A Remote Sensing Land Surface Temperature Interpolation Method for Cloud-Covered Areas Integrating Spatial Auxiliary Variables and Conditional Generative Adversarial Networks

Abstract

The present invention proposes a method for interpolating the remote sensing land surface temperature in cloud-covered areas by integrating spatial auxiliary variables and conditional generative adversarial networks. First, the CGAN model is pre-trained on monthly composite land surface temperature data with good spatio-temporal integrity, so that the model can evenly consider different regional characteristics during the training process. Then, transfer training is carried out based on daily-scale data to make it applicable to the interpolation of daily land surface temperature. And during the interpolation process, a spatial progressive interpolation strategy is adopted. By considering the background information of the interpolation area, interpolation is carried out gradually from the cloud cover edge to the cloud cover center to eliminate the edge effect, so as to achieve accurate interpolation of the land surface temperature in cloud-covered areas.

Description

Technical Field

[0001] The present invention relates to the technical field of thermal infrared remote sensing surface temperature, and particularly relates to a method for interpolating remote sensing surface temperature in cloud-covered areas by integrating spatial auxiliary variables and conditional generative adversarial networks. Background Art

[0002] The surface temperature retrieved by thermal infrared remote sensing has high accuracy and spatio-temporal resolution, and is widely used in many fields such as meteorology, hydrology, environment, and public health. However, the thermal infrared signal cannot penetrate clouds, and about two-thirds of the land on the earth is covered by clouds every day, resulting in a large number of data missing in the surface temperature data retrieved by thermal infrared. Therefore, reconstructing the remote sensing surface temperature in cloud-covered areas is of great significance for improving its application value.

[0003] Currently, the interpolation research on remote sensing surface temperature can be divided into two categories according to the information sources used for interpolation. One category is to use the information of cloud-free pixels adjacent to the cloud-covered area in the remote sensing surface temperature data to interpolate the cloud-covered area by using spatial interpolation or deep learning models such as generative adversarial networks (GANs); the other category is to construct a model based on environmental variables (such as vegetation index, elevation, surface albedo, etc.) that are closely related to the surface temperature of cloud-free pixels. Since these environmental variables change little over time, a spatially non-empty distribution map can be generated through multi-temporal synthesis, and then the model is applied to estimate the surface temperature of the environmental variables in the cloud-covered area. However, considering that the surface temperature has strong spatio-temporal variability, when simply relying on the correlation information with environmental variables to model and switching to different regions, the filling effect is often not stable due to the lack of support from background information; and only relying on the information of cloud-free pixels adjacent to the cloud area of the surface temperature for interpolation, the interpolation effect is also not ideal when facing large-scale spatio-temporal missing. Currently, there are still few studies on interpolating the missing area of surface temperature based on remote sensing data by combining the adjacent information and relevant information of the surface temperature missing area. Currently, the research on interpolating surface temperature using GAN does not use auxiliary variables or uses less auxiliary variable data, and does not fully utilize the multi-source auxiliary variable information. At the same time, when training GAN, it is necessary to construct a surface temperature image pair by adding a cloud mask based on spatially continuous non-missing surface temperature. Due to the spatio-temporal differences in cloud cover distribution, the number of spatially continuous image patches cropped from daily surface temperature data is large in arid regions and very small in cloudy and rainy regions. The non-uniformity of the data will lead to a large difference in the interpolation effect of the trained model in different regions. In addition, the interpolation research of deep learning models is to first segment into small images for interpolation, and then splice the interpolated small images into a large image. When splicing, not considering the spatial background information will lead to boundary effects, resulting in mosaic textures in the interpolated large-scale cloud-covered area, that is, incorrect surface temperature textures. Summary of the Invention

[0004] The purpose of the present invention is to propose a method for interpolating remotely sensed land surface temperature in cloud-covered areas by integrating spatial auxiliary variables and conditional generative adversarial networks. By combining the land surface temperature, vegetation index, water body index, elevation, and other auxiliary spatial variables of clear-sky pixels adjacent to cloud-covered areas, and based on conditional generative adversarial networks, full pre-training is carried out on a monthly scale and then transferred to a daily scale. The missing values in the daily land surface temperature in cloud-covered areas are gradually interpolated from the missing edge to the missing center, eliminating the edge effect caused by stitching after small-map interpolation by the conditional generative adversarial network (CGAN), and generating seamless daily remotely sensed land surface temperature.

[0005] To achieve the above technical objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for interpolating remotely sensed land surface temperature in cloud-covered areas by integrating spatial auxiliary variables and conditional generative adversarial networks, the method comprising the following steps:

[0007] Construct a monthly-scale interpolation model based on conditional generative adversarial networks. Use the monthly-scale land surface temperature data and spatially auxiliary variables for interpolation as the pre-training dataset, and pre-train the monthly-scale interpolation model. The spatial variables include at least the enhanced vegetation index EVI, the modified normalized difference water index MNDWI, and elevation. The monthly-scale interpolation model includes a generator and a discriminator. Among them, use the monthly composite land surface temperature data with simulated cloud cover, monthly composite EVI, monthly composite MNDWI, elevation, and cloud mask data as the input variables of the generator, and the cloud-free monthly composite land surface temperature as the target variable of the generator to train the generator. Use the monthly composite land surface temperature interpolated by the generator and the cloud-free monthly composite land surface temperature data as the input variables of the discriminator respectively, and the labels of the two input variables as the target variables of the discriminator to train the discriminator. Through the adversarial training of the generator and the discriminator, select the weight parameters with the highest accuracy tested by the validation set and use them as the best weight parameters of the monthly-scale interpolation model;

[0008] Load the best weight parameters, use the daily-scale land surface temperature data and spatially auxiliary variables for interpolation as the transfer training dataset, transfer train the generator and the discriminator to obtain the best generator parameters corresponding to the daily scale, and use the generator loaded with the best parameters to gradually interpolate the missing values in the daily land surface temperature in cloud-covered areas from the missing edge to the missing center.

[0009] Furthermore, the process of obtaining training data includes the following steps:

[0010] Collect daily land surface temperature data and spatially auxiliary variables for interpolation. The spatial variables include at least the enhanced vegetation index EVI, the modified normalized difference water index MNDWI, and elevation; extract the missing value area from the daily land surface temperature data as the cloud mask data;

[0011] Monthly synthesis and cloud removal processing are performed on daily surface temperature data, EVI, and MNDWI to obtain monthly synthesized surface temperature, monthly synthesized EVI, and monthly synthesized MNDWI;

[0012] The daily surface temperature, monthly synthesized surface temperature, monthly synthesized EVI, monthly synthesized MNDWI, and elevation data are cropped into sub-images according to a preset overlap ratio; then randomly extract cloud mask data to perform cloud masking processing on the monthly synthesized surface temperature to generate monthly synthesized surface temperature data with simulated cloud cover;

[0013] The cropped sub-images are combined according to the imaging time and imaging area, divided into daily-scale data and monthly-scale data, and the daily-scale data and monthly-scale data are respectively divided into training sets, validation sets, and test sets according to a preset ratio.

[0014] Furthermore, the Enhanced Vegetation Index EVI is calculated using the following formula:

[0015] ;

[0016] In the formula, EVI is the Enhanced Vegetation Index, is the reflectance of the near-infrared band, is the reflectance of the red light band, is the reflectance of the blue light band;

[0017] The Modified Normalized Difference Water Index MNDWI is calculated using the following formula:

[0018] ;

[0019] In the formula, MNDWI is the Modified Normalized Difference Water Index, is the reflectance of the green light band, is the reflectance of the short-wave infrared 1 band;

[0020] Extract elevation data from the digital elevation model data.

[0021] Furthermore, the monthly-scale data is used as the pre-training dataset to pre-train the monthly-scale interpolation model, and the training objective loss consists of adversarial loss and content loss; the adversarial loss uses the MSE loss function, and the content loss uses the L1 loss function; among them, the formula of the MSE loss function is as follows:

[0022] ;

[0023] In the formula is and the MSE loss between, is the predicted value, is the target value, and N represents the total number;

[0024] The formula of the L1 loss function is as follows:

[0025] ;

[0026] Wherein is and the L1 loss between them.

[0027] Furthermore, the formula for the monthly synthetic land surface temperature interpolated by the generator is as follows:

[0028] ;

[0029] Wherein G is the generator, represents the land surface temperature interpolated by the generator G, is the land surface temperature of the simulated cloud cover, Mask is the cloud mask, is the monthly synthetic EVI, is the monthly synthetic MNDWI, and EL is the elevation.

[0030] Furthermore, the formula for the discriminator to discriminate between the interpolated monthly synthetic land surface temperature and the true clear-sky monthly synthetic land surface temperature is as follows:

[0031] ;

[0032] Wherein D is the discriminator, is the land surface temperature interpolated by the generator the discrimination result after passing through the discriminator, is the true cloud-free land surface temperature the discrimination result after passing through the discriminator.

[0033] Furthermore, the process of selecting the weight parameters with the highest accuracy through the adversarial training of the generator and the discriminator and tested by the validation set includes the following steps:

[0034] The target loss function of the generator is composed of the adversarial loss and the content loss. The adversarial loss uses the MSE loss, and the content loss uses the L1 regularization. The specific formula is as follows:

[0035] ;

[0036] The target loss function of the discriminator is the adversarial loss. The specific formula is as follows:

[0037] ;

[0038] The target loss of the conditional generative adversarial network is as follows:

[0039] ;

[0040] Through the adversarial training of the generator and the discriminator, the weight parameters with the highest accuracy after being tested by the validation set are selected, and then its accuracy is tested with the test set to obtain the optimal weight parameters of the conditional generative adversarial network.

[0041] Furthermore, the process of using the generator loaded with the optimal parameters to perform progressive interpolation of the missing values in the cloud-covered area of the daily surface temperature from the missing edge to the missing center includes the following steps:

[0042] Taking the boundary pixels of the entire study area as the center, obtaining the surface temperature data, monthly composite EVI, monthly composite MNDWI, elevation, and cloud mask data within a 256×256 grid containing cloud cover, inputting the obtained data into the generator loaded with the optimal parameters for interpolation of cloudy pixels. After the interpolation is completed, the theoretical clear-sky surface temperature within the central 192×192 grid is spliced into the large image, and then the next cycle is carried out until all cloudy pixels within the entire study area are interpolated.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] First, the method for interpolating remotely sensed surface temperature in cloud-covered areas of the present invention that integrates spatial auxiliary variables and conditional generative adversarial networks interpolates the surface temperature of cloud cover by combining multi-source spatial auxiliary information and CGAN, taking into account the advantages of the CGAN model in utilizing spatial neighborhood context information and the ability of relevant environmental variables to characterize the surface characteristics of cloud-covered areas.

[0045] Second, the method for interpolating remotely sensed surface temperature in cloud-covered areas of the present invention that integrates spatial auxiliary variables and conditional generative adversarial networks first pre-trains the model with monthly-scale data, and then performs transfer training with daily-scale data, ensuring sufficient training samples and that there will be no significant differences in the interpolation effect in arid and cloudy and rainy areas.

[0046] Third, the method for interpolating remotely sensed surface temperature in cloud-covered areas of the present invention that integrates spatial auxiliary variables and conditional generative adversarial networks uses a spatial progressive method to perform interpolation from the cloud cover edge to the cloud cover center, considering the background surface temperature around the large image and eliminating the edge effect caused by stitching small images into a large image. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flowchart of the method for interpolating remotely sensed surface temperature in cloud-covered areas of the present invention that integrates spatial auxiliary variables and conditional generative adversarial networks;

[0048] Figure 2It is a diagram of the spatial progressive interpolation process;

[0049] Figure 3 It is a schematic diagram for training the interpolation model in a typical 256×256 area;

[0050] Figure 4 It is a schematic diagram of the interpolation result of the surface temperature of the entire area. Specific implementation manners

[0051] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0052] The present invention discloses a method for interpolating the remote sensing surface temperature in cloud-covered areas by integrating spatial auxiliary variables and conditional generative adversarial networks. The method includes the following steps:

[0053] Based on the conditional generative adversarial network, a monthly-scale interpolation model is constructed. Using the monthly-scale surface temperature data and the spatial variables for auxiliary interpolation as the pre-training data set, the monthly-scale interpolation model is pre-trained. The spatial variables include at least the enhanced vegetation index EVI, the modified normalized difference water index MNDWI, and elevation. The monthly-scale interpolation model includes a generator and a discriminator. Among them, the monthly synthesized surface temperature data with simulated cloud cover, the monthly synthesized EVI, the monthly synthesized MNDWI, elevation, and cloud mask data are used as the input variables of the generator, and the cloud-free monthly synthesized surface temperature is used as the target variable of the generator to train the generator. The monthly synthesized surface temperature interpolated by the generator and the monthly synthesized surface temperature data of the true clear sky are used as the input variables of the discriminator respectively, and the labels of the two input variables are used as the target variables of the discriminator to train the discriminator. Through the adversarial training of the generator and the discriminator, the weight parameters with the highest accuracy after being tested by the validation set are selected and used as the best weight parameters of the monthly-scale interpolation model;

[0054] Load the best weight parameters, use the daily-scale surface temperature data and the spatial variables for auxiliary interpolation as the transfer training data set, perform transfer training on the generator and the discriminator to obtain the best generator parameters corresponding to the daily scale, and use the generator loaded with the best parameters to perform progressive interpolation on the missing values in the cloud-covered areas of the daily surface temperature from the missing edge to the missing center.

[0055] The purpose of the present invention is to propose a method for interpolating the remote sensing surface temperature in cloud-covered areas based on multi-source spatial auxiliary variables and deep learning models. The present invention is based on the conditional generative adversarial network model (CGAN), combines the clear-sky surface temperature adjacent to the cloud-covered area of the remote sensing surface temperature, and auxiliary spatial variables such as vegetation index, water index, and altitude, and performs progressive interpolation on the missing values of the surface temperature in the cloud-covered area to generate seamless remote sensing surface temperature.

[0056] 1) Extraction and calculation of spatial variables

[0057] The spatial variables used in the present invention for assisting interpolation are the Enhanced Vegetation Index (EVI), the Modified Normalized Difference Water Index (MNDWI), and elevation.

[0058] The formula for calculating EVI is as follows:

[0059] (1);

[0060] In the formula, EVI is the Enhanced Vegetation Index, is the reflectance in the near-infrared band, is the reflectance in the red band, is the reflectance in the blue band.

[0061] The formula for calculating MNDWI is as follows:

[0062] (2);

[0063] In the formula, MNDWI is the Modified Normalized Difference Water Index, is the reflectance in the green band, is the reflectance in the shortwave infrared 1 band.

[0064] The elevation data is extracted from digital elevation model data.

[0065] The missing value areas are extracted from the daily surface temperature data, i.e., the cloud mask data.

[0066] 2) Monthly composite processing

[0067] In order to obtain large-area spatially continuous cloud-free surface temperature data for training the CGAN model, variables such as surface temperature, EVI, and MNDWI are subjected to monthly composite cloud removal processing.

[0068] The calculation formula for monthly mean synthesis is as follows:

[0069] (3);

[0070] In the formula can represent the monthly mean synthesized surface temperature (LST), EVI, and MNDWI, n represents the number of days in each month, can represent the daily LST, EVI, and MNDWI, represents the daily cloud mask data, represents cloud interference when represents clear sky when.

[0071] 3) Preparation of the deep learning dataset

[0072] The daily surface temperature, monthly composite surface temperature, monthly composite EVI, monthly composite MNDWI, and elevation data are cropped into small images of 256×256 according to a certain overlap ratio. Then, cloud mask data is randomly selected to perform cloud masking on the monthly-scale surface temperature to generate monthly composite surface temperature data with simulated cloud cover.

[0073] The 256×256-sized images are combined according to the imaging time and imaging area, and divided into daily-scale data and monthly-scale data. The difference between the two lies in whether the surface temperature has been monthly synthesized, while other data remains unchanged.

[0074] Then, the daily-scale data and monthly-scale data are divided into training sets, validation sets, and test sets according to the ratio of 7:2:1.

[0075] 4) Pre-training of the monthly-scale interpolation model

[0076] Using the monthly-scale data as the pre-training dataset, the model is pre-trained. The target loss for training consists of adversarial loss and content loss. The adversarial loss uses the MSE loss function, and the content loss uses the L1 loss function.

[0077] The formula for the MSE loss function is as follows:

[0078] (4);

[0079] In the formula is and the MSE loss between, is the predicted value, is the target value, and N represents the total number.

[0080] The formula for the L1 loss function is as follows:

[0081] (5);

[0082] In the formula is and the L1 loss between, is the predicted value, is the target value, and N represents the total number.

[0083] The monthly composite land surface temperature, EVI, MNDWI, elevation, and cloud mask that simulate cloud cover are used as input variables for the generator, and the cloud-free monthly composite land surface temperature is used as the target variable for the generator to train the generator. It is divided into a training set, a validation set, and a test set according to the ratio of 7:2:1. The land surface temperature interpolated by the generator and the cloud-free clear-sky land surface temperature data are used as input variables for the discriminator respectively. The label of the interpolated land surface temperature is set to 0, while the label of the cloud-free land surface temperature is set to 1. The label is used as the target variable for the discriminator to train the discriminator.

[0084] The formula for the land surface temperature interpolated by the generator is as follows:

[0085] (6);

[0086] In the formula, G is the generator, represents the land surface temperature interpolated by the generator G, is the land surface temperature that simulates cloud cover, Mask is the cloud mask, is the monthly composite EVI, is the monthly composite MNDWI, and EL is the elevation.

[0087] The formula for the discriminator to discriminate the interpolated land surface temperature and the true clear-sky land surface temperature is as follows:

[0088] (7);

[0089] In the formula, D is the discriminator, is the land surface temperature interpolated by the generator is the discrimination result after passing through the discriminator, is the true cloud-free land surface temperature is the discrimination result after passing through the discriminator. In the present invention, the label of the true land surface temperature is represented by 1, and the label of the interpolated land surface temperature is represented by 0.

[0090] The target loss function of the generator consists of an adversarial loss and a content loss. The adversarial loss uses the MSE loss, and the content loss is the L1 regularization. The specific formula is as follows:

[0091] (8);

[0092] The target loss function of the discriminator is the adversarial loss, and the specific formula is as follows:

[0093] (9);

[0094] The target loss of CGAN , and the specific formula is as follows:

[0095] (10);

[0096] Through the adversarial training of the generator and discriminator, select the weight parameters with the highest accuracy after being tested by the validation set, and then use the test set to test whether there are excessive fluctuations in its accuracy. If the accuracy evaluation in the test set is similar to the accuracy evaluation in the validation set, use it as the weight parameters of the best CGAN.

[0097] 5) Transfer training from monthly scale to daily scale

[0098] First, load the weights of the best CGAN model obtained by pre-training, then use the daily-scale data as the transfer training dataset, adjust some hyperparameters, and perform transfer training on the generator and discriminator. Then, select the best generator weight parameters according to the test results of the validation set, and use the test set to test whether there are excessive fluctuations in its accuracy. If the accuracy evaluation in the test set is similar to the accuracy evaluation in the validation set, it is used as the best generator parameters of the imputation model.

[0099] 6) Spatial progressive imputation of daily surface temperature cloud-covered areas

[0100] First, load the best weight parameters of the generator, and then detect the pixels containing cloud cover and being the boundary between the missing area and the complete area row by row and column by column for the surface temperature of the entire study area. Taking the boundary pixels as the center, obtain the surface temperature data, monthly composite EVI, monthly composite MNDWI, elevation, and cloud mask data within the 256×256 grid containing cloud cover. Then, input these data into the best generator for imputation of cloudy pixels. After the imputation is completed, splice the theoretical clear-sky surface temperature within the central 192×192 grid into the large image, and then perform the next cycle until all cloudy pixels in the large image are imputed.

[0101] Example

[0102] The present invention interpolates the surface temperature with cloud cover based on multi-source spatial auxiliary variables and conditional generative adversarial networks. The following are the specific implementation steps. The technical process is shown in Figure 1 .

[0103] 1) According to the MOD11A1 surface temperature product, perform quality control on it, assign the pixels interfered by clouds as NAN, and obtain the daily surface temperature.

[0104] 2) Binarize the daily surface temperature, assign the NAN value pixels as 1, and the non-NAN value pixels as 0, to obtain the cloud mask data.

[0105] 3) According to the MOD09GA surface reflectance product, perform quality control on it, obtain the pixels without cloud interference, and then calculate EVI and MNDWI. The calculations are shown in formulas (1) and (2).

[0106] 4) Extract elevation data from the GTOPO30 DEM product.

[0107] 5) Perform monthly mean synthesis on the MOD11A1 land surface temperature data, EVI, and MNDWI data to obtain cloud-free complete data. The calculation is shown in Equation (3). Then, resample the monthly synthesized EVI and monthly synthesized MNDWI to 1000 m, which is consistent with the MODIS land surface temperature.

[0108] 6) Crop the daily land surface temperature data, monthly synthesized land surface temperature data, monthly synthesized EVI, monthly synthesized MNDWI, elevation, and cloud mask data into 256×256 images according to a certain overlap ratio.

[0109] 7) Randomly extract cloud mask data from the cloud mask dataset to perform cloud masking operation on the cropped monthly-scale land surface temperature data. The value is -50 in the cloud-covered area, and the value remains unchanged in the remaining areas.

[0110] 8) Prepare the dataset for model training. Use time and location as the basis for matching, and combine the simulated cloud-covered land surface temperature, cloud-free land surface temperature, EVI, MNDWI, elevation, and cloud mask data.

[0111] 9) Pre-train the interpolation model using monthly-scale data. The hyperparameters for model pre-training are set as follows: Epoch is 100, and the batch size is 12. Use Adam as the optimizer, and the learning rate change strategy is to first warm up with a smaller learning rate and then use cosine annealing to gradually decrease the learning rate as the training progresses. Among them, the maximum and minimum learning rates are 0.001 and 0.0001 respectively, and the epoch to stop warming up is 30. The above hyperparameters are the same in the discriminator and the generator. The adversarial loss uses the MSE loss, as shown in Equation (4), and the content loss uses the L1 loss, as shown in Equation (5).

[0112] Divide it into a training set, a validation set, and a test set according to the ratio of 7:2:1. During the training process of the model, use the simulated cloud-covered monthly synthesized land surface temperature, EVI, MNDWI, elevation, and cloud mask data as the input variables of the generator of the conditional generative adversarial network, and the cloud-free monthly synthesized land surface temperature as the target variable of the generator to perform interpolation training on the generator. The land surface temperature interpolated by the generator is shown in Equation (6); Input the land surface temperature interpolated by the generator and the cloud-free monthly synthesized land surface temperature into the discriminator for discriminative training respectively. The discriminative result of the discriminator is shown in Equation (7).

[0113] The target loss function during the training of the generator consists of an adversarial loss function and a content loss function, as shown in Equation (8); the target loss function during the training of the discriminator is the adversarial loss function, as shown in Equation (9). Among them, the target loss function of CGAN is shown in Equation (10).

[0114] After each epoch is completed, the trained weight parameters are saved, and the accuracy of each epoch is tested using the validation set. After the model is trained, the weight parameters of the epoch with the highest accuracy are selected according to the results of the validation set test, and the test set is used to test whether there are large fluctuations in its accuracy. If the accuracy tested by the test set has a small difference from the accuracy in the validation set, these parameters are used as the initial parameters for the model transfer training.

[0115] 10) The pre-trained model at the monthly scale is transferred and trained using daily-scale data. The hyperparameters for the model pre-training are set as follows: Epoch is 50, batch size is 12, SGD is used as the optimizer, and the learning rate is fixed at 0.0001. These hyperparameters are the same for both the discriminator and the generator. The adversarial loss uses the MSE loss, and the regularization method uses L1 regularization. Only the number of training epochs, the optimizer, and the learning rate are modified during the transfer training, and the rest of the hyperparameters remain unchanged.

[0116] First, the weight parameters with the highest accuracy trained in the pre-trained model are loaded. Then, the daily-scale land surface temperature simulating cloud cover, monthly composite EVI, MNDWI, elevation, and cloud mask data are used as the target variables of the generator, and the cloud-free daily-scale land surface temperature is used as the target variable of the generator. The generator is transferred and trained from monthly-scale data to daily-scale data using daily-scale data. The land surface temperature interpolated by the generator and the cloud-free daily-scale land surface temperature are respectively input into the discriminator for the transfer training of the discriminator.

[0117] After the transfer training is completed, the weight parameters of the generator with the best training effect are selected in the same way as the best parameter selection during the pre-training. Figure 3 The figures of the pre-training interpolation and transfer training interpolation results of the land surface temperature in several typical 256×256 regions are given. According to the accuracy test, the R 2 (coefficient of determination) of the land surface temperature interpolated by the pre-training is 0.96, the MAE (mean absolute error) is 1.16, and the RMSE (root mean square error) is 1.68; after the transfer training, the MAE becomes 1.08 and the RMSE becomes 1.58.

[0118] 11) Before interpolating the land surface temperature with cloud cover over the entire region, the land surface temperature of all pixels with cloud interference is assigned a value of -50. Next, the land surface temperature data of this region is binarized, and the pixels with a value of -50 are marked as 1, and the rest of the pixels are marked as 0, serving as the cloud mask.

[0119] Then, a 3×3 convolution summation operation is performed on each pixel in the cloud mask. After convolution processing, pixels with values of 0 and 9 do not belong to the boundary pixels of the complete area and the missing area, while pixels with values between 1 and 8 are boundary pixels.

[0120] Next, these boundary pixels are retrieved pixel by pixel in row-column order. When a boundary pixel is detected, a 256×256 area centered on this pixel is selected, and the land surface temperature, EVI, MNDWI, elevation, and cloud mask data within this area are extracted.

[0121] Finally, the extracted data are input into the best generator to generate the interpolation result within this 256×256 area. The pixels within the range of 192×192 at the center of this area are selected and inserted into the corresponding positions. Repeat this process (as shown in Figure 2 ), until all the areas to be interpolated are successfully interpolated, and the seamless land surface temperature after interpolation for the entire study area is obtained (as shown in Figure 4 ).

[0122] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications once they know the basic creative concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0123] Obviously, those skilled in the art can make various changes and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. A method for interpolating remote sensing land surface temperature in cloud-covered areas that integrates spatial auxiliary variables and conditional generative adversarial networks, characterized in that The method includes the following steps: Based on a conditional generative adversarial network, construct a monthly-scale interpolation model. Use the monthly-scale land surface temperature data and auxiliary interpolation spatial variables as the pre-training dataset to pre-train the monthly-scale interpolation model. The spatial variables at least include the Enhanced Vegetation Index (EVI), the Modified Normalized Difference Water Index (MNDWI), and elevation. The monthly-scale interpolation model includes a generator and a discriminator. Among them, use the monthly composite land surface temperature data with simulated cloud cover, monthly composite EVI, monthly composite MNDWI, elevation, and cloud mask data as the input variables of the generator, and the cloud-free monthly composite land surface temperature as the target variable of the generator to train the generator. Use the monthly composite land surface temperature interpolated by the generator and the real clear-sky monthly composite land surface temperature data as the input variables of the discriminator respectively, and the labels of the two input variables as the target variables of the discriminator to train the discriminator. Through the adversarial training of the generator and the discriminator, select the weight parameters with the highest accuracy after being tested by the validation set, and use them as the best weight parameters of the monthly-scale interpolation model; Load the best weight parameters. Use the daily-scale land surface temperature data and auxiliary interpolation spatial variables as the transfer training dataset to perform transfer training on the generator and the discriminator to obtain the best generator parameters corresponding to the daily scale. Use the generator loaded with the best parameters to perform progressive interpolation on the missing values in the cloud-covered areas of the daily land surface temperature from the missing edge to the missing center.

2. The method for interpolating the remote sensing surface temperature in the cloud-covered area by integrating the spatial auxiliary variable and the conditional generative adversarial network according to claim 1, characterized in that The process of obtaining training data includes the following steps: Collect daily land surface temperature data and spatial variables for auxiliary interpolation. The spatial variables at least include the Enhanced Vegetation Index (EVI), the Modified Normalized Difference Water Index (MNDWI), and elevation. Extract the missing value areas from the daily land surface temperature data as the cloud mask data; Perform monthly composite cloud removal processing on the daily land surface temperature data, EVI, and MNDWI to obtain monthly composite land surface temperature, monthly composite EVI, and monthly composite MNDWI; Crop the daily land surface temperature, monthly composite land surface temperature, monthly composite EVI, monthly composite MNDWI, and elevation data into subgraphs according to a preset overlap ratio. Then randomly extract cloud mask data to perform cloud masking processing on the monthly composite land surface temperature to generate monthly composite land surface temperature data with simulated cloud cover; Combine the cropped subgraphs according to the imaging time and imaging area, divide them into daily-scale data and monthly-scale data, and divide the daily-scale data and monthly-scale data into training sets, validation sets, and test sets according to a preset ratio.

3. The remote sensing surface temperature interpolation method for cloud-covered areas integrating spatial auxiliary variables and conditional generative adversarial networks according to claim 1, characterized in that, Calculate the Enhanced Vegetation Index (EVI) using the following formula: where EVI is the Enhanced Vegetation Index, ρ NIR is the reflectance in the near-infrared band, ρ Red is the reflectance in the red band, R Blue is the reflectance in the blue band; Calculate the Modified Normalized Difference Water Index (MNDWI) using the following formula: In the formula, MNDWI is the Modified Normalized Difference Water Index, and ρ Green is the reflectance of the green light band, and ρ SWIR1 is the reflectance of the short-wave infrared 1 band; Extract elevation data from digital elevation model data.

4. The method for interpolating the remote sensing surface temperature in the cloud-covered area by integrating the spatial auxiliary variable and the conditional generative adversarial network according to claim 1, wherein Use the monthly-scale data as the pre-training dataset to pre-train the monthly-scale interpolation model. The training objective loss consists of adversarial loss and content loss. The adversarial loss uses the MSE loss function, and the content loss uses the L1 loss function. Among them, the formula of the MSE loss function is as follows: In the formula is y i and is the MSE loss between y i is the predicted value, is the target value, and N represents the total quantity; The formula of the L1 loss function is as follows: where is the L1 loss between y i and .

5. The method for interpolating the remote sensing surface temperature in the cloud-covered area by integrating the spatial auxiliary variable and the conditional generative adversarial network according to claim 1, wherein The formula of the monthly composite land surface temperature interpolated by the generator is as follows: LST G = LST masked × (1 - Mask) + G(LST masked , EVI Monthly , MNDWI monthly , EL, Mask) × Mask; where G is the generator, and LST G represents the land surface temperature interpolated by the generator G, LST masked is the land surface temperature simulating cloud cover, Mask is the cloud mask, and EVI Monthly is the monthly composite EVI, and MNDWI Monthly is the monthly composite MNDWI, and EL is the elevation.

6. The method for interpolating the remote sensing surface temperature in the cloud-covered area by integrating the spatial auxiliary variable and the conditional generative adversarial network according to claim 1, wherein The formula for the discriminator to discriminate between the interpolated monthly composite land surface temperature and the true clear-sky monthly composite land surface temperature is as follows: Class G , Class True = D(LST G ), D(LST true ); where D is the discriminator, and Class G is the land surface temperature LST interpolated by the generator G is the discrimination result by the discriminator, and Class True is the true cloudless land surface temperature LST True is the discrimination result by the discriminator.

7. The method for interpolating the remote sensing surface temperature in cloud-covered areas by integrating spatial auxiliary variables and conditional generative adversarial networks according to claim 6, wherein The process of selecting the weight parameters with the highest accuracy after being tested by the validation set through the adversarial training of the generator and the discriminator includes the following steps: The target loss function Loss of the generator G It consists of adversarial loss and content loss. The adversarial loss uses the MSE loss, and the content loss uses L1 regularization. The specific formula is as follows: Loss G = MSELoss(Class G , 1) + L1Loss(LST G , LST true ); The target loss function Loss of the discriminator D is the adversarial loss, and the specific formula is as follows: Objective Loss of Conditional Generative Adversarial Network CGAN The specific formula is as follows: Loss CGAN = Loss G + Loss D ; Through the adversarial training of the generator and the discriminator, select the weight parameters with the highest accuracy after being tested by the validation set, and then test its accuracy with the test set to obtain the best weight parameters of the conditional generative adversarial network.

8. The method for interpolating the remote sensing surface temperature in the cloud-covered area by integrating the spatial auxiliary variable and the conditional generative adversarial network according to claim 1, wherein The process of using the generator loaded with the best parameters to perform progressive interpolation of the missing values in the cloud-covered area of the daily land surface temperature from the missing edge to the missing center includes the following steps: Taking the boundary pixels of the entire study area as the center, obtain the land surface temperature data, monthly composite EVI, monthly composite MNDWI, elevation, and cloud mask data within the 256×256 grid containing cloud cover. Input the obtained data into the generator loaded with the best parameters for interpolation of cloudy pixels. After interpolation, splice the theoretical clear-sky land surface temperature within the central 192×192 grid into the large image, and then proceed to the next cycle until all cloudy pixels within the entire study area are interpolated.

Images