Night light remote sensing image super-division method and system using multi-source remote sensing data

Abstract

The invention discloses a nighttime light remote sensing image super-division method and system using multi-source remote sensing data, and the method comprises the steps: collecting an NPP-VIIRS low-resolution nighttime light image, a Landsat-8 image and a Luoma No.1 high-resolution nighttime light image, carrying out the preprocessing, feature texture extraction, resampling, and data set construction. Constructing a U-Net neural network model for increasing a channel attention mechanism to carry out super-resolution processing from a low-resolution image to a high-resolution image; according to the method, detail information in the low-resolution image is effectively supplemented, the image super-resolution precision and the image integrity are improved, compared with the prior art, the high-quality and richer night light super-resolution image is provided, and the method has the advantages of being advanced in algorithm, high in precision and high in intelligent degree and is suitable for popularization and application. And a high-quality data basis is provided for subsequent remote sensing image analysis and data extraction.

Description

Technical Field

[0001] The present invention relates to an image super-resolution method and system, and in particular to a nighttime light remote sensing image super-resolution method and system using multi-source remote sensing data. Background Art

[0002] Since the 1990s, nighttime light data has become a crucial tool for studying human activity, urbanization, and economic development. Satellite sensors capture the artificial light emitted by human activity at night on Earth's surface. This data provides a direct picture of the intensity and distribution of nocturnal activity. Especially for regions lacking traditional statistical data, nighttime light data offers a unique perspective, revealing regional economic disparities, urbanization, and energy consumption.

[0003] However, due to the scientific and technological levels of different eras and the iteration of satellite technology, the night light data obtained by satellite sensors launched at different times have obvious differences in spatial resolution, radiometric accuracy and time coverage, which brings challenges to data comparison and fusion. For example, the spatial resolution of Luojia-1 is 130 meters. This is also the original resolution of its night light image data. Compared with the original 750-meter resolution of NPP-VIIRS, Luojia-1 provides high-precision night light information that far exceeds NPP-VIIRS, and can provide more details for urban and regional analysis that requires detailed data. However, due to the limitations of satellite technology, the data that Luojia-1 can provide is only a short one-year time scale, far less than the time length that NPP-VIIRS can provide, which can provide data for every year since 2012.

[0004] Given this background, how to unify nighttime light data, resolution, and radiation data from different sources, and how to leverage the strengths of both Luojia-1 and NPP-VIIRS data, have become important issues worthy of research. In recent years, with the substantial increase in computing power, deep learning neural network algorithms have been widely used. They exhibit powerful feature extraction and nonlinear fitting capabilities, and can accurately extract the deep relationships between low-resolution and high-resolution images in image super-resolution technology. Neural network algorithms can quickly and accurately identify the potential connections between low-resolution and high-resolution images, thereby achieving high-resolution reconstruction of low-resolution images, providing a new possibility for high-precision reconstruction of nighttime light data. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a method for super-resolution of night light remote sensing images using multi-source remote sensing data to provide higher quality and more detailed night light super-resolution images, thereby improving the accuracy and image integrity during image super-resolution. On the other hand, a system for super-resolution of night light remote sensing images using multi-source remote sensing data is provided.

[0006] Technical solution: The remote sensing image surface temperature downscaling method of the present invention comprises the following steps:

[0007] S1. Acquire low-resolution NPP-VIIRS nighttime light images, high-resolution Luojia-1 nighttime light images, and Landsat-8 images;

[0008] S2. Preprocess the low-resolution NPP-VIIRS nighttime light image and the high-resolution Luojia-1 nighttime light image, synthesize the image daily data into monthly data, and convert the DN grayscale value into a radiance value;

[0009] S3. Calculate remote sensing indices based on each band of Landsat-8 images, obtain characteristic texture information, synthesize daily remote sensing index data into monthly data, and resample;

[0010] S4. Merge the resampled monthly remote sensing index data with the low-resolution NPP-VIIRS nighttime light image data to form multi-band remote sensing data. Compare the multi-band remote sensing data with the high-resolution Luojia-1 nighttime light image to construct a low-resolution data set.

[0011] S5. Randomly sampling data from the low-resolution data set to construct a control group;

[0012] S6. Build a U-Net neural network model with added channel attention and set model parameters;

[0013] S7, loading the control group data, inputting low-resolution monthly remote sensing index data and NPP-VIIRS night light image data, outputting high-resolution Luojia-1 night light image data, and training the U-Net neural network model;

[0014] S8. Input the low-resolution NPP-VIIRS night light image and its corresponding remote sensing index data into the trained U-Net neural network model, and output the high-resolution night light image to complete the reconstruction.

[0015] Preferably, step S2 includes: preprocessing night light images, splicing images of different regions at the same time to construct night light images of the Chinese region, resampling NPP-VIIRS night light images and Luojia-1 night light images, synthesizing image daily data into monthly data, unifying the coordinate system, eliminating negative image values ​​and converting them into 0 values, and using a raster calculator to convert the image DN grayscale value of the original data into actual radiation brightness value.

[0016] Preferably, step S3 includes: calculating remote sensing index data using each band of the Landsat-8 image to obtain characteristic texture information, calculating the remote sensing index of each month based on the coastal blue B1, blue B2, green B3, red B4, near infrared B5, shortwave infrared 1B6 and shortwave infrared 2B7 bands of the Landsat-8 sensor, synthesizing the daily data of each remote sensing index into monthly data, and resampling the monthly remote sensing index data to the same size scale as the low-resolution night light image.

[0017] Preferably, the remote sensing index includes the Normalized Difference Vegetation Index (NDVI), the Normalized Water Index (NDWI) and the Normalized Building Index (NDBI).

[0018] Preferably, step S5 includes: using the longitude and latitude of the Chinese regional range as the restriction range, using the random function to generate random longitude and latitude within the Chinese scale to construct random points, creating a grid rectangular frame with the specific grid position corresponding to the generated random point as the center point, sampling low-resolution and high-resolution images within the rectangular frame, using a mask tool to save the sampled image data as an image file named with the center point number and image type; pairing the data obtained by separate sampling according to the position coordinates, the position coordinates are the center position data in the file name, calling the OS module to select the corresponding file and generate a low-resolution image data pair to construct a control group.

[0019] Preferably, the network architecture of the U-Net neural network model with increased channel attention in step S6 consists of two modules, an encoder and a decoder, the encoder has a 5-layer structure, the 1-2 layer structure consists of 2 convolutional layers, 1 ReLU layer and 1 maximum pooling layer, the 3-5 layer structure consists of 1 convolutional layer, 1 ReLU layer and 1 maximum pooling layer, and each layer structure has a channel attention module; the decoder has a 7-layer structure, the 1-5 layer structure consists of 1 upsampling layer, 2 convolutional layers and 1 ReLU layer, and the 6-7 layer structure consists of 1 upsampling layer, 1 convolutional layer and 1 ReLU layer.

[0020] Preferably, the working process of the SENET module of the U-Net neural network model is expressed as follows:

[0021]

[0022] Where Z represents the output after the Squeeze operation, W and H represent the size of the input matrix, and x(i,j) represents the element in the i-th row and j-th column of the matrix;

[0023] S=F x (Z,W)=σ(g(Z,W));

[0024] Among them, S represents the output after the Excitation operation, Z represents the output after the Squeeze operation, W represents the size of the exponential dimension, σ represents the sigmoid function, and g represents the linear layer;

[0025]

[0026] in, Represents the output after tensor U operation, U represents a two-dimensional matrix, and S represents the learned weight.

[0027] The nighttime light remote sensing image super-resolution system of the present invention comprises:

[0028] Data acquisition module, used to obtain low-resolution NPP-VIIRS night light images, high-resolution Luojia-1 night light images, and Landsat-8 images;

[0029] A data preprocessing module, configured to preprocess the low-resolution NPP-VIIRS nighttime light image and the high-resolution Luojia-1 nighttime light image and perform image synthesis operations;

[0030] The feature extraction module is used to calculate remote sensing indices based on each band of the Landsat-8 image and obtain feature texture information; d) the resampling and band merging module is used to merge the resampled monthly remote sensing index data with the low-resolution NPP-VIIRS night light image data to form multi-band remote sensing data;

[0031] A dataset construction module is used to compare the multi-band remote sensing data with the high-resolution Luojia-1 night light image to construct a low-resolution dataset and a control group;

[0032] The model building module is used to build a U-Net neural network model with added channel attention and set model parameters, load control group data, and train the U-Net neural network model;

[0033] The reconstruction module is used to input the low-resolution NPP-VIIRS night light image and its corresponding remote sensing index data into the trained U-Net neural network model, and output the high-resolution night light image to complete the reconstruction.

[0034] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: 1. By introducing multi-source remote sensing data and improving the neural network model, an image super-resolution dataset based on multi-source remote sensing data and night light images of different resolutions is created, thereby providing night light images with higher resolution and more image details, and improving the accuracy and image integrity during image super-resolution; 2. By applying the improved deep learning neural network super-resolution algorithm, it is possible to better extract the deep texture details of the constructed low-resolution night light images and remote sensing indexes, and use the high-resolution image data generated by these incomplete details to generate night light images with high temporal and spatial resolution levels. Compared with other night light image sampling methods and super-resolution methods, it not only has higher accuracy, but also has better image texture features. It is suitable for many fields such as ecological environment monitoring, social and economic development, and energy resource management, and has significant application value and social benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the process of the present invention;

[0036] Figure 2 Schematic diagram of the SENET module structure of the U-Net neural network model of the present invention;

[0037] Figure 3 This is a structural diagram of the present invention's method for super-resolution of nighttime light remote sensing data using multi-source remote sensing data. DETAILED DESCRIPTION

[0038] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0039] The method for super-resolution of nighttime light remote sensing images using multi-source remote sensing data includes the following steps:

[0040] S1. Low-resolution NPP-VIIRS nighttime light images from 2018 to 2019 were obtained by the Earth Observation Group (EOG) at the Colorado School of Mines for resolution pairing. The data includes a Visible and Infrared Imaging Suite (VIIRS) Day-Night Band (DNB) with a spatial resolution of 750 m.

[0041] Based on the Luojia-1 official website, high-resolution Luojia-1 nighttime light images from 2018 to 2019 were obtained for resolution matching. This data includes a low-light detector band with a spatial resolution of 130m.

[0042] Landsat-8 images from 2018 to 2019 were acquired for feature texture extraction using the Google Earth Engine data cloud platform. The data includes five visible light bands, one near-infrared band, and two shortwave infrared bands from the OLI sensor, as well as two thermal infrared bands from the TIRS sensor, with a spatial resolution of 30 meters.

[0043] Download these images.

[0044] S2. Preprocess the low-resolution NPP-VIIRS nighttime light image and the high-resolution Luojia-1 nighttime light image, synthesize the image daily data into monthly data, and convert the DN grayscale value into the radiance value. The specific steps are as follows:

[0045] Preprocessing operations were performed on low-resolution NPP-VIIRS night light images and high-resolution Luojia-1 night light images respectively;

[0046] The NPP-VIIRS night light images from different regions at the same time were stitched together to construct the NPP-VIIRS night light image of China and resampled to 500m resolution. The Luojia-1 night light images from different regions at the same time were stitched together to construct the Luojia-1 night light image of China and resampled to 100m resolution.

[0047] The image daily data is synthesized into monthly data, the coordinate system is changed to WGS1984, the negative values ​​of the image are removed and converted to 0, and the image DN grayscale value of the original data is converted into the actual radiation brightness value using the raster calculator. The calculation formula is as follows:

[0048]

[0049] Where L represents the radiance value after absolute radiation correction, and the unit is W / (m 2 ·sr·μm), DN represents the image grayscale value;

[0050] Using the raster calculator, the dimensions of the two nighttime light remote sensing images (based on NPP-VIIRS) are calculated as follows:

[0051] L LJ =L×0.52×10 5 ;

[0052] Among them, L LJ It represents the radiance value after unified dimension, and the unit is NW / (cm 2 ·sr), L represents the grayscale value of the image after absolute radiation correction.

[0053] S3. Calculate the remote sensing index based on each band of the Landsat-8 image, obtain characteristic texture information, synthesize the daily remote sensing index data into monthly data, and resample. The specific steps are as follows:

[0054] Remote sensing index data were calculated using the various bands of Landsat-8 images to obtain characteristic texture information. Based on the coastal blue B1, blue B2, green B3, red B4, near-infrared B5, shortwave infrared 1B6, and shortwave infrared 2B7 bands of the Landsat-8-OLI sensor, the normalized difference vegetation index (NDVI), normalized water index (NDWI), and normalized building index (NDBI) for each month were calculated. The calculation formulas are as follows:

[0055]

[0056] Among them, NIR represents the reflectivity of the near-infrared band, and RED represents the reflectivity of the red band;

[0057]

[0058] Among them, GREEN represents the reflectivity of the green light band, and NIR represents the reflectivity of the near-infrared band;

[0059]

[0060] Among them, MIR stands for mid-infrared reflectance, and NIR stands for near-infrared reflectance;

[0061] The daily data of each index were synthesized into monthly data, and the monthly remote sensing index data were resampled to the same 500m pixel size as the low-resolution night light image.

[0062] S4. Merge the bands of the resampled monthly remote sensing index data with the same resolution and the low-resolution NPP-VIIRS night light image data to form multi-band remote sensing data to supplement the image feature texture information. The multi-band remote sensing data has four bands including the night light band, normalized difference vegetation index NDVI, normalized water index NDWI and normalized building index NDBI; compare the multi-band remote sensing data with the high-resolution Luojia-1 night light image of the same location to construct a low-resolution data set.

[0063] S5. Randomly sample data from the low-resolution data set to construct a control group. The specific steps are as follows:

[0064] Load pre-processed night light image data;

[0065] The longitude and latitude of China are used as the restriction range. The random function is used to generate random longitude and latitude within the scale of China to construct random points. The grid rectangle is created with the corresponding grid position of the generated random points as the center point. The size of the rectangle is 256000m*256000m. Within the rectangle, the image data of 512*512 size can be sampled for low-resolution images. The image data of 2560*2560 size can be sampled for high-resolution images within the rectangle. The mask tool is used to save the sampled image data as an image file named with the center point number and image type.

[0066] The sampled data are paired according to the position coordinates. The position coordinates are the center position data in the file name. The OS module is called to select the corresponding file and generate low-resolution image data pairs. According to the training requirements, a total of 1,000 sampled data pairs are generated to construct a control group.

[0067] S6. Build a U-Net neural network model with added channel attention and set the model parameters. The specific steps are as follows:

[0068] The pytorch module is used to build a U-Net neural network model with increased channel attention. The U-Net network architecture consists of two modules: the encoder and the decoder. The encoder includes a 5-layer structure. The 1-2 layer structure consists of two convolutional layers, a ReLU layer, and a maximum pooling layer. The 3-5 layer structure consists of a convolutional layer, a ReLU layer, and a maximum pooling layer. Each layer has a channel attention module to better extract channel feature information. The decoder includes a 7-layer structure. The 1-5 layer structure consists of an upsampling layer, two convolutional layers, and a ReLU layer. The 6-7 layer structure consists of an upsampling layer, a convolutional layer, and a ReLU layer. The working process of the SENET module is expressed as follows:

[0069]

[0070] Where Z represents the output after the Squeeze operation, W and H represent the size of the input matrix, and x(i,j) represents the element in the i-th row and j-th column of the matrix;

[0071] S=F x (Z,W)=σ(g(Z,W));

[0072] Among them, S represents the output after the Excitation operation, Z represents the output after the Squeeze operation, W represents the size of the exponential dimension, σ represents the sigmoid function, and g represents the linear layer;

[0073]

[0074] in, Represents the output after tensor U operation, U represents a two-dimensional matrix, and S represents the learned weight.

[0075] Set the loss function for model training and select L2 (MSE) loss as the loss function of the model to train the model to narrow the gap between the predicted value and the true value:

[0076]

[0077] Among them, n represents the sample size of the training data set, y sr and y hr They represent the predicted value and true value obtained after model training respectively.

[0078] The SSIM (Structural Similarity Index) and RMSE were calculated based on the deep learning super-resolution night light images simulated by the model and the night light images actually observed by the Luojia-1 satellite to verify the accuracy of the model. The calculation formula is:

[0079]

[0080] Among them, μ x and μ y represents the average brightness of the super-resolution night light image and the real high-resolution night light image, c1 and c2 represent constant terms for stable operation, and n represents the total number of image grids; and represents the variance contrast between the two, σ xy represents the covariance between the two, I x (i) and I y (i) Represents the value of each grid point in the super-resolved night light image and the real high-resolution night light image.

[0081] S7. Load the control group data, input low-resolution monthly remote sensing index data and NPP-VIIRS night light image data, output high-resolution Luojia-1 night light image data, and train the U-Net neural network model. The specific steps are as follows:

[0082] Load 1000 sets of sampled control image data, according to the ratio of 8:1:1, 800 sets of data are used as training sets, 100 sets of data are used as validation sets, and 100 sets of data are used as test sets;

[0083] Using low-resolution remote sensing index data and NPP-VIIRS night light image data as input and high-resolution Luojia-1 night light image data as output, the U-Net neural network model is trained using the data to learn the features between image transitions. The upper limit of training rounds is set to 200 rounds, the batch size is set to 64, the learning rate is set to 0.001, and the SGD optimizer is used. During the training round, the model parameters are continuously recorded and saved according to the reduction of the loss function until the end of the training round. The U-Net neural network model training is completed and exported.

[0084] S8. Input the processed lower-resolution NPP-VIIRS night light image and its corresponding remote sensing index data into the trained U-Net neural network model, and output the ultra-high-resolution night light image to complete the reconstruction.

[0085] The nighttime light remote sensing image super-resolution system using multi-source remote sensing data includes:

[0086] Data acquisition module, used to obtain low-resolution NPP-VIIRS night light images, high-resolution Luojia-1 night light images, and Landsat-8 images;

[0087] A data preprocessing module, configured to preprocess the low-resolution NPP-VIIRS nighttime light image and the high-resolution Luojia-1 nighttime light image and perform image synthesis operations;

[0088] The feature extraction module is used to calculate remote sensing indices based on each band of the Landsat-8 image and obtain feature texture information; d) the resampling and band merging module is used to merge the resampled monthly remote sensing index data with the low-resolution NPP-VIIRS night light image data to form multi-band remote sensing data;

[0089] A dataset construction module is used to compare the multi-band remote sensing data with the high-resolution Luojia-1 night light image to construct a low-resolution dataset and a control group;

[0090] The model building module is used to build a U-Net neural network model with added channel attention and set model parameters, load control group data, and train the U-Net neural network model;

[0091] The reconstruction module is used to input the low-resolution NPP-VIIRS night light image and its corresponding remote sensing index data into the trained U-Net neural network model, and output the high-resolution night light image to complete the reconstruction.

Claims

1. A method for super-resolution of nighttime light remote sensing images using multi-source remote sensing data, characterized in that: The following steps are involved: S1. Acquire low-resolution NPP-VIIRS nighttime light images, high-resolution Luojia-1 nighttime light images, and Landsat-8 images; S2. Preprocess the low-resolution NPP-VIIRS nighttime light image and the high-resolution Luojia-1 nighttime light image, synthesize the image daily data into monthly data, and convert the DN grayscale value into a radiance value; S3. Calculate remote sensing indices based on each band of Landsat-8 images, obtain characteristic texture information, synthesize daily remote sensing index data into monthly data, and resample; S4. Merge the resampled monthly remote sensing index data with the low-resolution NPP-VIIRS nighttime light image data to form multi-band remote sensing data. Compare the multi-band remote sensing data with the high-resolution Luojia-1 nighttime light image to construct a low-resolution data set. S5. Randomly sampling data from the low-resolution data set to construct a control group; S6. Build a U-Net neural network model with added channel attention and set model parameters; S7, loading the control group data, inputting low-resolution monthly remote sensing index data and NPP-VIIRS night light image data, outputting high-resolution Luojia-1 night light image data, and training the U-Net neural network model; S8. Input the low-resolution NPP-VIIRS night light image and its corresponding remote sensing index data into the trained U-Net neural network model, and output the high-resolution night light image to complete the reconstruction.

2. The nighttime light remote sensing image super-resolution method according to claim 1, characterized in that: Step S2 includes: preprocessing night light images, splicing images from different regions at the same time to construct night light images of the Chinese region, resampling the NPP-VIIRS night light images and the Luojia-1 night light images, synthesizing the image daily data into monthly data, unifying the coordinate system, eliminating negative image values ​​and converting them to zero values, and using a raster calculator to convert the image DN grayscale value of the original data into actual radiant brightness values.

3. The nighttime light remote sensing image super-resolution method according to claim 1, characterized in that: Step S3 includes: calculating remote sensing index data using each band of the Landsat-8 image to obtain characteristic texture information; calculating the remote sensing index for each month based on the coastal blue B1, blue B2, green B3, red B4, near-infrared B5, shortwave infrared 1B6, and shortwave infrared 2B7 bands of the Landsat-8 sensor; synthesizing the daily data of each remote sensing index into monthly data; and resampling the monthly remote sensing index data to the same size scale as the low-resolution night light image.

4. The nighttime light remote sensing image super-resolution method according to claim 3, characterized in that: The remote sensing index includes the normalized difference vegetation index NDVI, the normalized water index NDWI and the normalized building index NDBI.

5. The nighttime light remote sensing image super-resolution method according to claim 1, characterized in that: Step S5 includes: using the longitude and latitude of the Chinese region as the restriction range, using the random function to generate random longitude and latitude within the Chinese scale to construct random points, creating a grid rectangular frame with the specific grid position corresponding to the generated random point as the center point, sampling low-resolution and high-resolution images within the rectangular frame, using the mask tool to save the sampled image data as image files named with the center point number and image type; pairing the data obtained by separate sampling according to the position coordinates, the position coordinates are the center position data in the file name, calling the OS module to select the corresponding file and generate a low-resolution and high-resolution image data pair to construct a control group.

6. The nighttime light remote sensing image super-resolution method according to claim 1, characterized in that: The network architecture of the U-Net neural network model with added channel attention in step S6 consists of two modules: an encoder and a decoder. The encoder has a 5-layer structure. The 1-2 layer structure consists of 2 convolutional layers, 1 ReLU layer, and 1 maximum pooling layer. The 3-5 layer structure consists of 1 convolutional layer, 1 ReLU layer, and 1 maximum pooling layer. Each layer structure has a channel attention module. The decoder has a 7-layer structure, the 1-5 layers consist of 1 upsampling layer, 2 convolution layers and 1 ReLU layer, and the 6-7 layers consist of 1 upsampling layer, 1 convolution layer and 1 ReLU layer.

7. The nighttime light remote sensing image super-resolution method according to claim 6, characterized in that: The working process of the SENET module of the U-Net neural network model is expressed as follows: Where Z represents the output after the Squeeze operation, W and H represent the size of the input matrix, and x(i,j) represents the element in the i-th row and j-th column of the matrix; S=F x (Z,W)=σ(g(Z,W)); Among them, S represents the output after the Excitation operation, Z represents the output after the Squeeze operation, W represents the size of the exponential dimension, σ represents the sigmoid function, and g represents the linear layer; in, Represents the output after tensor U operation, U represents a two-dimensional matrix, and S represents the learned weight.

8. A nighttime light remote sensing image super-resolution system using multi-source remote sensing data, characterized in that: include: Data acquisition module, used to obtain low-resolution NPP-VIIRS night light images, high-resolution Luojia-1 night light images, and Landsat-8 images; A data preprocessing module, configured to preprocess the low-resolution NPP-VIIRS nighttime light image and the high-resolution Luojia-1 nighttime light image and perform image synthesis operations; The feature extraction module is used to calculate remote sensing indices based on each band of the Landsat-8 image and obtain feature texture information; d) the resampling and band merging module is used to merge the resampled monthly remote sensing index data with the low-resolution NPP-VIIRS night light image data to form multi-band remote sensing data; A dataset construction module is used to compare the multi-band remote sensing data with the high-resolution Luojia-1 night light image to construct a low-resolution dataset and a control group; The model building module is used to build a U-Net neural network model with added channel attention and set model parameters, load control group data, and train the U-Net neural network model; The reconstruction module is used to input the low-resolution NPP-VIIRS night light image and its corresponding remote sensing index data into the trained U-Net neural network model, and output the high-resolution night light image to complete the reconstruction.

9. A computer device, characterized in that: The method comprises one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the nighttime light remote sensing image super-resolution method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the nighttime light remote sensing image super-resolution method according to any one of claims 1 to 7 are implemented.

Images