A soil moisture product downscaling method based on two-dimensional convolutional neural network
Through the soil moisture reduction method based on two-dimensional convolutional neural network, the problem of information loss in traditional methods is solved, the accuracy of soil moisture products is significantly improved, and the acquisition of soil moisture data with higher resolution is achieved.
Patent Information
- Application Number
- CN202011622851.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-12-31
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2040-12-31
AI Technical Summary
Traditional soil moisture reduction method cannot effectively consider the information between cells in a scene image, resulting in information loss and it is difficult to meet the needs of refined applications.
Using a downscale method based on two-dimensional convolutional neural network (CNN), the corresponding relationship between auxiliary data and passive microwave soil moisture product images is constructed, training data is established, and training is used to improve spatial resolution using CNN network.
The accuracy of soil moisture products is significantly improved, and the average correlation coefficient, root mean square error and unbiased root mean square error are increased by 0.02-0.05 respectively, meeting the application needs of higher resolution.
Smart Images

Figure CN114694033B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of remote sensing technology, and in particular relates to a soil moisture product downscaling method based on a two-dimensional convolutional neural network. Background Art
[0002] Soil moisture (SM) is an important parameter in hydrology, meteorology, and agricultural applications. Its temporal and spatial distribution plays an important role in studying climate change, agricultural drought, and water resources management. The traditional method of monitoring SM through stations is difficult to obtain SM data over a large area, and its application is very limited. However, researchers have found that microwave remote sensing data has a good relationship with SM, so they proposed to use microwave remote sensing data to invert SM.
[0003] So far, great breakthroughs have been made in the research of SM inversion using passive microwave remote sensing data, and corresponding SM products have been released one after another, such as SMAP (Soil Moisture Active Passive) and SMOS (Soil Moisture and Ocean Salinity) global SM products. However, the spatial resolution of these products is mostly at the level of tens of kilometers, which is difficult to meet more sophisticated application requirements. Therefore, these products need to be downscaled to improve their spatial resolution.
[0004] There have been many studies on SM downscaling methods. In summary, they can be divided into: (1) downscaling methods based on empirical models; (2) downscaling methods based on physical models; (3) downscaling methods based on spatial interpolation; and (4) downscaling methods based on machine learning. Although methods (1)-(3) can also introduce new information through auxiliary data, they generally use less auxiliary data and rely on the one-to-one correspondence between a single pixel in the auxiliary data image and a single pixel in the target image after downscaling, that is, the "pixel-pixel" concept. Although method (4) can use machine learning methods and add a large amount of auxiliary data for training, the machine learning model currently used is still only based on the "pixel-pixel" concept, ignoring the internal connection between pixels in the same scene image. But obviously, each pixel in a scene image is not independent of each other. For example, due to the influence of slope and aspect, the solar energy received by the surface is different, thus affecting the evapotranspiration of the surface; at the same time, the height difference of the surface will also affect the distribution of soil particles and the infiltration of soil water. Different soil particles have different adsorption effects on SM, which will further lead to the spatial variation of SM. Therefore, the traditional downscaling method essentially loses some useful information.
[0005] Two-dimensional convolutional neural networks (CNN) were originally used for image recognition and classification. They can set a receptive field (i.e., convolution kernel) of a certain size (e.g., 3*3, 5*5, 7*7, etc.) to learn the entire image. In this way, not only can the information contained in each pixel be learned, but also the interaction between pixels in each receptive field can be considered, which can make up for the shortcomings of the traditional SM downscaling method. However, two-dimensional CNN is difficult to apply to the field of remote sensing inversion. The main reason is that the two-dimensional CNN network needs to use a large amount of spatially continuous data for training, and such data is difficult to obtain in remote sensing.
[0006] Therefore, the present invention aims to address the shortcomings of the traditional SM downscaling method that cannot consider the information between pixels in a scene image, and fully considers the solution to the problem of applying two-dimensional CNN to remote sensing inversion. A new SM downscaling method based on two-dimensional CNN technology is proposed, and the SMAP SM product is used as an example to illustrate it.
[0007] A search of domestic and foreign patents and literature shows that no similar downscaling method has been invented or published yet. Summary of the invention
[0008] The present invention aims at the deficiency that the above-mentioned traditional SM downscaling method cannot consider the information between pixels in an image, and fully considers the solution to the problem of applying two-dimensional CNN to remote sensing inversion, and proposes a SM downscaling method based on two-dimensional CNN.
[0009] The technical solution of the present invention is:
[0010] A soil moisture product downscaling method based on a two-dimensional convolutional neural network comprises the following steps:
[0011] 1) Obtain the passive microwave soil moisture product image and the corresponding auxiliary data image of the target area; and establish the corresponding relationship between the auxiliary data image and the passive microwave soil moisture product image;
[0012] 2) constructing training data according to the corresponding relationship; and then using the training data to train the CNN network;
[0013] 3) Use the trained CNN network to downscale the passive microwave soil moisture product image to be processed in the target area to obtain the target downscaled product.
[0014] Furthermore, the method for establishing the correspondence between the auxiliary data image and the passive microwave soil moisture product image is as follows: first, the auxiliary data image is resampled to the spatial resolution of the passive microwave soil moisture product image so that the pixels of the auxiliary data image correspond one-to-one with the pixels of the passive microwave soil moisture product image; a window of size m×m is set to crop the auxiliary data image, where m is an odd number; then each window of the auxiliary data image is taken as a "picture", and the passive microwave soil moisture product image corresponding to the central pixel of the "picture" is taken as a "pixel", thereby establishing a "picture-pixel" relationship as the corresponding relationship.
[0015] Furthermore, the larger the spatial variability of the area, the larger the window used for the passive microwave soil moisture product image; for the pixels at the edge of the auxiliary data image, new pixels are introduced around the edge of the auxiliary data image, and the values of the new pixels are set to the average value of all pixels in the auxiliary data image; then the pixels at the edge of the auxiliary data image are cropped.
[0016] Furthermore, a weight layer is set for the auxiliary data image after the new pixel is introduced, in which the original pixel weight of the auxiliary data image is set to H, and the weight of the introduced new pixel is h; wherein H is greater than h; and then the weight layer is combined with the auxiliary data image after the new pixel is introduced as a feature layer.
[0017] Furthermore, the method for constructing the training data is: determining a training data according to each "image-pixel" relationship, wherein the "pixel" in the training data is a label, and the feature layer corresponding to the "image" is used as feature data.
[0018] Furthermore, the method for constructing the training data is:
[0019] 11) Down-sampling the passive microwave soil moisture product image corresponding to the “pixel” in the training data determined according to each “image-pixel” relationship to obtain a low-resolution passive microwave soil moisture product image;
[0020] 12) Copy and expand each pixel in the low-resolution passive microwave soil moisture product image to obtain an image with the same resolution as the passive microwave soil moisture product image;
[0021] 13) Subtract the passive microwave soil moisture product image corresponding to the “pixel” from the image obtained in step 12) to obtain a ΔSM image, which is used as a label for training data;
[0022] 14) resampling the auxiliary data images corresponding to the “image” in the training data determined by the “image-pixel” relationship respectively to obtain the auxiliary data image T at the resolution of the passive microwave soil moisture product image and the auxiliary data image t at the low resolution;
[0023] 15) Copying and expanding the auxiliary data image t so that its resolution is the same as that of the auxiliary data image T;
[0024] 16) Weights are set for the pixels in the auxiliary data image t processed in step 15) to generate corresponding feature layers as feature data.
[0025] Furthermore, the n passive microwave soil moisture product images adjacent to and before the i-th passive microwave soil moisture product image are used to eliminate null value points of the i-th passive microwave soil moisture product image.
[0026] Furthermore, the value of n is determined based on how fast the passive microwave soil moisture product image changes over time; for each null point in the i-th passive microwave soil moisture product image, the average value of the pixels corresponding to the null point position in the n passive microwave soil moisture product images is taken as the pixel value of the null point position in the i-th passive microwave soil moisture product image.
[0027] Furthermore, in the process of training the CNN network using the training data, when the passive microwave soil moisture product image corresponding to the area with greater spatial variability is used as input, the CNN convolution kernel is set to be larger.
[0028] Furthermore, the auxiliary data include DEM data provided by SRTM, reflectivity data provided by MODIS and surface temperature product data; the method for obtaining the target downscaled product is: using the trained CNN network to predict the passive microwave soil moisture product image to be processed in the target area, and then according to SM D =SM O +SM CNN Get the target downscaling product SM D ; Among them, SM o represents the passive microwave soil moisture product image to be processed, SM CNN Represents the difference image ΔSM predicted by the CNN network.
[0029] The steps of the SM downscaling method based on two-dimensional CNN of the present invention are as follows:
[0030] S1. Download the passive microwave SM product images and corresponding auxiliary data images of the study area.
[0031] S2. All data were synthesized for 6 days to eliminate null points in the image.
[0032] S3. Use the data products obtained after 6 days of synthesis processing to establish the correspondence between the auxiliary data images and the SM images.
[0033] S4. Use the corresponding relationship established in S3 to construct corresponding training data.
[0034] S5. Build a CNN network and bring the training data obtained in S4 into the network for training.
[0035] S6. Use the trained network to perform SM downscaling to obtain the target downscaled product.
[0036] S7. Verify the accuracy of the downscaled products.
[0037] Furthermore, in step S1, the auxiliary data includes the digital elevation model (DEM) data provided by SRTM (Shuttle Radar Topography Mission), the reflectance data provided by the Moderate-resolution Imaging Spectroradiometer (MODIS), and the surface temperature product data. In addition, all the data need to be preprocessed to obtain the GeoTIFF format remote sensing image under the standard plane projection.
[0038] Furthermore, in step S2, different data synthesis methods are used for different data according to how fast they change over time. For example, for data that changes quickly over time, such as SM, average value synthesis is used, that is, for a specific pixel, the average value of SM of this pixel in 6 days is taken as the final synthesis value of 6 days. For data that changes slowly over time, such as vegetation reflectance, NDVI (Normalized Difference Vegetation Index) maximum value synthesis is used, that is, for a specific pixel, the day with the maximum NDVI value in 6 days is taken as the final synthesis value of 6 days.
[0039] Furthermore, in step S3, for establishing the correspondence between the auxiliary data image and the SM image, in order to distinguish it from the "pixel-to-pixel" concept used in traditional methods, the present invention adopts an "image-to-pixel" method, and in order to improve the accuracy, a weight layer is introduced in the auxiliary data.
[0040] First, the method of establishing the “image-pixel” relationship is as follows Figure 1 As shown:
[0041] First, the auxiliary data synthesized over 6 days needs to be resampled to the spatial resolution of the SM product so that the pixels of the auxiliary image can correspond one-to-one with the pixels of the SM image. Then a window of size m×m is designed ( Figure 1The window size in is 3×3), which is used to crop the auxiliary product image to obtain multiple windows; each window of the auxiliary data image is taken as a "picture", and the passive microwave soil moisture product image corresponding to the central pixel of the "picture" is taken as a "pixel" to establish the "picture-pixel" relationship as a corresponding relationship. Among them, m must be an odd number, and its size can be determined based on experience. For example, for plain areas with uniform surface, m can be set to 3 or 5, and for mountainous areas with large spatial variability, m can be set to 7 or 9; multiple groups of m can also be set, and finally the optimal value is determined by comparing with the verification data. In addition, the SM "pixel" corresponding to the central pixel of the auxiliary "picture" is taken and connected with the auxiliary "picture" to establish the "picture-pixel" relationship. As Figure 1 As shown in the left figure in , a window is composed of 9 pixels, 1, 2, 3, 5, 6, 7, 9, 10, and 11, as the "image". Pixel 6 is the central pixel of the "image", and the SM corresponding to pixel 6 is the "pixel". In this way, the "image-pixel" relationship is established as the corresponding relationship, so that the corresponding relationship between a 3×3 auxiliary "image" and the SM "pixel" can be determined.
[0042] For pixels at the edge of the image, such as pixel 1, the above cropping method cannot be used because there are no other pixels on its left and top. In this case, the auxiliary product image needs to be expanded, that is, new pixels are introduced around the auxiliary product image, such as Figure 1 As shown in the right figure of . The values of these pixels are set to the average value of all pixels in the auxiliary product image. In order to allow the outermost pixels to be cropped into the corresponding m×m window, the expanded width L should satisfy:
[0043]
[0044] Secondly, the method of introducing the weight layer is as follows Figure 2 shown.
[0045] According to the above, the “image-pixel” method will introduce extended pixels around the auxiliary image that do not actually exist. In order to weaken the influence of these extended pixels, the present invention also introduces a weight layer. Figure 2 As shown in FIG. 1 , the size of the weight layer is the same as the size of the auxiliary image after expansion, and the weight of the original pixel is set to 0.95, and the weight of the expanded pixel is set to 0.05. Next, the weight layer is combined with the auxiliary image to form a feature layer.
[0046] For convenience, in the following description, the SM product will be referred to as a label, and the feature layer formed by combining the weight layer and the auxiliary image will be referred to as a feature.
[0047] Furthermore, in step S4, the construction of training data is based on the “image-pixel” relationship. Figure 2 As shown in the lower right corner, after the feature and label images are determined, the "image-pixel" relationship established in step S3 can be used to obtain a corresponding auxiliary feature "image" for each SM "pixel". The "image" is used as the input of CNN and the "pixel" is used as the output of CNN, so each set of "image-pixel" constitutes a set of training data for CNN.
[0048] Furthermore, in step S5, the structure of CNN is as follows: Figure 3 The settings of the main parameters are shown in Table 1.
[0049] Table 1 shows the main parameters of CNN
[0050]
[0051] *R represents the resolution of the feature image (km)
[0052] Moreover, the size of the CNN convolution kernel can be changed according to different terrain conditions. For example, for plain areas with relatively uniform surface, the convolution kernel can be set to a smaller 3×3, while for mountainous areas with large spatial variability, the convolution kernel can be set to 5×5 or 7×7 to obtain more information between pixels in the same image. At the same time, in order to improve the accuracy of the downscaling results, the present invention trains and predicts the residual error (ΔSM) of SM. On this basis, the method for generating training data is as follows: Figure 4 As shown:
[0053] In order to obtain ΔSM, the present invention uses two sets of SM images with different resolutions, such as Figure 3 As shown, it is assumed that the size of the higher resolution image is 4*4 and the size of the lower resolution image is 2*2. For a specific passive microwave sensor, usually only one coarse resolution SM image can be provided. Therefore, the original 4*4 SM image is first downsampled to 2*2, and then each pixel in the 2*2 image is copied and expanded to obtain a 4*4 image again. Finally, the original 4*4 SM image is subtracted from the image after downsampling and copying and expansion to obtain a ΔSM image, which is used as the label of the CNN network. At the same time, the original high-resolution auxiliary data image is resampled to two resolutions corresponding to SM, and auxiliary data images of sizes 4*4 and 2*2 are obtained. Similarly, the auxiliary data image of 2*2 is copied and expanded to 4*4, and a 4*4 weight layer can be obtained according to the method of step S3. These two are combined with the 4*4 image as the feature of the CNN network. Finally, the training data is constructed using the method described in step S4.
[0054] Furthermore, in step S6, the model predicts the difference between the downscaled SM image and the original SM image. In order to obtain the downscaled SM image, another step of calculation is required, namely:
[0055] SM D =SM O +SM CNN
[0056] Where SM D represents the target downscaled SM image, SM o Indicates the original resolution SM image, SM CNN Represents the ΔSM image predicted by CNN.
[0057] Furthermore, in step S7, some statistical indicators need to be used to verify the downscaling result. The present invention uses the correlation coefficient R, the root mean square error RMSE and the unbiased root mean square error ubRMSE.
[0058] Compared with the prior art, the present invention has the following positive effects:
[0059] CNN effectively considers the information between pixels in the auxiliary image, thus significantly improving the accuracy of the downscaled soil moisture product. The ground station verification results show that when the original 36km SMAP SM product is downscaled to 3km, the average R, RMSE, and ubRMSE of the present invention are 0.6728, 0.0834, and 0.0497, respectively, while the R, RMSE, and ubRMSE of the traditional downscaling method using support vector regression are 0.6523, 0.0856, and 0.0525, respectively. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 It is a diagram of the method of establishing the “image-pixel” relationship;
[0061] Figure 2 It is the design diagram of the weight layer.
[0062] Figure 3 This is the basic structure diagram of CNN.
[0063] Figure 4 This is a diagram of a method for determining a label image and a corresponding feature image. DETAILED DESCRIPTION
[0064] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below by taking Oklahoma, USA as the study area and the 36km SM product provided by the SMAP satellite from January 1, 2018 to December 30, 2018 as an example, with reference to the accompanying drawings. The goal of this example is to downscale the 36km product of SMAP through auxiliary data to obtain a 9km SM downscaled image.
[0065] S1. There are two SMAP SM products that need to be downloaded: SMAP L3 Radiometer Global Daily 36kmEASE-Grid Soil Moisture V006 (L3_SM_P) and SMAP Enhanced L3 Radiometer GlobalDaily 9km EASE-Grid Soil Moisture V003 (L3_SM_P_E). L3_SM_P_E is used to verify the downscaled image. Considering the response of surface reflectivity and temperature to SM content, and the influence of terrain on SM spatial variation, the present invention uses DEM (Digital Elevation Model) data provided by SRTM (Shuttle Radar Topography Mission), reflectivity data provided by MODIS (Moderate-resolution Imaging Spectroradiometer), and surface temperature product data as auxiliary data. Among them, the resolution of DEM data is 30m; MODIS surface reflectance product has 7 bands, the resolution is 500m, and its product number is MOD09GA.006; the resolution of surface daytime temperature product is 1km, and its product number is MOD11A1.006. The spatial range of the downloaded L3_SM_P product is a regular 9*9 grid, covering a uniform grassland, and the spatial range of other products corresponds to it.
[0066] S2. Synthesize the SM data, DEM data and surface daytime temperature products according to the average value for 6 days, and synthesize the surface reflectance products according to the NDVI value and the maximum value of NDVI for 6 days. In addition, if there are null points in the synthesized surface temperature and surface reflectance products, the average value of the 4 nearest pixels around the null point will be used instead.
[0067] S3. The establishment of the “image-pixel” relationship is as follows Figure 2 As shown, the specific steps have been described in the content of the invention.
[0068] S4. The method for determining the label image and the corresponding feature image is as follows Figure 4As shown, the specific steps have also been described in the content of the invention. The first thing that needs to be determined is the features and labels in this example. The L3_SM_P product of 36km is scaled to 72km, and the image of 72km is copied and expanded to make it the same size as the product of 36km, and then the two images are subtracted to obtain ΔSM, which is used as the label of CNN; then the DEM data, the surface reflectance data of 7 bands in MODIS, and the MODIS surface daytime temperature data are resampled to 36km and 72km respectively, and the image of 72km is also copied and expanded, and the corresponding weight layer is introduced, and the three are combined as the features of CNN. After determining the features and label images, the training data set is generated by the method of establishing the "image-pixel" relationship described in step 3. The specific steps have been described in the content of the invention.
[0069] S5. The CNN built in this example is as described in the invention content. Since the research area selected this time is relatively uniform, a convolution kernel of size 3×3 is used. After the CNN is built, the training data set obtained in S4 is input to train the network.
[0070] S6. Based on the trained CNN, in order to obtain the 9km SM product, it is also necessary to first determine the features of the prediction process. First, the DEM data, the surface reflectance data of the 7 bands in MODIS, and the MODIS surface daytime temperature data are resampled to 36km and 9km, and the 36km image is copied and expanded to the same size as the 9km image; then the two resolution images are combined with the weight layer to form the features of the prediction process; then the "image-pixel" relationship establishment method in S3 is used to generate the feature data set of the prediction process; finally, these feature data sets are input into the trained CNN to obtain the prediction results. The predicted result is the difference between the 9km SM and the 36km SM. Therefore, the 9km SM product can be obtained by the following formula:
[0071] SM 9km =SM 36km +SM CNN
[0072] Where SM 36km The SMAP SM product (L3_SM_P) at 36 km, SM CNN is the result obtained by CNN during the prediction process, SM 9km It is a 9 km downscaled SM product.
[0073] S7. This paper uses the 9km SM product (L3_SM_P_E) provided by SMAP to verify the downscaled product. The results show that the 9km downscaled SM product has an R of 0.9509, RMSE of 0.0225, and ubRMSE of 0.0224 compared with the L3_SM_P_E product.
[0074] Finally, it should be noted that the above embodiments are only preferred embodiments of the present invention, which are used to illustrate the technical solution of the present invention rather than to limit the present invention. Those skilled in the art should understand that the technical solution of the present invention can be improved or changed without departing from the principle of the present invention, and they should all be covered by the protection scope of the claims attached to the present invention.
Claims
1. A soil moisture product downscaling method based on a two-dimensional convolutional neural network, the steps of which are include: 1) Obtain the passive microwave soil moisture product image and the corresponding auxiliary data image of the target area; and establish the corresponding relationship between the auxiliary data image and the passive microwave soil moisture product image; the method for establishing the corresponding relationship between the auxiliary data image and the passive microwave soil moisture product image is as follows: first, resample the auxiliary data image to the spatial resolution of the passive microwave soil moisture product image, so that the pixels of the auxiliary data image correspond to the pixels of the passive microwave soil moisture product image one by one; set a window of size m×m to crop the auxiliary data image, where m is an odd number; then take each window of the auxiliary data image as a "picture", and take the passive microwave soil moisture product image corresponding to the central pixel of the "picture" as a "pixel", so as to establish the "picture-pixel" relationship as the corresponding relationship; 2) constructing training data according to the corresponding relationship; then using the training data to train the CNN network; the method for constructing the training data is as follows: 11) downsampling the passive microwave soil moisture product image corresponding to the "pixel" in the training data determined according to each "image-pixel" relationship to obtain a low-resolution passive microwave soil moisture product image; 12) copying and expanding each pixel in the low-resolution passive microwave soil moisture product image to obtain an image with the same resolution as the passive microwave soil moisture product image again; 13) downsampling the passive microwave soil moisture product image corresponding to the "pixel" The image is subtracted from the image obtained in step 12) to obtain a ΔSM image, which is used as a label for training data; 14) the auxiliary data images corresponding to the "image" in the training data determined by the "image-pixel" relationship are resampled respectively to obtain an auxiliary data image T at the resolution of the passive microwave soil moisture product image and an auxiliary data image t at the low resolution; 15) the auxiliary data image t is copied and expanded to make its resolution the same as that of the auxiliary data image T; 16) the pixels in the auxiliary data image t processed in step 15) are weighted to generate a corresponding feature layer as feature data; 3) Use the trained CNN network to downscale the passive microwave soil moisture product image to be processed in the target area to obtain the target downscaled product.
2. The method according to claim 1, It is characterized in that The larger the spatial variability of the area, the larger the window used for the passive microwave soil moisture product image; for the pixels at the edge of the auxiliary data image, new pixels are introduced around the edge of the auxiliary data image, and the values of the new pixels are set to the average value of all pixels in the auxiliary data image; then the pixels at the edge of the auxiliary data image are cropped.
3. The method according to claim 2, It is characterized in that A weight layer is set for the auxiliary data image after the new pixel is introduced. The original pixel weight of the auxiliary data image in the weight layer is set to H, and the weight of the introduced new pixel is h; wherein H is greater than h; and then the weight layer is combined with the auxiliary data image after the new pixel is introduced as a feature layer.
4. The method according to claim 1, It is characterized in that Using the n passive microwave soil moisture product images adjacent to and before the ith passive microwave soil moisture product image, null value points of the ith passive microwave soil moisture product image are eliminated.
5. The method according to claim 4, It is characterized in that The value of n is determined based on how fast the passive microwave soil moisture product image changes over time; for each null point in the i-th passive microwave soil moisture product image, the average value of the pixels corresponding to the null point position in the n passive microwave soil moisture product images is taken as the pixel value of the null point position in the i-th passive microwave soil moisture product image.
6. The method according to claim 1, It is characterized in that In the process of training the CNN network using the training data, when the passive microwave soil moisture product image corresponding to the area with greater spatial variability is used as input, the CNN convolution kernel is set to be larger.
7. The method according to claim 1, It is characterized in that The auxiliary data include DEM data provided by SRTM, reflectivity data provided by MODIS and surface temperature product data; the method for obtaining the target downscaled product is: using the trained CNN network to predict the passive microwave soil moisture product image to be processed in the target area, and then according to SM D =SM O +SM CNN Get the target downscaling product SM D ; Among them, SM o represents the passive microwave soil moisture product image to be processed, SM CNN Represents the difference image ΔSM predicted by the CNN network.
