Method for estimating residential population by fusing remote sensing image data and grid population data

Abstract

The application discloses a residential population estimation method fusing remote sensing image data and grid population data, and relates to the technical field of geographic space artificial intelligence. The method comprises the following steps: acquiring night light image data, remote sensing image data and grid population data, and performing feature fusion on the night light image data and the remote sensing image data through a convolutional neural network; judging residential buildings by using a deep learning model according to the fused image feature data, the remote sensing image data and the grid population data, so as to obtain residential building data; dividing regional building levels by using the residential building data and the grid population data, and performing data sampling on buildings in each level, so as to obtain training samples; training a double-path convolutional neural network by using the fused image feature data and the grid population data of the training samples, so as to obtain a population prediction model; and predicting all buildings by using the population prediction model, so as to obtain a population prediction value of each building.

Description

Technical Field

[0001] This invention relates to the field of geospatial artificial intelligence technology, and more specifically, to a method for estimating resident population by fusing remote sensing image data and grid population data. Background Technology

[0002] High-precision resident population estimation plays an important supporting role in refined social management and land space governance. Its significance is mainly reflected in the optimization of social management and public services, dynamic supplementation of population census, service to urban planning and smart city construction, and support for economic development and business decision-making.

[0003] For social management and public service optimization, as well as for dynamic supplementation of the population census, high-resolution population distribution helps identify differences in demand in different regions, provides a scientific basis for the allocation of social resources such as education, medical care, and elderly care, and can also plan the location and capacity of public facilities (such as schools and hospitals) based on population density to avoid resource waste or shortage. At the same time, the traditional population census cycle is long and it is difficult to reflect the dynamic changes in the population in a timely manner. High-resolution population forecast can provide real-time updated supplementary data to provide a basis for adjustments.

[0004] For urban planning and smart city construction, land use and housing planning, and economic development and business decision support, population forecasting helps optimize infrastructure such as transportation networks, water and electricity supply systems, and waste disposal facilities. This avoids resource waste or overloading caused by uneven population distribution, improves urban operational efficiency, reduces energy consumption, and builds smart cities. Meanwhile, accurate building-level population estimation helps optimize land use planning, balance the proportion of residential, commercial, and industrial areas, and provide a basis for real estate development. High-resolution population data can also provide businesses with market potential assessment tools, accurately analyze the market, and help determine store locations, advertising placements, and logistics network layouts. Especially in industries such as retail, e-commerce, and express delivery, detailed population distribution data directly impacts decision-making.

[0005] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0006] To address the problems in related technologies, this invention proposes a method for estimating resident population by fusing remote sensing image data and grid population data, thereby overcoming the aforementioned technical problems existing in existing related technologies.

[0007] Therefore, the specific technical solution adopted by the present invention is as follows:

[0008] A method for estimating the resident population by fusing remote sensing image data and gridded population data, comprising the following steps:

[0009] S1. Acquire nighttime light image data, remote sensing image data, and grid population data. Use a convolutional neural network to fuse the features of the nighttime light image data and the remote sensing image data to obtain fused image feature data.

[0010] S2. Based on the fused image feature data, remote sensing image data, and grid population data, a deep learning model is used to determine residential buildings and obtain residential building data.

[0011] S3. Using residential building data and grid population data, divide the regional building hierarchy and sample the data of buildings in each hierarchy to obtain training samples.

[0012] S4. Using the fused image feature data and grid population data of the training samples, train the dual-path convolutional neural network to obtain the population prediction model.

[0013] S5. Use a population prediction model to predict the population of all buildings, obtain the predicted population value for each building, and use Kriging interpolation to verify the population prediction value.

[0014] Furthermore, nighttime light imagery data, remote sensing imagery data, and grid-based population data are acquired. A convolutional neural network is then used to fuse the features of the nighttime light imagery data and the remote sensing imagery data to obtain fused image feature data. This process includes the following steps:

[0015] S11. Acquire nighttime light image data, remote sensing image data, and grid population data, and preprocess the nighttime light image data and remote sensing image data to obtain standardized nighttime light image data and standardized remote sensing image data.

[0016] S12. Based on standardized remote sensing image data, extract remote sensing image features using a pre-trained residual network model to obtain remote sensing image feature data; simultaneously, based on standardized nighttime light image data, extract light features using a lightweight deep neural network model to obtain light feature data.

[0017] S13. Based on remote sensing image feature data and light feature data, feature fusion is performed using a convolutional neural network to obtain fused image feature data.

[0018] Furthermore, acquiring nighttime light imagery data, remote sensing imagery data, and grid-based population data, and preprocessing the nighttime light imagery data and remote sensing imagery data to obtain standardized nighttime light imagery data and standardized remote sensing imagery data includes the following steps:

[0019] S111. Acquire nighttime light image data, remote sensing image data, and grid population data. Align the nighttime light image data and remote sensing image data using bilinear interpolation to obtain aligned nighttime light image data and remote sensing image data.

[0020] S112. Normalize the aligned remote sensing image data to obtain normalized remote sensing image data, and perform logarithmic transformation on the aligned night light image data to obtain transformed night light image data.

[0021] S113. The normalized remote sensing image data and the transformed nighttime light image data are processed into blocks to obtain standardized nighttime light image data and standardized remote sensing image data.

[0022] Furthermore, based on the fused image feature data, remote sensing image data, and grid population data, a deep learning model is used to determine residential buildings, resulting in residential building data including the following steps:

[0023] S21. Align the grid population data with the remote sensing image data to obtain the aligned grid population data, and extract the building features from the remote sensing image data to obtain the building feature data.

[0024] S22. Using the aligned grid population data, fused image feature data, remote sensing image data and building feature data, the deep learning model is trained to obtain the initial residential building recognition model.

[0025] S23. Based on building feature data and aligned grid population data, construct a total loss function using prior constraints, optimize the initial residential building identification model, and obtain the residential building identification model.

[0026] S24. Identify residential buildings using a residential building identification model to obtain residential building data, and verify the residential building data.

[0027] Furthermore, based on building feature data and aligned grid population data, a total loss function is constructed using prior constraints to optimize the initial residential building identification model. The resulting residential building identification model includes the following steps:

[0028] S231. Based on the building feature data, generate a building area classification loss function using the cross-entropy loss function and the Dice loss function;

[0029] S232. Based on building feature data and aligned grid population data, generate a population distribution prior loss function using a prior constraint loss function.

[0030] S233. Using the building area classification loss function and the population distribution prior loss function, construct the total loss function to optimize the initial residential building identification model and obtain the residential building identification model.

[0031] Further validation of the residential building data includes: crossover ratio (CR) validation and F1 score validation.

[0032] Furthermore, by utilizing residential building data and grid population data, the regional building hierarchy is divided, and data sampling is performed on buildings at each hierarchy to obtain training samples, including the following steps:

[0033] S31. Based on residential building data and grid population data, determine the population density of buildings in the area, and classify the building levels in the area according to the population density of buildings in the area.

[0034] S32. Based on the divided regional building hierarchy, calculate the sample size of buildings in each level of the region;

[0035] S33. Samples are extracted according to the sample extraction quantity to obtain training samples, and consistency verification is performed on the training samples.

[0036] Furthermore, based on the divided regional building hierarchy, the formula for calculating the sample size of buildings in each hierarchy is as follows:

[0037] ;

[0038] In the formula, N h n represents the total number of buildings on the h-th floor; h This represents the number of samples to be drawn from the h-th layer; n represents the total number of samples; N represents the total number of buildings.

[0039] The formula for calculating the total sample size is as follows:

[0040] ;

[0041] In the formula, Z represents the score corresponding to the confidence level; p represents the proportion estimate of the target variable; and e represents the allowable sampling error.

[0042] Furthermore, by using the fused image feature data and grid population data from the training samples, a dual-path convolutional neural network is trained to obtain a population prediction model, including the following steps:

[0043] S41. Preprocess the fused image feature data and grid population data of the training samples to obtain multimodal fused data and preprocessed grid population data;

[0044] S42. Based on the multimodal fusion data and the preprocessed grid population data, extract the fused image features and grid population features respectively through a convolutional neural network, and generate the fused image feature weights and grid population feature weights using an attention mechanism.

[0045] S43. Based on the weights of the fused image features and the weights of the grid population features, calculate the weighted fused image features and grid population features, and then stitch and fuse the weighted fused image features and grid population features to obtain the stitched features.

[0046] S44. Input the spliced ​​features into the dual-path convolutional neural network, and optimize and train the dual-path convolutional neural network using the mean squared error loss function and regularization constraints to obtain the population prediction model.

[0047] Furthermore, population prediction models are used to forecast the population of all buildings, resulting in a predicted population value for each building. Kriging interpolation is then used to validate these population predictions, including the following steps:

[0048] S51. Input the pre-acquired nighttime light image data, remote sensing image data and grid population data of all buildings into the population prediction model, make predictions for all buildings, and obtain the population prediction value for each building.

[0049] S52. Based on the population forecast values ​​of all buildings, obtain the predicted population distribution and use the semivariogram function to calculate the unknown point values.

[0050] S53. Use the Kriging interpolation method to perform population prediction interpolation on unknown points, generate population distribution maps and building predicted population value distribution maps, verify the spatial continuity of the building population prediction values, and use the power law distribution to verify the power law characteristics of the population distribution.

[0051] The beneficial effects of this invention are as follows:

[0052] 1. This invention fully utilizes the complementarity of nighttime light imagery and remote sensing imagery to enhance the joint expressive power of the data. Nighttime light imagery can reflect regional economic activities and population density; remote sensing imagery can reflect building forms (area, shape) and environmental characteristics (green coverage, surrounding building density), details that are usually difficult to capture through traditional statistical data. At the same time, by using high-resolution imagery for prediction, compared with the generally low resolution of traditional population models (such as one-kilometer raster), this scheme significantly improves the accuracy and resolution of population estimation. Furthermore, compared with traditional models that are based on only a single data source (such as census data or low-resolution raster population data), this method is more advanced in its data usage, provides richer information, and has higher spatial resolution. In short, the detailed utilization of the data and the extraction of diverse features far exceed those of traditional models.

[0053] 2. This invention provides prior constraints on population distribution for residential building identification models by using grid population data, guiding the models to focus on buildings that are more likely to have residents; compared with traditional models that are mostly based on aggregated administrative divisions or grid population data, it improves the fine-grained characterization of individual buildings.

[0054] 3. This invention employs a dual-path convolutional neural network. One path processes fused image feature data to extract architectural and optical features, while the other path processes grid-based population data to provide prior constraints. Features are fused through a cross-modal attention mechanism to uncover interactive information between different modalities. Compared to traditional models that typically rely on statistical regression or simple machine learning methods, this invention is better suited for handling multimodal high-dimensional data. The cross-modal attention fusion enhances information interaction between different data sources and improves the feature representation capability of key areas. Compared to traditional models that typically perform simple weighted fusion, this invention focuses on the correlation between data.

[0055] 4. This invention refines the population estimation problem down to each building by fusing image feature data and grid population data, and jointly modeling at the pixel and building levels. It generates a population distribution map at the building level, which has high spatial resolution. Compared with traditional models that mostly perform coarse-grained estimation at the regional (such as zoning or grid) level, it can more accurately locate the resident population of each building. In particular, the deep learning model can capture the complex nonlinear relationship between population and building form, light intensity, etc., while traditional regression models mostly assume that the variables are linear or simple nonlinear, which is difficult to reflect the actual complexity. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart of a method for estimating the resident population by fusing remote sensing image data and grid population data according to an embodiment of the present invention;

[0058] Figure 2 This is one of the framework diagrams of a method for estimating the resident population by fusing remote sensing image data and grid population data according to an embodiment of the present invention;

[0059] Figure 3 This is the second framework diagram of a method for estimating the resident population by fusing remote sensing image data and grid population data according to an embodiment of the present invention. Detailed Implementation

[0060] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0061] According to embodiments of the present invention, a method for estimating the resident population by fusing remote sensing image data and grid population data is provided.

[0062] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, the method for estimating the resident population by fusing remote sensing image data and grid population data according to an embodiment of the present invention includes the following steps:

[0063] S1. Acquire nighttime light image data, remote sensing image data, and grid population data. Use a convolutional neural network to fuse the features of the nighttime light image data and the remote sensing image data to obtain fused image feature data.

[0064] Specifically, the process involves acquiring nighttime light imagery data, remote sensing imagery data, and grid-based population data. Then, a convolutional neural network is used to fuse the features of the nighttime light imagery data and the remote sensing imagery data to obtain fused image feature data. This includes the following steps:

[0065] S11. Acquire nighttime light image data, remote sensing image data, and grid population data, and preprocess the nighttime light image data and remote sensing image data to obtain standardized nighttime light image data and standardized remote sensing image data.

[0066] S12. Based on standardized remote sensing image data, extract remote sensing image features using a pre-trained residual network model to obtain remote sensing image feature data; simultaneously, based on standardized nighttime light image data, extract light features using a lightweight deep neural network model to obtain light feature data.

[0067] S13. Based on remote sensing image feature data and light feature data, feature fusion is performed using a convolutional neural network to obtain fused image feature data.

[0068] Specifically, acquiring nighttime light imagery data, remote sensing imagery data, and grid-based population data, and preprocessing the nighttime light imagery data and remote sensing imagery data to obtain standardized nighttime light imagery data and standardized remote sensing imagery data includes the following steps:

[0069] S111. Acquire nighttime light image data, remote sensing image data, and grid population data. Align the nighttime light image data and remote sensing image data using bilinear interpolation to obtain aligned nighttime light image data and remote sensing image data.

[0070] S112. Normalize the aligned remote sensing image data to obtain normalized remote sensing image data, and perform logarithmic transformation on the aligned night light image data to obtain transformed night light image data.

[0071] S113. The normalized remote sensing image data and the transformed nighttime light image data are processed into blocks to obtain standardized nighttime light image data and standardized remote sensing image data.

[0072] It should be noted that the alignment process between nighttime light image data and remote sensing image data using bilinear interpolation specifically involves resolution alignment between the nighttime light image data and remote sensing image data. Bilinear interpolation or resampling methods can be used to align the nighttime light image data and remote sensing image data to the same resolution.

[0073] It should be noted that bilinear interpolation is a method based on the weighted average of adjacent pixel values, which is applicable to any point in a two-dimensional plane; given the pixel value I(x,y) of the target point (x,y), its estimated value is calculated based on the four known pixel values ​​around it.

[0074] Assuming the pixel coordinates in the original nighttime light imagery and remote sensing imagery are integers, the coordinates of the four surrounding pixels of the target point (x, y) are the top left (x1, y1), top right (x2, y1), bottom left (x1, y2), and bottom right (x2, y2); the pixel value of the target point (x, y) is estimated by the following formula:

[0075] ;

[0076] The formulas for calculating u and v are:

[0077] ;

[0078] ;

[0079] In the formula, u represents the positional proportion of x between x1 and x2; v represents the positional proportion of y between y1 and y2.

[0080] The following steps are taken in the specific calculation:

[0081] The first step is to determine the target resolution. Since the resolutions of nighttime light imagery and remote sensing imagery differ, a target resolution is first selected (this approach uses the higher resolution as the target resolution). Assume the remote sensing imagery resolution is R. s (Pixel size, unit: meters); Nighttime light image data resolution is R n .

[0082] Set the target resolution to a higher resolution (such as R). s The target image size is:

[0083] ;

[0084] In the formula, H target W represents the height of the target image size. target H represents the width of the target image. n Indicates the raw height of the nighttime light image data; W n This indicates the raw width of the nighttime light image data.

[0085] The second step is coordinate mapping; that is, during the resampling process, it is necessary to map the coordinates from the target image back to the original image coordinates. The pixel coordinates of the target image are (x', y'), and their corresponding coordinates in the original image are calculated. The formula for calculating the corresponding coordinates is:

[0086] .

[0087] The third step is to find the surrounding pixels, that is, to map the target coordinates (x', y') back to the original coordinates (x, y); determine the coordinates of the four pixels surrounding (x, y), and the formula for calculating the coordinates of the four mapped pixels is:

[0088] .

[0089] The fourth step is to apply the bilinear interpolation formula; that is, to calculate the interpolation weights of (x, y). The formula for calculating the interpolation weights is:

[0090] ;

[0091] Calculate the target pixel value using four known pixel values:

[0092] ;

[0093] It should be noted that the remote sensing image data is high-resolution satellite image data.

[0094] It should be noted that the normalization process specifically involves normalizing each pixel value of the high-resolution satellite image data to [0,1].

[0095] ;

[0096] In the formula, X1 represents the normalized remote sensing image data; X1 represents the pixel value of the remote sensing image.

[0097] It should be noted that performing a logarithmic transformation on the aligned nighttime light image data can enhance the ability to distinguish faint lights.

[0098] ;

[0099] In the formula, X1 represents the transformed nighttime light image data; X2 represents the nighttime light image data.

[0100] It should be noted that the normalized remote sensing image data and the transformed nighttime light image data are processed by dividing them into blocks, that is, dividing the image into windows of a fixed size (such as 256×256) to facilitate input into the convolutional neural network (CNN).

[0101] It should be noted that, based on standardized remote sensing image data, remote sensing image features are extracted using a pre-trained residual network model. Specifically, the remote sensing image feature data is obtained by using a pre-trained residual network (ResNet) model as the backbone network, removing the fully connected layers, and outputting convolutional remote sensing image features. The backbone network uses ResNet-18; the kernel size of the first convolutional layer is 3×3, the stride is two, and the output channels are 64; the residual module contains four residual blocks with channels of 64, 128, 256, and 512, respectively.

[0102] The formula for calculating remote sensing image features is:

[0103] ;

[0104] In the formula, X1 represents the input standardized remote sensing image data, and (RGB image); F1 represents the output remote sensing image feature data, and .

[0105] It should be noted that the calculation formula for extracting light features using a lightweight deep neural network model based on standardized nighttime light image data is as follows:

[0106] ;

[0107] In the formula, X2 represents the input standardized nighttime light image data, and (Single-channel light intensity); F2 represents the output light characteristic data, and .

[0108] The backbone network of the lightweight deep neural network model is the lightweight deep neural network model (MobileNet). The kernel size in the first convolutional layer is 3×3, the stride is one, and the output channel is 32. Depthwise separable high convolution can increase nonlinearity and reduce computation.

[0109] It should be noted that the process of fusing features based on remote sensing image feature data and light feature data using a convolutional neural network to obtain fused image feature data specifically includes the following steps:

[0110] The first step is direct concatenation, which involves concatenating F1 and F2 along the channel dimension to obtain the concatenated feature F3. The calculation formula for concatenation is as follows:

[0111] ;

[0112] The second step is feature fusion, which involves introducing additional convolutional layers and non-linear activation functions to fuse features. The formula for feature fusion is as follows:

[0113] ;

[0114] In the formula, X fusion This indicates fused image feature data.

[0115] S2. Based on the fused image feature data, remote sensing image data, and grid population data, a deep learning model is used to determine residential buildings and obtain residential building data.

[0116] Specifically, based on fused image feature data, remote sensing image data, and grid population data, a deep learning model is used to determine residential buildings, resulting in residential building data, which includes the following steps:

[0117] S21. Align the grid population data with the remote sensing image data to obtain the aligned grid population data, and extract the building features from the remote sensing image data to obtain the building feature data.

[0118] S22. Using the aligned grid population data, fused image feature data, remote sensing image data and building feature data, the deep learning model is trained to obtain the initial residential building recognition model.

[0119] S23. Based on building feature data and aligned grid population data, construct a total loss function using prior constraints, optimize the initial residential building identification model, and obtain the residential building identification model.

[0120] S24. Identify residential buildings using a residential building identification model to obtain residential building data, and verify the residential building data.

[0121] Specifically, based on building feature data and aligned grid population data, a total loss function is constructed using prior constraints to optimize the initial residential building identification model. The resulting residential building identification model includes the following steps:

[0122] S231. Based on the building feature data, generate a building area classification loss function using the cross-entropy loss function and the Dice loss function;

[0123] S232. Based on building feature data and aligned grid population data, generate a population distribution prior loss function using a prior constraint loss function.

[0124] S233. Using the building area classification loss function and the population distribution prior loss function, construct the total loss function to optimize the initial residential building identification model and obtain the residential building identification model.

[0125] Specifically, the validation of residential building data includes: crossover ratio validation and F1 score validation.

[0126] It should be noted that by fusing image feature data, remote sensing image data, and grid population data, the building feature data obtained are specifically building form and environmental feature data (such as building density, greening rate, and surrounding environment). Combined with the population distribution prior loss function, residential buildings are accurately identified.

[0127] It should be noted that aligning the grid population data with the remote sensing image data to obtain the aligned grid population data is specifically a preprocessing step. This involves aligning the resolution of the grid population data with that of the remote sensing image data and resampling using bilinear interpolation to ensure spatial consistency between the two datasets.

[0128] It should be noted that extracting building features from remote sensing image data and obtaining building feature data specifically involves calculating multiple feature layers on the remote sensing image data to enhance building form and environmental feature information, such as building form features, greening rate, and texture features. These features can be automatically learned through deep learning networks, among which building form features include building area, aspect ratio, and edge density.

[0129] The greening rate is estimated using the Normalized Difference Vegetation Index (NDVI), and the formula for calculating the greening rate is:

[0130] ;

[0131] In the formula, NIR represents the near-infrared band, and RED represents the red band.

[0132] Among them, texture features are extracted using the gray-level co-occurrence matrix (GLCM) to extract information such as contrast, energy, and homogeneity.

[0133] It should be noted that training a deep learning model using aligned grid population data, fused image feature data, remote sensing image data, and building feature data requires generating a joint input by combining the aligned grid population data, fused image feature data, remote sensing image data, and building feature data into a multi-channel input.

[0134] The formula for calculating multi-channel input is:

[0135] ;

[0136] In the formula, X inputX represents the multi-channel input of a deep learning model; grid This represents the normalized value of the population grid.

[0137] It should be noted that the data needs to be divided into training and validation sets, assuming a total height * total width (H) total *W total ) pixels are labeled as residential buildings, of which a certain proportion (e.g., 70%) are included in the training set, with their row and column numbers representing height (H) and width (W), respectively; the remaining portion is included in the validation set, with its row and column numbers representing test height (H). test ) and test width (W test H+H test =H total W+W test =W total The final training set has an input dimension of H×W×C, where C is the number of channels.

[0138] It should be noted that the deep learning model chosen is the U-Net model with multi-feature fusion. The U-Net model structure is improved by using aligned grid population data, fused image feature data, remote sensing image data and building feature data, that is, by combining the building and environmental features of the fused image with the population grid prior, to comprehensively determine whether it is a residential building.

[0139] The optimized U-Net structure is as follows:

[0140] First, the input data is high-resolution image data, i.e., fused image feature data. .

[0141] Among them, auxiliary features (such as NDVI and X) grid Each of these is used as an independent channel input, with a total input dimension of (H×W×C); for example, C=5 includes RGB, NDVI, and population grid.

[0142] Second, the encoder part, namely the backbone network, extracts multi-scale spatial features; each convolutional block includes a 3×3 convolution, an activation function (ReLU), batch normalization, and 2×2 max pooling, outputting multi-scale features for skip connections.

[0143] Third, the decoder part, where each decoded block recovers resolution through deconvolution; skip connections preserve high-resolution features and are concatenated with decoder features.

[0144] Fourth, the feature fusion module, which adds an attention mechanism to each decoding layer to highlight the features of key regions, including channel attention and spatial attention. Channel attention uses squeeze-and-excitation; spatial attention uses a self-attention module to enhance the feature weights of key regions.

[0145] Fifth, the output layer uses 1×1 convolutions to generate a single-channel prediction image representing the probability of residential buildings; the activation function chosen is the sigmoid function.

[0146] It should be noted that the formula for calculating the building area classification loss function is as follows:

[0147] ;

[0148] In the formula, L segmentation L represents the loss function for classifying building regions; CE L represents the cross-entropy loss; Dice α represents the Dice loss function; α represents L CE The weight value; β represents L Dice The weight values ​​are: α + β = 1.

[0149] The formula for calculating cross-entropy loss is as follows:

[0150] ;

[0151] In the formula, y i Let represent the true label of the i-th sample, and (0 represents non-building or 1 represents building); This represents the predicted probability that the i-th sample is a building (output by the sigmoid function), and N represents the total number of buildings.

[0152] The formula for calculating Dice loss is as follows:

[0153] ;

[0154] In the formula, L Dice Indicates Dice loss, and The smaller the Dice loss, the more accurate the prediction.

[0155] It should be noted that the prior loss function for population distribution specifically establishes the constraint between the predicted building area and the grid population distribution; the formula for calculating the prior loss function for population distribution is:

[0156] ;

[0157] In the formula, L priorThis represents the prior loss function for population distribution. Let represent the predicted value of pixel (or finer unit) (i,j) by the residential building recognition model, and represent the probability that the pixel belongs to a building. If the value is close to 1, it means the model considers the pixel to be part of a residential building; if... If the value is close to 0, it means that the model considers the pixel to be neither a building nor a residential building. This represents the normalized population value corresponding to a pixel (or a grid cell of the same resolution) after logarithmic transformation or normalization of the population grid data. If a certain area has a large population, then... Relatively large; if the population is sparse, then Smaller; This represents a hyperparameter used to adjust the importance of the prior constraint in the overall loss function. If the population size is large, the model will place greater emphasis on meeting the prior requirements of population distribution; if... If the prior constraints are smaller, then the influence of the prior constraints on the model is weaker.

[0158] It should be noted that the formula for calculating the total loss function is:

[0159] ;

[0160] In the formula, L represents the total loss function; α represents L segmentation The weight value; β represents L prior The weight values; and α+β=1.

[0161] It should be noted that the training and validation of the residential building recognition model involves special enhancement of the training data, namely, adding data samples simulating sparsely populated or densely populated areas.

[0162] During training, the optimizer sets the initial learning rate of the Adam optimization algorithm to 10. -4 Batch size is set according to video memory limits (e.g., 16); and overfitting is prevented through an early stopping mechanism.

[0163] It should be noted that the validation and testing of the residential building recognition model involves using the validation set to evaluate the segmentation results, based on the division of the training and validation sets. This means using the trained residential building recognition model, inputting fused image feature data and building feature data into the model; specifically, inputting high-resolution image features and auxiliary features from the validation set into the model to obtain the prediction results for residential buildings in the validation set. and compared it with the real label Y test Compare and calculate test metrics, including intersection-union ratio (LOU), F1 score, etc.

[0164] It should be noted that, for a single class, the Intersection over Union (IoU) calculation includes: intersection and union. The intersection is the set of pixels in the predicted result and the set of pixels in the ground truth that are classified as belonging to that class, i.e., the number of pixels that are both predicted and ground truth. The union is the set of pixels in the predicted result and the set of pixels in the ground truth that are classified as belonging to that class, i.e., the total number of pixels belonging to that class (whether predicted or ground truth).

[0165] The formula for calculating the average crossover ratio is:

[0166] ;

[0167] In the formula, P represents the set of pixels predicted to be of this category; G represents the set of pixels actually labeled as this category; the value of IoU ranges from [0,1], and the larger the value, the better the prediction result matches the actual label.

[0168] It should be noted that the F1-score assumes that for a binary classification problem, the positive class is "Positive" and the negative class is "Negative", and the prediction results can be divided into the following four categories: True Positive (TP), False Positive (FP), True Negative (TN), and False Negative (FN); TP is the number of positive samples that are predicted as positive, FP is the number of negative samples that are incorrectly predicted as positive, TN is the number of negative samples that are predicted as negative, and FN is the number of positive samples that are incorrectly predicted as negative.

[0169] Precision measures the percentage of instances predicted as positive by a residential building identification model that are actually positive. The formula for calculating precision is:

[0170] ;

[0171] Recall measures how many instances that are actually positive are correctly predicted as positive by a residential building identification model. The formula for recall is:

[0172] ;

[0173] The F1 score is the harmonic mean of precision and recall. The formula for calculating the harmonic mean is:

[0174] ;

[0175] In the formula, F1 is the F1-score value.

[0176] It should be noted that the residential building identification model identifies residential buildings by using a trained deep learning model to analyze and identify all residential buildings within the research scope.

[0177] S3. Using residential building data and grid population data, divide the regional building levels and sample the data of buildings in each level to obtain training samples.

[0178] Specifically, using residential building data and grid population data, the area's building hierarchy is divided, and data sampling is performed on buildings at each hierarchy to obtain training samples, including the following steps:

[0179] S31. Based on residential building data and grid population data, determine the population density of buildings in the area, and classify the building levels in the area according to the population density of buildings in the area.

[0180] S32. Based on the divided regional building hierarchy, calculate the sample size of buildings in each level of the region;

[0181] S33. Samples are extracted according to the sample extraction quantity to obtain training samples, and consistency verification is performed on the training samples.

[0182] Specifically, based on the divided regional building hierarchy, the formula for calculating the sample size of buildings at each hierarchy level is as follows:

[0183] ;

[0184] In the formula, N h n represents the total number of buildings on the h-th floor; h This represents the number of samples to be drawn from the h-th layer; n represents the total number of samples; N represents the total number of buildings.

[0185] The formula for calculating the total sample size is as follows:

[0186] ;

[0187] In the formula, Z represents the score corresponding to the confidence level; p represents the proportion estimate of the target variable; and e represents the allowable sampling error.

[0188] It should be noted that, using residential building data and grid population data, the division of regional building levels is specifically done by dividing the region into high, medium, and low density levels based on the population density distribution of the grid where the building is located, and randomly sampling from each level proportionally; a stratified random sampling method is used to select samples from the extracted residential buildings to ensure coverage of building types and population density distributions in different areas.

[0189] The first principle of stratification is to stratify according to the population density distribution of the building, specifically into low-density areas, medium-density areas, and high-density areas; in low-density areas (such as rural areas), the population per grid unit is relatively small; in medium-density areas (such as small towns), the population per grid unit is moderate; and in high-density areas (such as the city center), the population per grid unit is relatively large. The second principle of stratification is to further subdivide according to the characteristics of the building form (such as area, number of floors, and building type).

[0190] It should be noted that the total sample size in the stratified sample size calculation is estimated using the sampling error formula; where the confidence level corresponds to the Z-score (e.g., 95% corresponds to Z=1.96); the proportion estimate of the target variable (e.g., the probability of living in a residential building, usually set to 0.5); and the allowable sampling error (e.g., 5%).

[0191] It should be noted that stratified sampling is a form of random sampling, which uses a pseudo-random number generator to select samples from each stratum according to n... h The number of buildings is randomly selected.

[0192] A field survey was conducted on the selected sample buildings to obtain data on their actual resident population. The survey methods included door-to-door surveys, community records, and supplementary data. Door-to-door surveys involved interviewing residents within the buildings. Community records were obtained from community and property management records. Supplementary data included information such as electricity meter readings, water meter readings, and the number of internet users. Data was meticulously recorded, with the actual resident population P recorded for each building. actual .

[0193] The acquired grid population data is associated with the corresponding building feature data for model training. The building feature data includes building morphology features (such as area and height) and environmental features (such as greening rate and surrounding traffic density) extracted from the building feature data, which are associated with the corresponding residential population density values ​​in the grid population data.

[0194] Construct training samples, where the features of each sample are represented as follows:

[0195] ;

[0196] In the formula, x i X represents the features of each sample; shape Indicates architectural form characteristics; X enviroument This indicates environmental characteristics, namely NDVI.

[0197] Among them, the actual population value is y i =P actual .

[0198] The preprocessing of grid population data and corresponding building feature data includes the following steps:

[0199] First, normalization: normalize all eigenvalues ​​to the interval [0,1]. The normalization formula is as follows:

[0200] ;

[0201] In the formula, represents all normalized eigenvalues; v represents the set of eigenvalues.

[0202] Second, the dataset is divided into training set (80%), validation set (10%) and test set (10%).

[0203] Third, data augmentation includes adding noise and data expansion; adding noise is to simulate measurement errors; data expansion is to add weights to samples in low-density areas to prevent imbalance.

[0204] Data visualization and verification: Visualize the sample distribution, check whether the sampling results cover different regions and building features, verify data consistency, and ensure that the actual population value does not deviate from the feature data.

[0205] S4. Using the fused image feature data and grid population data of the training samples, a dual-path convolutional neural network is trained to obtain a population prediction model.

[0206] Specifically, the population prediction model is obtained by training a dual-path convolutional neural network using fused image feature data and grid population data from training samples, including the following steps:

[0207] S41. Preprocess the fused image feature data and grid population data of the training samples to obtain multimodal fused data and preprocessed grid population data;

[0208] S42. Based on the multimodal fusion data and the preprocessed grid population data, extract the fused image features and grid population features respectively through a convolutional neural network, and generate the fused image feature weights and grid population feature weights using an attention mechanism.

[0209] S43. Based on the weights of the fused image features and the weights of the grid population features, calculate the weighted fused image features and grid population features, and then stitch and fuse the weighted fused image features and grid population features to obtain the stitched features.

[0210] S44. Input the spliced ​​features into the dual-path convolutional neural network, and optimize and train the dual-path convolutional neural network using the mean squared error loss function and regularization constraints to obtain the population prediction model.

[0211] It should be noted that the population prediction model is trained by using the fused image feature data and grid population data of the training samples to train the dual-path convolutional neural network. Specifically, the dual-path convolutional neural network processes the fused image feature data and grid population data separately, and performs comprehensive processing in the feature fusion layer.

[0212] Data preparation includes the following steps:

[0213] First, the input data, for grid population data, represents the distribution of the resident population in each grid, and is interpolated and aligned to the spatial resolution of the high-resolution image, with a size of (H×W×1); for fused image feature data, it combines the features of the high-resolution image and the nighttime light remote sensing data, and is fused through multimodal fusion (such as feature stitching or deep feature fusion), with a size of (H×W×C), where C is the number of channels.

[0214] Second, the target data is the actual number of residents in each residential building, with a size of (N,1).

[0215] Third, feature fusion involves concatenating high-resolution images and nighttime light images using a convolutional network or a simple channel concatenation method to obtain a fused image feature vector F. fusion .

[0216] The design of the population prediction model architecture includes the following steps:

[0217] First, a dual-input design is used. One input path is the grid population data path, which uses a convolutional neural network (CNN) to extract grid population distribution features. The formula for calculating the extracted grid population distribution features is as follows:

[0218] ;

[0219] In the formula, X grid This represents the normalized value of the population grid, and It is grid-based population data;

[0220] F grid This represents the population distribution characteristics of the grid, and It is the extracted grid population depth feature.

[0221] Input path two is the fused image feature data path, which uses a convolutional neural network (CNN) to extract fused image features:

[0222] ;

[0223] In the formula, X fusion This indicates the fusion of image feature data, and ;F fusion Represents the feature vector of the fused image. It is the extracted fused image feature depth feature.

[0224] Second, the cross-modal attention fusion module first uses a cross-modal attention mechanism to calculate attention weights and learn the interaction relationship between the two inputs.

[0225] Generate attention weights α for grid population features grid :

[0226] ;

[0227] In the formula, F grid This represents the feature representation extracted from grid population data (e.g., feature vectors obtained through CNNs or multilayer perceptrons (MLPs), typically with dimensions (T, D1), where T is the number of samples and D1 is the feature dimension; W grid This represents a trainable parameter matrix used to incorporate grid population features F grid Mapping to a new representation space or generating "logits" of attention scores, with dimensions (D1→K) (where K is the output dimension, generally related to the structure of the attention mechanism); α grid This represents the attention weights for the grid population features.

[0228] After normalization using the softmax function, α grid This forms a probability distribution, representing the degree of attention the population prediction model pays to different characteristic components or different locations.

[0229] Attention weights are generated for the fused image features:

[0230] ;

[0231] In the formula, F fusion The fused image feature vector is a representation obtained from high-resolution satellite imagery and nighttime light imagery through convolution and feature extraction networks; W fusion The weight matrix that can be trained is usually . The mapping (reducing high-dimensional features into scalar or lower-dimensional representations); α fusion The weight distribution obtained by applying the Softmax function to the output after linear transformation has the same size as the number of samples and represents the degree of attention paid to each feature vector.

[0232] The Softmax operation is:

[0233] ;

[0234] In the formula, b fusionThe slope of the linear transformation, i.e. the bias of the linear transformation, is a scalar; j represents the j-th sample.

[0235] grid population characteristics (F) grid The grid population characteristics (F) are obtained by weighted fusion. fused-grid ):

[0236] ;

[0237] Fuse image feature vectors (F) fusion The image features (F) are obtained by weighted fusion. fused-fusion ):

[0238] ;

[0239] The weighted fused image features and grid population features are then stitched together to obtain the stitched features (F). final ):

[0240] ;

[0241] Finally, we return to the output layer, where the input to the fully connected layer is... It uses a two-layer fully connected network, where the first layer is (D1+D2)→256 with ReLU activation function; the second layer is 256→1 (outputting the pre-reduced population value).

[0242] The predicted population value is the predicted resident population (P) for each building, which is the final output of the population prediction model. pred ):

[0243] ;

[0244] The optimization training of a dual-path convolutional neural network using the mean squared error loss function and regularization constraints includes the following steps in the specific loss function design:

[0245] First, regression error, using mean squared error (MSE) as the loss function:

[0246] ;

[0247] In the formula, L MSE This represents the mean squared error loss function, used to measure the degree of deviation between the population value predicted by the model and the actual population value; represents the predicted value, i.e., the predicted population of the i-th building, output by the population prediction model; represents the actual value, i.e., the actual population of the i-th building, obtained from the field survey; N represents the total sample size, i.e., the number of residential buildings.

[0248] MSE is a measure of predicted value. and actual value The squared error is a commonly used regression loss function; calculating the squared error makes larger errors contribute more significantly to the total loss, helping the model converge to a low-error state more quickly; by minimizing L... MSE The model will continuously adjust its parameters to make the predicted values ​​closer to the actual values.

[0249] Second, regularization constraints: Add regularization:

[0250] ;

[0251] In the formula, L reg θ represents the regularization constraint; θ represents all trainable parameters in the population prediction model, including convolutional kernel weights and fully connected layer weights. λ represents the sum of squares of the parameters, i.e., the complexity of the model; λ represents the regularization weight hyperparameter.

[0252] Regularization is a technique to prevent model overfitting. Regularization improves the model's generalization ability by penalizing the sum of squares of the model parameters, making the values ​​of the model parameters smaller. By adding a regularization term, the model will not overfit the training data and can better predict unseen data.

[0253] Third, the overall loss function:

[0254] ;

[0255] In the formula, L a λ represents the overall loss function; λ represents the regularization weight hyperparameter, used to balance the effects of the main loss term and the regularization term.

[0256] When λ > 0, and as λ increases, the effect of the regularization term strengthens, the model parameters tend to be simpler, and some predictive ability may be lost; the value of λ is selected through cross-validation, typically around 10. -3 Up to 10 -5 Within the range.

[0257] If the prediction is close to the actual value, L MSE Approaching 0; L reg Used to control model complexity and prevent overfitting; a common example is the sum of squares of the weight parameters. By penalizing the sum of squares of the weight parameters, the model tends to maintain smaller parameter values, reducing overly complex function fitting.

[0258] The population prediction model training process includes the following steps:

[0259] First, data augmentation; for high-resolution image paths, augmentation is achieved through random rotation, cropping, and flipping; for grid population paths, augmentation is achieved by adding noise or performing logarithmic transformation.

[0260] Second, the optimizer, using the Adam optimizer, with an initial learning rate. .

[0261] Third, during the training process, set the batch size to 16 or 32, the number of training rounds to 50-100, and stop after the validation set loss converges.

[0262] Fourth, the validation metric is the Mean Absolute Error (MAE):

[0263] ;

[0264] Root Mean Square Error (RMSE):

[0265] ;

[0266] Coefficient of determination (R) 2 ):

[0267] ;

[0268] In the formula, This represents the average of all real label values.

[0269] S5. Use a population prediction model to predict the population of all buildings, obtain the predicted population value for each building, and use Kriging interpolation to verify the population prediction value.

[0270] Specifically, the population prediction model is used to predict the population of all buildings, and the population prediction value for each building is obtained. The population prediction value is then verified using Kriging interpolation, including the following steps:

[0271] S51. Input the pre-acquired nighttime light image data, remote sensing image data and grid population data of all buildings into the population prediction model, make predictions for all buildings, and obtain the population prediction value for each building.

[0272] S52. Based on the population forecast values ​​of all buildings, obtain the predicted population distribution and use the semivariogram function to calculate the unknown point values.

[0273] S53. Use the Kriging interpolation method to perform population prediction interpolation on unknown points, generate population distribution maps and building predicted population value distribution maps, verify the spatial continuity of the building population prediction values, and use the power law distribution to verify the power law characteristics of the population distribution.

[0274] It should be noted that the population prediction model is used to predict the population of all buildings, and the population prediction value of each building is obtained. Kriging interpolation is then used to verify the population prediction value. Specifically, the trained model is used to predict the population of all buildings, generate the population estimate of each building, and evaluate the prediction accuracy.

[0275] The global prediction includes the following steps:

[0276] First, prepare the input data by inputting the grid population data for the entire region, i.e. Fusion imagery feature data across the entire region (combining high-resolution imagery with nighttime light data) Extract the outline and center point (x) of all residential buildings. i ,y i The total number is N. building .

[0277] Second, building feature extraction: cropping the corresponding X for each building. grid and X fusion Extract the input features of the building ( ):

[0278] ;

[0279] The input feature dimension for each building is (H) b ×W b ×C total ), where H b Indicates the height of the clipping window; W b Indicates the width of the clipping window; C total This represents the total constant.

[0280] Third, batch prediction using the model, employing the trained model f θ Perform batch predictions for all buildings:

[0281] ;

[0282] Generate a spatial population distribution map, including the predicted population value for each building. Map the population distribution to the building's center point or outline to generate a building-level population distribution map.

[0283] The validation of population projections using Kriging interpolation includes a spatial continuity test. Kriging interpolation is a geostatistical spatial interpolation method that estimates unknown values ​​by modeling the spatial autocorrelation (variance function) of variables. By performing Kriging interpolation on the projected population values, the spatial continuity of the projection results can be verified. The specific steps are as follows:

[0284] First, obtain the predicted population distribution and the coordinates of the building's center point (x).i ,y i ) and the corresponding population projections .

[0285] Second, the semivariogram is calculated to determine the semivariogram of the predicted building population:

[0286] ;

[0287] In the formula, represents the value of the semivariogram; h represents the distance between building point pairs; N(h) represents the number of point pairs with a distance of h.

[0288] Third, fit the variation model, using common variation models (such as the spherical model, exponential model, and linear model) to fit the semivariogram function:

[0289] ;

[0290] In the formula, C0 represents the nugget effect; C represents the structural variance; and a represents the range.

[0291] Fourth, interpolation calculation: using the Kriging method to perform population prediction interpolation for unknown points within the region.

[0292] ;

[0293] In the formula, This represents the interpolation predicted by the Kriging method; This represents the weights calculated using the variation model.

[0294] Fifth, spatial continuity test: compare the population distribution map generated by Kriging interpolation with the building-level predicted population distribution map to verify whether its spatial continuity meets expectations.

[0295] The verification of its spatial continuity through the power-law distribution test of the predicted values ​​specifically includes the following steps:

[0296] Population distribution typically exhibits power-law properties, meaning that most buildings have a smaller population, while a few buildings have a larger population; this distribution can be described by a power-law formula:

[0297] ;

[0298] In the formula, P(k) is the building ratio for a population of k; α is the power law exponent.

[0299] First, the generation frequency distribution is based on the predicted building population. By statistically analyzing the frequency distribution, the number of buildings in each population interval can be obtained.

[0300] Second, logarithmic transformation: Perform a logarithmic transformation on P(k) and k to linearize the power-law distribution:

[0301] ;

[0302] In the formula, const is the intercept of the formula, which is a constant.

[0303] Third, fit the power-law model, for Using the least squares method to fit logk, the power law exponent α is estimated:

[0304] ;

[0305] In the formula, k i P(k) represents the population value (or a representative value for a population range); i ) represents the proportion of buildings with a population equal to (or falling within the corresponding interval) k; M represents the total number of samples.

[0306] Fourth, test the goodness of fit and calculate R. 2 Coefficient of determination, used to evaluate the goodness of fit of the power-law model:

[0307] ;

[0308] In the formula, P(k) represents all samples i The mean of ).

[0309] Fifth, the test for deviation from the power law, if R 2 A low value indicates that the prediction result may deviate from the power-law distribution, and the distribution characteristics of the predicted value need to be re-analyzed.

[0310] In summary, by utilizing the above-mentioned technical solution of this invention, the present invention fully leverages the complementarity of nighttime light imagery data and remote sensing imagery data to enhance the joint expressive power of the data. Nighttime light imagery data can reflect regional economic activities and population density; remote sensing imagery data can reflect building morphology (area, shape) and environmental characteristics (green coverage rate, surrounding building density), details that are usually difficult to capture through traditional statistical data. Simultaneously, by utilizing high-resolution imagery prediction, compared to the generally low resolution of traditional population models (such as one-kilometer grids), this scheme significantly improves the accuracy and resolution of population estimation. Furthermore, compared to traditional models based on only a single data source (such as census data or low-resolution raster population data), the data usage is more advanced, the information is richer, and the spatial resolution is higher. In short, the detailed utilization of data and the extraction of diverse features far surpass traditional models. Through grid population data, prior constraints on population distribution are provided for residential building identification models, guiding the model to focus on buildings more likely to have residents. Compared to traditional models that are mostly based on aggregated administrative divisions or raster population data, the fine-grained characterization of individual buildings is improved. Through dual... This system employs two convolutional neural networks (CNNs): one to process fused image feature data to extract building and optical features, and the other to process grid-based population data to provide prior constraints. Features are fused through a cross-modal attention mechanism to uncover interactive information between different modalities. Compared to traditional models that typically rely on statistical regression or simple machine learning methods, this system is better suited for handling multimodal high-dimensional data. Cross-modal attention fusion enhances information interaction between different data sources and improves the feature representation capability of key areas. Compared to traditional models that typically perform simple weighted fusion, this system focuses on the correlation between data. By fusing image feature data and grid-based population data, pixel-level and building-level joint modeling refines the population estimation problem to each building, generating a population distribution map at the building level with high spatial resolution. Compared to traditional models that often perform coarse-grained estimation at the regional (e.g., zoning or grid) level, this system can more accurately locate the resident population of each building. The deep learning model can capture complex nonlinear relationships between population and building morphology, light intensity, etc., while traditional regression models mostly assume linear or simple nonlinear relationships between variables, making it difficult to reflect actual complexity.

[0311] The above description is only a preferred 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 within the protection scope of the present invention.

Claims

1. A method of residential population estimation fusing remote sensing image data and gridded population data, characterized by, The method includes the following steps: S1. Acquire nighttime light image data, remote sensing image data, and grid population data. Use a convolutional neural network to fuse the features of the nighttime light image data and the remote sensing image data to obtain fused image feature data. S2. Based on the fused image feature data, remote sensing image data, and grid population data, a deep learning model is used to determine residential buildings and obtain residential building data. S2 includes: S21. Align the grid population data with the remote sensing image data to obtain the aligned grid population data, and extract the building features from the remote sensing image data to obtain the building feature data. S22. Using the aligned grid population data, fused image feature data, remote sensing image data and building feature data, the deep learning model is trained to obtain the initial residential building recognition model. S23. Based on building feature data and aligned grid population data, construct a total loss function using prior constraints, optimize the initial residential building identification model, and obtain the residential building identification model. S24. Identify residential buildings using a residential building identification model, obtain residential building data, and verify the residential building data; S3. Using residential building data and grid population data, divide the regional building hierarchy and sample the data of buildings in each hierarchy to obtain training samples. S4. Using the fused image feature data and grid population data of the training samples, train the dual-path convolutional neural network to obtain the population prediction model. S4 includes: S41. Preprocess the fused image feature data and grid population data of the training samples to obtain multimodal fused data and preprocessed grid population data; S42. Based on the multimodal fusion data and the preprocessed grid population data, extract the fused image features and grid population features respectively through a convolutional neural network, and generate the fused image feature weights and grid population feature weights using an attention mechanism. S43. Based on the weights of the fused image features and the weights of the grid population features, calculate the weighted fused image features and grid population features, and then stitch and fuse the weighted fused image features and grid population features to obtain the stitched features. S44. Input the spliced ​​features into the dual-path convolutional neural network, and use the mean squared error loss function and regularization constraints to optimize and train the dual-path convolutional neural network to obtain the population prediction model. S5. Use a population prediction model to predict the population of all buildings, obtain the predicted population value for each building, and use Kriging interpolation to verify the population prediction value. 2.The method of claim 1, wherein, The process of acquiring nighttime light image data, remote sensing image data, and grid-based population data, and then fusing the features of the nighttime light image data and remote sensing image data using a convolutional neural network to obtain fused image feature data includes the following steps: S11. Acquire nighttime light image data, remote sensing image data, and grid population data, and preprocess the nighttime light image data and remote sensing image data to obtain standardized nighttime light image data and standardized remote sensing image data. S12. Based on standardized remote sensing image data, extract remote sensing image features using a pre-trained residual network model to obtain remote sensing image feature data; simultaneously, based on standardized nighttime light image data, extract light features using a lightweight deep neural network model to obtain light feature data. S13. Based on remote sensing image feature data and light feature data, feature fusion is performed using a convolutional neural network to obtain fused image feature data. 3.The method of claim 2, wherein, The process of acquiring nighttime light image data, remote sensing image data, and grid-based population data, and preprocessing the nighttime light image data and remote sensing image data to obtain standardized nighttime light image data and standardized remote sensing image data includes the following steps: S111. Acquire nighttime light image data, remote sensing image data, and grid population data. Align the nighttime light image data and remote sensing image data using bilinear interpolation to obtain aligned nighttime light image data and remote sensing image data. S112. Normalize the aligned remote sensing image data to obtain normalized remote sensing image data, and perform logarithmic transformation on the aligned night light image data to obtain transformed night light image data. S113. The normalized remote sensing image data and the transformed nighttime light image data are processed into blocks to obtain standardized nighttime light image data and standardized remote sensing image data. 4.The method of claim 1, wherein, The process of optimizing the initial residential building identification model based on building feature data and aligned grid population data, using prior constraints to construct a total loss function, and obtaining the residential building identification model includes the following steps: S231. Based on the building feature data, generate a building area classification loss function using the cross-entropy loss function and the Dice loss function; S232. Based on building feature data and aligned grid population data, generate a population distribution prior loss function using a prior constraint loss function. S233. Using the building area classification loss function and the population distribution prior loss function, construct the total loss function to optimize the initial residential building identification model and obtain the residential building identification model. 5.The method of claim 1, wherein, The verification of residential building data includes: performing average intersection-union verification and F1 score verification on the residential building data. 6.The method of claim 1, wherein, The process of using residential building data and grid population data to divide regional building levels and sampling data from buildings at each level to obtain training samples includes the following steps: S31. Based on residential building data and grid population data, determine the population density of buildings in the area, and classify the building levels in the area according to the population density of buildings in the area. S32. Based on the divided regional building hierarchy, calculate the sample size of buildings in each level of the region; S33. Samples are extracted according to the sample extraction quantity to obtain training samples, and consistency verification is performed on the training samples.

7. The method for estimating the resident population by fusing remote sensing image data and grid population data according to claim 6, characterized in that, The formula for calculating the sample size of buildings at each level of the regional building hierarchy is as follows: ； In the formula, N h represents the total number of buildings in the hth layer; n h represents the number of samples needed to be extracted in the hth layer; n represents the total number of samples; N represents the total number of buildings; The formula for calculating the total sample size is as follows: ； In the formula, Z represents the score corresponding to the confidence level; p represents the proportion estimate of the target variable; and e represents the allowable sampling error. 8.The method of claim 1, wherein, The process of predicting the population of all buildings using a population prediction model, obtaining the predicted population value for each building, and then verifying the population prediction value using Kriging interpolation includes the following steps: S51. Input the pre-acquired nighttime light image data, remote sensing image data and grid population data of all buildings into the population prediction model, predict the population of all buildings, and obtain the population prediction value of each building. S52. Based on the population forecast values ​​of all buildings, obtain the predicted population distribution and use the semivariogram function to calculate the unknown point values. S53. Use the Kriging interpolation method to perform population prediction interpolation on unknown points, generate population distribution maps and building predicted population value distribution maps, verify the spatial continuity of the building population prediction values, and use the power law distribution to verify the power law characteristics of the population distribution.

Images