A population spatialization method based on multi-source data and ensemble learning

Abstract

The present invention discloses a population spatialization method based on multi-source data and integrated learning, which includes the following steps: S1, obtaining multi-source data and fusing them to construct a population spatialization database; S2, constructing an index system for model fitting from the population spatialization database, and screening out effective indicators through the feature importance calculated by the integrated learning model; S3, combining the relationship between the effective indicators and the community population, constructing a Pop-XGBoost population spatialization model; S4, predicting the population spatial distribution, and aggregating the grid population simulation data to the community scale, and comparing it with the real community population statistics to verify the accuracy of the results. By combining multi-source data fusion technology, index screening technology and integrated learning technology to construct a population spatialization model, high-precision population spatialization prediction can be achieved accurately and efficiently.

Description

Technical Field

[0001] The present invention relates to the field of geoscience big data application technology, and in particular to a population spatialization method based on multi-source data and integrated learning. Background Art

[0002] Population information can provide scientific support for regional sustainable development and spatial planning. However, the current population data are usually statistical values ​​of each administrative district, with low spatial resolution, which cannot fully express the spatial differences in population distribution within administrative units. At the same time, statistical population data are difficult to match with some studies involving complex geographical boundaries, and are not convenient for integrating other multi-source data, such as remote sensing data. Population distribution data based on grids are not only more detailed, but also can better reveal the spatial heterogeneity of the population and realize the integration of resource, environmental and management information. Therefore, it is of great practical significance to carry out fine-scale population spatialization research.

[0003] According to the different research objectives and auxiliary data, the commonly used population distribution estimation methods can be summarized into two categories: regional interpolation method and statistical modeling method. The population distribution results obtained by regional interpolation method have low prediction accuracy and spatial resolution. The main reason is that it does not fully consider the various influencing factors of population distribution and the rationality of regional interpolation is poor. Compared with regional interpolation method, the use of statistical models can effectively improve the accuracy of population spatialization results. The statistical model method mainly constructs a linear relationship between population and various auxiliary variables through a specific statistical model, thereby realizing the spatial prediction of population. Commonly used auxiliary data include land use data, night light data, roads, coastlines, MODIS / EVI, digital elevation models (DEMs) and impervious surfaces.

[0004] Although the above auxiliary data are effective, they still cannot fully represent the economic, social and cultural factors of population spatial distribution. Therefore, more and more studies are applying open-access geospatial big data to draw population distribution maps to reflect the intensity of human activities. Commonly used geospatial big data include mobile phone signaling, behavioral trajectory data, points of interest (POIs), OpenStreetMap, social insurance accounts, housing prices and taxi trajectory data. In addition, building attribute data is also closely related to population distribution, such as building type, height, floor area ratio, building area and building area. However, a single auxiliary data may have problems such as incomplete data and data anomalies, resulting in a low correlation between the resulting population simulation data and the real population data. Therefore, integrating multi-source data to reflect the spatial distribution of population can reduce the prediction bias caused by a single type of auxiliary data. However, with the increase of data and its characteristics, data noise will appear when constructing a population spatialization model, making it difficult to achieve high-precision population spatialization prediction.

[0005] Statistical models are limited in dealing with complex and multivariate influencing factors due to their simple structure. Machine learning models can handle multi-source and multi-dimensional features well and accurately mine the relationship between multi-source auxiliary data and population. Among them, ensemble learning algorithms based on decision trees have high simulation accuracy, mainly including RF, GBDT, XGBoost, etc. At present, RF and XGBoost models have been widely used in the field of remote sensing, such as soil nutrient estimation, daily reference evapotranspiration calculation, land use classification, PM2.5 prediction, landslide susceptibility mapping, biomass estimation, etc. Some research results show that compared with RF and GBDT models, XGBoost model can not only reduce overfitting problems and computational complexity, but also improve prediction accuracy, making the optimal solution of the model more efficient. However, the application of XGBoost model in fine-scale population spatialization is still rare, especially for the research on integrating multi-source geographic big data and building information.

[0006] Therefore, current population spatialization research still has two major technical defects:

[0007] (1) From the data perspective, single auxiliary data will result in poor prediction accuracy.

[0008] (2) From the perspective of the model, the increase of auxiliary data will bring data noise and cause instability in model fitting, thus affecting the accuracy of population spatialization results. In addition, the application of XGBoost model in refined population spatialization is still rare. Summary of the invention

[0009] In view of the problems in the related technology, the present invention proposes a population spatialization method based on multi-source data and integrated learning to overcome the above-mentioned technical problems existing in the existing related technology.

[0010] To this end, the specific technical solution adopted by the present invention is as follows: the method comprises the following steps:

[0011] S1. Obtain and integrate multi-source data to build a population spatial database;

[0012] S2. constructing an indicator system for model fitting from the population spatialization database, and screening out effective indicators through the feature importance calculated by the ensemble learning model;

[0013] S3. Combining the relationship between the effective indicators and the community population, a Pop-XGBoost population spatialization model is constructed;

[0014] S4. Predict the spatial distribution of population and aggregate the grid population simulation data to the community scale, compare it with the actual community demographic data, and verify the accuracy of the results.

[0015] Furthermore, multi-source data is obtained and integrated to build a population spatial database, including the following steps:

[0016] S11, obtaining multi-source data;

[0017] S12, selecting several indicators from the multi-source data to construct an indicator system;

[0018] S13, resampling each of the indicators into a 100m×100m grid scale, and aggregating them to a community scale;

[0019] S14, counting building data at a grid scale, and dividing the buildings according to the grid;

[0020] S15. Construct a population spatial database with 100m×100m grid as the unit.

[0021] Furthermore, the multi-source data includes basic geographic data, remote sensing data, building data and geographic big data;

[0022] The building data includes the total building area, the building base area, the number of building floors, the building volume, and the building type.

[0023] Furthermore, the formula for segmenting the building according to the grid includes:

[0024] k=S(Building Ai ) / S(Building A )

[0025] GFA (Building Ai )=k×GFA(Building A )

[0026] GFA (Grid i )=∑GFA(Building)

[0027] In the formula, S(Building A ) represents the building area of ​​building A;

[0028] S(Building Ai ) indicates that building A is in the Grid i of building area;

[0029] k represents the division coefficient;

[0030] Grid i represents the i-th grid;

[0031] GFA (Building A ) represents the total area of ​​Building A;

[0032] GFA (Building Ai ) indicates Grid i Building area of ​​inner building A;

[0033] GFA (Grid i ) indicates Grid i The total floor area of ​​all buildings within the area.

[0034] Furthermore, an index system for model fitting is constructed from the population spatialization database, and effective indicators are screened out through the importance of features calculated by the ensemble learning model, including the following steps:

[0035] S21. Using the statistical values ​​of community scale indicators as input variables and the real population of the community as output targets, an integrated learning model is constructed;

[0036] S22. Calculate the feature importance of each indicator according to the constructed integrated learning model;

[0037] S23. According to the average value of the importance of the characteristics, select the indicator with the greatest impact on population distribution as the effective indicator.

[0038] Furthermore, the ensemble learning model includes random forest, gradient boosting decision tree and extreme gradient boosting decision tree.

[0039] Furthermore, the Pop-XGBoost population spatialization model is constructed by combining the relationship between the effective indicators and the community population, including the following steps:

[0040] S3 1. Based on the extreme gradient boosting decision tree, the GridSearchCV module of the sklearn library is used to automatically determine the optimal parameters of the Pop-XGBoost model within the specified range;

[0041] S32, using the effective indicators as input indicators and 75% of the communities and their real population data as training sets to build a Pop-XGBoost model;

[0042] S33. Using the remaining 25% of the communities and their population data as a test set, the Pop-XGBoost model is verified and analyzed.

[0043] Furthermore, the prediction of population spatial distribution, aggregating grid population simulation data to the community scale, and comparing with real community population statistics to verify the accuracy of the results, includes the following steps:

[0044] S41, introducing the XGB-Regressor module in the sklearn library to estimate the population of each grid;

[0045] S42. According to the proportion of the grid population obtained by the Pop-XGBoost model to the total population of all grids in the district or county, the population of each 100m×100m grid is redistributed using the allocation formula;

[0046] S43. Select evaluation factors to evaluate the accuracy of the predicted population.

[0047] Furthermore, the allocation formula is:

[0048]

[0049] In the formula, i represents each grid in the area;

[0050] j represents the administrative area where the grid is located;

[0051] P i Represents the population in each grid after correction;

[0052] D j Indicates the total population of the district or county where the grid is located;

[0053] M i Indicates the population of the grid estimated by the Pop-XGBoost model;

[0054] M j Represents the total model-estimated population of all grid cells in the administrative district where this grid cell is located.

[0055] Furthermore, the evaluation factors include relative mean absolute error, root mean square error and determination coefficient.

[0056] The beneficial effects of the present invention are as follows: by combining multi-source data fusion technology, indicator screening technology and integrated learning technology to construct a population spatialization model, high-precision population spatialization prediction can be achieved accurately and efficiently. First, basic geographic data, remote sensing data, building data (such as building type, floor area ratio, etc.), and geographic big data (housing price distribution, etc.) are integrated to construct a population spatialization database with 100m×100m grid as a unit, providing data support for fine-scale population spatialization. Secondly, in order to solve the problems of noise enhancement and unstable model fitting caused by the increase of auxiliary data, the importance of features is calculated through a variety of integrated learning models, and indicators are screened to reduce data noise and improve the accuracy of fine-scale population spatialization results; finally, the Pop-XGBoost population spatialization model is constructed to improve the accuracy of fine-scale population spatialization results. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0058] Figure 1 is a flow chart of a population spatialization method based on multi-source data and integrated learning according to an embodiment of the present invention;

[0059] Figure 2 is a schematic diagram of the administrative area of ​​Shenzhen in a population spatialization method based on multi-source data and integrated learning according to an embodiment of the present invention;

[0060] Figure 3 It is a technical roadmap of a population spatialization method based on multi-source data and integrated learning according to an embodiment of the present invention;

[0061] Figure 4 is the feature importance of different indicators in a population spatialization method based on multi-source data and ensemble learning according to an embodiment of the present invention;

[0062] Figure 5 It is a schematic diagram of the prediction results of Pop-XGBoost population spatialization in a population spatialization method based on multi-source data and integrated learning according to an embodiment of the present invention;

[0063] Figure 6 It is a schematic diagram comparing the results of Pop-XGBoost and Worldpop population spatialization (100m×100m) in a population spatialization method based on multi-source data and ensemble learning according to an embodiment of the present invention;

[0064] Figure 7 A population spatialization method based on multi-source data and integrated learning according to an embodiment of the present invention Figure 5 Schematic diagram comparing the population distribution in the four areas a, b, c, and d. DETAILED DESCRIPTION

[0065] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention and are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, ordinary technicians in the field should be able to understand other possible implementations and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0066] According to an embodiment of the present invention, a population spatialization method based on multi-source data and ensemble learning is provided.

[0067] In the embodiment of the present invention, the population of Shenzhen is selected as the analysis object. Shenzhen is a coastal city in southern China, adjacent to Hong Kong, located south of the Tropic of Cancer (113°46′~114°37′E, 22°27′~22°52′N). It is China's first special economic zone, including 10 administrative districts, including Futian District, Luohu District, Nanshan District, Yantian District, Bao'an District, Longgang District, Guangming District, Longhua District, Pingshan District, Dapeng New District, 74 streets, and 734 communities ( Figure 1 ). In 1979, Shenzhen's population was only 310,000, but by 2019, the permanent population had grown to 13.44 million. In 2019, Shenzhen's total area was 1,997.47 square kilometers, with a built-up area of ​​927.96 square kilometers.

[0068] Population prediction and verification are carried out based on the following five data and parameters.

[0069] (1) Demographic data

[0070] The study obtained demographic data of 270 communities in 2019 from the Shenzhen Municipal Land and Resources Commission as real data for model training and model validation, including 202 training sets (74.9%) and 68 test sets (25.1%). The test set, as an independent validation set, will not participate in the training of the model.

[0071] (2) Geographic Big Data

[0072] The geographic big data obtained in the study include: Baidu POIs, subway routes, housing prices, floor area ratio, greening rate, total number of households, parking spaces, OpenStreetMap road network data, etc.

[0073] The POIs data used in the study comes from Baidu Maps, the most widely used and largest Internet map service provider in China. The study quickly obtains POIs data from Baidu Maps by calling the Place API of Baidu Maps. The POIs obtained include attribute information such as latitude and longitude grid coordinates, categories, and names. The obtained data were sorted out, and some data with duplicates or missing attributes were screened out, and a total of 802,668 valid data were obtained. The kernel density analysis method was used to process various POIs data, and 10 types of raster data (100m×100m) were obtained, including government agencies, education, enterprises, shopping, finance, health, entertainment, transportation, catering, and residence.

[0074] The subway routes of the study area were obtained from the Baidu subway map JavaScript API, and the Euclidean distance tool in ArcGIS10.8 was used to calculate the distance between each grid (100m×100m) and the subway routes.

[0075] The data of 7714 residential communities in Shenzhen were obtained from the Anjuke website, including five attributes, including housing price, plot ratio, greening rate, total number of households, and parking spaces. Anjuke is the most commonly used online platform in China, mainly publishing real estate information for sale or rent. The five attribute data were interpolated to the entire study area using the Kriging interpolation method, and the information of housing price, plot ratio, greening rate, total number of households, and parking spaces for each grid (100m×100m) was obtained.

[0076] Obtain the road network data of the study area from OpenStreetMap. OpenStreetMap contains rich and accurate road network information, which can reflect the spatial distribution of the population within the city. OpenStreetMap road network types include 16 categories such as bicycle lanes, sidewalks, main roads, and highways. Road network density refers to the ratio of the total length of the road network within the study unit to the area of ​​the study unit.

[0077] (3) Remote sensing data

[0078] The remote sensing data obtained in the study include: digital elevation model (DEM), Landsat 8OLI data, impervious surface products, and night light data.

[0079] Landsat 8OLI (2018 / 04) and DEM were obtained from the Geospatial Data Cloud. Landsat 8OLI data were processed using ENVI5.3 to obtain the vegetation normalized index and surface temperature. DEM was processed using Acgis10.8 to extract aspect and slope data.

[0080] The global impervious surface product (MSMT_IS30) was obtained from the National Earth System Science Data Center. The overall accuracy of the data is 95.1% and the resolution is 30m. The impervious surface ratio of each grid in the study area is calculated. The impervious surface ratio refers to the ratio of the impervious surface area in the study unit to the area of ​​the study unit.

[0081] The night light data used is the Luojia-1 night light data with a spatial resolution of 130m.

[0082] Finally, the vegetation normalized index, elevation, percentage of impervious surface, aspect, slope, surface temperature and night light with a resolution of 30 m were resampled to 100 m.

[0083] (4) Basic geographic data

[0084] The study obtained basic geographic data such as land use and administrative district boundaries from the Shenzhen Municipal Planning and Land Resources Commission. The distance from the evaluation unit grid (100m×100m) to the water surface / coastline, the distance from the central urban area, and the distance from the green space / park were calculated based on the Euclidean distance method. The proportion of construction land refers to the ratio of the construction land area within the study unit to the area of ​​the study unit.

[0085] (5) Building data

[0086] The building data comes from the Shenzhen Municipal Planning and Land Resources Commission, which includes five attributes: total building area, building base area, number of building floors, building volume, and building type. Based on these five attributes, seven indicators, including building volume ratio, building density, total building area, building area, number of building floors, building volume, and building type, were calculated according to a 100m×100m grid (Table 1). The national standard "Uniform Standard for Design of Civil Buildings" (GB 50352-2019) divides building types into four categories, namely residential buildings, public buildings, industrial buildings, and agricultural buildings. Building types are quantified according to the distribution of population in different building types. The building type quantification equation is D7=Qi×D6, where Qi is the building type weight. The building weights of residential buildings, public buildings, industrial buildings, and agricultural buildings in the building type are 10, 5, 3, and 2, respectively. Building density refers to the ratio of the sum of the base areas of all buildings in the area to the area of ​​the area. The total building area ratio refers to the ratio of the sum of the total areas of all buildings in the study unit to the area of ​​the study unit.

[0087] The present invention is further described with reference to the accompanying drawings and specific embodiments. Figure 1-7 As shown, according to one embodiment of the present invention, a population spatialization method based on multi-source data and integrated learning is provided, and the method includes the following steps:

[0088] S1. Obtain and fuse multi-source data to build a population spatial database, including the following steps:

[0089] S11. Acquire multi-source data, including basic geographic data, remote sensing data, building data and geographic big data;

[0090] S12, selecting several indicators from the multi-source data to construct an indicator system;

[0091] S13, resampling each of the indicators into a 100m×100m grid scale, and aggregating them to a community scale;

[0092] S14. Count the building data at the grid scale and segment the buildings according to the grid. The segmentation formula includes:

[0093] k=S(Building Ai ) / S(Building A )

[0094] GFA (Building Ai )=k×GFA(Building A )

[0095] GFA (Grid i )=∑GFA(Building)

[0096] In the formula, S(Building A ) represents the building area of ​​building A;

[0097] S(Building Ai ) indicates that building A is in the Grid i of building area;

[0098] k represents the division coefficient;

[0099] Grid i represents the i-th grid;

[0100] GFA (Building A ) represents the total area of ​​Building A;

[0101] GFA (Building Ai ) indicates Grid i Building area of ​​inner building A;

[0102] GFA (Grid i ) indicates Grid i The total floor area of ​​all buildings in the area. S15. Construct a population spatial database with 100m×100m grid as the unit.

[0103] The multi-source data include four types of data: basic geographic data, remote sensing data, building data and geographic big data, and a total of 35 indicators are selected, as shown in Table 1;

[0104] The building data includes attributes such as the total building area, the building base area, the number of building floors, the building volume, and the building type.

[0105] Table 1: Population spatialization indicator system

[0106]

[0107]

[0108] S2. constructing an index system for model fitting from the population spatialization database, and screening out effective indicators through the importance of features calculated by the ensemble learning model, comprising the following steps:

[0109] S21. Using the statistical values ​​of community scale indicators as input variables and the real population of the community as output targets, an integrated learning model is constructed;

[0110] Ensemble learning completes learning tasks by building and combining multiple classifiers, also known as multi-classification systems. The purpose of ensemble learning is to improve the generalization ability and robustness of base learners by combining the prediction results of multiple machine learning classifiers. Among them, the ensemble learning model includes random forest (RF), gradient boosted decision tree (GBDT) and extreme gradient boosted decision tree (XGBoost). RF, GBDT and XGBoost models have advantages in evaluating feature importance and processing high-dimensional data.

[0111] Specifically, random forest (RF) is an ensemble algorithm based on decision trees. RF is an extended variant of bagging. It further introduces random feature selection in the training process of decision trees based on building bagging ensembles with decision trees as base learners. The implementation process of RF is as follows: First, from the original training data set (D1, D2, ..., D k ) to obtain k sampling sets; secondly, by training each sample set, k weak learners {h1, h2, …, h k}; Finally, use majority voting or the average of the prediction results to get the final classification result.

[0112] Gradient boosted decision tree (GBDT) is widely used in various tasks such as classification and regression. It is a combination of decision tree and gradient boosting. The main difference between GBDT and RF is that the decision tree in GBDT is fitted based on the residual of the previous decision tree. Therefore, GBDT can reduce bias and variance. In each iteration, the goal of GBDT is to build a weak classifier of the regression tree to reduce the iteration of the loss function, which can be expressed as The predicted value of the mth iteration, is the loss function, f m (x i ) is used to minimize the loss function.

[0113] Extreme Gradient Boosting Decision Tree (XGBoost) is a further optimization of GBDT. The loss function can be expressed as where y i is a true value, is the predicted value of the m-1th iteration. GBDT only uses the first-order derivative information during optimization. XGBoost performs a second-order Taylor expansion on the loss function to obtain the first-order and second-order derivatives, which can be expressed as g i and h i is the first-order and second-order gradient statistical loss function, which can be expressed as In addition, XGBoost adds a regularization term Ω(f m ) to control the complexity of the model. The smaller the value, the lower the computational complexity and the stronger the generalization ability. Therefore, XGBoost can not only reduce the variance and complexity of the model, but also avoid overfitting problems.

[0114] S22. Calculate the feature importance of each indicator according to the constructed integrated learning model;

[0115] S23. According to the average value of the importance of the characteristics, select the indicator with the greatest impact on population distribution as the effective indicator.

[0116] S3. Combining the relationship between the effective indicators and the community population, a Pop-XGBoost population spatialization model is constructed, including the following steps:

[0117] S31, based on the extreme gradient boosting decision tree, the GridSearchCV module of the sklearn library is used to automatically determine the optimal parameters of the Pop-XGBoost model within the specified range;

[0118] Among them, n_estimators is the maximum number of iterations of weak learners, or the maximum number of weak learners, which usually does not affect the complexity of the model. max_depth is the maximum depth of a tree. When there are many features, it is recommended to adjust the max_depth parameter to avoid overfitting. n_estimators and max_depth have a greater impact on the fitting accuracy of the Pop-XGBoost model. The learning rate can improve the robustness of the model by reducing the weight of each step. min_child_weight is the minimum leaf node sample weight sum. When the min_child_weight value is large, the model can be prevented from learning local special samples. However, if the min_child_weight is too high, it will lead to underfitting. Gamma specifies the minimum loss function drop required for node splitting. The larger the value of this parameter, the more conservative the algorithm. Subsample is the proportion of random sampling. Reducing the value of Subsample can avoid overfitting of the algorithm, but if this value is set too small, it may lead to underfitting. colsample_ubytree is the column sampling ratio when building each decision tree.

[0119] Therefore, finding the best learning_rate, min_child_weight, gamma, subsample, and colsample_bytree parameters is very important for the Pop-XGBoost model.

[0120] Table 2: Parameter ranges

[0121] parameter XGBoosT parameter value range Automatic search step N_ESTIMATORS 10-201 10 MAX_DEPTH 3-10 1 LEARNING_RATE 0-0.3 / MIN_CHILD_WEIGHT 1-6 1 GAMMA 0-0.5 / SUBSAMPLE 0.5-1 / COLSAMPLE_BYTREE 0.5-1 /

[0122] S32, using the effective indicators as input indicators and 75% of the communities and their real population data as training sets to build a Pop-XGBoost model;

[0123] S33. Using the remaining 25% of the communities and their population data as a test set, the Pop-XGBoost model is verified and analyzed.

[0124] S4. Predict the spatial distribution of population, aggregate the grid population simulation data to the community scale, and compare it with the real community population statistics to verify the accuracy of the results, including the following steps:

[0125] S41, introducing the XGB-Regressor module in the sklearn library to estimate the population of each grid;

[0126] Since the population of each grid estimated by the optimal model is an estimate generated based on the community population, it is necessary to use the statistical population data at the district and county scale as the basis for total control of the grid simulation population results.

[0127] S42. According to the proportion of the grid population obtained by the Pop-XGBoost model to the total population of all grids in the district or county, the population of each 100m×100m grid is redistributed using the allocation formula;

[0128] Wherein, the allocation formula is:

[0129]

[0130] In the formula, i represents each grid in the area;

[0131] j represents the administrative area where the grid is located;

[0132] P i Represents the population in each grid after correction;

[0133] D j Indicates the total population of the district or county where the grid is located;

[0134] M i Indicates the population of the grid estimated by the Pop-XGBoost model;

[0135] M j Represents the total model-estimated population of all grid cells in the administrative district where this grid cell is located.

[0136] S43. Select evaluation factors to evaluate the accuracy of the predicted population.

[0137] The evaluation factors include the relative mean absolute error (MAE), the root mean square error (RMSE) and the coefficient of determination (R2). The smaller the MAE and RMSE values ​​are, the more accurate the estimation results are; the value range of R2 is in the interval (0, 1), and the closer the value is to 1, the closer the predicted population is to the actual population.

[0138] The following is a specific example study and result analysis using the present invention:

[0139] (I) Index screening results

[0140] From the average value of the characteristic importance (relative value, expressed in percentage) of the 35 indicators, it can be seen that the sum of the characteristic importance of the five types of building data exceeds 69%, indicating that building data has a great influence on the spatial distribution of the population. Among them, the characteristic importance of building area (D3) is the largest, at 33.02%, and the characteristic importance of the number of floors (D5) and building type (D7) is relatively large, at 14.63% and 10.29%, respectively. The sum of the characteristic importance of geographic big data is about 21.73%. The sum of the characteristic importance of POIs density in geographic big data is about 16.57%, among which the contribution rate of medical facility core density (A6) is the largest, at 6.45%, and the contribution rate of restaurant density (A9) is the smallest, at 0.38%. In addition, the total number of households (A14) and subway distance (A16) also have high characteristic contribution rates, at 1.34% and 1.03%, respectively. The sum of the characteristic importance of remote sensing data is about 6.17%. DEM (B1), impervious surface ratio (B3), and night light intensity (B6) have high feature contribution rates, which are 1.27%, 1.30%, and 1.31%, respectively. The sum of the feature importance of basic geographic data is about 3.05%. Only the distance to the water / coastline (C1) has a large feature contribution rate (1%). In summary, 21 low-redundancy indicators with an importance score greater than 1% were selected. Among them, there are 7 valid indicators for building data, 10 valid indicators for geographic big data, 3 valid indicators for remote sensing data, and 1 valid indicator for basic geographic data.

[0141] (II) Constructing Pop-XGBoost population spatialization model based on effective indicators

[0142] The Pop-XGBoost population spatialization model was constructed with 21 valid indicators as input variables and the real population of 202 communities (accounting for 75% of the 270 communities) as output targets. The optimal parameters are shown in Table 3. The real population of 68 communities (accounting for 25% of the 270 communities) was used as the validation set to determine the accuracy of the constructed model. The results showed that the accuracy of the training set was as high as 99% and the accuracy of the test set was as high as 92%.

[0143] Table 3: Optimal parameters for different models

[0144]

[0145]

[0146] 3. Grid-scale population spatial distribution prediction results

[0147] The selected 21 indicators are counted to a 100m×100m grid scale as the input variables of the Pop-XGBoost model to predict the spatial distribution of Shenzhen’s population. Figure 5As shown in the figure, each grid represents the population within the grid area (1 hectare). Figure 5 It can be seen that the areas with high population density are mainly concentrated in the southeastern part of Futian District (Nanyuan Street, Fubao Street and Futian Street, etc.), the southwestern part of Luohu District (Nanhu Street, Dongmen Street, etc.), the central part of Longhua District (Longhua Street and Longgang District), and the southern part of Longgang District (Buji Street and Nanwan Street). In general, the population is concentrated in areas with superior location conditions and good economic conditions, which are convenient for people to carry out production and living activities. Therefore, the population spatial distribution prediction results based on Pop-XGBoost are consistent with the actual distribution characteristics of the population.

[0148] (IV) Accuracy verification

[0149] Obtain the Worldpop dataset of Shenzhen in 2019. This dataset is a population spatial grid dataset with wide application, high recognition and good accuracy. The spatial resolution of the dataset provided by it can reach 100m×100m. Compare with the prediction results of the Pop-XGBoost model ( Figure 6 ), the results show that the population spatial distribution prediction results of the Pop-XGBoost model and the Worldpop dataset have roughly the same trend, and the high-value population agglomeration areas are basically the same.

[0150] The 2019 Sentinel-2 remote sensing image data with a spatial resolution of 20 meters was selected to verify the population prediction results of the Pop-XGBoost model. Figure 7 From the detailed comparison results of the four regions (a, b, c, d) in Figure 1, it can be found that the prediction results of the Pop-XGBoost model and the Worldpop dataset in different regions are basically consistent. However, the simulation effect of Worldpop in low population density areas (such as region b) is not good and inconsistent with remote sensing images, mainly because the data used by Worldpop is not fine enough. The Pop-XGBoost model proposed in this study has a prediction result that is basically consistent with remote sensing images and can better reflect the population spatial distribution at a fine scale.

[0151] The grid population results obtained by the model are summarized to the community scale, compared with the statistical values ​​at the community scale, and the MAE, RMSE and R2 indicators are calculated. The results show that the RMSE value of the Pop-XGBoost model is 12783.82, the MAE value is 8006.07, and the R2 value is 80.54%, with good overall accuracy.

[0152] In summary, with the help of the above technical solutions of the present invention, a population spatialization model is constructed by combining multi-source data fusion technology, indicator screening technology and integrated learning technology to accurately and efficiently achieve high-precision population spatialization prediction. First, basic geographic data, remote sensing data, building data (such as building type, floor area ratio, etc.), and geographic big data (housing price distribution, etc.) are integrated to construct a population spatialization database with 100m×100m grid as a unit to provide data support for fine-scale population spatialization. Secondly, in order to solve the problems of noise enhancement and unstable model fitting caused by the increase of auxiliary data, the importance of features is calculated through a variety of integrated learning models, and indicators are screened to reduce data noise and improve the accuracy of fine-scale population spatialization results; finally, the Pop-XGBoost population spatialization model is constructed to improve the accuracy of fine-scale population spatialization results.

[0153] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0154] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A population spatialization method based on multi-source data and ensemble learning, characterized in that: The method comprises the following steps: S1. Obtain and integrate multi-source data to build a population spatial database; S2. constructing an indicator system for model fitting from the population spatialization database, and screening out effective indicators through the feature importance calculated by the ensemble learning model; S3. Combining the relationship between the effective indicators and the community population, a Pop-XGBoost population spatialization model is constructed; S4. Predict the spatial distribution of population and aggregate the grid population simulation data to the community scale, compare it with the actual community population statistics, and verify the accuracy of the results; Acquiring and integrating multi-source data to build a population spatial database includes the following steps: S11, obtaining multi-source data; S12, selecting several indicators from the multi-source data to construct an indicator system; S13, resampling each of the indicators into a 100m×100m grid scale, and aggregating them to a community scale; S14, counting building data at a grid scale, and dividing the buildings according to the grid; S15. Construct a population spatial database with 100m×100m grid as the unit; The multi-source data includes basic geographic data, remote sensing data, building data and geographic big data; the building data includes the total building area, building base area, number of building floors, building volume and building type; The formula for dividing the building into grids includes: k=S(Building Ai ) / S(Building A ) GFA(Building Ai )=k×GFA(Building A ) GFA(Grid i )=∑GFA(Building) In the formula, S(Building A ) represents the building area of ​​building A; S(Building Ai ) indicates that building A is in the Grid i of building area; k represents the division coefficient; Grid i represents the i-th grid; GFA (Building A ) represents the total area of ​​Building A; GFA (Building Ai ) indicates Grid i Building area of ​​inner building A; GFA (Grid i ) indicates Grid i The total floor area of ​​all buildings within the An indicator system for model fitting is constructed from the population spatialization database, and effective indicators are screened out through the importance of features calculated by the ensemble learning model, including the following steps: S21. Using the statistical values ​​of community scale indicators as input variables and the real population of the community as output targets, an integrated learning model is constructed; S22, calculating the feature importance of each indicator according to the constructed integrated learning model; S23. According to the average value of the importance of the characteristics, select the indicator with the greatest impact on population distribution as the effective indicator; Combining the relationship between the effective indicators and the community population, constructing a Pop-XGBoost population spatialization model includes the following steps: S31, based on the extreme gradient boosting decision tree, the GridSearchCV module of the sklearn library is used to automatically search for the optimal parameters of the Pop-XGBoost model within a specified range; S32, using the effective indicators as input indicators and 75% of the communities and their real population data as training sets to build a Pop-XGBoost model; S33, using the remaining 25% of the communities and their population data as a test set to verify and analyze the Pop-XGBoost model; The method of predicting population spatial distribution, aggregating grid population simulation data to the community scale, and comparing it with real community population statistics to verify the accuracy of the results, includes the following steps: S41, introducing the XGB-Regressor module in the sklearn library to estimate the population of each grid; S42. According to the proportion of the grid population obtained by the Pop-XGBoost model to the total population of all grids in the district or county, the population of each 100m×100m grid is redistributed using the allocation formula; S43, selecting evaluation factors to evaluate the accuracy of the predicted population; The allocation formula is: In the formula, i represents each grid in the area; j represents the administrative area where the grid is located; P i Represents the population in each grid after correction; D j Indicates the total population of the district or county where the grid is located; M i Indicates the population of the grid estimated by the Pop-XGBoost model; M j Represents the total model-estimated population of all grid cells in the administrative district where this grid cell is located.

2. A population spatialization method based on multi-source data and ensemble learning according to claim 1, characterized in that: The ensemble learning model includes random forest, gradient boosting decision tree and extreme gradient boosting decision tree.

3. A population spatialization method based on multi-source data and ensemble learning according to claim 2, characterized in that: The evaluation factors include relative mean absolute error, root mean square error and determination coefficient.

Images