Urban land utilization information extraction method based on mobile phone signaling data

Abstract

The invention discloses an urban land utilization information extraction method based on mobile phone signaling data, and particularly relates to the technical field of traffic planning and management, and the method comprises the following steps: S1, a user sample expansion coefficient, S2, sample expansion data calculation, S3, the recognition of a residence place, a working place and a travel mode of a user, S4, the analysis and prediction of employment density, s5, calculating the residential population density of the region; S6, calculating a joint density vector; and S7, predicting the utilization form of the mixed land. According to the method, the accuracy and efficiency of estimation are improved, powerful data support is provided for decision making of urban land utilization planning, meanwhile, the method can help to identify low-efficiency utilized land and optimize the layout of a traffic network and public service facilities, data support is further provided for urban planners, the accuracy and efficiency of estimation are improved, and the method is suitable for popularization and application. Efficient utilization of urban land resources and optimization and adjustment of functional layout are facilitated, and reasonable allocation of urban resources and balance of urban development are facilitated.

Description

Technical Field

[0001] This invention relates to the field of transportation planning and management technology, and in particular to a method for extracting urban land use information based on mobile phone signaling data. Background Technology

[0002] Against the backdrop of rapid economic growth and urbanization, exploring diversified land development models is crucial for enhancing economic vitality and improving the efficiency of land resource utilization. This diversified land development model is a key strategy for addressing urban land scarcity, as it not only rationally utilizes limited land resources but also maximizes their economic and social benefits, keeping pace with urbanization and meeting residents' demands for a high-quality living environment. In light of this, local governments have formulated and implemented diverse land development policies based on their specific circumstances. To enhance the scientific rigor and effectiveness of land planning decisions, forecasting future urban land use can improve decision-makers' understanding of future development and reduce uncertainty. This has a positive impact on alleviating urban housing problems, enhancing urban economic vitality, and promoting the long-term healthy development of cities. Through such forward-looking planning, it is possible to ensure that urban development is coordinated with the sustainable use of land resources, laying a solid foundation for the future development of cities. Summary of the Invention

[0003] The main objective of this invention is to provide a method for extracting urban land use information based on mobile phone signaling data, which can effectively solve the problem of not being able to accurately realize the mixed land use patterns in the research area.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] A method for extracting urban land use information based on mobile phone signaling data includes the following steps:

[0006] S1. User expansion coefficient: The expansion coefficient is calculated through a decision tree model, including cleaning user location and call data, extracting the average and maximum call duration and location center point features, filtering relevant features, training the decision tree model, and outputting the expansion coefficient.

[0007] S2. Extrapolation of expanded data: Extract core user residency information based on the expansion coefficient and extrapolate the full dataset;

[0008] S3. Identify users' residential and workplace locations and travel patterns: Analyze nighttime location data to identify the residential coordinate set, analyze daytime location data to identify the workplace coordinate set, and statistically analyze travel frequency to generate a commuting frequency matrix.

[0009] S4. Analysis and prediction of employment density: Employment types are classified by occupation, and employment density values ​​are output through a Poisson regression model;

[0010] S5. Estimating the regional population density: Based on the residential coordinate set data, the Gaussian kernel function is used to perform kernel density estimation, and the population density value is calculated by combining the regional area.

[0011] S6. Calculate the joint density vector: Based on the standardized population density and employment density values, generate the joint density vector through K-means clustering;

[0012] S7. Predict mixed land use patterns: Employ a multi-task learning model to simultaneously predict population density and employment density values, and output land use pattern classifications.

[0013] Preferably, in step S1, the user location and call data are provided by the operator. Cleaning the user location and call data specifically involves: cleaning the user location, call time and call duration data provided by the operator, extracting the average and maximum call duration and location center point features, filtering feature subsets through correlation analysis, training a decision tree model and calculating the root mean square error, and outputting the expansion coefficient.

[0014] The formula for calculating the mean square error is as follows:

[0015]

[0016] Where y i This is the actual value. This is the predicted value, and n is the number of samples;

[0017] The formula for calculating the root mean square error is as follows:

[0018]

[0019] Where y i This is the actual value. is the predicted value, and n is the number of samples.

[0020] Preferably, in step S2, the core user set is generated by extracting the residency records of individual users for more than 10 consecutive days, and the full dataset is extrapolated by applying the expansion coefficient.

[0021] Preferably, in step S3, the nighttime location dwell frequency is calculated to generate a residential coordinate set, the daytime location dwell frequency is calculated to generate a workplace coordinate set, and the number of moves from the residential location to the workplace is counted to generate a commuting frequency matrix.

[0022] Preferably, in step S4, seven occupational types are divided according to occupational categories, and the employment density value is output by Poisson regression maximum likelihood estimation using the level of regional economic activity as the explanatory variable.

[0023] Assume the response variable Y follows a Poisson distribution:

[0024]

[0025] Where λ is the expected number of employed persons (mean) in the given region, and y is the observed number of employed persons;

[0026] The number of employed persons is used as the response variable, and explanatory variables that may affect employment (including but not limited to the level of regional economic activity, transportation convenience, and education level) are used as predictor variables.

[0027] In Poisson regression, it is assumed that λ depends on a set of explanatory variables X:

[0028] λ=e (Xβ)

[0029] Where X is the vector of explanatory variables and β is the vector of regression coefficients, the model can be written as:

[0030] log(λ) = Xβ

[0031] The β coefficient is determined by maximum likelihood estimation, and the goal of the model is to find the β value that is most likely to be the observed data.

[0032] Suppose there are n regions, and the number of employed people in each region is y. i The corresponding explanatory variable is x. i The Poisson regression model can be written as:

[0033] log(λ i )=x i β

[0034] Where, λ i x is the expected number of employees in the i-th region. i is the vector of explanatory variables for the i-th region, including three factors: level of economic activity, transportation convenience, and education level; β is the vector of regression coefficients.

[0035] The parameter β of a Poisson regression model is usually determined by maximum likelihood estimation (MLE), where the likelihood function L(β) is the sum of all observed data y. i The product of probabilities:

[0036]

[0037] Wherein, the log-likelihood function Calculated using the following formula:

[0038]

[0039] Because log(λ) i )=x i β, then λ i Replace with

[0040]

[0041] To find the value of β that maximizes the log-likelihood function, for Take the derivative of β and set it to zero.

[0042]

[0043] Solve the equation and obtain the result;

[0044] The trained model is used to predict the number of employed persons on a dataset (which can be a training set or a test set). The predicted number of employed persons for each observation is obtained, and the sum of squared residuals (RSS) of the predicted number of employed persons for each observation is calculated. For each observation, the difference between the actual number of employed persons and the predicted number of employed persons is calculated, and these differences are squared and summed. The formula is as follows:

[0045]

[0046] To calculate the Total Sum of Squares (TSS), calculate the difference between actual employment and average employment, then square these differences and sum them. The formula is:

[0047]

[0048] Calculate the goodness of fit R 2 Using the RSS and TSS values ​​calculated above, substitute them into R 2 The formula for calculating the goodness of fit is as follows:

[0049]

[0050] Analysis of R 2 The value is used to evaluate the goodness of fit of the model, R. 2 The closer the value is to 1, the more variability the model explains and the better the fit.

[0051] Preferably, in step S5, the bandwidth parameter is calculated using the Silverman rule, kernel density estimation is performed based on the residential coordinate set data, the kernel density estimate is output, and the population density value is calculated in combination with the area of ​​the region.

[0052] When using kernel density to estimate the population density of a region, it is important to select an appropriate kernel function. The kernel function K(u) is a symmetric, smooth function with an area of ​​1. Commonly used kernel functions include, but are not limited to, Gaussian kernel, triangular kernel, and uniform kernel. Taking the Gaussian kernel as an example, its formula is as follows:

[0053]

[0054] Let the bandwidth be h. Bandwidth h is a key parameter of KDE, which determines the width of the kernel function and thus affects the smoothness of the estimation. The Silverman rule is used to select the bandwidth, as shown in the following formula:

[0055]

[0056] Where σ is the sample standard deviation, IQR is the interquartile range, and n is the sample size;

[0057] Given a dataset X = {x1, x2, ..., x...} n The kernel density estimate can be expressed as:

[0058]

[0059] in, It is the kernel function scaled according to bandwidth h;

[0060] Using kernel density estimates, the population density of each region can be calculated. For each region, first calculate the sum of the kernel density estimates of all settlements within that region, then divide by the area of ​​the region to obtain the population density of that region, as shown in the formula below:

[0061]

[0062] Where, x i It refers to the geographical location of the residence. It is in x i The estimated kernel density at that location.

[0063] Preferably, in step S6, a standardized feature vector is generated based on the standardized population density value and employment density value, and the cluster centers are calculated to generate a joint density vector by minimizing the sum of squared distances through K-means clustering.

[0064] The standardized formula above is:

[0065]

[0066] Where x is the original feature value, μ is the mean, and σ is the standard deviation;

[0067] When using the K-means algorithm to group regions, the objective function of K-means is to minimize the sum of squared distances from each point to its cluster center:

[0068]

[0069] Where k is the number of clusters, S i It is the set of points in the i-th cluster, μ i It is the center of the i-th cluster;

[0070] Calculate the population and employment density of each cluster center as a joint density vector;

[0071] Let μ i =(μ i1 ,μ i2 ) represents the center of the i-th cluster, where μ i1 It is population density, μ i2 It is employment density;

[0072] The detailed algorithm for the joint density vector is as follows:

[0073] First, population and employment data for each region are collected, and this data is used as a feature vector, as shown in the following formula:

[0074] x = (p, e)

[0075] Where p is population density and e is employment density;

[0076] Standardize the population density p and employment density e:

[0077]

[0078] The standardized feature vector is:

[0079] x ' =(p ' ,e ' )

[0080] Determine the number of clusters, perform K-means clustering, initialize k cluster centers, and assign the feature vector x of each region to each cluster. ' Assign the cluster to the nearest cluster center, update the cluster center to the average value of the region feature vector assigned to it, and repeat the above steps until the cluster center no longer changes significantly.

[0081] Calculate the joint density vector, and for each cluster i, calculate the cluster center μ. i =(μ i1 ,μ i2 ),in:

[0082]

[0083] Where |S i | represents the number of regions in the i-th cluster, and μ represents the number of cluster centers. i This represents the joint density vector of the regions in the cluster.

[0084] Preferably, in step S7, a multi-task learning model sharing a lower-level network is constructed, and task layer functions are defined;

[0085] In the process of constructing the above multi-task learning model: firstly, the underlying network is shared, and then the task-specific layers are separated. The model can be represented as:

[0086]

[0087] Where x is the input feature vector;

[0088] Secondly, the shared network portion can be represented as a series of shared layers h:

[0089] h = H(x)

[0090] Where H is a function of the shared network, and h is the shared feature representation;

[0091] Each task has its own output layer, which can be represented as:

[0092] f1(x) = W1h + b1

[0093] f2(x) = W2h + b2

[0094] f3(x) = W3h + b3

[0095] Where W1, W2, and W3 are task-specific weight matrices, and b1, b2, and b3 are task-specific bias vectors;

[0096] The loss function for multi-task learning is a weighted sum of the losses for each task:

[0097]

[0098] in It is a task-specific loss function, where y1, y2, and y3 are the corresponding true labels, and λ1, λ2, and λ3 are the weights that balance the importance of different tasks.

[0099] Using gradient descent to train all tasks simultaneously, the gradient updates for the shared layer and task-specific layer for each task can be represented as:

[0100]

[0101] Where θ represents the model parameters and α is the learning rate;

[0102] Predict mixed land use patterns in the target area using a trained model:

[0103]

[0104] in is the prediction result, and x is the feature vector of the target region.

[0105] Preferably, the land use pattern classification is achieved by mapping a joint density vector to a preset functional type range.

[0106] Compared with the prior art, the present invention has the following beneficial effects:

[0107] 1. This invention not only improves the accuracy and efficiency of estimation but also provides strong data support for urban land use planning decisions. For example, in the development of new urban areas, planners can use this method to accurately predict the distribution of future population and employment density, thereby rationally planning the location and scale of residential, commercial, and industrial areas, avoiding resource waste and functional imbalances. Simultaneously, in urban renewal, this method can help identify inefficiently used land, optimize the layout of transportation networks and public service facilities, and improve residents' quality of life. Furthermore, in responding to public health emergencies, this method can quickly assess population flow and aggregation, providing a scientific basis for emergency resource allocation and the formulation of prevention and control measures. In summary, this invention can provide data support for urban planners, improve the accuracy and efficiency of estimation, facilitate the efficient use of urban land resources and the optimized adjustment of functional layouts, and contribute to the rational allocation of urban resources and balanced urban development.

[0108] 2. This invention systematically solves the problem of sample bias in mobile signaling data by calculating user expansion coefficients and residency information, and transforms sparse individual data into a reliable profile of the entire population's activities. By combining residential / workplace identification with travel pattern analysis, this approach not only accurately depicts the spatial behavior patterns of urban residents (such as commuting routes and peak hours) but also provides a behavioral basis for subsequent density calculations. This integration of multi-dimensional activity data overcomes the lag of traditional static census data, enabling real-time capture of the dynamic distribution and mobility characteristics of the urban population. Furthermore, by using employment density prediction and residential population density estimation, the urban spatial carrying capacity is quantified from the perspectives of economic activity and housing demand, respectively. These two methods form a joint density vector through standardization and clustering, revealing the intrinsic correlation of regional functions (such as the job-housing balance). Finally, a multi-task learning model utilizes the synergistic effect of the aforementioned density vectors to simultaneously optimize the prediction accuracy of population and employment density, and outputs a mixed land use pattern classification accordingly. This progressive analysis chain of "behavior-density-function" not only significantly improves the spatial resolution and timeliness of land use identification but also provides decision support for urban planning through a data-driven modeling approach that balances scientific rigor (such as the statistical rigor of Poisson regression) and practicality (such as the interpretability of K-means clustering). Attached Figure Description

[0109] Figure 1 This is a schematic diagram illustrating the estimation of mixed land use patterns according to the present invention;

[0110] Figure 2This is a schematic diagram of the process of the present invention. Detailed Implementation

[0111] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0112] like Figure 1 and Figure 2 As shown, this case provides a method for extracting urban land use information based on mobile phone signaling data, including the following steps:

[0113] S1. User expansion coefficient: The expansion coefficient is calculated through a decision tree model, including cleaning user location and call data, extracting the average and maximum call duration and location center point features, filtering relevant features, training the decision tree model, and outputting the expansion coefficient.

[0114] S2. Extrapolation of expanded data: Extract core user residency information based on the expansion coefficient and extrapolate the full dataset;

[0115] S3. Identify users' residential and workplace locations and travel patterns: Analyze nighttime location data to identify the residential coordinate set, analyze daytime location data to identify the workplace coordinate set, and statistically analyze travel frequency to generate a commuting frequency matrix.

[0116] S4. Analysis and prediction of employment density: Employment types are classified by occupation, and employment density values ​​are output through a Poisson regression model;

[0117] S5. Estimating the regional population density: Based on the residential coordinate set data, the Gaussian kernel function is used to perform kernel density estimation, and the population density value is calculated by combining the regional area.

[0118] S6. Calculate the joint density vector: Based on the standardized population density and employment density values, generate the joint density vector through K-means clustering;

[0119] S7. Predict mixed land use patterns: Employ a multi-task learning model to simultaneously predict population density and employment density values, and output land use pattern classifications.

[0120] S1, User Sampling Coefficient, the specific steps are as follows:

[0121] S1a. First, collect data, obtaining user location information, call time, and call duration data from the operator. Then, clean the obtained data, removing missing values, outliers, and noisy data. From this data, extract useful features, such as the average and maximum call duration, and the center point of the user's location.

[0122] S1b. Use correlation analysis to select the feature most relevant to the expansion coefficient.

[0123] S1c, Use decision trees to train the model;

[0124] S1d. Use mean squared error or root mean square error to evaluate the model's performance. The formula for calculating mean squared error is:

[0125]

[0126] Where y i This is the actual value. This is the predicted value, and n is the number of samples;

[0127] The formula for calculating the root mean square error is:

[0128]

[0129] Where y i This is the actual value. This is the predicted value, and n is the number of samples;

[0130] The root mean square error is the square root of the mean square error, providing the magnitude of the error in the same units as the original data.

[0131] S1e: Use the trained model to predict new data and obtain the expansion coefficient.

[0132] S2. Calculate the expanded sample data. The specific steps are as follows:

[0133] S2a. Extract individual user registration information from the mobile signaling registration information database;

[0134] S2b: Identify and extract the residency information of core users, who are those users who have a continuous record of more than ten days in the study area.

[0135] S2c applies user expansion coefficients to the statistical data of these core users to obtain the full amount of expanded data.

[0136] S3. Identify the user's residence, workplace, and travel patterns. The detailed steps are as follows:

[0137] S3a. For residential location identification: By analyzing the user's geographic location data during the night and morning periods, identify the locations where the user appears most frequently. These locations are likely to be the user's residence because most people are resting at night and at home before going out in the morning.

[0138] For workplace identification: Analyze users' geographical location data during daytime periods to find the locations where users appear most frequently. These locations are likely to be users' workplaces because most people are at their workplaces during the daytime. In order to determine the common locations of each user in different time periods, we can calculate the frequency of stay at each location.

[0139] Let L i Let F be the number of times a user stays at location i, and T be the total number of time periods. Then the frequency of stay at location i is F. i The following formula can be used to calculate:

[0140]

[0141] S3b. Identify the place of residence and place of work. The specific analysis method is as follows: identify the most frequently occurring location as the place of residence or place of work. A threshold for the number of occurrences can be used to determine this.

[0142] Let k be the threshold, F res For the frequency of the place of residence, F work For the frequency of the working location; if F res If F ≥ k, then it is identified as a place of residence; if F work If the value is ≥k, then it is identified as the workplace;

[0143] Analyze and identify travel frequency, travel time, and travel routes;

[0144] Trip frequency analysis: Statistics on how often users move from their identified residences to their workplaces help to understand users' travel habits, such as whether users travel between their residences and workplaces every day or have different travel patterns.

[0145] Let M ij Let P be the number of trips from residence i to workplace j, and T be the total number of observation days. Then the trip frequency P ij The following formula can be used to calculate:

[0146]

[0147] Travel time analysis: Analyzing the time distribution of users from their residence to their workplace helps identify peak travel periods and provides data support for traffic planning and management;

[0148] Let T ijt Let N be the travel time of a user from their residence i to their workplace j on day t, and N be the total number of travel days. Then the average travel time is... The following formula can be used to calculate:

[0149]

[0150] Travel route identification: By analyzing users' movement trajectories between their residence and workplace, common travel routes are identified;

[0151] This can help city planners optimize traffic routes, reduce congestion, and improve the efficiency of public transportation services;

[0152] Let P path P represents the number of trips along a specific route. total The frequency C of a specific route is the sum of the number of trips along all routes. path The following formula can be used to calculate:

[0153]

[0154] The above formula can be used to extract useful information from mobile signaling data and perform quantitative analysis to identify the user's place of residence, place of work and travel patterns.

[0155] These analyses can provide a better understanding of urban residents' activity patterns and offer data support for urban planning and traffic management.

[0156] S3c uses an efficient database to store and query large amounts of mobile signaling data, processes and analyzes the data, and uses GIS tools to visualize and analyze users' geographic location data, helping to identify residences and workplaces, as well as analyze travel routes.

[0157] S3d obtains residential and workplace data through existing census data, business registration data, or in cooperation with relevant government departments, compares the results obtained by the model with known residential and workplace data, and evaluates the accuracy of pattern recognition.

[0158] S4. Analyze and predict employment density, with the following specific steps:

[0159] S4a. Use statistical models to analyze and predict employment density, and use Poisson regression to estimate regional employment density.

[0160] First, employment is categorized into seven types: management occupations, professional and technical occupations, skilled occupations, service and sales occupations, health and medical occupations, creative arts occupations, and trade, logistics and commerce occupations. Then, a Poisson regression model is used to predict the employment density of each region, which can predict the relationship between the number of employed people and various influencing factors.

[0161] Assume the response variable Y follows a Poisson distribution:

[0162]

[0163] Where λ is the expected number of employed persons (mean) in the given region, and y is the observed number of employed persons;

[0164] S4b. Using the number of employed persons as the response variable, and explanatory variables that may affect employment (such as the level of regional economic activity, transportation convenience, education level, etc.) as predictor variables:

[0165] In Poisson regression, it is assumed that λ depends on a set of explanatory variables X:

[0166] λ=e (Xβ)

[0167] Where X is the vector of explanatory variables and β is the vector of regression coefficients, the model can be written as:

[0168] log(λ) = Xβ

[0169] The β coefficient is determined by maximum likelihood estimation, and the goal of the model is to find the β value that is most likely to be the observed data.

[0170] Suppose there are n regions, and the number of employed people in each region is y. i The corresponding explanatory variable is x. i The Poisson regression model can be written as:

[0171] log(λ i )=x i β

[0172] Where, λ i x represents the expected number of employees in the i-th region. i is the vector of explanatory variables for the i-th region, which may include factors such as the level of economic activity, transportation convenience, and education level; β is the vector of regression coefficients.

[0173] In S4c, the parameter β of the Poisson regression model is usually determined by maximum likelihood estimation (MLE), and the likelihood function L(β) is the sum of all observed data y. i The product of probabilities:

[0174]

[0175] Wherein, the log-likelihood function Calculated using the following formula:

[0176]

[0177] Because log(λ) i )=x i β, then λ i Replace with

[0178]

[0179] To find the value of β that maximizes the log-likelihood function, for Take the derivative with respect to β and set it to zero:

[0180]

[0181] Solve the equation and obtain the result;

[0182] S4d. First, a set of cleaned observational data without missing or outliers, including the dependent variable (number of employed persons) and a set of explanatory variables (such as the level of regional economic activity, transportation convenience, education level, etc.), is fitted to the data using a Poisson regression model. The trained model is then used to predict the number of employed persons for each observation. The sum of squared residuals (RSS) of the predicted number of employed persons for each observation is calculated. For each observation, the difference between the actual number of employed persons and the predicted number of employed persons is calculated, and these differences are squared and summed. The formula is as follows:

[0183]

[0184] To calculate the Total Sum of Squares (TSS), calculate the difference between actual employment and average employment, then square these differences and sum them. The formula is:

[0185]

[0186] Calculate the goodness of fit R 2 Using the RSS and TSS values ​​calculated above, substitute them into R 2 The formula for calculating the goodness of fit is as follows:

[0187]

[0188] Analysis of R 2 The value is used to evaluate the goodness of fit of the model, R. 2 The closer the R value is to 1, the more variability the model explains and the better the fit. 2 A low value may require reconsidering the model choice or checking whether the data and model assumptions are appropriate.

[0189] S5. Estimate the regional population density. The specific steps are as follows:

[0190] S5a. To estimate the population density of a region using kernel density estimation (KDE), first collect the geographic coordinate data of residential locations and use this data as the basis for analysis.

[0191] S5b. Choose a suitable kernel function. The kernel function K(u) is a symmetric smooth function with an area of ​​1. Commonly used kernel functions include Gaussian kernel, triangular kernel, and uniform kernel. Taking the Gaussian kernel as an example, its formula is as follows:

[0192]

[0193] S5c. Determine the bandwidth. Let the bandwidth be h. Bandwidth h is a key parameter of KDE, which determines the width of the kernel function, thus affecting the smoothness of the estimation. Use the Silverman rule to select the bandwidth, as shown in the following formula:

[0194]

[0195] Where σ is the sample standard deviation, IQR is the interquartile range, and n is the sample size;

[0196] S5d, kernel density estimation, for a given dataset X = {x1, x2, ..., x...} n The kernel density estimate can be expressed as:

[0197]

[0198] in, It is the kernel function scaled according to bandwidth h;

[0199] S5e. Calculate the residential population density. Using kernel density estimates, the residential population density of each area can be calculated. For each area, first calculate the sum of the kernel density estimates of all settlements within that area, then divide by the area of ​​the area to obtain the residential population density of that area. The formula is as follows:

[0200]

[0201] Where, x i It refers to the geographical location of the residence. It is in x i The estimated kernel density at that location.

[0202] S6. Calculate the joint density vector. The specific steps are as follows:

[0203] S6a. Construct a joint density vector using cluster analysis. First, standardize the features to ensure the accuracy of the cluster analysis. The standardization formula is:

[0204]

[0205] Where x is the original feature value, μ is the mean, and σ is the standard deviation;

[0206] S6b. Perform clustering, using the K-means algorithm to group the regions. The objective function of K-means is to minimize the sum of squared distances from each point to its cluster center.

[0207]

[0208] Where k is the number of clusters, S i It is the set of points in the i-th cluster, μ i It is the center of the i-th cluster;

[0209] S6c. Calculate the joint density vector by calculating the population and employment density of each cluster center, which are then used as the joint density vector.

[0210] Let μ i =(μ i1 ,μ i2 ) represents the center of the i-th cluster, where μ i1 It is population density, μ i2 Employment density:

[0211] The detailed algorithm for the joint density vector is as follows:

[0212] First, population and employment data for each region are collected, and this data is used as a feature vector, as shown in the following formula:

[0213] x = (p, e)

[0214] Where p is population density and e is employment density;

[0215] Standardize the population density p and employment density e:

[0216]

[0217] The standardized feature vector is:

[0218] x ' =(p ' ,e ' )

[0219] Determine the number of clusters, perform K-means clustering, initialize k cluster centers, and assign the feature vector x of each region to each cluster. ' Assign the cluster to the nearest cluster center, update the cluster center to the average of the region feature vector assigned to it, and repeat the above steps until the cluster centers no longer change significantly. Finally, calculate the joint density vector. For each cluster i, calculate the cluster center μ. i =(μ i1 ,μ i2 ),in:

[0220]

[0221] Where |S i | represents the number of regions in the i-th cluster, and μ represents the number of cluster centers. i This represents the joint density vector of the regions in the cluster.

[0222] S7. Predict mixed land use patterns. The specific steps are as follows:

[0223] S7a. Use multi-task learning (MTL) to estimate mixed land use patterns. Improve the generalization ability of the model by training multiple related tasks at the same time. Estimate mixed land use patterns by predicting population density and employment density. That is, predict population density and employment density at the same time, use shared low-level feature representation, and use these prediction results to estimate mixed land use patterns.

[0224] Model building, i.e., constructing a multi-task learning model, shares the underlying network but separates task-specific layers. The model can be represented as:

[0225]

[0226] Where x is the input feature vector;

[0227] S7b, shared layers, and shared feature representation: The shared network part can be represented as a series of shared layers h:

[0228] h = H(x)

[0229] Where H is a function of the shared network, and h is the shared feature representation;

[0230] S7c, Task Layer Representation: Each task has its own output layer, which can be represented as:

[0231] f1(x) = W1h + b1

[0232] f2(x) = W2h + b2

[0233] f3(x) = W3h + b3

[0234] Where W1, W2, and W3 are task-specific weight matrices, and b1, b2, and b3 are task-specific bias vectors;

[0235] S7d, Loss Function: The loss function for multi-task learning is a weighted sum of the losses for each task.

[0236]

[0237] in It is a task-specific loss function, where y1, y2, and y3 are the corresponding true labels, and λ1, λ2, and λ3 are the weights that balance the importance of different tasks.

[0238] S7e, Model Training and Prediction: Gradient descent is used to train all tasks simultaneously. The gradient update shared layer and task-specific layer for each task can be represented as:

[0239]

[0240] Where θ represents the model parameters and α is the learning rate;

[0241] Predict mixed land use patterns in the target area using a trained model:

[0242]

[0243] in is the prediction result, and x is the feature vector of the target region.

[0244] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for extracting urban land use information based on mobile phone signaling data, characterized in that, Includes the following steps: S1. User expansion coefficient: The expansion coefficient is calculated through a decision tree model, including cleaning user location and call data, extracting the average and maximum call duration and location center point features, filtering relevant features, training the decision tree model, and outputting the expansion coefficient. S2. Extrapolation of expanded data: Extract core user residency information based on the expansion coefficient and extrapolate the full dataset; S3. Identify users' residential and workplace locations and travel patterns: Analyze nighttime location data to identify the residential coordinate set, analyze daytime location data to identify the workplace coordinate set, and statistically analyze travel frequency to generate a commuting frequency matrix. S4. Analysis and prediction of employment density: Employment types are classified by occupation, and employment density values ​​are output through a Poisson regression model; S5. Estimating the regional population density: Based on the residential coordinate set data, the Gaussian kernel function is used to perform kernel density estimation, and the population density value is calculated by combining the regional area. S6. Calculate the joint density vector: Based on the standardized population density and employment density values, generate the joint density vector through K-means clustering; S7. Predict mixed land use patterns: Employ a multi-task learning model to simultaneously predict population density and employment density values, and output land use pattern classifications.

2. The method for extracting urban land use information based on mobile phone signaling data according to claim 1, characterized in that: In S1, the user location and call data are provided by the operator. Cleaning the user location and call data specifically involves: cleaning the user location, call time and call duration data provided by the operator, extracting the mean, maximum and location center point features of the call duration, filtering feature subsets through correlation analysis, training the decision tree model and calculating the root mean square error, and outputting the expansion coefficient. The formula for calculating the mean square error is as follows: Where y i This is the actual value. This is the predicted value, and n is the number of samples; The formula for calculating the root mean square error is as follows: Where y i This is the actual value. is the predicted value, and n is the number of samples.

3. The method for extracting urban land use information based on mobile phone signaling data according to claim 1, characterized in that: In step S2, the core user set is generated by extracting the residence records of individual users for more than 10 consecutive days, and the full dataset is extrapolated by applying the expansion coefficient.

4. The method for extracting urban land use information based on mobile phone signaling data according to claim 1, characterized in that: In S3, the nighttime location dwell frequency is calculated to generate a residential coordinate set, the daytime location dwell frequency is calculated to generate a workplace coordinate set, and the number of moves from the residential location to the workplace is counted to generate a commuting frequency matrix.

5. The method for extracting urban land use information based on mobile phone signaling data according to claim 1, characterized in that: In S4, seven occupational types are divided according to occupational categories. The level of regional economic activity is used as the explanatory variable, and the employment density value is output by Poisson regression maximum likelihood estimation. Assume the response variable Y follows a Poisson distribution: Where λ is the expected number of employed persons (mean) in the given region, and y is the observed number of employed persons; The number of employed persons is used as the response variable, and explanatory variables that may affect employment (including but not limited to the level of regional economic activity, transportation convenience, and education level) are used as predictor variables. In Poisson regression, it is assumed that λ depends on a set of explanatory variables X: λ=e (Xβ) Where X is the vector of explanatory variables and β is the vector of regression coefficients, the model can be written as: log(λ) = Xβ The β coefficient is determined by maximum likelihood estimation, and the goal of the model is to find the β value that is most likely to be the observed data. Suppose there are n regions, and the number of employed people in each region is y. i The corresponding explanatory variable is x. i The Poisson regression model can be written as: log(λ i )=x i b Where, λ i x represents the expected number of employees in the i-th region. i is the vector of explanatory variables for the i-th region, including three factors: level of economic activity, transportation convenience, and education level; β is the vector of regression coefficients. The parameter β of a Poisson regression model is usually determined by maximum likelihood estimation (MLE), where the likelihood function L(β) is the sum of all observed data y. i The product of probabilities: Wherein, the log-likelihood function Calculated using the following formula: Because log(λ) i )=x i β, then λ i Replace with To find the value of β that maximizes the log-likelihood function, for Take the derivative of β and set it to zero. Solve the equation and obtain the result; The trained model is used to predict the number of employed persons on a dataset (which can be a training set or a test set). The predicted number of employed persons for each observation is obtained, and the sum of squared residuals (RSS) of the predicted number of employed persons for each observation is calculated. For each observation, the difference between the actual number of employed persons and the predicted number of employed persons is calculated, and these differences are squared and summed. The formula is as follows: To calculate the Total Sum of Squares (TSS), calculate the difference between actual employment and average employment, then square these differences and sum them. The formula is: Calculate the goodness of fit R 2 Using the RSS and TSS values ​​calculated above, substitute them into R 2 The formula for calculating the goodness of fit is as follows: Analysis of R 2 The value is used to evaluate the goodness of fit of the model, R. 2 The closer the value is to 1, the more variability the model explains and the better the fit.

6. The method for extracting urban land use information based on mobile phone signaling data according to claim 1, characterized in that: In S5, the bandwidth parameter is calculated using the Silverman rule, kernel density estimation is performed based on the residential coordinate set data, the kernel density estimate is output, and the population density value is calculated by combining the area of ​​the region. When using kernel density to estimate the population density of a region, it is important to select an appropriate kernel function. The kernel function K(u) is a symmetric, smooth function with an area of ​​1. Commonly used kernel functions include, but are not limited to, Gaussian kernel, triangular kernel, and uniform kernel. Taking the Gaussian kernel as an example, its formula is as follows: Let the bandwidth be h. Bandwidth h is a key parameter of KDE, which determines the width of the kernel function and thus affects the smoothness of the estimation. The Silverman rule is used to select the bandwidth, as shown in the following formula: Where σ is the sample standard deviation, IQR is the interquartile range, and n is the sample size; Given a dataset X = {x1, x2, ..., x...} n The kernel density estimate can be expressed as: in, It is the kernel function scaled according to bandwidth h; Using kernel density estimates, the population density of each region can be calculated. For each region, first calculate the sum of the kernel density estimates of all settlements within that region, then divide by the area of ​​the region to obtain the population density of that region, as shown in the formula below: Where, x i It refers to the geographical location of the residence. It is in x i The estimated kernel density at that location.

7. The method for extracting urban land use information based on mobile phone signaling data according to claim 1, characterized in that: In S6, a standardized feature vector is generated based on the standardized population density value and employment density value. The cluster center is calculated to generate a joint density vector by minimizing the sum of squared distances through K-means clustering. The standardized formula above is: Where x is the original feature value, μ is the mean, and σ is the standard deviation; When using the K-means algorithm to group regions, the objective function of K-means is to minimize the sum of squared distances from each point to its cluster center: Where k is the number of clusters, S i μ is the set of points in the i-th cluster. i It is the center of the i-th cluster; Calculate the population and employment density of each cluster center as a joint density vector; Let μ i =(μ i1 ,μ i2 ) represents the center of the i-th cluster, where μ i1 It is population density, μ i2 It is employment density; The detailed algorithm for the joint density vector is as follows: First, population and employment data for each region are collected, and this data is used as a feature vector, as shown in the following formula: x = (p, e) Where p is population density and e is employment density; Standardize the population density p and employment density e: The standardized feature vector is: x ' =(p ' ,And ' ) Determine the number of clusters, perform K-means clustering, initialize k cluster centers, and assign the feature vector x of each region to each cluster. ' Assign the cluster to the nearest cluster center, update the cluster center to the average value of the region feature vector assigned to it, and repeat the above steps until the cluster center no longer changes significantly. Calculate the joint density vector, and for each cluster i, calculate the cluster center μ. i =(μ i1 ,μ i2 ),in: Where |S i | represents the number of regions in the i-th cluster, and μ represents the number of cluster centers. i This represents the joint density vector of the regions in the cluster.

8. The method for extracting urban land use information based on mobile phone signaling data according to claim 1, characterized in that: In step S7, a multi-task learning model with a shared underlying network is constructed, and task layer functions are defined. In the process of constructing the above multi-task learning model: firstly, the underlying network is shared, and then the task-specific layers are separated. The model can be represented as: Where x is the input feature vector; Secondly, the shared network portion can be represented as a series of shared layers h: h = H(x) Where H is a function of the shared network, and h is the shared feature representation; Each task has its own output layer, which can be represented as: f1(x) = W1h + b1 f2(x) = W2h + b2 f3(x) = W3h + b3 Where W1, W2, and W3 are task-specific weight matrices, and b1, b2, and b3 are task-specific bias vectors; The loss function for multi-task learning is a weighted sum of the losses for each task: in It is a task-specific loss function, where y1, y2, and y3 are the corresponding true labels, and λ1, λ2, and λ3 are the weights that balance the importance of different tasks. Using gradient descent to train all tasks simultaneously, the gradient updates for the shared layer and task-specific layer for each task can be represented as: Where θ represents the model parameters and α is the learning rate; Predict mixed land use patterns in the target area using a trained model: in is the prediction result, and x is the feature vector of the target region.

9. The method for extracting urban land use information based on mobile phone signaling data according to claim 1, characterized in that: The land use pattern classification is achieved by mapping the joint density vector to a preset functional type range.

Images