Cloud coverage area remote sensing surface temperature interpolation method integrating space auxiliary variables and conditional generative adversarial network

Abstract

The invention provides a cloud coverage area remote sensing surface temperature interpolation method integrating space auxiliary variables and a conditional generative adversarial network, and the method comprises the steps: firstly carrying out the CGAN model pre-training of monthly synthetic surface temperature data with better space-time integrity, and enabling the model to consider the characteristics of different regions in a balanced manner in the training process; migration training is carried out based on daily scale data to enable the data to be suitable for daily surface temperature interpolation, a spatial progressive interpolation strategy is adopted in the interpolation process, interpolation is carried out gradually from the cloud coverage edge to the cloud coverage center by considering background information of an interpolation area, and the edge effect is eliminated; therefore, accurate interpolation of the surface temperature of the cloud coverage area is realized.

Description

Technical Field

[0001] The invention relates to the technical field of thermal infrared remote sensing surface temperature, and in particular to a cloud coverage area remote sensing surface temperature interpolation method integrating spatial auxiliary variables and conditional generative adversarial networks. Background Art

[0002] The surface temperature inverted by thermal infrared remote sensing has high accuracy and temporal and spatial resolution, and is widely used in many fields such as meteorology, hydrology, environment, and public health. However, thermal infrared signals cannot penetrate clouds, and about two-thirds of the world's land is covered by clouds every day, resulting in a lot of missing data in the surface temperature data inverted by thermal infrared. Therefore, reconstructing the remote sensing surface temperature in cloud-covered areas is of great significance to enhance its application value.

[0003] At present, the interpolation research on remote sensing land surface temperature can be divided into two categories according to the information source used for interpolation. One is to use the information of cloud-covered areas in the remote sensing land surface temperature data adjacent to cloud-free pixels to interpolate the cloud-covered areas using spatial interpolation or deep learning models such as generative adversarial networks (GAN); the other is to build a model based on the land surface temperature values ​​of cloud-free pixels and environmental variables closely related to land surface temperature (such as vegetation index, elevation, surface albedo, etc.). Since these environmental variables change little over time, a spatial distribution map without gaps can be generated through multi-temporal synthesis, and then the model is applied to the environmental variables in the cloud-covered area to estimate the land surface temperature. However, considering that the land surface temperature has a strong spatiotemporal variability, the filling effect is often not stable enough when the model is transferred to different regions due to the lack of support from background information; and the interpolation effect is not ideal when facing large-scale spatiotemporal missing only by relying on the information of cloud-free pixels adjacent to the land surface temperature cloud area. At present, there are still few studies on interpolating the land surface temperature missing areas based on remote sensing data and combining the adjacent information and related information of the land surface temperature missing areas. At present, the research on surface temperature interpolation using GAN does not use auxiliary variables or uses less auxiliary variable data, and does not make full use of multi-source auxiliary variable information. At the same time, when training GAN, it is necessary to add cloud masks based on spatially continuous and non-missing surface temperatures to construct surface temperature image pairs. Due to the spatiotemporal differences in cloud distribution, the number of spatially continuous image blocks cropped from daily surface temperature data is large in arid areas, while the number is small in cloudy and rainy areas. The unevenness of the data will cause the interpolation effect of the trained model to vary greatly in different regions. In addition, the interpolation research of the deep learning model is to first divide it into small images for interpolation, and then use the interpolated small images to splice into large images. Failure to consider the spatial background information during splicing will lead to boundary effects, resulting in mosaic textures in large cloud-covered areas of interpolation, which leads to incorrect surface temperature textures. Summary of the invention

[0004] The purpose of the present invention is to propose a remote sensing surface temperature interpolation method for cloud-covered areas that integrates spatial auxiliary variables and conditional generative adversarial networks. At the same time, the surface temperature of the clear sky pixels adjacent to the cloud-covered area and auxiliary spatial variables such as vegetation index, water index, and altitude are combined. The conditional generative adversarial network is fully pre-trained on a monthly scale and then migrated to a daily scale. The missing values ​​of the cloud-covered area in the daily surface temperature are progressively interpolated from the missing edge to the missing center, eliminating the edge effect caused by the splicing of the conditional generative adversarial network (CGAN) small image interpolation, and generating daily seamless remote sensing surface temperature.

[0005] In order to achieve the above technical objectives, the technical solution adopted by the present invention is: A method for interpolating remote sensing surface temperature in cloud-covered areas by integrating spatial auxiliary variables and conditional generative adversarial networks, the method comprising the following steps: A monthly-scale interpolation model is constructed based on a conditional generative adversarial network. The monthly-scale surface temperature data and auxiliary interpolated spatial variables are used as pre-training data sets to pre-train the monthly-scale interpolation model. The spatial variables at least include the enhanced vegetation index EVI, the improved normalized water index MNDWI and elevation. The monthly-scale interpolation model includes a generator and a discriminator. The monthly synthetic surface temperature data with simulated cloud cover, the monthly synthetic EVI, the monthly synthetic MNDWI, the elevation and the cloud mask data are used as the input variables of the generator, and the cloudless monthly synthetic surface temperature is used as the target variable of the generator to train the generator. The monthly synthetic surface temperature interpolated by the generator and the monthly synthetic surface temperature data of the real clear sky are used as the input variables of the discriminator, 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 parameter with the highest accuracy tested by the validation set is selected and used as the optimal weight parameter of the monthly-scale interpolation model. The optimal weight parameters are loaded, and the daily surface temperature data and auxiliary interpolation spatial variables are used as migration training data sets. The generator and discriminator are migrated and trained to obtain the optimal generator parameters corresponding to the daily scale. The generator loaded with the optimal parameters is used to progressively interpolate the missing values ​​of the cloud coverage area in the daily surface temperature from the missing edge to the missing center.

[0006] Furthermore, the process of obtaining training data includes the following steps: Collect daily surface temperature data and spatial variables for auxiliary interpolation, the spatial variables at least include enhanced vegetation index EVI, modified normalized difference water index MNDWI and elevation; extract missing value areas from daily surface temperature data as cloud mask data; Perform monthly composite cloud removal on daily surface temperature data, EVI and MNDWI to obtain monthly composite surface temperature, monthly composite EVI and monthly composite MNDWI. The daily surface temperature, monthly composite surface temperature, monthly composite EVI, monthly composite MNDWI, and elevation data are cropped into sub-images according to the preset overlap ratio; then the cloud mask data are randomly extracted to perform cloud mask processing on the monthly composite surface temperature to generate monthly composite surface temperature data with simulated cloud cover; The cropped sub-images are combined according to the imaging time and imaging area and divided into daily scale data and monthly scale data. The daily scale data and monthly scale data are divided into training set, validation set and test set respectively according to preset proportions.

[0007] Furthermore, the enhanced vegetation index EVI is calculated using the following formula: ; Where EVI is the enhanced vegetation index, is the reflectivity in the near-infrared band, is the reflectivity in the red light band, is the reflectivity of the blue light band; The improved normalized water index MNDWI is calculated using the following formula: ; Where, MNDWI is the improved normalized water index, is the reflectivity of the green light band, is the reflectivity of the shortwave infrared 1 band; Extract elevation data from digital elevation model data.

[0008] Furthermore, the monthly scale data is used as a pre-training data set to pre-train the monthly scale interpolation model. The target loss of the 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. The formula of the MSE loss function is as follows: ; In the formula for and The MSE loss between is the predicted value, is the target value, and N represents the total number; The formula of L1 loss function is as follows: ; In the formula for and L1 loss between them.

[0009] Furthermore, the formula for the monthly synthetic surface temperature interpolated by the generator is as follows: ; Where G is the generator, represents the surface temperature interpolated by generator G, To simulate the surface temperature of cloud cover, Mask is the cloud mask. Synthesize EVI for the month, is the monthly composite MNDWI and EL is the elevation.

[0010] Furthermore, the discriminator uses the following formula to discriminate the interpolated monthly synthetic surface temperature and the real clear sky monthly synthetic surface temperature: ; Where D is the discriminator, Interpolated surface temperature for the generator After the discriminator's judgment result, is the true cloud-free surface temperature The discrimination result of the discriminator.

[0011] Furthermore, through adversarial training of the generator and the discriminator, the process of selecting the weight parameter with the highest accuracy tested on the validation set includes the following steps: The objective loss function of the generator It consists of adversarial loss and content loss. The adversarial loss uses MSE loss and the content loss uses L1 regularization. The specific formula is as follows: ; The objective loss function of the discriminator To combat the loss, the specific formula is as follows: ; Target loss for conditional generative adversarial networks The specific formula is as follows: ; Through adversarial training of the generator and the discriminator, the weight parameters with the highest accuracy tested on the validation set are selected, and then their accuracy is tested on the test set to obtain the optimal weight parameters of the conditional generative adversarial network.

[0012] Furthermore, the process of using the generator loaded with the optimal parameters to progressively interpolate the missing values ​​of the cloud coverage area in the daily 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, the surface temperature data, monthly composite EVI, monthly composite MNDWI, elevation and cloud mask data in the 256×256 grid containing cloud cover are obtained, and these data are input into the generator loaded with the optimal parameters to interpolate the cloud pixels. After the interpolation is completed, the theoretical clear sky surface temperature in the central 192×192 grid is spliced ​​into the large picture, and then the next cycle is performed until all the cloud pixels in the entire study area are interpolated.

[0013] Compared with the prior art, the present invention has the following beneficial effects: First, the remote sensing surface temperature interpolation method for cloud-covered areas that integrates spatial auxiliary variables and conditional generative adversarial networks in the present invention combines multi-source spatial auxiliary information and CGAN to interpolate the surface temperature of cloud coverage, taking into account the advantages of the CGAN model in utilizing spatial neighborhood contextual information and the ability of related environmental variables to characterize the surface characteristics of cloud-covered areas.

[0014] Second, the remote sensing surface temperature interpolation method for cloud-covered areas that integrates spatial auxiliary variables and conditional generative adversarial networks of the present invention first uses monthly scale data to pre-train the model, and then uses daily scale data for migration training, ensuring sufficient training samples and that the interpolation effect will not have too large differences in drought and cloudy areas.

[0015] Third, the remote sensing surface temperature interpolation method for cloud-covered areas that integrates spatial auxiliary variables and conditional generative adversarial networks of the present invention adopts a spatial progressive method to gradually interpolate from the edge of cloud coverage to the center of cloud coverage, taking into account the background surface temperature around the large image, and eliminating the edge effect caused by splicing small images into a large image. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flow chart of a method for interpolating remotely sensed surface temperature in a cloud-covered area using an integrated spatial auxiliary variable and conditional generation adversarial network according to the present invention; Figure 2 It is a diagram of the spatial progressive interpolation process; Figure 3 This is a typical 256×256 region interpolation model training diagram; Figure 4 Schematic diagram of the surface temperature interpolation results for the entire region. DETAILED DESCRIPTION

[0017] The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0018] The present invention discloses a cloud coverage area remote sensing surface temperature interpolation method integrating spatial auxiliary variables and conditional generative adversarial networks, the method comprising the following steps: A monthly-scale interpolation model is constructed based on a conditional generative adversarial network. The monthly-scale surface temperature data and auxiliary interpolated spatial variables are used as pre-training data sets to pre-train the monthly-scale interpolation model. The spatial variables at least include the enhanced vegetation index EVI, the improved normalized water index MNDWI and elevation. The monthly-scale interpolation model includes a generator and a discriminator. The monthly synthetic surface temperature data with simulated cloud cover, the monthly synthetic EVI, the monthly synthetic MNDWI, the elevation and the cloud mask data are used as the input variables of the generator, and the cloudless monthly synthetic surface temperature is used as the target variable of the generator to train the generator. The monthly synthetic surface temperature interpolated by the generator and the monthly synthetic surface temperature data of the real clear sky are used as the input variables of the discriminator, 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 parameter with the highest accuracy tested by the validation set is selected and used as the optimal weight parameter of the monthly-scale interpolation model. The optimal weight parameters are loaded, and the daily surface temperature data and auxiliary interpolation spatial variables are used as migration training data sets. The generator and discriminator are migrated and trained to obtain the optimal generator parameters corresponding to the daily scale. The generator loaded with the optimal parameters is used to progressively interpolate the missing values ​​of the cloud coverage area in the daily surface temperature from the missing edge to the missing center.

[0019] The purpose of this invention is to propose a method for interpolating remote sensing surface temperature in cloud-covered areas based on multi-source spatial auxiliary variables and deep learning models. Based on the conditional generative adversarial network model (CGAN), this invention combines the clear sky surface temperature and auxiliary spatial variables such as vegetation index, water index, and altitude adjacent to the remote sensing surface temperature cloud-covered area to progressively interpolate the missing values ​​of the surface temperature in the cloud-covered area and generate seamless remote sensing surface temperature.

[0020] 1) Spatial variable extraction and calculation The spatial variables used for auxiliary interpolation in the present invention include enhanced vegetation index EVI, improved normalized water index MNDWI and elevation.

[0021] The formula for calculating EVI is as follows: (1); Where EVI is the enhanced vegetation index, is the reflectivity in the near-infrared band, is the reflectivity in the red light band, is the reflectivity of the blue light band.

[0022] The formula for calculating MNDWI is as follows: (2); Where, MNDWI is the improved normalized water index, is the reflectivity of the green light band, is the reflectivity of the shortwave infrared 1 band.

[0023] Elevation data is extracted from digital elevation model data.

[0024] Extract missing value areas, i.e., cloud mask data, from daily surface temperature data.

[0025] 2) Monthly synthesis processing In order to obtain large-area continuous cloud-free surface temperature data for training the CGAN model, variables such as surface temperature, EVI and MNDWI are processed by monthly synthesis and cloud removal.

[0026] The calculation formula for the monthly average composite is as follows: (3); In the formula It can represent the monthly mean composite land surface temperature (LST), EVI and MNDWI, where n represents the number of days in each month. It can represent daily LST, EVI and MNDWI. represents the daily cloud mask data, Time represents cloud interference, Time represents clear sky.

[0027] 3) Preparation of deep learning dataset The daily surface temperature, monthly composite surface temperature, monthly composite EVI, monthly composite MNDWI, and elevation data were cropped into 256 × 256 small images according to a certain overlap ratio. Then, the cloud mask data were randomly extracted to perform cloud mask processing on the monthly scale surface temperature to generate monthly composite surface temperature data with simulated cloud cover.

[0028] The 256×256 size images are combined according to the imaging time and imaging area, and are divided into daily scale data and monthly scale data. The difference between the two is whether the surface temperature is synthesized monthly, and other data remain unchanged.

[0029] Then, the daily scale data and monthly scale data are divided into training set, validation set and test set respectively according to the ratio of 7:2:1.

[0030] 4) Monthly scale interpolation model pre-training The monthly scale data is used as the pre-training dataset to pre-train the model. The target loss of the 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.

[0031] The formula of the MSE loss function is as follows: (4); In the formula for and The MSE loss between is the predicted value, is the target value, and N represents the total number.

[0032] The formula of L1 loss function is as follows: (5); In the formula for and L1 loss between is the predicted value, is the target value, and N represents the total number.

[0033] The synthetic monthly surface temperature, EVI, MNDWI, elevation and cloud mask of simulated cloud cover are used as the input variables of the generator, and the synthetic monthly surface temperature without cloud is used as the target variable of the generator to train the generator. It is divided into training set, validation set and test set in a ratio of 7:2:1. The surface temperature interpolated by the generator and the surface temperature data of cloudless sky are used as the input variables of the discriminator respectively. The label of the interpolated surface temperature is set to 0, and the label of the cloudless surface temperature is set to 1. The label is used as the target variable of the discriminator to train the discriminator.

[0034] The formula for the surface temperature interpolated by the generator is as follows: (6); Where G is the generator, represents the surface temperature interpolated by generator G, To simulate the surface temperature of cloud cover, Mask is the cloud mask. Synthesize EVI for the month, is the monthly composite MNDWI and EL is the elevation.

[0035] The discriminator uses the following formula to distinguish the interpolated surface temperature from the real clear sky surface temperature: (7); Where D is the discriminator, Interpolated surface temperature for the generator After the discriminator's judgment result, is the true cloud-free surface temperature The discriminant result of the discriminator is shown in the present invention as follows: In the present invention, the label of the real surface temperature is represented by 1, and the label of the interpolated surface temperature is represented by 0.

[0036] The objective loss function of the generator It is composed of adversarial loss and content loss. The adversarial loss uses MSE loss and the content loss is L1 regularization. The specific formula is as follows: (8); The objective loss function of the discriminator To combat the loss, the specific formula is as follows: (9); CGAN’s target loss , the specific formula is as follows: (10); Through adversarial training of the generator and the discriminator, the weight parameter with the highest accuracy tested on the validation set is selected, and then the test set is used to test whether its accuracy fluctuates too much. If the accuracy evaluation in the test set is similar to that in the validation set, it is used as the weight parameter of the optimal CGAN.

[0037] 5) Migration training from monthly scale to daily scale First, load the weights of the best CGAN model obtained by pre-training, then use the daily scale data as the transfer training data set, adjust some hyperparameters, and perform transfer training on the generator and discriminator. Then select the best generator weight parameters based on the test results of the validation set, and then use the test set to test whether its accuracy fluctuates too much. If the accuracy evaluation in the test set is similar to that in the validation set, it is used as the best generator parameter for the interpolation model.

[0038] 6) Daily surface temperature and cloud cover area spatial progressive interpolation First, the optimal weight parameters of the generator are loaded, and then the surface temperature of the entire study area is detected row by row and column by column for pixels with cloud coverage and at the boundary between the missing area and the complete area. With the boundary pixel as the center, the surface temperature data, monthly synthetic EVI, monthly synthetic MNDWI, elevation and cloud mask data in the 256×256 grid with cloud coverage are obtained. Then these data are input into the optimal generator for interpolation of cloud pixels. After the interpolation is completed, the theoretical clear sky surface temperature in the central 192×192 grid is spliced ​​into the large image, and the next cycle is performed until all cloud pixels in the large image are interpolated.

[0039] Examples The present invention interpolates the surface temperature of cloud cover based on multi-source spatial auxiliary variables and conditional generative adversarial networks. The following are the specific implementation steps. Figure 1 .

[0040] 1) Based on the MOD11A1 surface temperature product, quality control is performed, and the pixels with cloud interference are assigned NAN values ​​to obtain the daily surface temperature.

[0041] 2) Binarize the daily surface temperature, assign 1 to NAN value pixels and 0 to non-NAN value pixels, and obtain cloud mask data.

[0042] 3) Perform quality control on the MOD09GA surface reflectance product to obtain pixels without cloud interference, and then calculate EVI and MNDWI. The calculations are shown in formulas (1) and (2).

[0043] 4) Extract elevation data from GTOPO30DEM products.

[0044] 5) The MOD11A1 surface temperature data, EVI and MNDWI data are synthesized by monthly mean to obtain cloud-free complete data. The calculation is shown in formula (3). Then, the monthly synthesized EVI and monthly synthesized MNDWI are resampled to 1000m to be consistent with the MODIS surface temperature.

[0045] 6) The daily surface temperature data, monthly composite surface temperature data, monthly composite EVI, monthly composite MNDWI, elevation and cloud mask data are cropped into 256×256 images according to a certain overlap ratio.

[0046] 7) Cloud mask data are randomly extracted from the cloud mask dataset to perform cloud mask operation on the cropped monthly-scale surface temperature data. The value in the cloud-covered area is -50, and the values ​​in other areas remain unchanged.

[0047] 8) Prepare the dataset for model training. Combine the simulated cloud-covered surface temperature, cloud-free surface temperature, EVI, MNDWI, elevation, and cloud mask data based on time and location.

[0048] 9) Use monthly data to pre-train the interpolation model. The hyperparameters for model pre-training are set as follows: Epoch is 100, batch size is 12. Adam is used as the optimizer, and the learning rate change strategy is to first use a smaller learning rate for preheating, and then use cosine annealing to gradually reduce 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 for stopping preheating is 30. The above hyperparameters are consistent in the discriminator and the generator. The adversarial loss uses MSE loss, see formula (4), and the content loss uses L1 loss, see formula (5).

[0049] The training set, validation set, and test set are divided into a ratio of 7:2:1. In the model training process, the synthetic surface temperature of the month with simulated cloud cover, EVI, MNDWI, elevation, and cloud mask data are used as the input variables of the generator of the conditional generative adversarial network, and the synthetic surface temperature of the cloudless month is used as the target variable of the generator. The generator is interpolated and trained. The surface temperature interpolated by the generator is shown in formula (6); the surface temperature interpolated by the generator and the synthetic surface temperature of the cloudless month are respectively input into the discriminator for discriminant training. The discriminant result of the discriminator is shown in formula (7).

[0050] The target loss function when training the generator consists of the adversarial loss function and the content loss function, see formula (8); the target loss function when training the discriminator does not include the adversarial loss function, see formula (9). Among them, the target loss function of CGAN is shown in formula (10).

[0051] After each epoch is completed, the trained weight parameters are saved, and the accuracy of each epoch is tested with the validation set. After the model is trained, the epoch weight parameters with the highest accuracy are selected based on the results of the validation set test, and the test set is used to test whether there is excessive fluctuation in its accuracy. If the accuracy of the test set is slightly different from that of the validation set, this parameter is used as the initial parameter for model migration training.

[0052] 10) The monthly pre-trained model is transferred and trained with daily data. The hyperparameters of the model pre-training are as follows: 50 epochs, 12 batch sizes, SGD as the optimizer, and a fixed learning rate of 0.0001. These hyperparameters are consistent in the discriminator and the generator. MSE loss is used as the adversarial loss, and L1 regularization is used as the regularization method. Transfer training only modifies the number of training epochs, optimizer, and learning rate, and the rest of the hyperparameters remain unchanged.

[0053] First, load the weight parameters trained to the highest accuracy in the pre-trained model. Then use the simulated cloud-covered daily-scale surface temperature, monthly synthetic EVI, MNDWI, elevation, and cloud mask data as the target variable of the generator, and the cloud-free daily-scale surface temperature as the target variable of the generator. Use the daily-scale data to perform migration training on the generator from monthly-scale data to daily-scale data. The interpolated surface temperature of the generator and the cloud-free daily-scale surface temperature are respectively input into the discriminator for discriminator migration training.

[0054] After the transfer training is completed, the generator weight parameters with the best training effect are selected in the same way as the pre-training optimal parameter selection. Figure 3The pre-training interpolation and transfer training interpolation results of several typical 256×256 regional surface temperatures are given. According to the accuracy test, the R 2 (coefficient of determination) is 0.96, MAE (mean decision error) is 1.16, and RMSE (root mean square error) is 1.68; after transfer training, MAE becomes 1.08 and RMSE becomes 1.58.

[0055] 11) Before interpolating the surface temperature of the cloud-covered area, the surface temperature of all pixels with cloud interference is assigned a value of -50. Next, the surface temperature data of the area is binarized, and the pixels with a value of -50 are marked as 1, and the remaining pixels are marked as 0, as a cloud mask.

[0056] Then, a 3×3 convolution summation operation is performed on each pixel in the cloud mask. After the convolution process, pixels with values ​​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.

[0057] Next, these boundary pixels are retrieved pixel by pixel in row and column order. When a boundary pixel is detected, a 256×256 region centered on the pixel is selected to extract the surface temperature, EVI, MNDWI, elevation, and cloud mask data within the region.

[0058] Finally, the extracted data is input into the optimal generator to generate the interpolation result in the 256×256 area. The pixels in the 192×192 range in the center of the area are selected and inserted into the corresponding position. Repeat this process (the process is as follows Figure 2 ), until all the areas to be interpolated are successfully interpolated, and the interpolated seamless surface temperature of the entire study area is obtained (as shown in Figure 4 as shown).

[0059] Although the preferred embodiments of the present application 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 falling within the scope of the present application.

[0060] Obviously, those skilled in the art can make various changes and modifications 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 equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for interpolating remote sensing surface temperature in cloud-covered areas by integrating spatial auxiliary variables and conditional generative adversarial networks, characterized in that: The method comprises the following steps: A monthly-scale interpolation model is constructed based on a conditional generative adversarial network. The monthly-scale surface temperature data and auxiliary interpolated spatial variables are used as pre-training data sets to pre-train the monthly-scale interpolation model. The spatial variables at least include the enhanced vegetation index EVI, the improved normalized water index MNDWI and elevation. The monthly-scale interpolation model includes a generator and a discriminator. The monthly synthetic surface temperature data with simulated cloud cover, the monthly synthetic EVI, the monthly synthetic MNDWI, the elevation and the cloud mask data are used as the input variables of the generator, and the cloudless monthly synthetic surface temperature is used as the target variable of the generator to train the generator. The monthly synthetic surface temperature interpolated by the generator and the monthly synthetic surface temperature data of the real clear sky are used as the input variables of the discriminator, 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 parameter with the highest accuracy tested by the validation set is selected and used as the optimal weight parameter of the monthly-scale interpolation model. The optimal weight parameters are loaded, and the daily surface temperature data and auxiliary interpolation spatial variables are used as migration training data sets. The generator and discriminator are migrated and trained to obtain the optimal generator parameters corresponding to the daily scale. The generator loaded with the optimal parameters is used to progressively interpolate the missing values ​​of the cloud coverage area in the daily surface temperature from the missing edge to the missing center.

2. The cloud coverage area remote sensing surface temperature interpolation method integrating spatial auxiliary variables and conditional generative adversarial networks according to claim 1 is characterized in that: The process of obtaining training data includes the following steps: Collect daily surface temperature data and spatial variables for auxiliary interpolation, the spatial variables at least include enhanced vegetation index EVI, modified normalized difference water index MNDWI and elevation; extract missing value areas from daily surface temperature data as cloud mask data; Perform monthly composite cloud removal on daily surface temperature data, EVI and MNDWI to obtain monthly composite surface temperature, monthly composite EVI and monthly composite MNDWI. The daily surface temperature, monthly composite surface temperature, monthly composite EVI, monthly composite MNDWI, and elevation data are cropped into sub-images according to the preset overlap ratio; then the cloud mask data are randomly extracted to perform cloud mask processing on the monthly composite surface temperature to generate monthly composite surface temperature data with simulated cloud cover; The cropped sub-images are combined according to the imaging time and imaging area and divided into daily scale data and monthly scale data. The daily scale data and monthly scale data are divided into training set, validation set and test set respectively according to preset proportions.

3. The cloud coverage area remote sensing surface temperature interpolation method integrating spatial auxiliary variables and conditional generative adversarial networks according to claim 1 is characterized in that: The enhanced vegetation index (EVI) is calculated using the following formula: ； Where EVI is the enhanced vegetation index, is the reflectivity in the near-infrared band, is the reflectivity in the red light band, is the reflectivity of the blue light band; The improved normalized water index MNDWI is calculated using the following formula: ； Where, MNDWI is the improved normalized water index, is the reflectivity of the green light band, is the reflectivity of the shortwave infrared 1 band; Extract elevation data from digital elevation model data.

4. The cloud coverage area remote sensing surface temperature interpolation method integrating spatial auxiliary variables and conditional generative adversarial networks according to claim 1 is characterized in that: The monthly scale data is used as the pre-training data set to pre-train the monthly scale interpolation model. The target loss of the 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. The formula of the MSE loss function is as follows: ； In the formula for and The MSE loss between is the predicted value, is the target value, and N represents the total number; The formula of L1 loss function is as follows: ； In the formula for and The L1 loss between .

5. The cloud coverage area remote sensing surface temperature interpolation method integrating spatial auxiliary variables and conditional generative adversarial networks according to claim 1 is characterized in that: The formula for the monthly composite surface temperature interpolated by the generator is as follows: ； Where G is the generator, represents the surface temperature interpolated by generator G, To simulate the surface temperature of cloud cover, Mask is the cloud mask. Synthesize EVI for the month, is the monthly composite MNDWI and EL is the elevation.

6. The cloud coverage area remote sensing surface temperature interpolation method integrating spatial auxiliary variables and conditional generative adversarial networks according to claim 1 is characterized in that: The discriminator uses the following formula to distinguish the interpolated monthly synthetic surface temperature from the real clear sky monthly synthetic surface temperature: ； Where D is the discriminator, Interpolated surface temperature for the generator After the discriminator's judgment result, is the true cloud-free surface temperature The discrimination result of the discriminator.

7. The cloud coverage area remote sensing surface temperature interpolation method integrating spatial auxiliary variables and conditional generative adversarial networks according to claim 1 is characterized in that: Through adversarial training of the generator and the discriminator, the process of selecting the weight parameter with the highest accuracy tested on the validation set includes the following steps: The objective loss function of the generator It consists of adversarial loss and content loss. The adversarial loss uses MSE loss and the content loss uses L1 regularization. The specific formula is as follows: ； The objective loss function of the discriminator To combat the loss, the specific formula is as follows: ； Target loss for conditional generative adversarial networks The specific formula is as follows: ； Through adversarial training of the generator and the discriminator, the weight parameters with the highest accuracy tested on the validation set are selected, and then their accuracy is tested on the test set to obtain the optimal weight parameters of the conditional generative adversarial network.

8. The cloud coverage area remote sensing surface temperature interpolation method integrating spatial auxiliary variables and conditional generative adversarial networks according to claim 1 is characterized in that: The process of using the generator loaded with the optimal parameters to progressively interpolate the missing values ​​of the cloud coverage area in the daily surface temperature from the missing edge to the missing center includes the following steps: Taking the boundary pixel of the entire study area as the center, the surface temperature data, monthly composite EVI, monthly composite MNDWI, elevation and cloud mask data in the 256×256 grid containing cloud cover are obtained, and these data are input into the generator loaded with the optimal parameters to interpolate the cloud pixels. After the interpolation is completed, the theoretical clear sky surface temperature in the central 192×192 grid is spliced ​​into the large picture, and then the next cycle is carried out until all the cloud pixels in the entire study area are interpolated.

Images