Semi-supervised multi-source data derivative feature fusion population spatial distribution model construction method

Abstract

The invention provides a semi-supervised multi-source data derivative feature fusion population spatial distribution model construction method. The method specifically comprises the following steps: acquiring multi-source data; preprocessing multi-source data, extracting a feature factor cluster, and storing the data; constructing and training a meta-regression population distribution model; verifying the precision of the population distribution model; and performing semi-supervised population distribution prediction on different target area data based on the trained model to obtain a rasterized model prediction result. According to the method, the multi-dimensional feature factors of the multi-source data are extracted, the feature factors are screened according to the data features, and the population prediction model is trained for population prediction. And part of key feature factors are selected in a semi-supervised manner, so that the model compatibility and the accuracy of population distribution prediction are greatly improved.

Description

Technical Field

[0001] The present invention relates to geographic space analysis, and in particular to a method for constructing a population spatial distribution model by fusing multi-source data derived features under semi-supervision. Background Art

[0002] Population is the number of people engaged in production and activities within a region. Population data is one of the most important basic geographic data. Mastering accurate and specific population data information is of great significance for solving various population-related research and applications. Population spatial distribution refers to the spatial distribution of human beings within a certain time and geographical location. It is the result of the comprehensive effect of historical natural conditions and social and economic environment. Refined population distribution is the hot spot and difficulty of current population geography research. With the continuous development of my country's urbanization process, the gradually growing population has brought huge challenges to the sustainable utilization of urban resources, the carrying capacity of the ecological environment, and the reasonable planning of infrastructure. Cities are facing various problems such as comprehensive utilization of resources, ecological and environmental protection, disaster risk assessment and rescue, regional planning and development, etc. At present, China's urbanization development is moving from extensive development to refined development. Refined urban management is synonymous with new urbanization. Refined urban management urgently needs refined population data as basic data, thereby effectively improving the quality and level of urban development.

[0003] In the existing technology, the research on population model mainly uses multi-source data such as city annotation data, night light data, and night light data for feature extraction and fusion to simulate and calculate the distribution of population model to the greatest extent. Chen Hongxing of East China Normal University (Building-scale population estimation based on multi-source data and random forest algorithm, Chen Hongxing) uses random forest machine learning method to calculate population model. This method uses building layout as the distribution base map of population model, which can accurately reflect the population distribution pattern to a certain extent. However, there are still some shortcomings:

[0004] (1) Although the multi-source data population distribution model constructed by Chen Hongxing et al. can accurately reflect the population distribution to a certain extent, the irregular population distribution model is not conducive to processing, storage and application. In addition, the feature extraction of night light data by this method cannot fully reflect the characteristics of the population.

[0005] (2) Night light data is the most direct feature reflecting the dense population distribution. However, current research work cannot achieve a semi-supervised model for feature extraction and preprocessing of night light data, and it is difficult to flexibly change it according to the region.

[0006] (3) Existing population model building technology does not make good use of geospatial data. Geospatial data is highly correlated with population distribution. The terrain undulation, direction, slope, etc. in different regions determine the nature, attributes and direction of buildings on the surface to a certain extent. In current research, common parameters related to terrain undulation, such as slope, slope direction, elevation, and terrain undulation, are used as influencing parameters to build models. However, their adaptability is poor. The models vary greatly for different regional ranges and their universality is poor. Different models need to be built for different regions, and repeated training and experimental verification are required. Summary of the invention

[0007] In view of the above-mentioned deficiencies and defects in the prior art, the present invention provides a method and device for constructing a population spatial distribution model by fusing multi-source data-derived features under semi-supervision. The method extracts multi-dimensional feature factors from multi-source data, screens the features, and trains a population prediction model for population prediction. The selection of some key feature factors is carried out in a semi-supervised manner, which greatly improves the compatibility of the model.

[0008] The present invention provides a method for constructing a population spatial distribution model by fusing features derived from multi-source data under semi-supervision, comprising:

[0009] S1, obtain multi-source data: obtain the latest census data of the target area, the geographic space data corresponding to the target area, the night light data corresponding to the target area, the land use corresponding to the target area, the city annotation data corresponding to the target area, and the urban building data corresponding to the target area;

[0010] S2, preprocessing multi-source data, extracting characteristic factor clusters, and storing data: preprocessing the census data; extracting characteristic factors and constructing characteristic factor clusters for the geographic spatial data corresponding to the target area, the night light data corresponding to the target area, the land use corresponding to the target area, the city annotation data corresponding to the target area, and the urban building data corresponding to the target area; storing the constructed characteristic factor cluster data in the population spatial characteristic database of the study area.

[0011] S3, construction and training of multivariate regression population distribution model;

[0012] S4, Verification of the accuracy of the population distribution model.

[0013] S5, based on the trained model, semi-supervised population distribution prediction is performed on the data of different target areas to obtain the rasterized model prediction results.

[0014] Further, in step S1, the geographic spatial data corresponding to the population census data of the region includes: first spatial data, second spatial data, and third spatial data having physical world location information and identification;

[0015] Furthermore, the first spatial data carries attribute field information of the administrative district code and administrative district name of the target area;

[0016] Furthermore, the third spatial data is extracted based on the SRTM data obtained after splicing and cropping;

[0017] Furthermore, in step S2, the preprocessing of multi-source data and the extraction of characteristic factor clusters specifically include:

[0018] Gridding the first spatial data to obtain fifth spatial data, matching the population census data to the fifth spatial data to obtain fused model basic sample data; wherein the grid size of the fifth spatial data is set to 100m×100m;

[0019] Performing spatial matching and gridding on the third spatial data to obtain sixth spatial data, and extracting characteristic factors based on the sixth spatial data to obtain a first characteristic factor cluster, wherein the sixth spatial data has the same grid size and spatial coordinate reference as the basic sample data;

[0020] Acquire first nighttime light data, perform spatial matching and gridding on a spatial range covered by the first nighttime light data, and cut the data range based on the basic sample data to obtain second nighttime light data, wherein the second nighttime light data has the same grid size and spatial coordinate reference as the basic sample data;

[0021] Acquire the third nighttime light data, perform data splicing on the third nighttime light data, perform spatial matching and gridding on the third nighttime light data, and cut the data range based on the basic sample data to obtain the fourth nighttime light data; wherein the fourth nighttime light data has the same grid size and spatial coordinate reference as the basic sample data;

[0022] A second characteristic factor cluster is obtained based on the fusion calculation of the second nighttime light data and the fourth nighttime light data;

[0023] Classification processing is performed based on the currently acquired sentinel image data to obtain a number of first land use data after classification;

[0024] Performing spatial matching and gridding on the plurality of blocks of first land use data, and cutting and splicing the plurality of blocks of first land use data based on the basic sample data to obtain second land use data;

[0025] Based on the statistics and calculation of the second land use data, a third characteristic factor cluster is obtained;

[0026] Obtain the first city's labeled data, perform preprocessing and data reconstruction, and obtain reconstructed second city's labeled data;

[0027] Based on the reconstructed second city labeled data, a corresponding plurality of characteristic factors are obtained to form a fourth characteristic factor cluster;

[0028] Acquire first city building data, perform spatial matching and gridding on the first city building data, and cut the data range based on the basic sample data to obtain second city building data; wherein the second city building data has the same grid size and spatial coordinate reference as the basic sample data;

[0029] Based on the second city building data, a corresponding plurality of characteristic factors are obtained to form a fifth characteristic factor cluster.

[0030] Furthermore, matching the census data to the fifth data to obtain the fused fifth data specifically includes:

[0031] The census data is matched to the corresponding grid according to administrative attributes, and the corresponding population data in each grid is calculated according to the grid size and actual population density, so as to obtain the initial sample data of the population distribution model after preliminary fusion, wherein the sample data formed includes population data fields.

[0032] Furthermore, the sixth spatial data has the same grid size and spatial coordinate reference as the basic sample data.

[0033] Furthermore, the first characteristic factor cluster includes:

[0034] Elevation characteristic factor, slope characteristic factor, aspect characteristic factor, and hyperparameter terrain characteristic factor.

[0035] Furthermore, the extracting characteristic factors based on the sixth spatial data to obtain a first characteristic factor cluster specifically includes:

[0036] (1) Obtaining an elevation factor by performing data interpolation on the sixth spatial data; specific data interpolation methods may be linear interpolation, bilinear interpolation, inverse distance weighted interpolation, weighted average method, moving fitting method, Kriging interpolation, natural neighbor interpolation method, spline interpolation method, and other methods and their improvements.

[0037] (2) extracting slope grid values ​​from the sixth spatial data; performing partition statistics on the extracted slope grid values ​​to calculate the average slope value of each area, and using it as a slope characteristic factor.

[0038] (3) extracting aspect grid values ​​from the sixth spatial data; performing partition statistics on the extracted aspect grid values ​​to calculate the average slope value of each area and use it as the aspect characteristic factor.

[0039] (4) Based on the elevation characteristic factors obtained above, the hyperparameter terrain characteristic factors are further calculated.

[0040] (5) Obtaining the first characteristic factor cluster based on the fusion of the elevation factor, the slope characteristic factor, the aspect characteristic factor, and the hyperparameter terrain characteristic factor.

[0041] Furthermore, on the basis of the elevation characteristic factors obtained above, the hyperparameter terrain characteristic factors are further calculated, specifically including:

[0042] ① Based on the elevation characteristic factor and the slope characteristic factor extracted from the sixth spatial data, the mean elevation Hp and the difference between the highest and lowest elevations Hc in the target area are calculated; and,

[0043] ② Calculate the proportion β of the area in the target area where the slope characteristic factor meets the specified threshold α to the area of ​​the target area;

[0044] ③ Based on the following algorithm, the hyperparameter terrain characteristic factor A is calculated:

[0045] δ 1 A=Hp+Hc*β*δ 2

[0046] In the above formula, δ 1 is the plane scale parameter, which is generally determined according to the average elevation Hp, δ 2 is the terrain feature influencing parameter, representing the importance of the ground area elevation fluctuation on the overall impact, where δ 2 >0.

[0047] ④Set the analysis radius of the target area;

[0048] ⑤ Based on the analysis radius set, the plane scale parameter δ 1 , terrain characteristics affect parameter δ 2 Make adaptive settings;

[0049] ⑥ Calculate the hyperparameter terrain characteristic factor A based on the set parameters.

[0050] Furthermore, the calculation to obtain the ratio β of the area range in the target area where the slope characteristic factor meets the specified threshold α to the range of the target area includes:

[0051] The slope characteristic factor data in the target area are statistically analyzed; based on the statistical results, the slope characteristic factors are clustered based on the threshold α to obtain the slope characteristic factor classification that meets the threshold α, and further obtain the proportion β of the area range that meets the specified threshold α to the range of the target area.

[0052] Furthermore, the second characteristic factor cluster is obtained by fusion calculation based on the second nighttime light data and the fourth nighttime light data, specifically including:

[0053] (1) Extraction of DN values ​​of night light data: Extract the DN values ​​of the second night light data and the fourth night light data respectively, and map the data processed with DN values ​​to the light data that has undergone matching, spatial matching, gridding, and cropping operations through linear interpolation to form characteristic factors of the two types of night light data.

[0054] (2) Extracting brightness values ​​of nighttime light data: Calculating and converting the brightness value L of the fourth nighttime light data, and mapping the calculated corresponding brightness value to the fourth nighttime light data to form a characteristic factor of the nighttime light data.

[0055] (3) Binarization of night light data: Based on the data characteristics of the second night light data and the fourth night light data and the characteristics of the target area, feature screening is performed, and the selected feature values ​​of the brightness value of the second night light data are set to a1, a2, and a3; the selected feature values ​​of the brightness value of the fourth night light data are set to a1%, a2%, and a3% of the highest value. When each type of data corresponds to the corresponding threshold, if it meets the corresponding threshold, the pixel value of the corresponding target area is assigned to 255; if it does not meet the corresponding threshold, the pixel value of the corresponding target area is assigned to 0. After traversing each pixel assignment, the binary data of the three types of night light data generated by the second night light data and the fourth night light data are obtained as the feature factors of the corresponding dimensions, totaling 6 types of feature factors of night light data.

[0056] (4) A total of nine characteristic factors of the corresponding night light data obtained by combining the DN value of the night light data, the brightness value L of the night light data, and the binary data of the night light data, form a second characteristic factor cluster.

[0057] Furthermore, the classification processing is performed based on the currently acquired sentinel image data to obtain a plurality of first land use data after classification, specifically including:

[0058] (1) Acquire sentinel image data corresponding to the target area;

[0059] (2) classifying land cover types based on a deep learning algorithm to obtain a plurality of first land use data after classification;

[0060] The grid size of the sentinel image data is 10m, and each piece of the classified first land use data contains 10 land use types;

[0061] Furthermore, based on the statistics and calculation of the second land use data, a third characteristic factor cluster is obtained, which specifically includes:

[0062] extracting 10 land use types from the second land use data as 10-dimensional land use characteristic factors;

[0063] The 10 land use types are merged into a new characteristic factor for assignment, so that the new characteristic factor contains all land use types and is represented by different codes;

[0064] The 10-dimensional land use characteristic factors and the combined new characteristic factors are used together as the third characteristic factor cluster;

[0065] Furthermore, the first city annotation data is data obtained from an open source platform;

[0066] Furthermore, the preprocessing and data reconstruction of the acquired first city label data to obtain the reconstructed second city label data specifically includes:

[0067] (1) Log in to the local data acquisition system and configure the API interface between the local data acquisition system and the open source platform;

[0068] (2) After the login identity is verified, select the first city annotation data in the target area for download;

[0069] (3) preprocessing the acquired first city labeled data, including data cleaning, data correction, and data reconstruction;

[0070] Furthermore, the data cleaning specifically includes:

[0071] ① Matching the spatial location information and name identifiers of each specific point of the first city annotation data, filtering out the first city annotation data that is duplicated or has too small a data volume and deleting it;

[0072] ②Manually recheck the data after the above deletion;

[0073] Furthermore, the data correction specifically includes:

[0074] Check and fill in the gaps in the first city’s labeled data;

[0075] Furthermore, the data reconstruction specifically includes:

[0076] According to the current data analysis requirements, redundant attribute field data in the first city's labeled data is deleted, and only the location information, identification information, and classification information in the first city's labeled data are retained to achieve data reconstruction and lightweighting;

[0077] (4) reclassifying and format-converting the first city annotated data; wherein the reclassified first city annotated data includes 18 characteristic factors; and the format-converted first city annotated data has the same format as the first spatial data;

[0078] (5) performing spatial matching and gridding on the first city annotation data after reclassification and format conversion, and cutting the data range based on the basic sample data to obtain second city annotation data; wherein the second city annotation data has the same grid size and spatial coordinate reference as the basic sample data;

[0079] Furthermore, based on the reconstructed second city annotated data, characteristic sub-factors corresponding to the 18 characteristic factors are obtained, and finally a fourth characteristic factor cluster is formed, which specifically includes obtaining the characteristic sub-factors of the 18 characteristic factors:

[0080] (1) Calculate the number of city annotation data points that fall within each grid and obtain the first characteristic sub-factor of the city annotation data;

[0081] (2) Calculate the distance between each grid and the nearest city annotation data point, obtain the minimum distance between each grid and the non-internal city annotation data point, and obtain the second characteristic sub-factor of the city annotation data;

[0082] (3) Calculation of kernel density related factors: According to the characteristics of the target area, through feature screening, the grid size r1 for large-scale kernel density calculation and the grid size r2 for small-scale kernel density calculation are obtained respectively.

[0083] Calculate the kernel density of the city annotation data points in the entire domain with the distance of r1 grids as the radius, obtain the corresponding large-scale kernel density, and obtain the third characteristic sub-factor of the city annotation data;

[0084] Calculate the kernel density of the city annotation data points in the whole domain with the distance of r2 grids as the radius, obtain the corresponding small-scale kernel density, and obtain the fourth characteristic sub-factor of the city annotation data;

[0085] (4) A reference range with a radius of 2,500 meters is established for the city annotation data points in a fusion mode. If a point falls within the reference range, it is deemed to fall within the service range of the city annotation data point of this type. All the corresponding geographic space area pixels within the range are assigned a value of 1, and the geographic space area pixels outside the reference range are assigned a value of 0. The data after the assignment is used as the fifth characteristic sub-factor of the city annotation data;

[0086] (5) The characteristic sub-factors of the above five types of city annotation data are integrated to obtain the characteristic factors of each type of city annotation data. The characteristic factor data of the 18 types of city annotation data together form the fourth characteristic factor cluster.

[0087] Furthermore, the first urban building data at least includes building outline data, building height data, and unique identification data of each building;

[0088] Furthermore, the first city building data can be directly downloaded from an open source platform database, or obtained based on further processing of existing data;

[0089] Furthermore, the further processing based on the existing data may be:

[0090] Obtaining corresponding first city building data based on the fusion processing of the second spatial data, the third spatial data and the corresponding building outline data;

[0091] Furthermore, the building outline data can be directly obtained from an existing open source database, or can be automatically extracted from image data corresponding to the study area based on a deep neural network model;

[0092] Furthermore, the obtaining of a plurality of corresponding characteristic factors based on the second city building data to form a fifth characteristic factor cluster specifically includes:

[0093] (1) Calculate the area of ​​the building based on the building outline data, and calculate the volume of the building based on the building height data;

[0094] (2) The area and volume attributes of the building are used as two characteristic factors of the second city building data to form the fifth characteristic factor cluster.

[0095] Furthermore, the feature screening specifically includes:

[0096] (1) vectorizing the data to be feature screened having multiple feature factors to obtain vector data of the data to be feature screened respectively;

[0097] (2) Traversing each feature factor in the vector data of the data to be feature screened, randomly shuffling the feature values ​​in each feature factor, and inputting the feature factors and other feature factors that have not been randomly shuffled into a subsequent processing model, comparing the accuracy of the model output result with the verification data to determine whether it meets a preset threshold;

[0098] (3) Repeat the process in step (2) above until all characteristic factors have been subjected to eigenvalue shuffling and model processing to obtain corresponding accuracy verification results;

[0099] (4) The characteristic factors whose errors in the accuracy verification results during the processing of steps (2)-(3) are retained as input characteristic factors for the subsequent population prediction model.

[0100] Furthermore, the construction of the multivariate regression population distribution model specifically includes:

[0101] S3-1, Sample data preparation: Population survey data is used as the basic sample, and the multidimensional features formed by the corresponding five characteristic factor clusters are used as sample data for model construction.

[0102] S3-2, sample data generation: Calculate population density according to the geographic space grid, use this data as the sample dependent variable, and number the grid ID of the target area as the grid identifier. The multidimensional modeling features extracted from the grid are used as sample independent variables to form the training sample.

[0103] S3-3, model training: The grid pixels of the target area with multi-dimensional data features and unique ID numbers are used as model input, and the population model data of the same pixel size is used as output. The population distribution model is constructed and trained based on the random forest model. In the process of model training, a random sampling method with replacement is adopted. During the training process, some original samples are selected as training data, and the remaining sample data constitute out-of-bag data.

[0104] S3-4, model optimization: evaluate the accuracy of the random forest model through out-of-bag samples; optimize the parameters of the random forest framework and decision tree during training.

[0105] Furthermore, in step S4, the model accuracy verification specifically includes:

[0106] The mean relative error (MRE), root mean square error (RMSE) and %RMSE are used as evaluation indicators.

[0107] MRE represents the mean absolute error between the population forecast value and the true population value, reflecting the actual situation of the population forecast value error, which is expressed as:

[0108]

[0109] Where n is the number of grids containing real population. Pa represents the predicted population of the ath grid, and Ra represents the real population of the ath grid. Since MRE can only estimate grids containing real population, it cannot show the error impact of grids that originally have no population but are overestimated. Therefore, the root mean square error (RMSE) is used to evaluate the accuracy of grid-level population estimation:

[0110]

[0111] %RMSE represents the root mean square error between the population prediction value and the population true value divided by the mean of the population true value, reflecting the overall accuracy of the model simulation. The calculation formula is as follows:

[0112]

[0113] Compared with the existing technical solutions, the technical solution of the present invention has the following advantages: the technical solution of the present invention details the factors affecting the population molecules by extracting multidimensional characteristic factors from multi-source data, controls the effective data from the source of the data, and improves the accuracy of the population distribution model prediction; in the multidimensional feature extraction of the technical solution of the present invention, a semi-supervised method is used to extract the characteristic influencing factors according to the characteristics of some key data themselves, so that the subsequent population distribution prediction is more in line with reality and has higher accuracy; in the technical solution of the present invention, the final population distribution data results are rasterized, which is conducive to the analysis, processing and storage of population model data. BRIEF DESCRIPTION OF THE DRAWINGS

[0114] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are 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.

[0115] Figure 1 A schematic diagram of the process of constructing a population spatial distribution model by fusing features derived from multi-source data under semi-supervision provided in an embodiment of the present disclosure.

[0116] Figure 2 A schematic diagram of the feature factor extraction process of corresponding nighttime light data provided in an embodiment of the present disclosure.

[0117] Figure 3A schematic diagram of the feature factor extraction process of city annotation data provided in an embodiment of the present disclosure.

[0118] Figure 4 Schematic diagram of the target research area involved in the embodiments of the present disclosure.

[0119] Figure 5 It is a schematic diagram of the cropping effect of the second spatial data involved in the embodiment of the present disclosure.

[0120] Figure 6 It is a schematic diagram of the model construction effect without adding the hyperparameter terrain feature factor to the population construction model involved in the embodiments of the present disclosure.

[0121] Figure 7 A schematic diagram of the model construction effect of adding a hyperparameter terrain feature factor to the population construction model involved in the embodiments of the present disclosure. DETAILED DESCRIPTION

[0122] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0123] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are configured to distinguish similar objects, and are not necessarily configured to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0124] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a product or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such a product or system. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the product or system including the elements.

[0125] The disclosed embodiment of the present invention discloses a method for constructing a population spatial distribution model by fusing features derived from multi-source data under semi-supervision, which is described below.

[0126] The disclosed embodiment is mainly divided into five key steps: target area selection and processing, multi-source data preprocessing and feature extraction, multivariate regression population distribution model construction, model accuracy assessment, and population distribution prediction. Among them, the target area selection and processing mainly involves rasterizing the obtained census data to construct a population spatialization data set of a geographic grid, which can be used as sample data. Multi-source data preprocessing and feature extraction mainly involve aligning the geographic coordinate system and grid of various types of data to extract feature dimension information that contributes positively. The multivariate regression population distribution model construction is to train the sample data, so as to obtain the optimal hyperparameters of the model for data in different regions, and finally obtain an accurate population distribution model. The model accuracy assessment mainly involves comparing and analyzing the estimated population obtained by deducing the population distribution model with the actual population, and quantifying the accuracy of the model, so as to verify the practicality of the model.

[0127] Figure 1 A flowchart of a first embodiment of a method for constructing a population spatial distribution model by integrating features derived from multi-source data under semi-supervision provided by an embodiment of the present invention, the method comprising the following steps:

[0128] S1, multi-source data acquisition: obtain the latest census data of the target area, the corresponding geographic spatial data of the target area, night light data, land use data, city annotation data, and urban building data;

[0129] S2, preprocessing multi-source data, extracting characteristic factor clusters, and storing data: preprocessing the census data; extracting characteristic factors and constructing characteristic factor clusters for the geographic spatial data corresponding to the target area, the night light data corresponding to the target area, the land use corresponding to the target area, the city annotation data corresponding to the target area, and the urban building data corresponding to the target area; storing the constructed characteristic factor cluster data in the population spatial characteristic database of the study area.

[0130] S3, construction and training of multivariate regression population distribution model;

[0131] S4, Verification of the accuracy of the population distribution model.

[0132] S5, based on the trained model, semi-supervised population distribution prediction is performed on the data of different target areas to obtain the rasterized model prediction results.

[0133] In some embodiments, in step S1, the geospatial data corresponding to the population census data of the region includes: first spatial data, second spatial data, and third spatial data having physical world location information and identification;

[0134] In some embodiments, the first spatial data carries attribute field information of the administrative district code and administrative district name of the target area;

[0135] The first spatial data of the target research area and its spatial scope can be obtained from online public data and related map websites. Population statistics can be obtained from local statistical bureaus or the official website of the National Bureau of Statistics. These two types of data constitute the base map data of the embodiment of the present application.

[0136] In some embodiments, the third spatial data is extracted based on the SRTM (Shuttle Radar Topography Mission) data obtained after splicing and cropping;

[0137] In some embodiments, in step S2, the preprocessing of multi-source data and the extraction of characteristic factor clusters specifically include:

[0138] Gridding the first spatial data to obtain fifth spatial data, matching the population census data to the fifth spatial data to obtain fused model basic sample data; wherein the grid size of the fifth spatial data is set to 100m×100m;

[0139] In the embodiment of the present application, the population sample data is matched to the corresponding grid according to the administrative area code or name attribute field, and the corresponding population data in each grid is calculated based on the grid size and actual population density, thereby obtaining the initial sample data of the population distribution model.

[0140] According to the experimental needs, the first spatial data is gridded in the data preparation stage to obtain a corresponding grid as a subsequent reference, so that the final result generated by the present invention is a gridded population distribution model, which is convenient for post-processing and application of the model data.

[0141] Performing spatial matching and gridding on the third spatial data to obtain sixth spatial data, and extracting characteristic factors based on the sixth spatial data to obtain a first characteristic factor cluster, wherein the sixth spatial data has the same grid size and spatial coordinate reference as the basic sample data;

[0142] The third spatial data may be digital elevation model data. Digital elevation model data is highly correlated with population distribution and is therefore usually used as an important characteristic molecular factor in population distribution prediction models.

[0143] Acquire first nighttime light data, perform spatial matching and gridding on a spatial range involved in the first nighttime light data, and cut the data range based on the basic sample data to obtain second nighttime data, wherein the second nighttime data has the same grid size and spatial coordinate reference as the basic sample data;

[0144] Acquire the third nighttime light data, perform data splicing on the third nighttime light data, perform spatial matching and gridding on the third nighttime light data, and cut the data range based on the basic sample data to obtain the fourth nighttime data; wherein the fourth nighttime data has the same grid size and spatial coordinate reference as the basic sample data;

[0145] A second characteristic factor cluster is obtained based on the fusion calculation of the second nighttime light data and the fourth nighttime light data.

[0146] The first night light data may be the night light data acquired by the "Luojia-1" satellite, and the third night light data may be the NPP-VIIRS night light data. Among them, the "Luojia-1" night light data requires multiple scenes to fully cover the study area, so the data needs to be spliced ​​and then cropped, and the NPP-VIIRS night light data has a larger coverage area, so it is cropped according to the scope of the study area. Secondly, the two night light data with different resolutions are matched from different original resolutions to the same 100m×100m grid as the basic sample data. Then the original spatial reference of the two night light data used is converted to the same spatial reference framework as the basic sample data. Finally, the processed night light data is extracted to obtain the second characteristic factor cluster. The night light data processing process is as follows Figure 2 shown.

[0147] The nighttime light data of "Luojia-1" has high spatial resolution and image quality. Due to the storage problem of floating-point data, the absolute radiation-corrected "Luojia-1" data needs to be converted when using brightness. The formula is as follows:

[0148]

[0149] Where L is the radiation brightness value after absolute radiation correction, and DN is the grayscale value of the night light data of "Luojia-1".

[0150] Classification processing is performed based on the currently acquired sentinel image data to obtain a number of first land use data after classification;

[0151] Performing spatial matching and gridding on the plurality of blocks of first land use data, and cutting and splicing the plurality of blocks of first land use data based on the basic sample data to obtain second land use data;

[0152] Based on the statistics and calculation of the second land use data, a third characteristic factor cluster is obtained.

[0153] Acquire the first city annotated data, perform spatial matching and gridding on the first city annotated data, and perform preprocessing and data reconstruction to obtain reconstructed second city annotated data;

[0154] Based on the reconstructed second city labeled data, a corresponding plurality of characteristic factors are obtained to form a fourth characteristic factor cluster.

[0155] Acquire first city building data, perform spatial matching and gridding on the first city building data, and cut the data range based on the basic sample data to obtain second city building data; wherein the second city building data has the same grid size and spatial coordinate reference as the basic sample data;

[0156] Based on the second city building data, a corresponding plurality of characteristic factors are obtained to form a fifth characteristic factor cluster.

[0157] In some embodiments, matching the census data to the fifth data to obtain the fused fifth data specifically includes:

[0158] The census data is matched to the corresponding grid according to administrative attributes, and the corresponding population data in each grid is calculated according to the grid size and actual population density, so as to obtain the initial sample data of the population distribution model after preliminary fusion, wherein the sample data formed includes population data fields.

[0159] In some embodiments, the sixth spatial data has the same grid size and spatial coordinate reference as the basic sample data.

[0160] Furthermore, the first characteristic factor cluster includes:

[0161] Elevation characteristic factor, slope characteristic factor, aspect characteristic factor, and hyperparameter terrain characteristic factor.

[0162] Since topography has a great influence on population distribution, and the original population aggregation in a region is often determined by topography, elevation, slope, aspect, and hyperparameter terrain features are extracted based on the third spatial data reflecting this feature. These four types of data will be used as four characteristic factors derived from the third spatial data to construct a population spatial distribution model.

[0163] Elevation: Elevation is an important indicator that affects the comfort of human life and can be directly obtained through interpolation of corresponding raster data.

[0164] Slope: Slope affects the population spatial pattern by affecting the cost and convenience of various human constructions. Based on the third spatial data, the present invention uses the ArcGIS slope command to obtain the slope raster value, and further applies the partition statistics in the spatial analysis module to calculate the average slope of each area, and uses it as the slope value feature.

[0165] Slope aspect: Slope aspect affects the spatial layout of population by affecting the orientation and planning of human buildings. Based on DEM data, the present invention uses ArcGIS slope aspect command to obtain the slope aspect raster value and uses it as the slope aspect value feature.

[0166] In addition, there are also a large number of derivative features in the prior art that are obtained by further processing and analyzing elevation, slope, slope direction, etc., such as terrain undulation. The research on terrain undulation has a long history, and its calculation method has been continuously improved, such as the method of only calculating the elevation difference, the calculation method proposed by Niu Wenyuan et al., etc. However, the corresponding evaluation factors obtained by this type of calculation method for different types of terrain areas have insufficient accuracy when used for subsequent population distribution predictions. To address this problem, the present application innovatively invented a calculation method of hyperparameter terrain characteristic factor based on multiple experiments. Through this method, the operator can adjust the set influencing parameters according to the experimental target data for different areas, so as to obtain relevant terrain influencing factors in a semi-supervised manner for the construction of subsequent population prediction models.

[0167] Furthermore, the extracting characteristic factors based on the sixth spatial data to obtain a first characteristic factor cluster specifically includes:

[0168] (1) Obtaining an elevation factor by performing data interpolation on the sixth spatial data; specific data interpolation methods may be linear interpolation, bilinear interpolation, inverse distance weighted interpolation, weighted average method, moving fitting method, Kriging interpolation, natural neighbor interpolation method, spline interpolation method, and other methods and their improvements.

[0169] (2) extracting slope grid values ​​from the sixth spatial data; performing partition statistics on the extracted slope grid values ​​to calculate the average slope value of each area, and using it as a slope characteristic factor.

[0170] (3) extracting aspect grid values ​​from the sixth spatial data; performing partition statistics on the extracted aspect grid values ​​to calculate the average slope value of each area and use it as the aspect characteristic factor.

[0171] (4) Based on the elevation characteristic factors obtained above, the hyperparameter terrain characteristic factors are further calculated.

[0172] (5) Obtaining the first characteristic factor cluster based on the fusion of the elevation factor, the slope characteristic factor, the aspect characteristic factor, and the hyperparameter terrain characteristic factor.

[0173] In some embodiments, based on the elevation characteristic factor obtained above, a hyperparameter terrain characteristic factor is further calculated, specifically including:

[0174] ① Based on the elevation characteristic factor and the slope characteristic factor extracted from the sixth spatial data, the mean elevation Hp and the difference between the highest and lowest elevations Hc in the target area are calculated; and,

[0175] ② Calculate the proportion β of the area in the target area where the slope characteristic factor meets the specified threshold α to the area of ​​the target area;

[0176] ③ Based on the following algorithm, the hyperparameter terrain characteristic factor A is calculated:

[0177] δ 1 A=Hp+Hc*β*δ 2

[0178] In the above formula, δ 1 is the plane scale parameter, which is generally determined according to the average elevation Hp, δ 2 is the terrain feature influencing parameter, representing the importance of the ground area elevation fluctuation on the overall impact, where δ 2 >0.

[0179] ④Set the analysis radius of the target area;

[0180] ⑤ Based on the analysis radius set, the plane scale parameter δ 1 , terrain characteristics affect parameter δ 2 Make adaptive settings;

[0181] ⑥ Calculate the hyperparameter terrain characteristic factor A based on the set parameters.

[0182] In some embodiments, the calculation of the ratio β of the area range of the target area where the slope characteristic factor meets the specified threshold α to the area range of the target area includes:

[0183] The slope characteristic factor data in the target area are statistically analyzed; based on the statistical results, the slope characteristic factors are clustered based on the threshold α to obtain the slope characteristic factor classification that meets the threshold α, and further obtain the proportion β of the area range that meets the specified threshold α to the range of the target area.

[0184] When obtaining the hyperparameter terrain characteristic factor based on the above formula, it is necessary to set the analysis window, traverse the target analysis area based on the set analysis window, merge all the hyperparameter terrain characteristic factor values ​​obtained in the traversal and solution process, and obtain the final hyperparameter terrain characteristic factor of the target analysis area. In the above process, the analysis window uses a rectangular window, and searches and extracts within the range of x km×x km with each grid as the center.

[0185] The selection of the rectangular window size x can be set according to the specific target analysis object. When x takes different values, in order to ensure the model simulation effect, the corresponding δ 1 and δ 2 The value is adjusted accordingly; for example, the x value can be set to 0.9. After experimental verification of the model prediction results, the value is set to δ 1 =1000,δ 2 =2 has the best model prediction effect.

[0186] In the prior art, there are many algorithms based on factors such as elevation, slope, aspect and various characteristic factors derived from them, all of which have achieved considerable technical application effects. Based on the research of predecessors, this invention innovatively proposes a method for calculating hyperparameter terrain characteristic factors in order to realize flexible construction of population models under specific conditions, combined with current scientific research projects. The hyperparameters in the calculation formula can be set manually according to different characteristics, realizing semi-supervised extraction of the characteristic factors and improving the accuracy of population model calculation. Experiments show that after adding hyperparameter terrain characteristic factors, the accuracy of population model construction can be significantly improved for areas with large elevation changes.

[0187] In some embodiments, the second characteristic factor cluster is obtained by fusion calculation based on the second nighttime light data and the fourth nighttime light data, specifically including:

[0188] (1) Extraction of DN values ​​of night light data: Extract the DN values ​​of the second night light data and the fourth night light data respectively, and map the data processed with DN values ​​to the light data that has undergone matching, spatial matching, gridding, and cropping operations through linear interpolation to form characteristic factors of the two types of night light data.

[0189] (2) Extracting brightness values ​​of nighttime light data: Calculating and converting the brightness value L of the fourth nighttime light data, and mapping the calculated corresponding brightness value to the fourth nighttime light data to form a characteristic factor of the nighttime light data.

[0190] (3) Binarization of night light data: Based on the data characteristics of the second night light data and the fourth night light data and the characteristics of the target area, feature screening is performed, and the selected feature values ​​of the brightness value of the second night light data are set to a1, a2, and a3; the selected feature values ​​of the brightness value of the fourth night light data are set to a1%, a2%, and a3% of the highest value. When each type of data corresponds to the corresponding threshold, if it meets the corresponding threshold, the pixel value of the corresponding target area is assigned to 255; if it does not meet the corresponding threshold, the pixel value of the corresponding target area is assigned to 0. After traversing each pixel assignment, the binary data of the three types of night light data generated by the second night light data and the fourth night light data are obtained as the feature factors of the corresponding dimensions, totaling 6 types of feature factors of night light data.

[0191] The stable light data brightness value is between 0-63. In this application, the values ​​of a1, a2, and a3 can be 80, 60, and 40 respectively, or they can be set to other numerical thresholds after feature screening according to the specific target area to obtain better model simulation effects. Night light data from two different data sources are used to extract multiple threshold features. By flexibly limiting the threshold, artificial parameters can be set for regions and cities with different characteristics. Compared with directly using the light brightness value data as data for population distribution prediction, this application uses artificial settings to perform semi-supervised intervention on the light data, which can speed up feature training efficiency and improve the accuracy of training results.

[0192] (4) A total of nine characteristic factors of the night light data obtained by combining the DN value of the night light data, the brightness value L of the night light data, and the binary data of the night light data, form a second characteristic factor cluster.

[0193] In some embodiments, in step S2-4, the classification processing is performed based on the currently acquired sentinel image data to obtain a plurality of classified first land use data, specifically including:

[0194] (1) Acquire sentinel image data corresponding to the target area;

[0195] (2) classifying land cover types based on a deep learning algorithm to obtain a plurality of first land use data after classification;

[0196] The grid size of the sentinel image data is 10m, and each piece of the classified first land use data contains 10 land use types;

[0197] In some embodiments, the third characteristic factor cluster is obtained based on statistics and calculations of the second land use data, specifically including:

[0198] extracting 10 land use types from the second land use data as 10-dimensional land use characteristic factors;

[0199] The 10 land use types are merged into a new characteristic factor for assignment, so that the new characteristic factor contains all land use types and is represented by different codes;

[0200] The 10-dimensional land use characteristic factors and the merged new characteristic factors are taken together as the third characteristic factor cluster.

[0201] The land use data used in the present invention is global land cover data produced by Esri based on 10-meter resolution Sentinel image data using deep learning methods, which includes ten types of land features, namely: water bodies, woodlands, grasslands, submerged vegetation, cultivated land, shrubs, buildings, bare land, snow / ice, and clouds. The land classification data is presented in the form of a spatial grid. When processing the data, first find one or more geographic grid data containing the study area and download them locally. Then use the vector boundary data of the study area as a mask to extract the downloaded data, and finally splice the extracted results.

[0202] In the process of population spatialization, land use data is often considered separately for different land use types in practical applications, so 10 land use types can generate 10-dimensional land use data features. At the same time, the present invention merges all land use types into a new feature for assignment. This feature contains all land use types and is represented by different codes. For example, the 10 land use types are expressed as labels using the numbers [0], [1]...[9], respectively. The 11th expression factor formed by the merger is the label [0,1,2,3,4,5,6,7,8,9]. The land use data derived features have a total of 11 dimensions. After extraction, they are distributed through spatial connection and added to the geographic grid. In the above process, in order to achieve data alignment, the labels of a single land use type are padded to facilitate data processing; for example, [0] is padded to [0,0,0,0,0,0,0,0,0,0], and [1] is padded to [1,0,0,0,0,0,0,0,0,0].

[0203] In some embodiments, the first city annotation data is data obtained from an open source platform;

[0204] In some embodiments, the preprocessing and data reconstruction of the acquired first city label data to obtain the reconstructed second city label data specifically includes:

[0205] (1) Log in to the local data acquisition system and configure the API interface between the local data acquisition system and the open source platform;

[0206] (2) After the login identity is verified, select the first city annotation data in the target area for download;

[0207] (3) preprocessing the acquired first city labeled data, including data cleaning, data correction, and data reconstruction;

[0208] In some embodiments, the data cleaning specifically includes:

[0209] ① Matching the spatial location information and name identifiers of each specific point of the first city annotation data, filtering out the first city annotation data that is duplicated or has too small a data volume and deleting it;

[0210] ②Manually recheck the data after the above deletion;

[0211] In some embodiments, the data modification specifically includes:

[0212] Check and fill in the gaps in the first city’s labeled data;

[0213] In some embodiments, the data reconstruction specifically includes:

[0214] According to the current data analysis requirements, redundant attribute data in the first city's labeled data is deleted, and only the location information, identification information, and classification information in the first city's labeled data are retained to achieve data reconstruction and lightweighting;

[0215] (4) reclassifying and format-converting the first city annotated data; wherein the reclassified first city annotated data includes 18 characteristic factors; and the format-converted first city annotated data has the same format as the first spatial data.

[0216] The first city annotation data can be obtained from an open source platform or database, such as the AutoNavi platform or the Baidu platform. After the initial first city annotation data is obtained, it is necessary to clean and eliminate redundant information in the data. It mainly includes data reconstruction and correction. Data reconstruction is to unify the acquired city annotation data, extract text information, retain the spatial attributes of the city annotation data such as longitude, latitude, name, and classification, and clear other data. At this time, a multidimensional table composed of all city annotation data is obtained, and data reconstruction is completed.

[0217] Finally, the processed city annotation data is cleared of duplicate data, that is, the spatial location information and name information of all city annotation data points are matched, and possible duplicate city annotation data points are removed. Then, all the city annotation data to be removed are output in csv format and manually reviewed to avoid erroneous data removal.

[0218] The cleaned city annotation data also needs to be reclassified and coordinate converted. Since the original city annotation data has too many categories and the amount of some types of city annotation data is too small, the original city annotation data is reclassified, mainly to merge some city annotation data types, and remove some city annotation data with too small data in the category. Finally, the city annotation data is batch format converted. The specific acquisition and processing process of city annotation data is as follows: Figure 3 shown.

[0219] (5) performing spatial matching and gridding on the first city annotation data after reclassification and format conversion, and cutting the data range based on the basic sample data to obtain the second city annotation data; wherein the second city annotation data has the same grid size and spatial coordinate reference as the basic sample data.

[0220] In some embodiments, the obtaining of characteristic sub-factors corresponding to the 18 characteristic factors based on the reconstructed second city annotated data and finally forming a fourth characteristic factor cluster specifically includes obtaining the characteristic sub-factors of the 18 characteristic factors:

[0221] (1) Calculate the number of city annotation data points that fall within each grid and obtain the first characteristic sub-factor of the city annotation data;

[0222] (2) Calculate the distance between each grid and the nearest city annotation data point, obtain the minimum distance between each grid and the non-internal city annotation data point, and obtain the second characteristic sub-factor of the city annotation data;

[0223] (3) Calculation of kernel density related factors: According to the characteristics of the target area, through feature screening, the grid size r1 for large-scale kernel density calculation and the grid size r2 for small-scale kernel density calculation are obtained respectively.

[0224] Calculate the kernel density of the city annotation data points in the entire domain with the distance of r1 grids as the radius, obtain the corresponding large-scale kernel density, and obtain the third characteristic sub-factor of the city annotation data;

[0225] Calculate the kernel density of the city annotation data points in the whole domain with the distance of r2 grids as the radius, obtain the corresponding small-scale kernel density, and obtain the fourth characteristic sub-factor of the city annotation data;

[0226] (4) A reference range with a radius of 2,500 meters is established for the city annotation data points in a fusion mode. If a point falls within the reference range, it is deemed to fall within the service range of the city annotation data point of this type. All the corresponding geographic space area pixels within the range are assigned a value of 1, and the geographic space area pixels outside the reference range are assigned a value of 0. The data after the assignment is used as the fifth characteristic sub-factor of the city annotation data;

[0227] (5) The characteristic sub-factors of the above five types of city annotation data are integrated to obtain the characteristic factors of each type of city annotation data. The characteristic factor data of the 18 types of city annotation data together form the fourth characteristic factor cluster.

[0228] The acquired city annotation data can be classified according to its function. Each city annotation data has multidimensional features, which provide information such as the spatial distribution location of the city annotation data, the type of city annotation data, administrative district, name, address, etc. The city annotation data reflects the intensity of basic service construction and economic activities. In urban areas, the city annotation data is highly concentrated, the population density is high, the basic service facilities are relatively complete, and the economic activities are strong. Extending outward from the urban area, the aggregation phenomenon of the city annotation data gradually weakens and the population density gradually decreases. After a certain range away from the urban area, the city annotation data is obviously discrete and the population density is obviously lower. This data and the building data play an important role in the population distribution model. Therefore, the present invention combines the city annotation data with the building data, and realizes the reasonable expression of the population distribution model by extracting the spatial distance between the building and the city annotation data.

[0229] The present invention extracts multi-dimensional features of city annotation data for population spatial estimation. After reclassification, the city annotation data has N categories (the number of reclassification categories is determined according to the city annotation data of the target research area). In addition, all types of city annotation data are regarded as a separate derivative feature extraction. That is, the above five feature sub-factors are extracted for N+1 types of city annotation data, totaling (N+1)-dimensional features. After all features are extracted, their attributes are matched and fused into the corresponding grids of the basic sample data.

[0230] In some embodiments, the first urban building data includes at least building outline data, building height data, and unique identification data of each building;

[0231] In some embodiments, the first city building data may be directly downloaded from an open source platform database, or obtained based on further processing of existing data;

[0232] In some embodiments, the further processing based on the existing data may be:

[0233] Obtaining corresponding first city building data based on the fusion processing of the second spatial data, the third spatial data and the corresponding building outline data;

[0234] In some embodiments, the building outline data may be directly obtained from an existing open source database, or may be automatically extracted from image data corresponding to the study area based on a deep neural network model;

[0235] In some embodiments, the obtaining of a corresponding plurality of characteristic factors based on the second city building data to form a fifth characteristic factor cluster specifically includes:

[0236] (1) Calculate the area of ​​the building based on the building outline data, and calculate the volume of the building based on the building height data;

[0237] (2) The area and volume attributes of the building are used as two characteristic factors of the second city building data to form the fifth characteristic factor cluster.

[0238] The first city building data can be downloaded through the Internet, but there will generally be missing information. Therefore, the present invention also uses the collected high-precision image data and the existing second spatial data to extract the building boundaries. If the second spatial data of the study area cannot be obtained, a deep neural network model such as Mask RCNN and HTC can be used to directly extract the building instance objects on the high-precision image data. If the second spatial data of the study area can be obtained, it can be combined with the high-precision image to extract the outline and height of the building through information such as shadows to improve the accuracy of the constructed population spatial distribution model.

[0239] After the building outline is extracted, the area of ​​the building is calculated based on the building outline data, and the maximum value of the second spatial data within the building outline is used as the building height to calculate the building volume. The area and volume attributes of the building are used as two features in the feature library.

[0240] In some embodiments, the feature screening specifically includes:

[0241] (1) vectorizing the data to be feature screened having multiple feature factors to obtain vector data of the data to be feature screened respectively;

[0242] (2) Traversing each feature factor in the vector data of the data to be feature screened, randomly shuffling the feature values ​​in each feature factor, and inputting the feature factors and other feature factors that have not been randomly shuffled into a subsequent processing model, comparing the accuracy of the model output result with the verification data to determine whether it meets a preset threshold;

[0243] (3) Repeat the process in step (2) above until all characteristic factors have been subjected to eigenvalue shuffling and model processing to obtain corresponding accuracy verification results;

[0244] (4) The characteristic factors whose errors in the accuracy verification results during the processing of steps (2)-(3) are retained as input characteristic factors for the subsequent population prediction model.

[0245] In some embodiments, the multivariate regression population distribution model construction specifically includes:

[0246] S3-1, Sample data preparation: Population survey data is used as the basic sample, and the multidimensional features formed by the corresponding five characteristic factor clusters are used as sample data for model construction.

[0247] The invention of this application introduces a multi-source data feature factor cluster to comprehensively train the model for population distribution prediction. Some feature factor clusters contain a large number of feature sub-factors. The purpose of feature screening is to retain the feature sub-factors that have a positive effect on the training accuracy of the population distribution prediction model, and to eliminate the feature sub-factors that have a small or no positive effect, thereby achieving data dimensionality reduction and reducing the amount of data, thereby avoiding unnecessary calculations and the adverse effects of noise data.

[0248] The feature screening operation may further perform feature screening after obtaining multi-dimensional features according to the target area and data conditions. The specific steps may include:

[0249] (1) vectorizing the first characteristic factor cluster, the second characteristic factor cluster, the third characteristic factor cluster, the fourth characteristic factor cluster, and the fifth characteristic factor cluster to obtain their corresponding vector data respectively;

[0250] (2) retaining all the features of the fourth and fifth characteristic factor clusters, and correspondingly increasing the weights of the vector data of the fourth and fifth characteristic factor clusters in subsequent model construction;

[0251] (3) Traversing each characteristic factor in the first characteristic factor cluster, the second characteristic factor cluster, and the third characteristic factor cluster, randomly shuffling the characteristic values ​​in each characteristic factor, and inputting them together with other characteristic factors whose characteristic values ​​have not been randomly shuffled, into the subsequent processing model, and comparing the accuracy of the model output results with the verification data to determine whether they meet the pre-set threshold;

[0252] (4) Repeat the process in step (3) above until all characteristic factors have been subjected to eigenvalue shuffling and model processing to obtain corresponding accuracy verification results;

[0253] (5) The characteristic factors whose errors in the accuracy verification results during the processing of steps (3)-(4) are retained as input characteristic factors for the subsequent population prediction model.

[0254] S3-2, sample data generation: Calculate population density according to the geographic space grid, use this data as the sample dependent variable, and number the grid ID of the target area as the grid identifier. The multidimensional modeling features extracted from the grid are used as sample independent variables to form the training sample.

[0255] S3-3, model training: The grid pixels of the target area with multi-dimensional data features and unique ID numbers are used as model input, and the population model data of the same pixel size is used as output. The population distribution model is constructed and trained based on the random forest model. In the process of model training, a random sampling method with replacement is adopted. During the training process, some original samples are selected as training data, and the remaining sample data constitute out-of-bag data. The output population model data corresponds to the data representing the population value of the corresponding area, and the format is an M×N×1-dimensional array of positive integers.

[0256] S3-4, model optimization: evaluate the accuracy of the random forest model through out-of-bag samples; optimize the parameters of the random forest framework and decision tree during training.

[0257] The present invention uses a random forest model for sample training and population distribution model construction. Based on the above-constructed population spatial feature database of the research area, the population survey data is used as a sample, and the multidimensional features after preprocessing and screening are used to construct the model. The present invention first calculates the population density according to the geographic space grid, and the data is used as the sample dependent variable, and the grid ID of the target area is numbered as the grid identifier. The modeling features extracted from the grid are used as sample independent variables to jointly constitute the training sample.

[0258] In the process of model training, a random sampling method with replacement is adopted. During the training process, about 70% of the original samples are used as training data, and the remaining 30% of the sample data constitute out-of-bag data. The oob_score parameter of the random forest model is called, and the accuracy of the random forest model, i.e., the generalization performance of the model, is evaluated by out-of-bag samples. The training process of the random forest model is mainly the adjustment process of the model hyperparameters, which is generally referred to as model parameter adjustment. All kinds of machine learning algorithms involve model parameter adjustment problems. Many models have many parameters, and different parameter value combinations will produce different effects on different practical problems. Therefore, there is no set of parameters that can optimize the model effect in various problems, and it is necessary to achieve a parameter combination suitable for a certain type of problem through continuous training and adjustment. The random forest model parameters used in the present invention mainly include two categories: random forest framework parameters (also known as bagging framework parameters) and decision tree parameters.

[0259] The present invention uses the GridSearchCV function of the scikit-learn library to adjust the parameters of the model. The principle of this method is to set a parameter dictionary containing all candidate value ranges of all parameters, and try each parameter to be optimized one by one according to the set value range. The search is stopped when the parameter combination with the best verification result is obtained. Since the grid search determines the influence of each candidate value of each parameter on the accuracy of the model, the global optimal parameter combination can be obtained. This method is suitable for most data sets with small or moderate data volumes. If the data volume is huge, using grid search to adjust the parameters will result in a huge amount of calculation, so other parameter adjustment methods can be considered.

[0260] In some embodiments, other models may also be selected for population prediction.

[0261] Existing studies have shown that the random forest model itself has a high accuracy, and can adapt to the situation of more dimensional feature factors defined in this application, and can adapt to the purpose of this application to improve the prediction accuracy of the population prediction model. However, there are also some problems in itself: for example, the model itself is relatively poor in interpretability: the random forest is a black box model, and it is difficult to explain the decision-making process of a single tree. Although the overall model can be understood through the importance of features, the specific decision-making process is relatively opaque. For the study of population prediction models, researchers sometimes need to refine the specific mechanism of action to learn the influence mechanism of different dimensional research objects on the final prediction results during the research process, which is unfriendly at this time. For another example, the training method of the model itself causes it to require more resources. When the object data to be predicted is relatively large and involves data of more dimensions, the hardware resources required for training and the time spent are long, which is not suitable for some training needs of scientific researchers. For another example, the noise problem, in some cases, the random forest may be sensitive to data containing a lot of noise because it may overfit the noise. This may become a bottleneck for population prediction. The collection of population data and building vector data will inevitably have errors or mistakes, especially in areas with sparse population and buildings. The error impact will be greater. Although the relevant data has been cleaned in the data preprocessing stage, it is still impossible to completely avoid the retention of errors or wrong data. Based on this consideration, in the prediction after extracting multi-source data, the population prediction model can also be replaced by other models.

[0262] In some embodiments, when the grid size of the target study area is greater than the threshold d1 and the overall area of ​​the target study area is less than the threshold d2, or when the target study area is an important urban area, the random forest model is selected as the population prediction model. When the grid size of the target study area is less than the threshold d1 and the overall area of ​​the target study area is greater than the threshold d2, or when the target study area is a sparsely populated town or rural area, a convolutional neural network (CNN) is selected as the population prediction model.

[0263] As a neural network designed for processing image data, convolutional neural networks use convolutional layers to capture local features and pooling layers to reduce the dimension of data. The convolutional layer performs local convolution operations on the input data and uses a parameter sharing mechanism to reduce the number of parameters of the model. The pooling layer downsamples the output of the convolutional layer to reduce the dimension and computational complexity of the data. This structure is particularly suitable for processing image data. Therefore, it has a good effect on improving the efficiency of processing multi-dimensional feature parameters and spatial image data from multiple sources in this application.

[0264] Of course, as an early image processing model, convolutional neural networks also have some limitations, such as being unsuitable for processing time series data. When the target research area has relatively complete data for many consecutive years, historical data can be used as a characteristic factor for population distribution prediction.

[0265] In some embodiments, when the multi-source data of the target study area is greater than one time phase, such as having multi-source data of several consecutive or discontinuous years, a long short-term memory network (LSTM) is selected as the prediction model;

[0266] After feature screening or pre-processing of multi-period and multi-source data without feature screening, the data is used as input data of the long short-term memory network and enters the input layer of the long short-term memory network for processing;

[0267] It is then transmitted to the hidden layer of the long short-term memory network, which has multiple processing units, each of which includes an input gate, a forget gate, an update gate, and an output gate. During this processing, the multi-period and multi-source data that has passed through the input layer is multiplied by the output of the forget gate with the output result of the previous phase, thereby achieving feature elimination and retention, and further adding the result to the processing result of the input gate to obtain a new processing unit result, which is input into the next processing unit, and finally outputs the prediction result.

[0268] In some embodiments, the prediction results of the random forest model or the convolutional neural network model can be fused with the prediction results of the long short-term memory network model according to research needs. When the fusion result is verified to have a higher accuracy, the fused model is used as the entire population distribution model.

[0269] For example, the multiple time-phase data of the constructed random forest model or convolutional neural network model can be processed separately to obtain prediction results, and then fused with the model prediction results of the long short-term memory network obtained by training multi-phase multi-source data. In this way, multiple groups of fusion prediction results will appear, and the relationship between model selection and the characteristics of the target research area can be further studied and analyzed based on the accuracy verification of the fusion results.

[0270] In some embodiments, in step S5, the model accuracy verification specifically includes:

[0271] The mean relative error (MRE), root mean square error (RMSE) and %RMSE are used as evaluation indicators.

[0272] MRE represents the mean absolute error between the population forecast value and the true population value, reflecting the actual situation of the population forecast value error, which is expressed as:

[0273]

[0274] Where: n is the number of grids containing the real population. a represents the predicted population of the ath grid, R a Represents the actual population of the ath grid. Since MRE can only estimate grids with real populations, it cannot show the error impact of grids that have no population but are overestimated. Therefore, the root mean square error RMSE is used to evaluate the accuracy of grid-level population estimation:

[0275]

[0276] %RMSE represents the root mean square error between the population prediction value and the population true value divided by the mean of the population true value, reflecting the overall accuracy of the model simulation. The calculation formula is as follows:

[0277]

[0278] In order to verify the effectiveness of the method of the present invention, a population model was constructed in the Atlanta area, and its accuracy was compared with the public Atlanta population model. Figure 4 The Atlanta public population data set is used as the true value of the population model. The hyperparameter terrain feature factor in the present invention is used as a control variable to compare the change in population model accuracy and its impact before and after adding the hyperparameter terrain feature factor. Figure 5 is the third spatial data that can reflect the terrain characteristics of the Atlanta area. When the hyperparameter terrain characteristic factor is not added, the population model construction results for the Atlanta area are as follows: Figure 6As shown in Figure 1, the corresponding accuracy indicators are: mean absolute error MAE = 4183, root mean square error RMSE = 19804, relative root mean square error %RMSE = 88.7%. When the hyperparameter terrain feature factor is added, the population model construction results for the Atlanta area are as follows Figure 7 As shown, the corresponding accuracy indicators are as follows: mean absolute error MAE = 11.5, root mean square error RMSE = 20.4, relative root mean square error % RMSE = 71.8%. Through experimental data comparison, it can be found that when the hyperparameter terrain feature factor is added, the population model accuracy of the Atlanta area can be significantly improved by judging the terrain undulation, which proves the effectiveness of the method of the present invention.

[0279] Although the present invention provides a plurality of specific embodiments, it should be understood that the disclosed method can also be embodied in a variety of other specific forms without departing from the spirit or scope of the present invention. The examples of the present invention should be regarded as illustrative rather than restrictive, and the present invention is not limited to the details given herein. For example, various factors or spatial data can be combined or merged in another system, or certain features can be omitted or not implemented. That is, the above-mentioned specific implementation methods do not constitute limitations on the scope of protection of the present invention. The above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. Claim 1 seeks to protect a method for constructing a population spatial distribution model by fusing features derived from multi-source data under semi-supervision, characterized in that: include: S1, multi-source data acquisition: obtain the census data of the target area, the corresponding geographic space data, night light data, land use data, city annotation data, and urban building data of the target area; S2, preprocessing multi-source data, extracting characteristic factor clusters, and storing data: preprocessing the census data; extracting characteristic factors and constructing characteristic factor clusters for the geographic spatial data, night light data, land use data, city annotation data, and urban building data corresponding to the target area; storing the constructed characteristic factor cluster data in the population spatial characteristic database of the study area. For the city annotation data: obtaining the first city annotation data, performing spatial matching and gridding on the first city annotation data, and performing preprocessing and data reconstruction to obtain the reconstructed second city annotation data; Based on the reconstructed second city annotated data, the characteristic sub-factors corresponding to 18 characteristic factors are obtained, including: (1) Calculate the number of city annotation data points that fall within each grid and obtain the first characteristic sub-factor of the city annotation data; (2) Calculate the distance between each grid and the nearest city annotation data point, obtain the minimum distance between each grid and the non-internal city annotation data point, and obtain the second characteristic sub-factor of the city annotation data; (3) Calculation of kernel density related factors: According to the characteristics of the target area, through feature screening, the grid size r1 for large-scale kernel density calculation and the grid size r2 for small-scale kernel density calculation are obtained respectively. Calculate the kernel density of the city annotation data points in the entire domain with the distance of r1 grids as the radius, obtain the corresponding large-scale kernel density, and obtain the third characteristic sub-factor of the city annotation data; Calculate the kernel density of the city annotation data points in the whole domain with the distance of r2 grids as the radius, obtain the corresponding small-scale kernel density, and obtain the fourth characteristic sub-factor of the city annotation data; (4) A reference range with a radius of 2,500 meters is established for the city annotation data points in a fusion mode. If a point falls within the reference range, it is deemed to fall within the service range of the city annotation data point of this type. All the corresponding geographic space area pixels within the range are assigned a value of 1, and the geographic space area pixels outside the reference range are assigned a value of 0. The data after the assignment is used as the fifth characteristic sub-factor of the city annotation data; (5) The characteristic sub-factors of the above five types of city annotation data are integrated to obtain the characteristic factors of each type of city annotation data. The characteristic factor data of the 18 types of city annotation data together form the fourth characteristic factor cluster. S3, construction and training of multivariate regression population distribution model; S4, accuracy verification of population distribution model; S5, based on the trained model, semi-supervised population distribution prediction is performed on the data of different target areas to obtain the rasterized model prediction results.

2. The method for constructing a population spatial distribution model by fusion of multi-source data derived features under semi-supervision according to claim 1, characterized in that: The geographic spatial data includes: first spatial data, second spatial data, and third spatial data with physical world location information and identification; the first spatial data carries attribute field information of the administrative district code and administrative district name of the target area; the third spatial data is extracted based on the SRTM data obtained after splicing and cropping. The preprocessing of multi-source data and the extraction of characteristic factor clusters specifically include: Gridding the first spatial data to obtain fifth spatial data, matching the population census data to the fifth spatial data to obtain fused model basic sample data; wherein the grid size of the fifth spatial data is set to 100m×100m; Performing spatial matching and gridding on the third spatial data to obtain sixth spatial data, and extracting characteristic factors based on the sixth spatial data to obtain a first characteristic factor cluster, wherein the sixth spatial data has the same grid size and spatial coordinate reference as the basic sample data; Acquire first nighttime light data, perform spatial matching and gridding on a spatial range involved in the first nighttime light data, and cut the data range based on the basic sample data to obtain second nighttime data, wherein the second nighttime data has the same grid size and spatial coordinate reference as the basic sample data; Acquire the third nighttime light data, perform data splicing on the third nighttime light data, perform spatial matching and gridding on the third nighttime light data, and cut the data range based on the basic sample data to obtain the fourth nighttime data; wherein the fourth nighttime data has the same grid size and spatial coordinate reference as the basic sample data; A second characteristic factor cluster is obtained based on the fusion calculation of the second nighttime light data and the fourth nighttime light data; Classification processing is performed based on the currently acquired image data to obtain a number of first land use data after classification; Performing spatial matching and gridding on the plurality of blocks of first land use data, and cutting and splicing the plurality of blocks of first land use data based on the basic sample data to obtain second land use data; Based on the statistics and calculation of the second land use data, a third characteristic factor cluster is obtained; Acquire first city building data, perform spatial matching and gridding on the first city building data, and cut the data range based on the basic sample data to obtain second city building data; wherein the second city building data has the same grid size and spatial coordinate reference as the basic sample data; Based on the second city building data, a corresponding plurality of characteristic factors are obtained to form a fifth characteristic factor cluster.

3. The method for constructing a population spatial distribution model by fusion of multi-source data derived features under semi-supervision according to claim 2, characterized in that: The extracting characteristic factors based on the sixth spatial data to obtain a first characteristic factor cluster specifically includes: (1) obtaining an elevation factor by performing data interpolation on the sixth spatial data; (2) extracting slope grid values ​​from the sixth spatial data; performing partition statistics on the extracted slope grid values ​​to calculate the average slope value of each area and use it as a slope characteristic factor; (3) extracting aspect grid values ​​from the sixth spatial data; performing partition statistics on the extracted aspect grid values ​​to calculate the average slope value of each area and use it as the aspect characteristic factor; (4) Based on the elevation characteristic factors obtained above, the hyperparameter terrain characteristic factors are further calculated; (5) Obtaining the first characteristic factor cluster based on the fusion of the elevation factor, the slope characteristic factor, the aspect characteristic factor, and the hyperparameter terrain characteristic factor.

4. The method for constructing a population spatial distribution model by fusion of multi-source data derived features under semi-supervision according to claim 3, characterized in that: On the basis of the elevation characteristic factors obtained above, the hyperparameter terrain characteristic factors are further calculated, specifically including: ① Based on the elevation characteristic factor and the slope characteristic factor extracted from the sixth spatial data, the mean elevation Hp and the difference between the highest and lowest elevations Hc in the target area are calculated; and, ② Calculate the proportion β of the area in the target area where the slope characteristic factor meets the specified threshold α to the area of ​​the target area; ③ Based on the following algorithm, the hyperparameter terrain characteristic factor A is calculated: δ1A=Hp+Hc*β*δ2 In the above formula, δ1 is the plane scale parameter, which is generally determined according to the average elevation Hp, and δ2 is the terrain feature influence parameter, which represents the importance of the elevation fluctuation of the ground area on the overall impact, where δ2>0; ④Set the analysis radius of the target area; ⑤ Based on the set analysis radius, adaptively set the plane scale parameter δ1 and the terrain feature influence parameter δ2; ⑥ Calculate the hyperparameter terrain characteristic factor A based on the set parameters.

5. The method for constructing a population spatial distribution model by fusion of multi-source data derived features under semi-supervision according to claim 2, characterized in that: The step of obtaining a second characteristic factor cluster based on the fusion calculation of the second nighttime light data and the fourth nighttime light data specifically includes: (1) By extracting and processing the DN values ​​of night light data, two characteristic factors of night light data are formed; (2) By extracting and processing the brightness value L of the night light data, a characteristic factor of the night light data is formed; (3) By setting a threshold for feature screening, the night light data is binarized, and the binary data of three types of night light data generated by the second night light data and the fourth night light data are obtained as feature factors of corresponding dimensions, totaling six types of feature factors of night light data. (4) The DN value of the night light data, the brightness value of the night light data, and the nine characteristic factors of the night light data obtained by binarization of the night light data are collectively formed into a second characteristic factor cluster.

6. The method for constructing a population spatial distribution model by fusion of multi-source data derived features under semi-supervision according to claim 2, characterized in that: The third characteristic factor cluster is obtained based on the statistics and calculation of the second land use data, specifically including: (1) extracting 10 land use types from the second land use data as 10-dimensional land use characteristic factors; (2) The 10 land use types are merged into a new characteristic factor and assigned a value, so that the new characteristic factor contains all land use types and is represented by different codes; (3) The 10-dimensional land use characteristic factors and the merged new characteristic factors are taken together as the third characteristic factor cluster.

7. The method for constructing a population spatial distribution model by fusion of multi-source data derived features under semi-supervision according to claim 2, characterized in that: The obtaining of a plurality of corresponding characteristic factors based on the second city building data to form a fifth characteristic factor cluster specifically includes: (1) Calculate the area of ​​the building based on the building outline data, and calculate the volume of the building based on the building height data; (2) The area and volume attributes of the building are used as two characteristic factors of the second city building data to form the fifth characteristic factor cluster.

8. The method for constructing a population spatial distribution model by integrating features derived from multi-source data under semi-supervision according to claim 1 or 5, characterized in that: The feature screening specifically includes: (1) vectorizing the data to be feature screened having multiple feature factors to obtain vector data of the data to be feature screened respectively; (2) Traversing each feature factor in the vector data of the data to be feature screened, randomly shuffling the feature values ​​in each feature factor, and inputting the feature factors and other feature factors that have not been randomly shuffled into a subsequent processing model, comparing the accuracy of the model output result with the verification data to determine whether it meets a preset threshold; (3) Repeat the process in step (2) above until all characteristic factors have been subjected to eigenvalue shuffling and model processing to obtain corresponding accuracy verification results; (4) The characteristic factors whose errors in the accuracy verification results during the processing of steps (2)-(3) are retained as input characteristic factors for the subsequent population prediction model.

9. The method for constructing a population spatial distribution model by fusion of multi-source data derived features under semi-supervision according to claim 1, characterized in that: The construction of the multivariate regression population distribution model specifically includes: S3-1, Sample data preparation: Population survey data is used as the basic sample, and the multidimensional features formed by the corresponding five characteristic factor clusters are used as sample data for model construction. S3-2, sample data generation: Calculate population density according to the geographic space grid, use this data as the sample dependent variable, and number the grid ID of the target area as the grid identifier. The multidimensional modeling features extracted from the grid are used as sample independent variables to form the training sample. S3-3, model training: The grid pixels of the target area with multi-dimensional data features and unique ID numbers are used as model input, and the population model data of the same pixel size is used as output. The population distribution model is constructed and trained based on the random forest model. In the process of model training, a random sampling method with replacement is adopted. During the training process, some original samples are selected as training data, and the remaining sample data constitute out-of-bag data. S3-4, model optimization: evaluate the accuracy of the random forest model through out-of-bag samples; optimize the parameters of the random forest framework and decision tree during training.

10. The method for constructing a population spatial distribution model by fusion of multi-source data derived features under semi-supervision according to claim 1, characterized in that: In step S4, the model accuracy verification specifically includes: Mean relative error (MRE), root mean square error (RMSE) and %RMSE were used as evaluation indicators. MRE represents the mean absolute error between the population forecast value and the true population value, reflecting the actual situation of the population forecast value error, which is expressed as: Where: n is the number of grids containing the real population. a represents the predicted population of the ath grid, R a represents the actual population of the ath grid; The root mean square error (RMSE) is used to evaluate the accuracy of grid-level population estimation: %RMSE represents the root mean square error between the population prediction value and the population true value divided by the mean of the population true value, reflecting the overall accuracy of the model simulation. The calculation formula is as follows:

Images