Urban population statistical data spatialization method and device based on deep learning

Abstract

The invention discloses an urban population statistical data spatialization method and device based on deep learning, and belongs to the field of remote sensing science and technology. The method comprises the following steps: step 1, preprocessing acquired noctilucent data, and resampling global building area footprint data and urban building height data according to the resolution of the noctilucent data; step 2, obtaining a population distribution range based on multi-scale spatial constraint based on a data preprocessing result in the step 1; 3, taking the demographic statistical data as dependent variables, and taking the urban building height data and the noctilucent target area subjected to spatial constraint as independent variables to generate a spatialization fine training sample; and step 4, constructing a deep learning model, training the deep learning model by using the spatialized fine training sample, and performing population distribution mapping based on the trained model. According to the method, the resolution of population spatialization is improved, the calculation complexity is reduced, and the intelligence and precision of population spatialization are improved.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing science and technology, and specifically relates to a method and device for spatializing urban population data based on deep learning. Background Art

[0002] Population data, as a direct indicator of human activities, is closely related to various social elements. Accurate population distribution maps can support monitoring, evaluation, and decision-making. With the advancement of earth observation technology, the application of multi-source remote sensing data has promoted the development of spatial population distribution research. Based on multi-source remote sensing data and other geographic information data, global multi-year population spatial data products have been generated, which have made significant progress in filling global data gaps. However, due to the limitations of resolution (for example, the resolution of WorldPop is 100 meters) and accuracy, these datasets still face challenges in supporting in-depth applications to meet the needs of the current era for fine-grained population management, such as smart cities, rural construction, and accurate public resource allocation.

[0003] In the production of population distribution data, night light (NTL) data are widely used as the main data source because they capture the visible light emitted from the Earth's surface observed at night from space. However, the spatial resolution of these data (1 km for DMSP-OLS and 500 m for NPP-VIIRS) has been a limiting factor in achieving detailed and accurate population distribution maps.

[0004] The satellite SDGSAT-1 is equipped with an advanced Glimmer Imager (GLI), which innovatively uses a color band. With a spatial resolution of 40 meters, the GLI color image can accurately detect different categories of urban and rural lights and show more spatial and spectral details than most existing night light satellite data. Therefore, SDGSAT-1's GLI has greater potential for high-resolution, spatially detailed population distribution maps.

[0005] In the process of population distribution data production, count mapping methods are usually used to establish the relationship between population variables and covariates through independent samples of the census. When NTL data and auxiliary data are integrated into the model construction, intelligent count mapping builds a regression model to derive the population weight factor for each grid. Initially, the regression model of intelligent count mapping relies on traditional statistical methods such as multivariate linear regression, random forest (RF) regression, etc., but increases the dependence on the number of training samples.

[0006] Recently, deep learning has shown promising capabilities in various fields because of its ability to automatically learn complex features from large datasets. Currently, two main methods are used to generate training samples. One involves using demographic data from administrative units, but the spatial scale difference between the training and target data may affect the accuracy, especially in high-resolution population distribution mapping; the other method uses population grid data such as WorldPop. Although this is the main method used by deep learning in population distribution research, it may not always be consistent with the actual population distribution due to its low resolution and limited data years, which may reduce the accuracy of the results. When deep learning methods are used for population distribution mapping, the relationship between input features and population data is mainly established by using pixel-level data as training samples. Commonly used models include one-dimensional convolutional neural networks (CNNs) and multi-layer perceptrons (MLPs). Although these methods are simple and effective, they have limitations because they ignore the spatial information in population distribution data. However, the high heterogeneity of population distribution, coupled with the fact that expanding the neighborhood space dilutes the importance of central pixel information, may lead to poor performance in population estimation.

[0007] Therefore, in order to improve the accuracy of population distribution estimates, it is necessary to change the current model structure and training methods, and to integrate models that can take into account geographic unit information and spatial positioning characteristics into population distribution research. Summary of the invention

[0008] In order to solve the above technical problems, the present invention provides a method and device for spatializing urban population data based on deep learning, which adopts SDGSAT-1 GLI images and a high-resolution population mapping method based on deep learning. Under the multi-scale feature constraints of Open Street Map data (OSM) and World Building Footprint (WSF) data, the population distribution range is refined to reduce the impact of non-residential areas on the accuracy of population distribution mapping, and a fine sample set generation method based on multivariate regression and hierarchical screening is used to screen grid data with different population density levels. Inspired by the U-shaped network used in semantic segmentation, a fine population mapping network (PNet) with an enhanced efficient multi-scale attention module is proposed to achieve more accurate and high-resolution population spatial distribution results.

[0009] To achieve the above purpose, the technical solution adopted by the present invention is as follows: In one aspect, the present invention provides a method for spatializing urban population data based on deep learning, the method comprising the following steps: Step 1: pre-process the acquired night light data, and resample the global building area footprint data and urban building height data according to the resolution of the night light data; Step 2: Under the multi-scale feature constraints of the open street map data and the resampled global building area footprint data, the preprocessed night light data and the resampled urban building height data are spatially constrained at the regional scale and the building scale to obtain the urban building height data and night light target area with spatial constraints; Step 3: Using demographic data as the dependent variable, spatially constrained urban building height data and night light target areas as independent variables, generate spatially refined training samples based on a multivariate regression model and hierarchical screening method; Step 4: Build a deep learning model and use the spatialized fine training samples to train the deep learning model to obtain a trained weight model. Use the spatially constrained urban building height data and night light target areas as inputs to the trained weight model to perform population distribution mapping and obtain a fine population spatialization result.

[0010] Furthermore, the step 1 comprises: The acquired night light data is converted into the radiant brightness of each band, and the average brightness of the night light data image is calculated based on the brightness of each RGB channel; The bilinear resampling method is used to downsample the global built-up area footprint data and urban building height data to make the resolution of multi-source data consistent.

[0011] Furthermore, the step 3 comprises: Establish a multiple regression model, taking the average brightness and building height as independent variables and demographic data as dependent variables, derive regression weights of the average brightness and building height, and obtain population weight factors for each grid in the study area; Assign the population weight factor of each grid in the study area to obtain the population grid data in the study area; The study area is divided into different levels according to population density. For each population density level, the relative error is calculated based on the population grid data to obtain the study area with the smallest relative error at each population density level. The study area with the smallest relative error at each population density level is used as a spatial fine training sample.

[0012] Furthermore, the construction of the deep learning model in step 4 includes: the deep learning model includes a PNet model, the PNet model includes an encoder, a decoder and a connector, wherein: The encoder comprises a first shallow feature generation module and a second shallow feature generation module of the same structure, which are used to generate the first shallow feature and the second shallow feature respectively; The second shallow feature is connected to the connector in a second jump connection and then input into the decoder; The decoder is symmetrically designed relative to the encoder, and includes a first deep feature generation module and a first output module, wherein the EMA module of the first deep feature generation module is used to receive the input after the second jump connection to generate the first deep feature; the first shallow feature and the first deep feature are input into the EMA module of the first output module after the first jump connection via the connector to generate a fine population spatialization result.

[0013] On the other hand, the present invention provides a device for spatializing urban population data based on deep learning, the device comprising: A night light data preprocessing unit, used to preprocess the acquired night light data and resample the global building area footprint data and urban building height data according to the resolution of the night light data; A spatial constraint unit is used to perform spatial constraints at regional scales and building scales on the preprocessed night light data and the resampled urban building height data under the multi-scale feature constraints of the open street map data and the resampled global building area footprint data, so as to obtain the urban building height data and night light target area after spatial constraints; A fine training sample generating unit is used to generate spatial fine training samples based on a multivariate regression model and a hierarchical screening method by taking demographic data as dependent variables, and spatially constrained urban building height data and night light target areas as independent variables; The population spatialization result acquisition unit is used to build a deep learning model, train the deep learning model with spatialized fine training samples, obtain a trained weight model, use the spatially constrained urban building height data and night light target area as the trained weight model input, perform population distribution mapping, and obtain a fine population spatialization result.

[0014] In a third aspect, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0015] In a fourth aspect, the present invention provides a computer-readable storage medium, characterized in that executable instructions are stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for spatializing urban demographic data based on deep learning.

[0016] The beneficial effects of the present invention are: The present invention uses high-resolution night light data (SDGSAT-1) as the main data for population spatialization, thereby improving the resolution of population spatialization; uses building height data + high-resolution night light data as independent variables, thereby reducing the number of independent variables (or characteristic factors) and reducing the computational complexity; utilizes open street map data and global building area footprint data to eliminate the influence of non-building residential areas, thereby improving the rationality of population spatialization; generates a high-quality sample set based on a refined sample set generation method based on multiple regression and hierarchical screening; and improves the intelligence and accuracy of population spatialization based on a semantic segmentation deep network model. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a schematic diagram of a method for spatializing urban population data based on deep learning according to the present invention; Figure 2 This is a schematic diagram of the PNet framework structure. DETAILED DESCRIPTION

[0018] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0019] Firstly, the definitions of relevant data used in this method are introduced in detail, including: demographic data, night light data (SDGSAT-1), building height data, OpenStreetMap (OSM) data, and World Settlement Footprint (WSF) data.

[0020] Demographic data: Demographic data refers to the population census statistics based on administrative districts, or the demographic data based on administrative districts in the statistical yearbooks of provinces and cities.

[0021] Night light data: The present invention uses SDGSAT-1 GLI RGB data with a resolution of 40 meters.

[0022] Building height data: Building height reflects the level of urbanization and development, is closely related to population distribution, and is also used as an important influencing factor in population mapping. Therefore, building height data is used as another independent population variable factor alongside night light data.

[0023] Open Street Map Data: A collaborative mapping project that aims to create a freely editable global map database through an online community. The data is detailed and constantly updated, covering roads, buildings, natural features, facilities and infrastructure around the world. It is used to determine the type of built-up areas, eliminate non-residential buildings, and focus population spatialization processing on human-inhabited built-up areas, improving the accuracy of spatialization.

[0024] Global built-up area footprint data: A global dataset that depicts the extent of human settlement areas around the world, used to remove non-built-up areas and improve the accuracy of population spatialization.

[0025] As mentioned above, Figure 1 As shown, the present invention provides a method for spatializing urban demographic data based on deep learning, which adopts SDGSAT-1 GLI images and a high-resolution population mapping method of deep learning, pre-processes the acquired night light data, and resamples the global building area footprint data and urban building height data according to the resolution of the night light data, so that the resolution of the two is consistent with the night light data; under the multi-scale feature constraints of open street map data and the resampled global building area footprint data, the pre-processed night light data and the resampled urban building height data are spatially constrained at the regional scale and the building scale, and the night light target area and urban building height data with spatial constraints are obtained, so as to refine the population distribution range and reduce the influence of non-residential areas on the accuracy of population distribution mapping; the demographic data are used as the dependent variable, and the urban Building height data and night light target areas are used as independent variables, and a fine sample set generation method based on multivariate regression and hierarchical screening is used to screen grid data with different population density levels as fine training samples for population spatialization of the deep learning model; inspired by the U-shaped network used in semantic segmentation, a fine population mapping network (PNet) with an enhanced efficient multi-scale attention module is proposed. The deep learning model is trained with the constructed training sample set to obtain a trained weight model. The spatially constrained urban building height data and night light target areas are used as the input of the trained weight model to perform population distribution mapping and obtain a fine spatialization result of the population, achieving more accurate and high-resolution population spatial distribution results.

[0026] 1. Data preparation and preprocessing; The initial input data of the present invention is WSF data, OSM data, urban building height data and night light remote sensing data. The range of WSF data, OSM data and urban building height data is selected according to the study area, and the number of scenes of night light remote sensing data covering the study area is determined to prepare data. The present invention selects high-resolution SDGSAT-1 night light data as the main data. The SDGSAT-1 night light remote sensing data needs to be preprocessed before population processing.

[0027] The SDGSAT-1 GLI RGB data is converted into radiance, and then the average brightness value is calculated. Each pixel value (Digital Number, DN) of the SDGSAT-1 RGB band is converted into the radiance R of each band according to the parameters provided in the metadata file, as shown in formula (1): (1) The unit of radiant brightness R is W·m -2 ·sr -1 μm -1 , DN is the pixel value of the image after relative radiation calibration, Gain represents gain, and Bias represents bias.

[0028] Next, the average brightness BR of the SDGSAT-1 GLI image is calculated based on the brightness of each RGB channel, as shown in formula (2): (2) Among them, R R , R G and R B Represents the radiance of the R, G and B channels calculated according to formula (1).

[0029] Since the collected 10-meter resolution CNBH data, 10-meter resolution global built-up footprint (WSF) data and 40-meter resolution SDGSAT-1 GLI data need to be jointly analyzed, the bilinear resampling method is used to downsample the 10-meter CNBH data and WSF data to 40 meters to ensure the consistency of multi-source data resolution.

[0030] 2. Extraction of population distribution range based on multi-scale spatial constraints; By integrating information from OSM data and WSF data, accurate regional positioning and fine screening of population distribution ranges are achieved on the basis of spatial constraints of multi-scale features (regional scale and building scale), mainly by utilizing the functional zoning of OSM and the building range information provided by WSF.

[0031] OSM data contains accurate information on urban functional zoning, which is usually used to identify areas related to population residence and exclude interference from non-residential areas in population distribution studies. Therefore, using OSM functional zoning data can accurately identify non-residential areas such as industrial areas, farmlands, forests, etc. based on field attribute information, thereby excluding them from the background area and ensuring the reliability of subsequent analysis.

[0032] Despite the initial screening of OSM data, certain features such as transportation routes and rivers may still mask the true extent of residential areas. To refine the analysis, WSF building outline data was integrated to depict the morphology of residential structures. This approach combines OSM functional zoning and WSF built-up area data and applies feature constraints at the regional scale and building scale, thus ensuring the accuracy and granularity of population distribution analysis.

[0033] Through the above processing, the independent variables (i.e., characteristic factors) of the population spatialization in the present invention: urban building height and night light target area can be constrained to the range of buildings where the population lives, which can effectively exclude external factors, better establish the relationship with the dependent variable (i.e., demographic data), and improve the accuracy of spatial mapping of population distribution.

[0034] 3. Generation of refined sample sets based on multivariate regression and hierarchical screening; In order to solve the problems of scale mismatch, low precision and insufficient population training samples, the present invention introduces a new method for generating fine sample data sets. By using multiple regression and hierarchical screening techniques, a step-by-step screening process from county-level administrative division constraints to township-level verification is established, providing samples covering different population density scenarios and enhancing the generalization ability of the model.

[0035] First, a multivariate regression model was established with each county-level administrative district as the basic unit, and the average brightness BR and building height H of the regional statistical results were used as independent variables, and the demographic data as the dependent variable. Through the multivariate regression model, the regression weights of average brightness BR and building height H are derived. and .

[0036] Next, use formula (3) to obtain the population weight factor of the i-th grid in the study area: : (3) in, , Represent the average brightness and building height of the i-th grid respectively.

[0037] Then, taking each county-level administrative district as the basic unit and county-level demographic data as the constraint, we assign values ​​according to the population weight factors of each grid in the administrative district to obtain the population grid data in the study area. As shown in formula (4), the population gridding results in the study area are obtained: : (4) in, represents the census data of the mth county, Represents the population weight factor in the nth grid of the mth county.

[0038] Get the population grid results in the initial study area Finally, the population gridding results will be spatially summed within the area of ​​township-level administrative divisions as the basic unit to obtain the sum of the population gridding data within the township-level administrative unit. Based on township-level census data As the true value, the relative error RE between the two is calculated as shown in formula (5): (5) Define the population density of township-level administrative units in the study area As shown in formula (6), Area t Represents the area of ​​township-level administrative units.

[0039] (6) Taking into account the spatial variation of population distribution, the township administrative units in the study area are divided into three different levels according to different population densities: low, medium and high density. For different population density levels, one or several administrative units with the smallest relative error are selected respectively, and the preliminary results of populationization in these areas are used as the data source of the sample set, and they are sliced ​​to construct the sample set used in the deep learning method. The strategy of hierarchical screening of sample data sources adopted in the present invention aims to better capture the performance of the model under different population density scenarios, thereby improving the representativeness of the training samples and the generalization ability of the model. The representativeness of different population characteristics in the study area is ensured, which contributes to the robustness and accuracy of the deep learning model.

[0040] 4. High-resolution population distribution mapping based on PNet; Build a deep learning model, train the deep learning model based on the sample set established in the above steps, obtain a trained weight model, perform spatial distribution mapping on the demographic data, and finally obtain a detailed spatialization result of the population.

[0041] Compared with traditional fully connected networks (such as MLP) or global pooling operations (such as VGG), semantic segmentation networks adopt a multi-scale feature fusion mechanism, which enables them to utilize a wider range of contextual information. This ability enhances target perception, facilitates feature extraction, and improves prediction accuracy while preserving spatial details. Therefore, the present invention integrates the semantic segmentation network structure into population distribution research and proposes a PNet model equipped with an efficient multi-scale attention (EMA) module for high-resolution population distribution mapping.

[0042] like Figure 2As shown in the figure, the input data of the PNet model are night light data and building height data, and the output is the population spatialization result. The PNet model consists of three main parts: encoder, decoder and connector. The encoder includes two shallow feature generation modules, among which the structure of the first shallow feature generation module is the same as that of the second shallow feature generation module, both of which include two consecutive 3×3 convolutional layers (Conv), and the ReLU activation function is used in each convolutional layer to introduce nonlinear information processing; two consecutive 3×3 convolutional layers are followed by a 2×2 maximum pooling layer (Maxpool) to reduce the size of the feature map and effectively expand the receptive field; finally, the first shallow feature is output through a 1×1 convolution; the corresponding second shallow feature generation module outputs the second shallow feature. A jump connector is used between the encoder and the decoder to combine the convolutional layers. Jump connections play a vital role between the encoder and the decoder. These connections help retain the low-level feature details in the encoder and help the decoder to more accurately reconstruct the subtle details and local information of the population distribution. The connector contains two 3×3 convolutional layers and one 2×2 upsampling convolution. The second shallow feature is connected to the connector in a second jump connection and then input into the first deep feature generation module of the decoder.

[0043] The decoder is designed symmetrically with respect to the encoder, and includes a first deep feature generation module and a first output module, wherein the first deep feature generation module includes an EMA module, two consecutive 3×3 convolutional layers, which also use the ReLU activation function, followed by a 2×2 convolutional layer for upsampling, and the output first deep feature and the first shallow feature are sent to the first output module of the decoder after the first jump connection through the connector. The first output module includes an EMA module, two consecutive 3×3 convolutional layers, followed by a convolutional layer of size 1×1, in order to ensure that the output data size is the same as the input data size. Each subsequent decoder submodule has a similar structure to ensure a gradual reduction in the number of output channels, which is crucial for accurately reconstructing population regression results. The encoder, connector and decoder are connected through jump connections to splice and combine features, integrating the shallow features generated by the encoder with the deep features generated by the connector and decoder, making the network learning features richer.

[0044] To further improve PNet's ability to handle the complexity of population regression and effectively capture subtle changes and local features in population distribution, the model concatenates the shallow features generated by the encoder with the deep features generated by the connector and decoder, and then connects them to the EMA (efficient multi-scale attention) module of the first deep feature generation module. The EMA module processes the result of the fusion of the encoder and decoder output features. The EMA module is used to dynamically adjust the attention weights at different spatial scales, so that the network can more effectively focus on relevant features while suppressing irrelevant features. By integrating the EMA module into the architecture, PNet is able to effectively highlight important local patterns and detailed changes in population distribution data, thereby improving the accuracy and robustness of population regression tasks. EMA reduces computational costs while maintaining information validity, and achieves this goal by reconfiguring some channels to the batch dimension and balancing spatial features within feature groups. EMA encodes global information using a branching structure and enhances features through cross-dimensional interactions. EMA allows the model to adaptively adjust the importance of each location during training. It pays more attention to areas with higher population density and ignores areas that are less relevant to population distribution. This adaptive attention significantly improves the precision and accuracy of population regression analysis.

[0045] In addition, PNet sets the number of sub-modules for both downsampling and upsampling to two, which can reduce the number of parameters and computational complexity of the model, thereby speeding up the training and reasoning of the model and improving its efficiency. Secondly, PNet uses smaller-sized sample slices for training, which can allow the model to focus more on extracting local detailed features. In addition, by replacing the traditional copy and crop operations with 1×1 convolutional layers in the skip connection layer, the model benefits from enhanced feature information transmission, making the representation more discriminative and expressive.

[0046] The model uses night light data and building height data as input data, and the model output is the population spatialization result, which has the same size as the input data to ensure that the population distribution in different areas is accurately and detailedly mapped.

[0047] Accordingly, the present invention provides a device for spatializing urban population data based on deep learning, and the various units of the device can implement the various steps of the above method by cooperating with each other. Specifically, the device includes: A night light data preprocessing unit, used to preprocess the acquired night light data and resample the global building area footprint data and urban building height data according to the resolution of the night light data; A spatial constraint unit is used to perform spatial constraints at regional scales and building scales on the preprocessed night light data and the resampled urban building height data under the multi-scale feature constraints of the open street map data and the resampled global building area footprint data, so as to obtain the urban building height data and night light target area after spatial constraints; A fine training sample generating unit is used to generate spatial fine training samples based on a multivariate regression model and a hierarchical screening method by taking demographic data as dependent variables, and spatially constrained urban building height data and night light target areas as independent variables; The population spatialization result acquisition unit is used to build a deep learning model, train the deep learning model with spatialized fine training samples, obtain a trained weight model, use the spatially constrained urban building height data and night light target area as the trained weight model input, perform population distribution mapping, and obtain a fine population spatialization result.

[0048] On the other hand, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0049] It should be understood by those skilled in the art that the embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The solutions in the embodiments of the present invention may be implemented in various computer languages, for example, object-oriented programming language Java and interpreted scripting language JavaScript, etc.

[0050] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0051] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0052] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0053] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for spatializing urban population data based on deep learning, characterized in that: The method comprises the following steps: Step 1: pre-process the acquired night light data, and resample the global building area footprint data and urban building height data according to the resolution of the night light data; Step 2: Under the multi-scale feature constraints of the open street map data and the resampled global building area footprint data, the preprocessed night light data and the resampled urban building height data are spatially constrained at the regional scale and the building scale to obtain the urban building height data and night light target area with spatial constraints; Step 3: Using demographic data as the dependent variable, spatially constrained urban building height data and night light target areas as independent variables, generate spatially refined training samples based on a multivariate regression model and hierarchical screening method; Step 4: Build a deep learning model and use the spatialized fine training samples to train the deep learning model to obtain a trained weight model. Use the spatially constrained urban building height data and night light target areas as inputs to the trained weight model to perform population distribution mapping and obtain a fine population spatialization result.

2. The method for spatializing urban population data based on deep learning according to claim 1, characterized in that: The step 1 comprises: The acquired night light data is converted into the radiant brightness of each band, and the average brightness of the night light data image is calculated based on the brightness of each RGB channel; The bilinear resampling method is used to downsample the global built-up area footprint data and urban building height data to make the resolution of multi-source data consistent.

3. The method for spatializing urban population data based on deep learning according to claim 1, characterized in that: The step 3 comprises: Establish a multiple regression model, taking the average brightness and building height as independent variables and demographic data as dependent variables, derive regression weights of the average brightness and building height, and obtain population weight factors for each grid in the study area; Assign the population weight factor of each grid in the study area to obtain the population grid data in the study area; The study area is divided into different levels according to population density. For each population density level, the relative error is calculated based on the population grid data to obtain the study area with the smallest relative error at each population density level. The study area with the smallest relative error at each population density level is used as a spatial fine training sample.

4. The method for spatializing urban population data based on deep learning according to claim 1, characterized in that: The construction of the deep learning model in step 4 includes: the deep learning model includes a PNet model, and the PNet model includes an encoder, a decoder and a connector, wherein: The encoder comprises a first shallow feature generation module and a second shallow feature generation module of the same structure, which are used to generate the first shallow feature and the second shallow feature respectively; The second shallow feature is connected to the connector in a second jump connection and then input into the decoder; The decoder is symmetrically designed relative to the encoder, and includes a first deep feature generation module and a first output module, wherein the EMA module of the first deep feature generation module is used to receive the input after the second jump connection to generate the first deep feature; the first shallow feature and the first deep feature are input into the EMA module of the first output module after the first jump connection via the connector to generate a fine population spatialization result.

5. A device for spatializing urban population statistics based on deep learning, characterized in that: The device comprises: A night light data preprocessing unit, used to preprocess the acquired night light data and resample the global building area footprint data and urban building height data according to the resolution of the night light data; A spatial constraint unit is used to perform spatial constraints at regional scales and building scales on the preprocessed night light data and the resampled urban building height data under the multi-scale feature constraints of the open street map data and the resampled global building area footprint data, so as to obtain the urban building height data and night light target area after spatial constraints; A fine training sample generating unit is used to generate spatial fine training samples based on a multivariate regression model and a hierarchical screening method by taking demographic data as dependent variables, and spatially constrained urban building height data and night light target areas as independent variables; The population spatialization result acquisition unit is used to build a deep learning model, train the deep learning model with spatialized fine training samples, obtain a trained weight model, use the spatially constrained urban building height data and night light target area as the trained weight model input, perform population distribution mapping, and obtain a fine population spatialization result.

6. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs; Among them, when one or more programs are executed by the one or more processors, the one or more processors implement the deep learning-based spatialization method of urban demographic data as described in any one of claims 1-4.

7. A computer-readable storage medium, characterized in that: Executable instructions are stored thereon, which, when executed by a processor, enable the processor to implement a method for spatializing urban demographic data based on deep learning as described in any one of claims 1 to 4.

Images