Mesh population quantity determination method and device, electronic equipment and storage medium

Abstract

The present application relates to the technical field of data processing, and provides a grid population quantity determination method and device, electronic equipment and storage medium, wherein the method comprises: inputting a first data set of a target image block corresponding to a target grid in a target area into a target model, extracting global spatial relationship features of the target image block according to the first data set by a global spatial relationship extraction module in the target model, and extracting local spatial relationship features in the global spatial relationship features by a local spatial relationship extraction module in the target model; determining a population weight of the target grid according to the local spatial features, wherein the population weight is used to determine the population quantity in the target grid. Through the present application, the problem that the spatial correlation of the population is not considered when the population is gridded in the related art, thereby resulting in low accuracy of the population quantity in the determined grid, is solved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method, apparatus, electronic device, and storage medium for determining the population size of a grid. Background Technology

[0002] Population gridding is the process of dividing discrete population data into grids in geospatial space and distributing them into regular grids. This involves dividing the study area into multiple regular grids and allocating discrete population data to the corresponding grids according to the geographical location of the population. When allocating the population, it is necessary to calculate the number of people in each grid. Accurately determining the number of people in each grid can improve the accuracy of population gridding and provide a reliable foundation for subsequent spatial and statistical analysis.

[0003] Population distribution often exhibits a certain degree of spatial correlation, meaning that the population distribution of adjacent grids may be similar or mutually influential. Related technologies do not consider the spatial correlation of population when performing population gridding, resulting in low accuracy of the population count in the determined grid. Summary of the Invention

[0004] The problem addressed by this invention is that related technologies do not consider the spatial correlation of the population when performing population gridding, resulting in low accuracy of the population count in the determined grid.

[0005] To address the aforementioned problems, this invention provides a method for determining the population size of a grid. The method includes: inputting a first data set of target image blocks corresponding to target grids in a target region into a target model. The target image block comprises n×n grids, including the target grid and grids in the neighborhood of the target grid. The first data set includes n×n×C feature vectors, where C is the number of population influencing factors, n is an integer greater than 1, and C is an integer greater than or equal to 1.

[0006] The global spatial relationship extraction module in the target model extracts the global spatial relationship features of the target image patch based on the first data set, and the local spatial relationship extraction module in the target model extracts the local spatial relationship features from the global spatial relationship features.

[0007] The population weight of the target grid is determined based on the local spatial features, wherein the population weight is used to determine the population size in the target grid.

[0008] In an optional embodiment, before inputting the first set of data corresponding to the target image patches of the target grid in the target region into the target model, the method further includes: acquiring remote sensing data and POI data corresponding to the target region, wherein the remote sensing data includes data corresponding to C1 type population influence factors, and the POI data includes data corresponding to C2 type population influence factors, where C = C1 + C2; resampling the data corresponding to each type of population influence factor in the remote sensing data and the POI data to a first resolution to obtain a second set of data; acquiring a first population gridded dataset corresponding to the first resolution of the target region, wherein the first resolution is greater than the target resolution, and the target resolution is consistent with the size of the target grid; and using the second set of data and the first population gridded dataset as a training set to train the model to obtain the target model.

[0009] In an optional embodiment, after obtaining the first population gridded dataset of the first resolution corresponding to the target area, the method further includes: obtaining population census data of the target area;

[0010] The first gridded population dataset is corrected based on the census data to obtain the corrected first gridded population dataset; the second dataset and the corrected first gridded population dataset are used as the training set for model training to obtain the target model; wherein, the first gridded population dataset is corrected using the following formula:

[0011]

[0012] Among them, y i and y i '' represents the population size in the i-th grid before and after the correction, respectively, y j P represents the population within the j-th grid of the D-th street in the target area. D A represents the population of the Dth street in the census data. j This represents the area of ​​the j-th grid.

[0013] In an optional embodiment, before inputting the first data set of target image patches corresponding to the target grid in the target area into the target model, the method further includes: acquiring remote sensing data and POI data corresponding to the target area, wherein the remote sensing data includes data corresponding to C1 type population influence factors, and the POI data includes data corresponding to C2 type population influence factors, where C = C1 + C2; resampling the data corresponding to each type of population influence factor in the remote sensing data and the POI data to the target resolution to obtain a third data set, wherein the target resolution is consistent with the size of the target grid, and the third data set includes m1 × m2 × C feature vectors; and determining the first data set from the third data set.

[0014] In an optional embodiment, the step of extracting the global spatial relationship features of the target image patch by the global spatial relationship extraction module in the target model based on the first data set includes: mapping the feature vectors in the first data set to the target embedding space to obtain a target embedding sequence, wherein the target embedding sequence includes C target embedding features; and performing information interaction and transmission on the target embedding sequence through multiple attention units to obtain global spatial relationship features, wherein the global spatial relationship features include C layers of spatial features, and each layer of spatial features in the C layers is used to represent a global feature of one of the population influence factors.

[0015] In an optional embodiment, the local spatial relationship features are extracted from the global spatial relationship features by the local spatial relationship extraction module in the target model, including: determining the weights corresponding to each layer of spatial features in the global spatial relationship features to obtain C fusion weights; multiplying each of the C fusion weights with the corresponding layer of spatial features in the global spatial relationship features to obtain a fusion feature map, wherein the fusion feature map includes C layers of fusion features; and performing a convolution operation on the fusion feature map by the convolution block in the local spatial relationship extraction module to obtain the local spatial relationship features.

[0016] In an optional embodiment, determining the weights corresponding to each layer of spatial features in the global spatial relationship features to obtain C fusion weights includes: performing a global average pooling encoding operation on each layer of spatial features in the global spatial relationship features to obtain C encoded features; and obtaining the C fusion weights using the following formula:

[0017]

[0018] Where δ(x) is the ReLU function, σ(x) is the Sigmoid function, and W L It is the weight matrix of the first fully connected layer. It is the weight matrix of the second fully connected layer, l f Including the C encoded features, w f This includes the aforementioned C fusion weights.

[0019] The present invention also provides a grid population determination device, the device comprising: an input unit, configured to input a first data set of target image blocks corresponding to target grids in a target region into a target model, wherein the target image block comprises n×n grids, the n×n grids comprising the target grid and grids in the neighborhood of the target grid, the first data set comprising n×n×C feature vectors, wherein C is the number of types of population influencing factors, n is an integer greater than 1, and C is an integer greater than or equal to 1;

[0020] The extraction unit is used to extract the global spatial relationship features of the target image patch based on the first data set through the global spatial relationship extraction module in the target model, and to extract the local spatial relationship features from the global spatial relationship features through the local spatial relationship extraction module in the target model.

[0021] A determining unit is configured to determine the population weight of the target grid based on the local spatial features, wherein the population weight is used to determine the population size in the target grid.

[0022] The present invention also provides an electronic device, including a memory and a processor; the memory is used to store a computer program; the processor is used to implement the grid population determination method as described above when the computer program is executed.

[0023] The present invention also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the grid population determination method described above is performed.

[0024] This invention, when determining the population weight of a target grid, uses the data of multiple population factors corresponding to the target grid and multiple grids in its neighborhood—the first dataset—as input to the target model. Compared to inputting only the data of multiple population influencing factors corresponding to the target grid, the first dataset better reflects the spatial correlation of population distribution. The global spatial relationship extraction module in the target model extracts global spatial relationship features, and then extracts local spatial relationship features from the global spatial relationship features. Finally, the extracted local spatial relationship features aggregate global and local spatial relationship information, making full use of the spatial correlation of the population. This results in higher accuracy of the population weight determined based on the aforementioned local spatial relationship features, thereby enabling precise determination of the population size in the target grid and improving the accuracy of the determined population size in the grid. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the method for determining the grid population according to an embodiment of the present invention;

[0026] Figure 2 This is a schematic diagram of a target image block according to an embodiment of the present invention;

[0027] Figure 3 This is a schematic diagram of the process for generating a population weight factor distribution with a resolution of 25m according to an embodiment of the present invention;

[0028] Figure 4 This is a schematic diagram of an optional fusion feature map determination process according to an embodiment of the present invention;

[0029] Figure 5 This is a schematic diagram illustrating the process of determining population weights using an optional target model according to an embodiment of the present invention;

[0030] Figure 6 This is a structural block diagram of a grid population determination device according to an embodiment of the present invention. Detailed Implementation

[0031] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0032] While some embodiments of the invention are shown in the accompanying drawings, it should be understood that the invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the invention. It should be understood that the accompanying drawings and embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the invention.

[0033] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0034] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0035] The names of the messages or information exchanged between the multiple devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0036] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for determining the population size of a grid, the method comprising the following steps:

[0037] Step S101: Input the first data set of the target image block corresponding to the target grid in the target region into the target model, wherein the target image block includes n×n grids, the n×n grids include the target grid and the grids in the neighborhood of the target grid, and the first data set includes n×n×C feature vectors, where C is the number of population influencing factors, n is an integer greater than 1, and C is an integer greater than or equal to 1;

[0038] In this embodiment, the above method is applied to perform population gridding on a target area, which can be a geographical region of different levels, such as a city, a county-level administrative region, or a township. The target area is divided into grids according to the target resolution (i.e., the selected grid size), and the population count in each grid is calculated.

[0039] When performing population gridding on the target area, it is necessary to calculate the population in each grid. The target grid mentioned above is any one of the multiple grids obtained by dividing the target area. When calculating the population in the target grid, the first dataset is used as the input of the target model, and the target model outputs the population weight corresponding to the target grid. Then, the population in the target grid is determined based on the population weight.

[0040] Figure 2 This is a schematic diagram of a target image block according to an embodiment of the present invention, such as... Figure 2 As shown, the target area is divided into m1×m2 grids. Figure 2The shaded area represents the target image patch. The grid at the center of the shaded area corresponds to the aforementioned target grid. The target image patch is the image patch extracted from the target region that corresponds to the target grid. Specifically, when extracting the target image patch, in addition to including the target grid itself, the target image patch also extends n grids around grid A, i.e., the grids in the target grid's neighborhood, finally forming an image patch containing n×n grids (corresponding to...). Figure 2 (shaded area), where m1 and m2 are integers greater than or equal to 1, and both m1 and m2 are greater than or equal to n. m1 and m2 can be equal or not.

[0041] The first dataset includes data on C types of population impact factors corresponding to each grid in the target image patch. Specifically, the first dataset includes n×n×C feature vectors, where each feature vector represents the data of a population impact factor for a corresponding grid. These population impact factors are factors that affect population size. The population size within the corresponding target grid is estimated based on the population impact factor data within the target image patch. Population impact factors can be divided into C types. The following are some example types of population impact factors: nighttime light, DEM (Digital Elevation Model), NDVI (Normalized Difference Vegetation Index), and various categories of POIs, etc.

[0042] Optionally, before inputting the first data set of target image patches corresponding to the target grid in the target area into the target model, the method further includes: acquiring remote sensing data and POI data corresponding to the target area, wherein the remote sensing data includes data corresponding to C1 type population influence factors, and the POI data includes data corresponding to C2 type population influence factors, where C = C1 + C2; resampling the data corresponding to each type of population influence factor in the remote sensing data and the POI data to the target resolution to obtain a third data set, wherein the target resolution is consistent with the size of the target grid, and the third data set includes m1 × m2 × C feature vectors; and determining the first data set from the third data set.

[0043] Before inputting the first dataset into the target model, it is necessary to obtain the remote sensing data and POI data corresponding to the target area, and preprocess the remote sensing data and POI data to obtain the first dataset.

[0044] Specifically, remote sensing data is data about the Earth's surface acquired through satellites, aircraft, or other sensors. POI (Point of Interest) data refers to data identifying geographical locations with specific significance or value on maps or in geographic information systems. Remote sensing data includes data corresponding to C1-type population impact factors. For example, if C1-type population impact factors are nighttime light, DEM, and NDVI, then the remote sensing data includes data corresponding to nighttime light, DEM, and NDVI. POI data includes data corresponding to C2-type population impact factors. For example, if C2-type population impact factors are catering services, commercial facilities, cultural venues, tourist attractions, educational institutions, and transportation facilities, then the POI data includes data corresponding to catering services, commercial facilities, cultural venues, etc. Here, C1 and C2 are integers greater than or equal to 1, and both C1 and C2 are less than C. C1 and C2 may or may not be equal.

[0045] The remote sensing and POI data contain a total of C population impact factors, where C = C1 + C2. For each population impact factor in the remote sensing and POI data, it needs to be resampled to the target resolution. The target resolution refers to the smallest unit size, which is consistent with the size of the target grid. This involves calculating the data for each population impact factor in each grid within the target region. If the target region is divided into m1 × m2 grids, the data for each population impact factor in the remote sensing and POI data is resampled to the target resolution, resulting in m1 × m2 × C feature vectors. One feature vector indicates the data for a specific population impact factor within the corresponding grid. The third dataset includes population impact factor data for all grids in the target region. Therefore, when determining the first dataset for the target image patch within the target region, the aforementioned first dataset is extracted from the third dataset based on the position of each grid within the target image patch.

[0046] Optionally, for the collected C2 class POI data, it is converted into a hierarchical density heatmap tensor through kernel density analysis. In the density heatmap tensor, each layer corresponds to the spatial distribution of a class of POIs (a population influence factor), and each layer in the density heatmap tensor is resampled to the target resolution.

[0047] Optionally, before inputting the first dataset into the target model, model training is required to obtain the target model. That is, before inputting the first dataset of target image patches corresponding to the target grid in the target region into the target model, the method further includes: acquiring remote sensing data and POI data corresponding to the target region, wherein the remote sensing data includes data corresponding to C1 type population influence factors, and the POI data includes data corresponding to C2 type population influence factors, where C = C1 + C2; resampling the data corresponding to each type of population influence factor in the remote sensing data and the POI data to a first resolution to obtain a second dataset; acquiring a first population gridded dataset at the first resolution corresponding to the target region, wherein the first resolution is greater than the target resolution, and the target resolution is consistent with the size of the target grid; and using the second dataset and the first population gridded dataset as training sets to train the model to obtain the target model.

[0048] In this embodiment, the first population gridded dataset with a first resolution refers to a target area divided into multiple grids, where the first resolution refers to the size of a single grid, and the first population gridded dataset includes population statistics within each grid of the first resolution size. The first population gridded dataset is an existing dataset obtained by performing coarse-resolution (i.e., first-resolution) population gridding on the target area. The first population gridded dataset can be obtained by gridding the target area using relevant technologies. In this embodiment, a coarse-resolution population gridded dataset, i.e., the first population gridded dataset, is directly collected from a database or network.

[0049] Optionally, the obtained first population gridded dataset may have issues with data reliability and accuracy. Therefore, the first population gridded dataset is corrected. Specifically, it is corrected using census data from the target area. The corrected first population gridded dataset is obtained by using the following formula:

[0050]

[0051] Among them, y i and y′ i Let y represent the population size in the i-th grid before and after the correction, respectively. j P represents the population within the j-th grid of the D-th street in the target area. D A represents the population of the Dth street in the census data. j This represents the area of ​​the j-th grid.

[0052] In addition, the data corresponding to each population influence factor in the remote sensing data and POI data of the target area need to be resampled to the first resolution to obtain the second dataset. The second dataset and the first population gridded dataset, or the second dataset and the corrected first population gridded dataset, are used as the training set to train the model to obtain the target model. The second dataset is used as an independent variable input into the model to be trained, and the first population gridded dataset or the corrected first population gridded dataset is used as the training label, i.e., the target output value, to supervise the training process of the model.

[0053] The remote sensing data and POI data are resampled to a first resolution to obtain a second dataset. A first population grid dataset, obtained from coarse-resolution population gridding of the target area using other related techniques, is also acquired. The target model is trained using the second dataset and the first population grid dataset. The target model learns the mapping relationship between population influencing factors and population at the coarse resolution (first resolution). Based on this learned mapping relationship, the target model receives high-resolution (target resolution) remote sensing data and POI data (first dataset) as input. It determines the population weight of each grid in the target area at high resolution based on the learned mapping relationship, and then determines the population size within the corresponding grid based on the population weight. For example, ... Figure 3 As shown, the first resolution is 100m, and the target resolution is 25m. After resampling the acquired multi-source data (remote sensing data and POI data), the population distribution influencing factor data at 100m resolution was extracted, which is the third dataset. The population gridded data at 100m resolution was obtained. The model was trained using the third dataset and the population gridded data at 100m resolution to obtain the target model. The population distribution influencing factor data at 25m resolution was input into the target model, and the population weight factor distribution at 25m resolution was calculated using the trained target model.

[0054] Step S102: The global spatial relationship extraction module in the target model extracts the global spatial relationship features of the target image patch based on the first data set, and the local spatial relationship extraction module in the target model extracts the local spatial relationship features from the global spatial relationship features.

[0055] Optionally, the step of extracting the global spatial relationship features of the target image patch based on the first data set by the global spatial relationship extraction module in the target model includes: mapping the feature vectors in the first data set to the target embedding space to obtain a target embedding sequence; and exchanging and transmitting information between the target embedding sequence through multiple attention units to obtain global spatial relationship features, wherein the global spatial relationship features include C-layer spatial features, and each layer of spatial features in the C-layer spatial features is used to represent a global spatial relationship feature of one of the population influence factors.

[0056] In this embodiment, the global spatial relationship extraction module considers the spatial relationships between different grids within the target image patch to capture global spatial relationships. Specifically, the first set of data of the target image patch corresponding to the target grid is first used as input features, where a population influence factor corresponds to the input feature f of the target image patch. v Notation:

[0057]

[0058] in, This represents the characteristics of all grids in a neighborhood formed by extending n grids outward from grid V, centered at grid V. Each grid V... ij It can be represented as (Attri ij1 ,...,Attri ijc ), where i and j are the location attributes of grid V, and c is the information attribute, corresponding to a population influence factor.

[0059] The feature vectors in the first dataset are mapped to the target embedding space to obtain the target embedding sequence. Specifically, the input features of the target image patch corresponding to each population influencing factor can be mapped to the target embedding space, and the learnable embedding f can be... learn Add to the embedding sequence to enhance the original feature representation, adjust the feature space, and position embedding E pos Adding to this embedding space to preserve location information, the mapping from the target population influence factor to the target embedding space is completed using the following formula, resulting in a target embedding feature:

[0060]

[0061] Where z0 is the target embedding feature, and N = n 2 , Let E be the feature vector corresponding to the target population influence factor of the m-th grid in the first dataset, and let E be the mapping parameter.

[0062] By using multiple attention units to interact and transmit information about the target embedding sequence, global spatial relationship features are obtained. Specifically,

[0063]

[0064] in, For the first The output of each attention unit, where L is the number of attention units.

[0065] One population influence factor corresponds to one global feature, and the global feature corresponding to any one population influence factor is z. L That is, the above-mentioned global spatial relationship features include C-level spatial features, and each layer of spatial features in the C-level spatial features is used to represent a global feature of one of the population influence factors.

[0066] Optionally, the local spatial relationship features are extracted from the global spatial relationship features by the local spatial relationship extraction module in the target model, including: determining the weights corresponding to each layer of spatial features in the global spatial relationship features to obtain C fusion weights; multiplying each of the C fusion weights with the corresponding layer of spatial features in the global spatial relationship features to obtain a fusion feature map, wherein the fusion feature map includes C layers of fusion features; and performing a convolution operation on the fusion feature map by the convolution block in the local spatial relationship extraction module to obtain the local spatial relationship features.

[0067] In this embodiment, when extracting local spatial relationship features, it is first necessary to fuse multiple layers of spatial features in the global spatial relationship features to obtain a fused feature map, so as to learn the complementary features of multiple population influencing factor data. Convolution operation is performed on the fused feature map to obtain local spatial relationship features, so as to perceive local features more accurately.

[0068] Specifically, when fusing multiple layers of spatial features in the global spatial relationship features, it is necessary to calculate the weight of each layer of spatial features, that is, to obtain C fusion weights. The C fusion weights are then multiplied by the corresponding layer of spatial features in the global spatial relationship features to obtain C layers of fusion features. The C layers of fusion features constitute the fusion feature map.

[0069] Optionally, determining the weights corresponding to each layer of spatial features in the global spatial relationship features to obtain C fusion weights includes:

[0070] Perform global average pooling encoding on each layer of spatial features in the global spatial relationship features to obtain C encoded features;

[0071] The C fusion weights are obtained using the following formula:

[0072]

[0073] Where δ(x) is the ReLU function, σ(x) is the Sigmoid function, and W L It is the weight matrix of the first fully connected layer. It is the weight matrix of the second fully connected layer, l f Including the C encoded features, w f This includes the aforementioned C fusion weights.

[0074] In calculating the weights of each layer of spatial features, a global average pooling encoding operation is first performed on each layer of spatial features. The following formula is used to perform a global average pooling encoding operation on the c-th layer of spatial features in the global spatial relationship features to obtain the encoded feature l. c :

[0075]

[0076] Among them, F f (i,j,c) represents the grid V in the global spatial relation features. ij The corresponding c-th type of population influence factor features, where n is the side length of the target image patch, and i and j are the grid values ​​V. ij The positional attribute.

[0077] The global spatial relationship features are converted into C encoded features, i.e., l f =[l1,l2,…,l C ].

[0078] When calculating the fusion weights, a gating mechanism consisting of two fully connected layers: an FC layer (the first fully connected layer) and a sigmoid function (the second fully connected layer) is used. The C fusion weights are obtained through the following formula:

[0079]

[0080] Where δ(x) is the ReLU function, σ(x) is the Sigmoid function, and W L It is the weight matrix of the first fully connected layer. It is the weight matrix of the second fully connected layer, l f Including the C encoded features, w f This includes the aforementioned C fusion weights.

[0081] The first fully connected layer has learnable weights. The C encoded features are upsampled to enrich the fused features, and the second fully connected layer has learnable weights. Reduce the output back to C dimensions, where R1 is the rescaling ratio.

[0082] The process of determining the fused feature map is as follows: Figure 4As shown, global average pooling encoding is performed on each layer of spatial features. Then, C fusion weights are determined through two fully connected layers (the first fully connected layer and the second fully connected layer). The C fusion weights are multiplied layer by layer with the C layer spatial features in the global spatial relationship features to fuse the features corresponding to different population influence factors, thus obtaining the fused feature map.

[0083] After determining C fusion weights, the fusion feature map is obtained by multiplying the C fusion weights layer by layer with the C-layer spatial features in the global spatial relationship features to fuse the features corresponding to different population influencing factors.

[0084]

[0085] Among them, w f Including the aforementioned C fusion weights, w1 is the first fusion weight among the C fusion weights, w2 is the second fusion weight among the C fusion weights, and so on, w C Let F be the Cth fusion weight out of C fusion weights. f For global spatial relationship features, including C-level spatial features, F f1 F is the first layer spatial feature in the C-layer spatial features. f2 This is the second layer of spatial features in layer C, and so on, F fC This refers to the Cth layer spatial feature within the C-layer spatial features.

[0086] After obtaining the fused feature map, a convolution operation is performed on the fused feature map to obtain local spatial relationship features. For example, the fused feature map can be used to... Applying a 3×3 convolution and padding to obtain a new feature map of the same size as the fused feature map, namely the local spatial relation feature F. s :

[0087]

[0088] in, W represents convolution. s It is a 3×3 convolution kernel, b s It is F f Bias of the same dimension. δ(x) is the ReLU activation function.

[0089] Step S103: Determine the population weight of the target grid based on the local spatial features, wherein the population weight is used to determine the population size in the target grid.

[0090] In this embodiment, the target model outputs the population weights of the grid at the target resolution. However, the actual population distribution is controlled by the total population of each district. Based on the population weights of the generated target grid, the census data is decomposed from each district into the population of the target grid using a formula:

[0091]

[0092] Where w grid_V It is the population weight of grid V, w district_P POP represents the total population weight of all grids in region P to which grid V belongs. district_P It refers to the census population of area P. grid_V This represents the predicted population size for grid V, which corresponds to the target grid.

[0093] Through the above embodiments, when determining the population weight of the target grid, the data of multiple population factors corresponding to the target grid and multiple grids in the target grid's neighborhood, i.e., the first dataset, is used as the input to the target model. Compared with only inputting the data of multiple population influence factors corresponding to the target grid, the first dataset can better reflect the spatial correlation of population distribution. The global spatial relationship extraction module in the target model extracts global spatial relationship features, and local spatial relationship features are extracted from the global spatial relationship features. Finally, the extracted local spatial relationship features aggregate global and local spatial relationship information, making full use of the spatial correlation of the population. This makes the population weight determined based on the above local spatial relationship features more accurate, and thus can accurately determine the population size in the target grid.

[0094] Optionally, the process for determining population weights in the target model is as follows: Figure 5 As shown, the first dataset is divided into two parts: one part is remote sensing data of the target area, including data corresponding to the C1 type of population impact factor, and the other part is POI data of the target area, including data corresponding to the C2 type of population impact factor, where C = C1 + C2.

[0095] The first dataset is input into the target model, where the global spatial relationship extraction module extracts global spatial relationship features. Specifically, the global spatial relationship extraction module includes two global spatial relationship extraction units (the first global spatial relationship extraction unit and the second global spatial relationship extraction unit). The remote sensing data and POI data from the first dataset are input into the first and second global spatial relationship extraction units, respectively. The two global spatial relationship extraction units independently extract the global spatial relationships of the C-level population influence factors contained in the remote sensing data and POI data. The global spatial relationship extraction module outputs the aforementioned global spatial relationship features. The local spatial relationship extraction module includes a fusion unit and a convolutional block. The fusion unit determines the weights corresponding to the C-level spatial features in the global spatial relationship features and fuses the C-level spatial features according to the weights to obtain a fused feature map. The convolutional block performs a convolution operation on the fused feature map to obtain local spatial relationship features. Finally, the population weights corresponding to the target grid are determined based on the local spatial relationship features.

[0096] like Figure 6 As shown, in another embodiment of the present invention, a grid population determination device is provided, the device comprising:

[0097] The input unit 601 is used to input the first data set of the target image block corresponding to the target grid in the target region into the target model. The target image block includes n×n grids, and the n×n grids include the target grid and the grids in the neighborhood of the target grid. The first data set includes n×n×C feature vectors, where C is the number of types of population influencing factors, n is an integer greater than 1, and C is an integer greater than or equal to 1.

[0098] Extraction unit 602 is used to extract global spatial relationship features of the target image block according to the first data set through the global spatial relationship extraction module in the target model, and to extract local spatial relationship features from the global spatial relationship features through the local spatial relationship extraction module in the target model;

[0099] The determining unit 603 is used to determine the population weight of the target grid based on the local spatial features, wherein the population weight is used to determine the population size in the target grid.

[0100] In an optional embodiment, the above-described apparatus is further configured to, before inputting a first data set of target image patches corresponding to the target grid in the target region into the target model, acquire remote sensing data and POI data corresponding to the target region, wherein the remote sensing data includes data corresponding to C1 type population influence factors, and the POI data includes data corresponding to C2 type population influence factors, where C = C1 + C2; resample the data corresponding to each type of population influence factor in the remote sensing data and the POI data to a first resolution to obtain a second data set; acquire a first population gridded dataset corresponding to the first resolution of the target region, wherein the first resolution is greater than the target resolution, and the target resolution is consistent with the size of the target grid; and use the second data set and the first population gridded dataset as a training set to train the model to obtain the target model.

[0101] In an optional embodiment, the above-described apparatus is further configured to: after acquiring a first population gridded dataset of a first resolution corresponding to the target area, acquire census data of the target area; correct the first population gridded dataset according to the census data to obtain a corrected first population gridded dataset; and use the second dataset and the corrected first population gridded dataset as a training set to train a model to obtain the target model.

[0102] The first population gridded dataset is corrected using the following formula:

[0103]

[0104] Among them, y i and y′ i Let y represent the population size in the i-th grid before and after the correction, respectively. j P represents the population within the j-th grid of the D-th street in the target area. D A represents the population of the Dth street in the census data. j This represents the area of ​​the j-th grid.

[0105] In an optional embodiment, the above-described apparatus is further configured to, before inputting a first data set of target image patches corresponding to the target grid in the target area into the target model, acquire remote sensing data and POI data corresponding to the target area, wherein the remote sensing data includes data corresponding to C1 type population influence factors, and the POI data includes data corresponding to C2 type population influence factors, where C = C1 + C2; resample the data corresponding to each type of population influence factor in the remote sensing data and the POI data to the target resolution to obtain a third data set, wherein the target resolution is consistent with the size of the target grid, and the third data set includes m1 × m2 × C feature vectors; and determine the first data set from the third data set.

[0106] In an optional embodiment, the above-described apparatus is further configured to extract global spatial relationship features of the target image patch from the first data set through the global spatial relationship extraction module in the target model in the following manner: mapping the feature vectors in the first data set to the target embedding space to obtain a target embedding sequence, wherein the target embedding sequence includes C target embedding features; and performing information interaction and transmission on the target embedding sequence through multiple attention units to obtain global spatial relationship features, wherein the global spatial relationship features include C layers of spatial features, and each layer of spatial features in the C layers of spatial features is used to represent a global feature of one of the population influence factors.

[0107] In an optional embodiment, the above-described apparatus is further configured to extract local spatial relationship features from the global spatial relationship features through the local spatial relationship extraction module in the target model in the following manner: determining the weights corresponding to each layer of spatial features in the global spatial relationship features to obtain C fusion weights; multiplying each of the C fusion weights with a corresponding layer of spatial features in the global spatial relationship features to obtain a fusion feature map, wherein the fusion feature map includes C layers of fusion features; and performing a convolution operation on the fusion feature map through a convolution block in the local spatial relationship extraction module to obtain the local spatial relationship features.

[0108] In an optional embodiment, the above-described apparatus is further configured to determine the weights corresponding to each layer of spatial features in the global spatial relationship features, thereby obtaining C fusion weights, by performing a global average pooling encoding operation on each layer of spatial features in the global spatial relationship features to obtain C encoded features; and obtaining the C fusion weights using the following formula:

[0109]

[0110] Where δ(x) is the ReLU function, σ(x) is the Sigmoid function, and WL It is the weight matrix of the first fully connected layer. It is the weight matrix of the second fully connected layer, l f Including the C encoded features, w f This includes the aforementioned C fusion weights.

[0111] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0112] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for determining the population size of a grid, characterized in that, include: The first data set of the target image block corresponding to the target grid in the target region is input into the target model. The target image block includes n×n grids, and the n×n grids include the target grid and the grids in the neighborhood of the target grid. The first data set includes n×n×C feature vectors, where C is the number of population influencing factors, n is an integer greater than 1, and C is an integer greater than or equal to 1. The global spatial relationship extraction module in the target model extracts the global spatial relationship features of the target image patch based on the first data set, and the local spatial relationship extraction module in the target model extracts the local spatial relationship features from the global spatial relationship features. The population weight of the target grid is determined based on the local spatial relationship features, wherein the population weight is used to determine the population size in the target grid; before inputting the first data set of target image patches corresponding to the target grid in the target region into the target model, the method further includes: Acquire remote sensing data and POI data corresponding to the target area. The remote sensing data includes data corresponding to the C1 type population impact factor, and the POI data includes data corresponding to the C2 type population impact factor. C = C1 + C2. The C1 type population impact factors are nighttime light, DEM, and NDVI. Therefore, the remote sensing data includes data corresponding to nighttime light, DEM, and NDVI. The data corresponding to each type of population impact factor in the remote sensing data and the POI data are resampled to the first resolution to obtain the second dataset. Obtain a first population gridded dataset with a first resolution corresponding to the target region, wherein the first resolution is greater than the target resolution, and the target resolution is consistent with the size of the target grid. The second dataset and the first population gridded dataset are used as training sets to train the model, thereby obtaining the target model; the step of extracting the global spatial relationship features of the target image patch based on the first dataset using the global spatial relationship extraction module in the target model includes: The feature vectors in the first dataset are mapped to the target embedding space to obtain a target embedding sequence, wherein the target embedding sequence includes C target embedding features; By interacting and transmitting information on the target embedding sequence through multiple attention units, global spatial relationship features are obtained. The global spatial relationship features include C-layer spatial features, and each layer of spatial features in the C-layer spatial features is used to represent a global feature of one of the population influence factors. The local spatial relationship extraction module in the target model extracts local spatial relationship features from the global spatial relationship features, including: Determine the weights corresponding to each layer of spatial features in the global spatial relationship features to obtain C fusion weights; Each of the C fusion weights is multiplied by a corresponding layer of spatial features in the global spatial relationship features to obtain a fusion feature map, wherein the fusion feature map includes C layers of fusion features; The local spatial relationship features are obtained by performing a convolution operation on the fused feature map using the convolution block in the local spatial relationship extraction module.

2. The method for determining the population size of a grid according to claim 1, characterized in that, After obtaining the first population gridded dataset with a first resolution corresponding to the target region, the method further includes: Obtain population census data for the target area; The first population gridded dataset is corrected based on the census data to obtain the corrected first population gridded dataset; The second dataset and the corrected first population gridded dataset are used as training sets to train the model, thereby obtaining the target model; The first population gridded dataset is corrected using the following formula: in, and These represent the first and second digits before and after the correction, respectively. Population within each grid Indicates the first in the target region The street Population within each grid The first in the population census data Population of each street Indicates the first The area of ​​each grid cell.

3. The method for determining the population size of a grid according to claim 1, characterized in that, Before inputting the first set of data of the target image patches corresponding to the target grid in the target region into the target model, the method further includes: Acquire remote sensing data and POI data corresponding to the target area, wherein the remote sensing data includes data corresponding to the C1 type population impact factor, and the POI data includes data corresponding to the C2 type population impact factor, where C = C1 + C2; The data corresponding to each population influence factor in the remote sensing data and the POI data are resampled to the target resolution to obtain a third data set, wherein the target resolution is consistent with the size of the target grid, and the third data set includes m1×m2×C feature vectors; The first data set is determined from the third data set.

4. The method for determining the population size of a grid according to claim 1, characterized in that, The step of determining the weights corresponding to each layer of spatial features in the global spatial relationship features yields C fusion weights, including: Perform global average pooling encoding on each layer of spatial features in the global spatial relationship features to obtain C encoded features; The C fusion weights are obtained using the following formula: in, It is the ReLU function. It is the Sigmoid function. It is the weight matrix of the first fully connected layer. It is the weight matrix of the second fully connected layer. Including the C encoded features, This includes the aforementioned C fusion weights.

5. A grid population determination device, characterized in that, include: The input unit is used to input the first data set of the target image block corresponding to the target grid in the target region into the target model. The target image block includes n×n grids, and the n×n grids include the target grid and the grids in the neighborhood of the target grid. The first data set includes n×n×C feature vectors, where C is the number of types of population influencing factors, n is an integer greater than 1, and C is an integer greater than or equal to 1. The extraction unit is used to extract the global spatial relationship features of the target image patch based on the first data set through the global spatial relationship extraction module in the target model, and to extract the local spatial relationship features from the global spatial relationship features through the local spatial relationship extraction module in the target model. The determining unit is configured to determine the population weight of the target grid based on the local spatial relationship features, wherein the population weight is used to determine the population size in the target grid; before inputting the first data set of target image patches corresponding to the target grid in the target region into the target model, the device is further configured to: Acquire remote sensing data and POI data corresponding to the target area. The remote sensing data includes data corresponding to the C1 type population impact factor, and the POI data includes data corresponding to the C2 type population impact factor. C = C1 + C2. The C1 type population impact factors are nighttime light, DEM, and NDVI. Therefore, the remote sensing data includes data corresponding to nighttime light, DEM, and NDVI. The data corresponding to each type of population impact factor in the remote sensing data and the POI data are resampled to the first resolution to obtain the second dataset. Obtain a first population gridded dataset with a first resolution corresponding to the target region, wherein the first resolution is greater than the target resolution, and the target resolution is consistent with the size of the target grid. The second dataset and the first population gridded dataset are used as training sets to train the model, thereby obtaining the target model; the step of extracting the global spatial relationship features of the target image patch based on the first dataset using the global spatial relationship extraction module in the target model includes: The feature vectors in the first dataset are mapped to the target embedding space to obtain a target embedding sequence, wherein the target embedding sequence includes C target embedding features; By interacting and transmitting information on the target embedding sequence through multiple attention units, global spatial relationship features are obtained. The global spatial relationship features include C-layer spatial features, and each layer of spatial features in the C-layer spatial features is used to represent a global feature of one of the population influence factors. The local spatial relationship extraction module in the target model extracts local spatial relationship features from the global spatial relationship features, including: Determine the weights corresponding to each layer of spatial features in the global spatial relationship features to obtain C fusion weights; Each of the C fusion weights is multiplied by a corresponding layer of spatial features in the global spatial relationship features to obtain a fusion feature map, wherein the fusion feature map includes C layers of fusion features; The local spatial relationship features are obtained by performing a convolution operation on the fused feature map using the convolution block in the local spatial relationship extraction module.

6. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement the grid population determination method as described in any one of claims 1 to 4 when executing the computer program.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, it implements the grid population determination method as described in any one of claims 1-4.

Images