Night light data downscaling method and device and electronic equipment

Abstract

The invention provides a downscaling method and device for night light data and electronic equipment, relates to the technical field of night light remote sensing, and specifically comprises the following steps: acquiring auxiliary data describing geographical conditions in a target area and first night light image data with a first resolution in the target area; rasterizing the auxiliary data to obtain first raster data and second raster data, the first raster data being raster data matched with the first night light image data, the resolution of the second raster data being a second resolution, and the second resolution being higher than the first resolution; based on the first night light image data and the first raster data, training a pre-constructed geographical weighted neural network model to obtain a target model; and inputting the second raster data into the target model to obtain second night light image data of which the resolution is the second resolution in the target area.

Description

Technical Field

[0001] The present invention relates to the technical field of nighttime light remote sensing, and in particular to a method, device and electronic equipment for downscaling nighttime light data. Background Art

[0002] As an optical remote sensing technology that can detect nighttime dim light, nighttime light remote sensing is different from traditional remote sensing that focuses on daytime information. It can obtain information that is difficult to obtain during the day from different perspectives. Nighttime light remote sensing has the advantages of wide coverage, fast timeliness, and easy acquisition. Therefore, it is widely used in the estimation of socioeconomic indicators such as population, gross domestic product (GDP), electricity consumption, and carbon emissions. The data obtained by the Visible Infrared Imaging Radiometer Suite (VIIRS) on the National Polar-orbiting Partnership (NPP) (note that the data will be referred to as NPP-VIIRS below) has the characteristics of wide coverage, long duration, short re-entry period, and easy acquisition. It is currently the most widely used nighttime light remote sensing data. However, the spatial resolution of NPP-VIIRS data is only 500m at most. The coarse resolution makes it difficult to distinguish basic urban structural elements such as roads, residential areas, and industrial areas in NPP-VIIRS data, and it is impossible to analyze small-scale urban areas in detail. For downscaling of night light data, how to convert large-scale, low-resolution night light data into small-scale, high-resolution data has become a hot topic to be studied.

[0003] Mainstream remote sensing image downscaling methods can now be divided into three categories: data fusion, machine learning, and statistical regression. The data fusion method has a simple model and low time complexity, which is suitable for regional research. However, the physical true value of night light data is often eliminated during the data fusion process, which can easily lead to low accuracy of downscaled night light image data. Machine learning methods use machine learning models such as random forests to build relationships between data. This type of method can achieve good fitting effects without considering the complex spatial relationships between variables. The disadvantage is that the model is not interpretable enough and the effect is often poor. Statistical regression methods usually perform downscaling through interpolation. However, due to the existence of spatial heterogeneity, the image data before and after downscaling cannot be simply converted through interpolation, which can easily lead to low accuracy of downscaled night light image data. Summary of the invention

[0004] The present invention provides a method, device and electronic device for downscaling nighttime light data, which are used to solve the problem of low accuracy in downscaling nighttime light image data in the related art.

[0005] In a first aspect, the present invention provides a method, device and electronic device method for downscaling nighttime light data, comprising: Acquire auxiliary data describing geographical conditions within a target area and first nighttime light image data having a first resolution within the target area; Performing rasterization processing on the auxiliary data to obtain first raster data and second raster data, wherein the first raster data is raster data matching the first nighttime light image data, and the second raster data has a second resolution, which is higher than the first resolution; Based on the first nighttime light image data and the first raster data, a pre-built geographically weighted neural network model is trained to obtain a target model; The second raster data is input into the target model to obtain second nighttime light image data with a resolution of the second resolution in the target area.

[0006] Optionally, the training of a pre-built geographically weighted neural network model based on the first nighttime light image data and the first raster data to obtain a target model includes: Reprojecting the first nighttime light image data and the first raster data into the same coordinate system to obtain the training data set; The pre-built geographically weighted neural network model is trained using the training data set to obtain a target model.

[0007] Optionally, the geographically weighted neural network model includes an input layer, a hidden layer and an output layer, and the pre-constructed geographically weighted neural network model is trained by the training data to obtain a target model, including: Inputting the training data set into the input layer, the training data set includes A grid data, where A is a positive integer; Each data point in each grid data in the training data set is used as a prediction point in turn, and the following steps are performed for the prediction point: The input layer inputs the spatial distance between the prediction point and other sample points into the hidden layer, and the other sample points are data points in the training data set other than the prediction point; Performing convolution processing on the spatial distance between the prediction point and other sample points through the hidden layer to obtain the spatial weight of the prediction point; Fitting the spatial weight to the regression coefficient corresponding to the prediction point through the output layer to obtain a training fitting value of the prediction point, wherein the regression coefficient is obtained based on the first nighttime light image data and the first raster data; According to the error between the fitting image data and the first nighttime light image data, the hidden layer is iteratively optimized to obtain a target model, wherein the fitting image data includes the training fitting values ​​of all the prediction points on the same raster data.

[0008] Optionally, the iterative optimization of the hidden layer according to the error between the fitting image data and the first nighttime light image data to obtain the target model includes: Calculating a mean square error between the fitting image data and the first nighttime light image data; If the mean square error is less than the preset error value, or the number of training times is greater than the preset training value, the training is stopped and the geographically weighted neural network model is used as the target model; If the mean square error is not less than the preset error value, or the number of training times is not greater than the preset training value, the weight coefficients in the hidden layer are iteratively optimized, and the optimized geographically weighted neural network model is used as the target model.

[0009] Optionally, the auxiliary data includes multiple types of auxiliary data, and rasterizing the auxiliary data to obtain first raster data includes: Performing a first transformation process on each type of the auxiliary data to obtain B pieces of intermediate grid data with a resolution of the first resolution, where B is a positive integer; The B pieces of intermediate raster data are reprojected into the same coordinate system to obtain the first raster data.

[0010] Optionally, the multiple types of auxiliary data include at least one of surface coverage type data within the target area, surface temperature data within the target area, road network data within the target area, and point of interest data within the target area.

[0011] In a second aspect, the present invention provides a device for downscaling nighttime light data, the device comprising: An acquisition module, used for acquiring auxiliary data describing the geographical conditions in a target area and first nighttime light image data with a first resolution in the target area; A processing module, configured to perform rasterization processing on the auxiliary data to obtain first raster data and second raster data, wherein the first raster data is raster data matching the first nighttime light image data, and the second raster data has a second resolution, which is higher than the first resolution; A training module, used for training a pre-built geographically weighted neural network model based on the first nighttime light image data and the first raster data to obtain a target model; The input module is used to input the second raster data into the target model to obtain the second nighttime light image data with the second resolution in the target area.

[0012] In a third aspect, the present invention provides an electronic device, comprising: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the method as described in the first aspect.

[0013] In a fourth aspect, the present invention provides a computer-readable storage medium. When instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the method described in the first aspect.

[0014] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, wherein the computer program is executed by a processor as the method described in the first aspect.

[0015] The embodiment of the present invention provides a method for downscaling night light data, specifically, obtaining auxiliary data describing the geographical conditions in a target area and first night light image data with a first resolution in the target area; performing rasterization processing on the auxiliary data to obtain first raster data and second raster data, wherein the first raster data is raster data matching the first night light image data, and the second raster data has a second resolution, which is higher than the first resolution; training a pre-constructed geographically weighted neural network model based on the first night light image data and the first raster data to obtain a target model; inputting the second raster data into the target model to obtain second night light image data with a second resolution in the target area. In this way, the geographically weighted neural network model is trained by the first night light image data and the first raster data (matching the first night light image data), so that the geographically weighted neural network model fits the spatial nonlinear relationship of the target area in the downscaling process, and establishes a spatial regression relationship between the first night light image data with a lower resolution and the first raster data with a lower resolution. Compared with the machine learning method in the related art, the target model obtained according to the geographically weighted neural network model in the embodiment of the present invention has a better expression of complex spatial heterogeneity, and then according to the principle of invariance of the spatial relationship between ground physical elements and pixel values ​​at different scales, the second raster data with higher resolution is reconstructed with the target model to obtain second night light image data with higher resolution, and the downscaling processing of the first night light image data is completed, so that the obtained second night light image data has higher accuracy, which solves the problem of low accuracy of downscaling of night light image data in the related art to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0017] Figure 1 A flowchart of a method for downscaling nighttime light data provided by an embodiment of the present invention; Figure 2 A structural design diagram of a geographically weighted neural network model provided by an embodiment of the present invention; Figure 3 A conceptual diagram of a method for downscaling nighttime light data provided by an embodiment of the present invention; Figure 4A conceptual diagram of a training process of a geographically weighted neural network model provided by an embodiment of the present invention; Figure 5 A comparative reference diagram of a downscaling method for nighttime light data provided by an embodiment of the present invention; Figure 6 A structural block diagram of a device for downscaling nighttime light data provided by an embodiment of the present invention; Figure 7 A structural block diagram of an electronic device provided for the implementation of the present invention. DETAILED DESCRIPTION

[0018] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings, but it should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present disclosure.

[0019] As described in the background technology, night light remote sensing, as an optical remote sensing technology that can detect nighttime dim light, is different from traditional remote sensing that focuses on daytime information. It can obtain information that is difficult to obtain during the day from different perspectives. Night light remote sensing has the advantages of wide coverage, fast timeliness, and easy acquisition. Therefore, it is widely used in the estimation of social and economic indicators such as population, gross domestic product (GDP), electricity consumption, and carbon emissions. The data obtained by the Visible Infrared Imaging Radiometer Suite (VIIRS) on the National Polar-orbiting Partnership (NPP) (note that the data will be referred to as NPP-VIIRS below) has the characteristics of wide coverage, long duration, short re-entry period, and easy acquisition. It is currently the most widely used night light remote sensing data. However, the spatial resolution of NPP-VIIRS data is only 500m at most. (Resolution 500m‌ usually refers to images or data with a spatial resolution of 500 meters. In the field of remote sensing, a resolution of 500m means that each pixel of the image represents an area of ​​about 500 meters × 500 meters on the ground. In the following text, the resolution will also be expressed in this way, that is, the smaller the resolution value, the higher the resolution of the image data.) The coarser resolution makes it difficult to distinguish basic urban structural elements such as roads, residential areas, and industrial areas in the NPP-VIIRS data, and it is impossible to analyze small-scale urban areas in detail. Downscaling the night light data, how to convert large-scale, low-resolution night light data into small-scale, high-resolution data has become a hot topic to be studied.

[0020] Mainstream remote sensing image downscaling methods can now be divided into three categories: data fusion, machine learning, and statistical regression. The data fusion method has a simple model and low time complexity, which is suitable for regional research. However, the physical true value of night light data is often eliminated during the data fusion process, which can easily lead to low accuracy of downscaled night light image data. Machine learning methods use machine learning models such as random forests to build relationships between data. This type of method can achieve good fitting effects without considering the complex spatial relationships between variables. The disadvantage is that the model is not interpretable enough and the effect is often poor. Statistical regression methods usually perform downscaling through interpolation. However, due to the existence of spatial heterogeneity, the image data before and after downscaling cannot be simply converted through interpolation, which can easily lead to low accuracy of downscaled night light image data.

[0021] In this regard, an embodiment of the present invention provides a method for downscaling night light data. Specifically, auxiliary data describing the geographical conditions in a target area and first night light image data with a first resolution in the target area are obtained; the auxiliary data is rasterized to obtain first raster data and second raster data, the first raster data is raster data matching the first night light image data, and the second raster data has a second resolution, which is higher than the first resolution; based on the first night light image data and the first raster data, a pre-constructed geographically weighted neural network model is trained to obtain a target model; the second raster data is input into the target model to obtain second night light image data with a resolution of the second resolution in the target area. In this way, the geographically weighted neural network model is trained by the first night light image data and the first raster data (matching the first night light image data), so that the geographically weighted neural network model fits the spatial nonlinear relationship of the target area in the downscaling process, and establishes a spatial regression relationship between the first night light image data with a lower resolution and the first raster data with a lower resolution. Compared with the machine learning method in the related art, the target model obtained according to the geographically weighted neural network model in the embodiment of the present invention has a better expression of complex spatial heterogeneity, and then according to the principle of invariance of the spatial relationship between ground physical elements and pixel values ​​at different scales, the second raster data with higher resolution is reconstructed with the target model to obtain second night light image data with higher resolution, and the downscaling processing of the first night light image data is completed, so that the obtained second night light image data has higher accuracy, which solves the problem of low accuracy of downscaling of night light image data in the related art to a certain extent.

[0022] In addition, in the embodiment of the present invention, since the second night-time light image data of the second resolution is obtained based on the second raster data of the same second resolution, the interpolation processing involved in the statistical regression method in the related technology is avoided, and the second night-time light image data is obtained more reasonably, so that the second night-time light image data provided by the embodiment of the present invention has a better expression effect.

[0023] It should be understood that the downscaling method for nighttime light data provided in the embodiment of the present invention can be executed by a target device. The target device can be one electronic device or multiple electronic devices working together to execute the method. The electronic device can be a server, such as an independent physical server, a server cluster composed of multiple servers, and a cloud server capable of cloud computing.

[0024] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present invention will be described below in conjunction with the accompanying drawings.

[0025] Figure 1 A flowchart of a method for downscaling nighttime light data provided by an embodiment of the present invention. Figure 1 As shown, the method for downscaling nighttime light data provided by the embodiment of the present invention includes steps 110 to 140.

[0026] Step 110 , obtaining auxiliary data describing the geographical conditions in the target area and first nighttime light image data with a first resolution in the target area.

[0027] In the embodiment of the present invention, the target area can be an area of ​​any size in any place. For example, the target area can be the area where a cell is located, the area where a mountain is located, and the area where a lake and surrounding mountains are located. The embodiment of the present invention does not impose specific restrictions on the geographical location and size of the target area.

[0028] In an embodiment of the present invention, the geographical conditions in the target area may include natural geographical features such as climate, topography, landform, hydrology, soil, etc. in the target area, and may also include human geographical features such as population, economy, culture, etc. in the target area. Accordingly, the auxiliary data describing the geographical conditions in the target area may be data describing the natural geographical features in the target area, such as surface cover type data in the target area, and may also be data describing the human geographical features in the target area, such as road network data or point of interest (POI) data in the target area. The format of the auxiliary data may be any data format, such as raster format, table format, and vector data format, etc., and the size of the auxiliary data may also be any size. The embodiment of the present invention does not impose specific restrictions on the format, size, and other data attributes of the auxiliary data. The target device may obtain the auxiliary data through the Internet, and may also obtain the auxiliary data by receiving externally imported data from the user. The embodiment of the present invention does not impose specific restrictions on this.

[0029] In an embodiment of the present invention, the first nighttime light image data may be nighttime light image data having a first resolution in a target area. The target area is the same area covered by the auxiliary data explained above. The first resolution may be any resolution. In an embodiment of the present invention, the first resolution may be a resolution popular among current nighttime remote sensing image data, such as 500m. The nighttime light image data may be any image data in a preset time interval, such as any image data between 10pm and 12pm, or may be image data intelligently determined based on image brightness, such as any image data with lower brightness.

[0030] In an embodiment of the present invention, the target device can obtain the first night-time light image data by receiving external input. For example, the target device can obtain the first night-time light image data by receiving user input, or by receiving data sent by an external camera. The embodiment of the present invention does not impose any specific restrictions on this.

[0031] After the target device acquires the auxiliary data and the first nighttime light image data, step 120 may be executed to pre-process the auxiliary data based on the resolution of the first nighttime light image.

[0032] Step 120, rasterizing the auxiliary data to obtain first raster data and second raster data, wherein the first raster data is raster data matching the first nighttime light image data, and the second raster data has a second resolution, which is higher than the first resolution.

[0033] In an embodiment of the present invention, in order to facilitate the unified calculation and fitting processing of the auxiliary data and the first nighttime light image data, the data format of the first nighttime light image data can be raster data, and the auxiliary data is rasterized to obtain the first raster data and the second raster data respectively. Among them, the resolution of the first raster data is the same as the resolution of the first nighttime light image data, and the raster size of the first raster data is also the same as the raster size of the first nighttime light image data. When the above two conditions are met at the same time, the first raster data can be regarded as matching the first nighttime light image. If the auxiliary data includes multiple data, correspondingly, the first raster data also includes multiple raster data tables, then the resolution and raster size corresponding to each raster data table should meet the above two conditions at the same time. The second raster data can be raster data with a resolution of the second resolution, and accordingly, the raster size of the second raster data will also be the raster size corresponding to the second resolution.

[0034] In the embodiment of the present invention, the second resolution may be any resolution higher than the first resolution. For example, the second resolution is an ideal resolution for the downscaling process, 100 m.

[0035] In the embodiment of the present invention, since the resolution of the first raster data is different from the resolution of the second raster data, the target device can perform different rasterization processes on the auxiliary data. In addition, when the auxiliary data includes multiple types of data, the target device can also perform transformation processes corresponding to the data according to different data formats and / or data contents, thereby obtaining the first raster data and the second raster data, which is not specifically limited in the embodiment of the present invention.

[0036] After obtaining the first raster data with the same resolution as the first night light image data, step 130 can be executed to fit the first raster data representing the geographic conditions of the target area with the first night light image data through a pre-constructed geographically weighted neural network model, so that the pre-constructed geographically weighted neural network model can better express the complex spatial heterogeneity in the target area.

[0037] Step 130: training a pre-built geographically weighted neural network model based on the first nighttime light image data and the first raster data to obtain a target model.

[0038] In the embodiment of the present invention, the pre-constructed geographically weighted neural network model can be any neural network model. The pre-constructed geographically weighted neural network model is trained by the first nighttime light image data and the first raster data describing the geographical conditions in the target area, so that the pre-constructed geographically weighted neural network model can accurately learn the geographical conditions in the target area, and learn how to perform more reasonable weighted calculations on the geographical conditions in the target area and the nighttime light image data in the target area, thereby further improving the accuracy of the downscaling process.

[0039] In an embodiment of the present invention, when the number of training times of the geographically weighted neural network model reaches a certain value or the output of the geographically weighted neural network model achieves a certain effect during the training process, the geographically weighted neural network model can be regarded as trained and the geographically weighted neural network model at this time is used as the target model.

[0040] In an embodiment of the present invention, after obtaining a target model using first night-time light image data of a first resolution and first raster data of the same first resolution, the target model can well combine the night-time light image data with the geographical conditions in the target area, and then step 140 can be executed to obtain second night-time light image data of the second resolution using the target model and second raster data of the second resolution.

[0041] Step 140: input the second raster data into the target model to obtain second nighttime light image data in the target area with a resolution of the second resolution.

[0042] In an embodiment of the present invention, the second raster data can be input into the target model, and the second raster data can be processed by the fitting method of the geographical conditions and image data learned by the target model during the training process, so as to obtain the second nighttime light image data with the same resolution as the second resolution, and complete the downscaling processing of the first nighttime light image data. Since the second resolution is greater than the first resolution, the second nighttime light image data can be understood as the first light image data after downscaling processing.

[0043] In this way, the geographically weighted neural network model is trained by the first nighttime light image data and the first raster data (matching the first nighttime light image data), so that the geographically weighted neural network model fits the spatial nonlinear relationship of the target area in the downscaling process, and establishes a spatial regression relationship between the first nighttime light image data with a lower resolution and the first raster data with a lower resolution. Compared with the machine learning method in the related art, the target model obtained according to the geographically weighted neural network model in the embodiment of the present invention has a better expression of complex spatial heterogeneity, and then according to the principle of invariance of the spatial relationship between ground physical elements and pixel values ​​at different scales, the second raster data with a higher resolution and the target model are reconstructed to obtain the second nighttime light image data with a higher resolution, and the downscaling of the first nighttime light image data is completed, so that the obtained second nighttime light image data has a higher accuracy, which solves the problem of low accuracy in downscaling of nighttime light image data in the related art to a certain extent.

[0044] In addition, in the embodiment of the present invention, since the second night-time light image data of the second resolution is obtained based on the second raster data of the same second resolution, the interpolation processing involved in the statistical regression method in the related technology is avoided, and the second night-time light image data is obtained more reasonably, so that the second night-time light image data provided by the embodiment of the present invention has a better expression effect.

[0045] In order to better explain the downscaling method of night light data provided by an embodiment of the present invention, the training process of the geographically weighted neural network and the use process of the target model will be explained in turn below.

[0046] In the embodiment of the present invention, as described above, in steps 120 and 130, the auxiliary data is transformed to obtain the first raster data, and then the geographically weighted neural network is trained by the first nighttime light image data and the first raster data. Wherein, the auxiliary data includes multiple types of auxiliary data, and in order to obtain the first raster data, the following steps can be performed: for each type of the auxiliary data, a first transformation is performed to obtain B pieces of intermediate raster data with a resolution of the first resolution, where B is a positive integer; and the B pieces of the intermediate raster data are reprojected into the same coordinate system to obtain the first raster data.

[0047] In an embodiment of the present invention, the auxiliary data may include multiple types of data, for example, the auxiliary data may include the surface coverage type data and road network data described above, and a first transformation process is performed on each type of data, and each type of data is processed into corresponding raster data with a first resolution, thereby obtaining B sheets of intermediate raster data with a first resolution, and the value of B is not less than the number of data types included in the auxiliary data, that is, one type of auxiliary data corresponds to at least one sheet of intermediate raster data. In order to make all grids correspond and achieve a better fitting effect, the B sheets of intermediate raster data can be reprojected into the same coordinate system, and the reprojected result is used as the first raster data. Among them, the coordinate system of the reprojection of the B sheets of intermediate raster data can be any projection coordinate system, such as the Asia Lambert Conformal Conic projection coordinate system, and the embodiment of the present invention does not impose specific restrictions on this.

[0048] In an embodiment of the present invention, the multiple types of auxiliary data include at least one of surface coverage type data in the target area, surface temperature data in the target area, road network data in the target area, and point of interest data in the target area.

[0049] In the embodiment of the present invention, accordingly, if the auxiliary data includes the land cover type data in the target area, the land cover type data can be reclassified according to the idea of ​​"randomly setting a land type to 1 and other land types to 0" to obtain the binary distribution raster data of the randomly set land type. That is, if the land cover type data in the auxiliary data includes multiple land cover types, there is a corresponding binary distribution of the land type for each land cover type, and multiple raster data are used as intermediate raster data to complete the first transformation processing of the land cover type data.

[0050] In an embodiment of the present invention, if the auxiliary data includes surface temperature data in the target area, the surface temperature data in raster form can be directly used as intermediate raster data. If the auxiliary data includes road network data in the target area, the road network data in vector form can be converted to obtain intermediate raster data corresponding to the road network data. If the auxiliary data includes point of interest data in the target area, the point of interest data in vector form can also be converted to obtain intermediate raster data corresponding to the point of interest data. The raster data corresponding to each auxiliary data (in this case, B is the number of raster data corresponding to all auxiliary data) is reprojected into the same coordinate system to obtain the first raster data used to train the geographically weighted neural network model. Different transformation processing is performed on each type of auxiliary data, so that the obtained first raster data has stronger pertinence and accuracy.

[0051] In an embodiment of the present invention, after obtaining the first raster data, for better training effect, the first nighttime light image data and the first raster data can be reprojected into the same coordinate system to obtain the training data set. The projection coordinate system used in this reprojection process can be the same as the projection coordinate system where the first raster data is located, or it can be different, as long as the first nighttime light image data and the first raster data are located in the same coordinate system. The training data set obtained by this reprojection includes (B+1) raster data. The training data set can include not only all the values ​​corresponding to the (B+1) raster data, but also the coordinates corresponding to each raster point. Subsequently, the pre-constructed geographically weighted neural network model is trained by the training data set to obtain a target model. For example, the training data set can be divided into three data sets, namely, a training set, a test set and a validation set, according to a certain ratio, and they are input into the geographically weighted neural network model for training. By comparing the error between the actual output (such as the fitted image data described below) and the preset reference value (such as the first nighttime light image data described below), the geographically weighted neural network model is iteratively optimized using the gradient descent method to obtain the target model.

[0052] In an embodiment of the present invention, Figure 2 A structural design diagram of a geographically weighted neural network model provided by an embodiment of the present invention. Figure 2 As shown, the geographically weighted neural network model provided by the embodiment of the present invention includes an input layer, a hidden layer and an output layer. The specific process of training the pre-built geographically weighted neural network model with the training data to obtain the target model can refer to the following steps: The training data set is input into the input layer. The training data set includes A grid data, where A is a positive integer. Referring to the above description, the value of A is (B+1). Each data point in each grid data in the training data set is used as a prediction point in turn. That is, the prediction point can come from any grid data in the first grid data, or from the first nighttime light image data itself. For the prediction point, Figure 2 shown Perform the following steps: (1) If Figure 2 As shown in the input layer box, the input layer inputs the spatial distance between the prediction point and other sample points into the hidden layer. The other sample points are data points in the training data set other than the prediction point. The spatial distance between the prediction point and other sample points can be the Euclidean distance between two points, for example Figure 2 In the Input Layer box And other data.

[0053] (2) If Figure 2As shown in the hidden layer box, the spatial distance between the prediction point and other sample points is convolved by the hidden layer to obtain the spatial weight of the prediction point. The spatial weight of the prediction point can be a matrix. The specific processing process can refer to the following formula ①: In formula ①, Indicates the predicted point The spatial weight of Indicates the predicted point The coordinates on the corresponding raster data; represents the Euclidean distance between the prediction point and other sample points; SWNN represents the pre-built geographically weighted neural network model; the size of p is the number of the first raster data plus 1, because considering that the factors that affect the downscaling effect are not only auxiliary data, it is assumed here that there are other factors; For Figure 2 As shown, it represents the non-stationary weight array of the prediction point.

[0054] Before performing the fitting process, the regression coefficient corresponding to the training point can be obtained, and the estimated value of the regression coefficient can be calculated by the following formula ②: In formula ②, is a matrix composed of the values ​​on the first nighttime light image data; X is a matrix composed of the values ​​in the first grid data.

[0055] Specifically, The expression is as follows: In formula ③, represents a point value on the first nighttime light image data, and n is the assumed total point value of the first nighttime light image data.

[0056] X can be expressed in the following formula ④: In formula ④, Indicates the value of a point on the raster data corresponding to the first land cover type data; Indicates the value of a point on the grid data corresponding to the second land cover type data; Indicates the value of a point on the grid data corresponding to the surface temperature data; Indicates the value of a point on the raster data corresponding to the point of interest data; Represents the value of a point on the raster data corresponding to the road network data.

[0057] (3) If Figure 2 As shown in the output layer box, the spatial weight is fitted with the regression coefficient corresponding to the prediction point through the output layer to obtain the training fitting value of the prediction point, wherein the regression coefficient can be obtained by calculating the value of the first night light image data and the value of the first raster data using the ordinary least squares regression model (Ordinary Least Squares, referred to as OLS). The data form of the regression coefficient can also be a number set. The specific processing process can refer to the following formula: Before introducing the calculation formula of the training fitting value, let's first introduce the calculation formula of the true value corresponding to the training fitting value, as shown in the following formula ⑤: In formula ⑤, For Figure 2 As shown, it represents the non-stationary weight array of the prediction point; An array of regression coefficients representing the predicted points; Represents the value of the i-th row and k-th column in formula ④; Indicates calculation error.

[0058] Substituting the estimated value calculation formula of the regression coefficient ② into formula ⑤, we can get the following formula ⑥: Furthermore, substituting formula ① into formula ⑥, we can obtain the following formula ⑦ for calculation: (4) If Figure 2 As shown on the right side of the output layer box, according to the error between the fitted image data and the first night light image data, the fitted image data includes the training fitting values ​​of all the prediction points on the same grid data. For example, the determination coefficient between the fitted image data and the first night light data is calculated, the hidden layer is iteratively optimized, and the training is terminated after reaching the maximum number of iterations or when overfitting occurs, and the target model is obtained.

[0059] In the embodiment of the present invention, the specific implementation method of iteratively optimizing the hidden layer can refer to the following description: Figure 2 As shown, the mean square error between the fitting image data and the first night light image data is calculated; if the mean square error is less than the preset error value, or the number of training times is greater than the preset training value, the training is stopped and the geographically weighted neural network model is used as the target model; if the mean square error is not less than the preset error value, or the number of training times is not greater than the preset training value, the weight coefficients in the hidden layer are iteratively optimized, and the optimized geographically weighted neural network model is used as the target model.

[0060] The above is the training process of the geographically weighted neural network. Next, the use process of the target model will be introduced. Before using the target model, the auxiliary data is processed into the second raster data of the second resolution. The specific processing process can refer to the process of obtaining the first raster data described above, which will not be repeated here.

[0061] In the embodiment of the present invention, after obtaining the second grid data, the second grid data can be input into the target model and then calculated by the above-mentioned formula ⑦. It should be noted that in the process of obtaining the second nighttime light image data using formula ⑦, Correspondingly, it represents the i-th row data in the matrix constructed for the second raster data, It is also expressed as the Euclidean distance from a certain point to other points in the second night light image data. It is still the value calculated by formula ②. In this way, the second nighttime light image data with the same second resolution in the target area is obtained through the second grid data input with the second resolution, so that the second nighttime light image data has higher accuracy.

[0062] In order to better understand the downscaling method for nighttime light data provided by an embodiment of the present invention, an example is now given. Figure 3 A conceptual diagram of a method for downscaling nighttime light data provided by an embodiment of the present invention, such as Figure 3 As shown, the downscaling method for nighttime light data provided by an embodiment of the present invention may include steps S1 to S4.

[0063] S1) Obtain the original nighttime light image data with a first resolution that needs to be downscaled in the target area and the corresponding auxiliary data such as the surface coverage type, surface temperature, road network data, POI data, etc. in the target area. For example, in an embodiment of the present invention, NASA's Black Marble nighttime light dataset based on VIIRS-NPP can be selected as the original nighttime light image, the Landsat Collection 2 Level-2 product surface temperature dataset can be selected as the surface temperature data, the global 30-meter surface coverage fine classification dataset released by the Space and Air Information Innovation Research Institute of the Chinese Academy of Sciences can be selected as the surface coverage data, and the road network data of OpenStreetMap and the point of interest data of the AutoNavi open platform can be selected.

[0064] S2) preprocesses the auxiliary data obtained in S1, performs different transformations on the auxiliary data from different sources, forms raster data that matches the resolution of the original nighttime light image (i.e., the first resolution), and obtains second raster data with a second resolution. The specific steps for forming raster data that matches the resolution of the original nighttime light image (i.e., the first resolution) can be referred to as follows: S21) Convert vector data such as road network and POI into raster form. Use Gaussian kernel density estimation to convert POI data and road networks of various levels into density grids with a resolution of 500 meters. For road network data, OSM road networks of various levels are reclassified into primary, secondary and tertiary roads, and the density grids are superimposed according to different weights to obtain the final road network density distribution.

[0065] S22) Since there are many categories of land cover type data, in the example provided in the embodiment of the present invention, it is mainly urban areas, so the two land cover types of impervious surface and water body are mainly used, and other land types are reclassified into the third category. At the same time, in order to avoid uncertainty in the model caused by multicollinearity of variables during model fitting, only the first two categories are used in the example provided in the embodiment of the present invention. For each land type, the pixel of the land type is assigned to 1, and the pixels of other land types are assigned to 0 to obtain the binary distribution of the pixels of the land type. In addition, it should be noted that in the process of resampling, the pixels need to be aggregated according to the mean, so that the raster image obtained is actually the proportion of land types in each grid, avoiding information loss during the resampling process.

[0066] S23) All data are reprojected to the Asia Lambert Conformal Conic projection coordinate system to ensure that all rasters can be calculated accordingly.

[0067] S3) Construct a geographically weighted neural network model, divide the low-resolution nighttime light remote sensing data and the matching auxiliary data as the data set, and input the model for training after dividing it into training data set, verification data set and test data set. By comparing the error between the actual output (i.e., the fitted image data described above) and the expected output (i.e., the first nighttime light image data described above), the model is iteratively trained and optimized using the gradient descent method until the best model is obtained. The design of the geographically weighted neural network model is as follows: Figure 2 As shown, the model is connected in series in the form of a fully connected network and uses dropout technology to improve the generalization ability of the model. In addition, for the initial parameters and activation functions of the hidden layer, the model uses He parameter initialization and PreLU activation function strategy to improve the iterative optimization efficiency of the model. The batch normalization method is used in the middle level of the model to reduce the internal covariate offset. In the example provided in the embodiment of the present invention, the number of layers of the geographically weighted neural network model is set to 7 layers, including 1 input layer, 5 hidden layers and one output layer. The specific neural network parameter settings are shown in Table 1 below: Table 1 Neural network parameters refer to Figure 2 as well as Figure 4 As shown, the specific training and use steps of the model are: S31) The preprocessed night light data and auxiliary variable data are used as sample data sets, and are randomly divided into three parts: training set, test set and validation set in a ratio of 0.7:0.15:0.15.

[0068] S32) The distance between the predicted point P in the training set and other sample points is input into the neural network as the input layer.

[0069] S33) Use the OLR model to calculate the OLS coefficient of each variable in advance, and multiply and add the non-stationary weight of each independent variable derived from the output layer to obtain the fitted value.

[0070] S34) Calculate the error through the error function and implement iterative optimization of the network model through the back propagation algorithm.

[0071] S35) When the model training reaches 10,000 batches or 2,000 batches and the training error cannot be further reduced, the model training is terminated.

[0072] S4) The high-resolution auxiliary data is brought back into the optimal model in S3 to reconstruct a high-resolution night light image.

[0073] The final downscaling result is Figure 5 As shown in A-1 and A-2: While keeping the overall spatial distribution characteristics unchanged, the downscaled night light image ( Figure 5 A-2 in Figure 1) compared to the original VIIRS-NPP night light image ( Figure 5 A-1 in the figure has extremely rich detail features. At the same time, the downscaled image has similar minimum, maximum, mean and other raster statistics with the original image, which reflects the high consistency between the downscaled image and the original image. By introducing auxiliary variables such as road network, POI, and surface temperature, the highlight spillover effect of the original image is effectively suppressed, the image details are greatly enhanced, and the boundaries of buildings, roads, water bodies and other areas can be clearly and accurately displayed. The downscaled image well reveals the differences in night light distribution caused by differences in surface cover types, and objectively shows the refined texture characteristics of the city.

[0074] In addition, the example provided in the embodiment of the present invention also selects a row of pixels (such as Figure 5The blue line in Figure A-1 and Figure A-2 in Figure 1 is sampled to explore the characteristics of the downscaled image under different surface types. It is not difficult to see from the brightness change trend graph that the change trend between the image pixels before and after downscaling is completely consistent, while the sampling curve of the downscaled image has more fluctuation characteristics. Due to the existence of spatial heterogeneity, the two cannot be simply converted by interpolation. In complex urban areas, different types of buildings and roads greatly affect the changes in night lights. By downscaling the night light images, the more detailed local structural characteristics and regional distribution changes in the city are better displayed, which can be applied to more refined urban problem research.

[0075] In order to analyze the results of the examples provided in the embodiments of the present invention, three traditional methods, namely, geographically weighted regression (GWR) model, mixed geographically weighted regression (MGWR) model, and random forest model, are compared with them. The fitting ability and accuracy of the model are evaluated from five indicators, namely, the coefficient of determination (R2), root mean square error (RMSE), peak signal-to-noise ratio (PSNR), Akaike information criterion (AIC), and corrected Akaike information criterion (AICc). Among them, since the random forest method cannot be verified by the information criterion method, the AIC and AICc indicators cannot be calculated. Table 2 gives the evaluation results of the four downscaling algorithms.

[0076] Table 2 Evaluation results of four downscaling algorithms In general, the geographic neural network weighted regression (GNNWR) model (i.e., the target model provided by the embodiment of the present invention) achieved the best performance in all indicators, and its fitting performance was also relatively stable. For model fitting parameters such as determination coefficient, root mean square error, and peak signal-to-noise ratio, the gap between the GNNWR model and the MGWR model and the GWR model is small. However, for model information criterion indicators such as AIC and AICc, the GNNWR model has achieved a significant lead, which shows that the GNNWR model has achieved better model fit while using fewer parameters.

[0077] Figure 6 The structure block diagram of a nighttime light data downscaling device provided by an embodiment of the present invention. Figure 6 As shown, the apparatus for downscaling nighttime light data provided by the embodiment of the present invention includes an acquisition module 610 , a processing module 620 , a training module 630 and an input module 640 .

[0078] An acquisition module 610 is used to acquire auxiliary data describing the geographical conditions in a target area and first nighttime light image data with a first resolution in the target area; The processing module 620 is used to perform rasterization processing on the auxiliary data to obtain first raster data and second raster data, wherein the first raster data is raster data matching the first nighttime light image data, and the resolution of the second raster data is a second resolution, which is higher than the first resolution; A training module 630 is used to train a pre-built geographically weighted neural network model based on the first nighttime light image data and the first raster data to obtain a target model; The input module 640 is used to input the second raster data into the target model to obtain second nighttime light image data with a resolution of the second resolution in the target area.

[0079] It should be noted that the embodiment of the device for downscaling night light data in this specification and the embodiment of the method for downscaling night light data in this specification are based on the same inventive concept. Therefore, the specific implementation method of this embodiment can refer to the corresponding embodiment of the method for downscaling night light data in the previous text, and the repeated parts will not be repeated.

[0080] Figure 7 A structural block diagram of an electronic device provided for the implementation of the present invention. Figure 7 As shown, the electronic device provided by the embodiment of the present invention includes a processor 710 and a memory 720, wherein the memory is used to store instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the following Figure 1 Downscaling method for night-time light data shown.

[0081] In an exemplary embodiment, the electronic device may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above methods.

[0082] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory including instructions, and the instructions can be executed by a processor of a device to complete the above method. For example, the non-transitory computer-readable storage medium can be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device. The non-transitory computer-readable storage medium, when the instructions in the storage medium are executed by the processor of an electronic device, enables the electronic device to perform the following steps: Figure 1 Downscaling method for night-time light data shown.

[0083] The present invention also provides a computer program product, comprising a computer program, which when executed by a processor Figure 1 Downscaling method for night-time light data shown.

[0084] In the above description, the technical details such as the patterning of each layer are not described in detail. However, those skilled in the art should understand that various technical means can be used to form layers, regions, etc. of desired shapes. In addition, in order to form the same structure, those skilled in the art can also design methods that are not completely the same as the methods described above. In addition, although each embodiment is described separately above, this does not mean that the measures in each embodiment cannot be effectively combined and used.

[0085] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0086] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A downscaling method for nighttime light data, characterized in that: include: Acquire auxiliary data describing geographical conditions within a target area and first nighttime light image data having a first resolution within the target area; Performing rasterization processing on the auxiliary data to obtain first raster data and second raster data, wherein the first raster data is raster data matching the first nighttime light image data, and the second raster data has a second resolution, which is higher than the first resolution; Based on the first nighttime light image data and the first raster data, a pre-built geographically weighted neural network model is trained to obtain a target model; The second raster data is input into the target model to obtain second nighttime light image data with a resolution of the second resolution in the target area.

2. The method according to claim 1, characterized in that: The method of training a pre-built geographically weighted neural network model based on the first nighttime light image data and the first raster data to obtain a target model includes: Reprojecting the first nighttime light image data and the first raster data into the same coordinate system to obtain a training data set; The pre-built geographically weighted neural network model is trained using the training data set to obtain a target model.

3. The method according to claim 2, characterized in that The geographically weighted neural network model comprises an input layer, a hidden layer and an output layer. The pre-built geographically weighted neural network model is trained by the training data to obtain a target model, comprising: Inputting the training data set into the input layer, the training data set includes A grid data, where A is a positive integer; Each data point in each grid data in the training data set is used as a prediction point in turn, and the following steps are performed for the prediction point: The input layer inputs the spatial distance between the prediction point and other sample points into the hidden layer, and the other sample points are data points in the training data set other than the prediction point; Performing convolution processing on the spatial distance between the prediction point and other sample points through the hidden layer to obtain the spatial weight of the prediction point; Fitting the spatial weight to the regression coefficient corresponding to the prediction point through the output layer to obtain a training fitting value of the prediction point, wherein the regression coefficient is obtained based on the first nighttime light image data and the first raster data; According to the error between the fitting image data and the first nighttime light image data, the hidden layer is iteratively optimized to obtain a target model, wherein the fitting image data includes the training fitting values ​​of all the prediction points on the same raster data.

4. The method according to claim 3, characterized in that The step of iteratively optimizing the hidden layer according to the error between the fitted image data and the first nighttime light image data to obtain a target model comprises: Calculating a mean square error between the fitting image data and the first nighttime light image data; If the mean square error is less than the preset error value, or the number of training times is greater than the preset training value, the training is stopped and the geographically weighted neural network model is used as the target model; If the mean square error is not less than the preset error value, or the number of training times is not greater than the preset training value, the weight coefficients in the hidden layer are iteratively optimized, and the optimized geographically weighted neural network model is used as the target model.

5. The method according to claim 1, characterized in that: The auxiliary data includes multiple types of auxiliary data, and the auxiliary data is subjected to rasterization processing to obtain first raster data, including: Performing a first transformation process on each type of the auxiliary data to obtain B pieces of intermediate grid data with a resolution of the first resolution, where B is a positive integer; The B pieces of intermediate raster data are reprojected into the same coordinate system to obtain the first raster data.

6. The method according to claim 5, characterized in that The multiple types of auxiliary data include at least one of ground cover type data within the target area, ground surface temperature data within the target area, road network data within the target area, and point of interest data within the target area.

7. A downscaling device for nighttime light data, characterized in that: The device comprises: An acquisition module, used for acquiring auxiliary data describing the geographical conditions in a target area and first nighttime light image data with a first resolution in the target area; A processing module, configured to perform rasterization processing on the auxiliary data to obtain first raster data and second raster data, wherein the first raster data is raster data matching the first nighttime light image data, and the second raster data has a second resolution, which is higher than the first resolution; A training module, used for training a pre-built geographically weighted neural network model based on the first nighttime light image data and the first raster data to obtain a target model; The input module is used to input the second raster data into the target model to obtain the second nighttime light image data with the second resolution in the target area.

8. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the method according to any one of claims 1 to 6. 9 . A computer-readable storage medium, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the method according to claim 1 .

10. A computer program product, comprising a computer program, wherein the computer program is used by a processor to execute the method according to any one of claims 1 to 6.

Images