head self-attention mechanism is used to extract global semantic features and output the corresponding semantic feature vector. This establishes cross-modal associations between the physical environment in the visual and semantic spaces; For multiple images or multiple scene data within the same traffic analysis area, visual feature vectors will be used. With semantic feature vector By concatenating or adding the samples, a single-sample comprehensive feature vector can be obtained. .

8. The method for probabilistic quantification identification of mixed states of urban functional areas by fusing multi-source heterogeneous data according to claim 1, characterized in that, In S8, the specific implementation details are as follows: Traffic analysis area socio-semantic features Dynamic activity characteristics and physical scene features The fusion is performed to obtain the fused feature vector. : =[ ]； Subsequently, the feature vectors will be fused. Input distribution predictor; assuming the distribution predictor includes a fully connected layer and a Softmax output layer, then the traffic analysis area... The predicted probability distribution vector is: ； in, Traffic analysis area The predicted probability distribution vector for each functional category; This represents the output layer weight matrix. Indicates the bias term. Represents the fused feature vector; This represents the normalized exponential function.

9. The method for probabilistic quantification identification of mixed states of urban functional areas by fusing multi-source heterogeneous data according to claim 1, characterized in that, In S9, the specific implementation details are as follows: The true probability distribution labels generated in S4 are used as supervision signals to train the distribution predictor in S8; KL divergence is used as the loss function; for any traffic analysis area Its distribution fitting loss Defined as: ； in, Represents any traffic analysis area Distribution fitting loss; Traffic analysis area The true probability distribution label for the k-th function; This represents the traffic analysis area predicted by the model. Predicted probability for the k-th function; Indicates the total number of preset function categories; This represents a very small constant set to prevent the denominator from being zero or to prevent abnormal logarithmic calculations. The overall loss function over the entire traffic analysis area for: ； in, Indicates the total number of traffic analysis zones; Represents any traffic analysis area Distribution fitting loss; To enhance spatial continuity, a smoothing constraint term based on adjacency relationships is constructed. Let E be the set of adjacency edges in the traffic analysis area, then the smoothing loss function is... Defined as: ； in, Traffic analysis area The predicted probability distribution vector; Traffic analysis area Adjacent traffic analysis area The predicted probability distribution vector; Traffic analysis area Traffic Analysis Area There is a spatial adjacency relationship between them; This represents the squared L2 norm between the predicted probability distribution vectors of two adjacent traffic analysis zones; The total loss function is obtained by synthesis. Specifically: + ； in, Indicates the smoothing constraint weight parameters; Represents the smoothing loss function; This represents the overall loss function.

10. The method for probabilistic quantification identification of mixed states of urban functional areas by fusing multi-source heterogeneous data according to claim 1, characterized in that, In S10, the specific implementation details are as follows: For traffic analysis area Predicted probability distribution vector The category with the highest probability is determined as the dominant functional category. Then calculate the distribution entropy. The traffic analysis area is determined to be a single-function area, a dual-function mixed area, or a multi-function mixed area.

Images