Sea-land pollution error correction method for space-borne synthetic aperture microwave radiometer

Abstract

The application discloses a sea-land pollution error correction method for a spaceborne synthetic aperture microwave radiometer and belongs to the technical field of microwave remote sensing, and comprises the following steps: in a training stage, obtaining a sea-land surface radiation brightness temperature TB, taking the radiation brightness temperature TB as an input of a spaceborne synthetic aperture radiometer observation model, and calculating an observation brightness temperature TA; normalizing and matching the observation brightness temperature TA and the radiation brightness temperature TB into pairs as a data set of a CNN network, and taking a sea-land label on each brightness temperature image pixel as a label; taking the observation brightness temperature TA and the radiation brightness temperature TB as inputs of the CNN network, and training the CNN network until loss convergence; in an application stage, inputting a spaceborne synthetic aperture microwave radiometer observation brightness temperature image into the trained CNN network, and performing error correction. The application can efficiently and timely realize sea-land pollution error correction in sea-land salinity remote sensing of a spaceborne synthetic aperture microwave radiometer, and can reduce the complexity of error correction and improve the accuracy of error correction.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of microwave remote sensing, and more particularly relates to a sea-land pollution error correction method for a spaceborne synthetic aperture microwave radiometer. BACKGROUND

[0002] The synthetic aperture microwave radiometer is an effective payload for realizing all-weather and all-time satellite sea surface salinity remote sensing, and has been used in the SMOS ocean salinity remote sensing satellite mission of the European Space Agency. The sea surface salinity microwave remote sensing works in the L band, and the land brightness temperature (240-280K) is much higher than the ocean brightness temperature (80-140K), so a large brightness temperature gradient will be formed in the sea-land boundary area. The synthetic aperture microwave radiometer obtains the spatial frequency sampling of the brightness temperature distribution in the field of view region, i.e. the visibility function, and then uses a Fourier transform or other reconstruction algorithm to reconstruct the observed brightness temperature image from the visibility function. Due to the large loss of high-frequency component information, the large brightness temperature gradient will produce oscillation in the entire field of view in the reconstructed observed brightness temperature, which is called Gibbs effect, resulting in a significant increase in the observed brightness temperature error in the near-shore area (up to 200 kilometers range), and the accuracy of the inverted sea surface salinity cannot meet the requirements of remote sensing products, which is called sea-land pollution in the spaceborne synthetic aperture microwave radiometer ocean salinity remote sensing.

[0003] There are mainly three existing sea-land pollution error correction methods: brightness temperature compensation method, windowing method and node sampling method. The brightness temperature compensation method mainly estimates the error between the observed brightness temperature value and the ideal (without sea-land pollution) brightness temperature, and then uses the error to compensate the observed brightness temperature. Since this method mainly relies on historical experience data for estimation, it is very difficult to find the optimal error compensation value, and the correction effect is very limited, and many abnormal data cannot be processed. The windowing method mainly aims at the discrete Fourier transform process of brightness temperature reconstruction, and can reduce the amplitude of sea-land pollution in the brightness temperature reconstruction image to a certain extent, but at the cost of resolution, and the sea-land pollution effect still exists in the near-shore area, and the reconstructed brightness temperature error is still large. The node sampling method is to oversample the brightness temperature image at the point where the oscillation interference causes the least distortion to the geophysical signal, but the correction effect is good only for the Gibbs error caused by the point-like target of RFI, and the effect is not good for the Gibbs error caused by the sea-land boundary. SUMMARY

[0004] In view of the defects and improvement needs of the prior art, the application provides a sea-land pollution error correction method for a spaceborne synthetic aperture microwave radiometer, which aims to efficiently and timely correct the sea-land pollution error in the spaceborne synthetic aperture microwave radiometer ocean salinity remote sensing, and reduce the complexity and improve the accuracy of the error correction.

[0005] To achieve the above object, according to a first aspect of the present application, a sea-land pollution error correction method for a spaceborne synthetic aperture microwave radiometer is provided, comprising:

[0006] a training phase:

[0007] obtaining a sea-land surface radiance brightness temperature TB and taking the radiance brightness temperature TB as an input of a spaceborne synthetic aperture radiometer observation model to calculate a spaceborne synthetic aperture microwave radiometer observation brightness temperature TA;

[0008] normalizing and matching the observation brightness temperature TA and the radiance brightness temperature TB into pairs as a data set of a CNN network, and taking a sea-land label on each brightness temperature image pixel as a label;

[0009] taking the observation brightness temperature TA and the radiance brightness temperature TB as inputs of the CNN network, taking minimizing a feature difference loss between a predicted brightness temperature TB0 output by the CNN network and the radiance brightness temperature TB as a target, training the CNN network and reversely adjusting parameters of the CNN network to make the loss converge;

[0010] an application phase:

[0011] inputting a spaceborne synthetic aperture microwave radiometer observation brightness temperature image into the trained CNN network to output a sea-land pollution error corrected brightness temperature image.

[0012] Further, the CNN network comprises a first full connection layer, a first deconvolution layer, a convolution layer, a second deconvolution layer and a second full connection layer connected in sequence.

[0013] Further, obtaining the sea-land surface radiance brightness temperature TB comprises:

[0014] S11, obtaining ground surface and sea surface physical parameters;

[0015] S12, interpolating the obtained ground surface and sea surface physical parameters to spaceborne synthetic aperture microwave radiometer observation pixel positions, and dividing the interpolated data into sea area and land area data;

[0016] S13, inputting the sea area data into a FASTEM-5 model and inputting the land area data into a CMEM model to obtain the sea-land surface radiance brightness temperature TB.

[0017] Further, in S11, ERA-5 historical reanalysis data is taken as an initial value to drive a numerical weather prediction model WRF to obtain the ground surface and sea surface physical parameters.

[0018] Further, calculating the spaceborne synthetic aperture microwave radiometer observation brightness temperature TA comprises:

[0019] S21, based on the observation index of the spaceborne synthetic aperture microwave radiometer, the radiation brightness temperature TB coordinates on the latitude and longitude are converted into antenna coordinate grid points on the spatial cosine plane;

[0020] S22, the corrected brightness temperature is calculated by using the radiation brightness temperature TB after coordinate conversion, and the DFT calculation is performed on the corrected brightness temperature to obtain the spatial frequency spectrum sampling of the brightness temperature distribution;

[0021] S23, the spaceborne synthetic aperture microwave radiometer observation brightness temperature TA is obtained by performing inverse discrete Fourier transform IDFT on the spatial frequency spectrum sampling.

[0022] Further, before the inverse discrete Fourier transform IDFT is performed on the spatial frequency spectrum sampling in S23, it further includes: performing FTT flat target transform on the spatial frequency spectrum sampling.

[0023] Further, the observation brightness temperature TA and the radiation brightness temperature TB are normalized and matched in pairs, including:

[0024] The normalized radiation brightness temperature TB, observation brightness temperature TA and corresponding pixel points under the same polarization are arranged into groups in units of snapshots;

[0025] The data of each snapshot is cut to the same length.

[0026] Further, the training stage further includes the steps of:

[0027] The SMOS satellite historical real observation data is input into the trained CNN network to obtain the predicted brightness temperature image;

[0028] The root mean square error between the predicted brightness temperature image and the radiation brightness temperature TB is calculated to test the performance of the trained CNN network.

[0029] According to another aspect of the present application, a spaceborne synthetic aperture microwave radiometer sea-land pollution error correction system is provided for executing the method of any one of the above first aspect, comprising a training module and an application module;

[0030] The training module comprises:

[0031] The brightness temperature data construction unit is used for acquiring the sea-land surface radiation brightness temperature TB, and taking the radiation brightness temperature TB as the input of the spaceborne synthetic aperture radiometer observation model to calculate the spaceborne synthetic aperture microwave radiometer observation brightness temperature TA;

[0032] The data set construction unit is used for normalizing and matching the observation brightness temperature TA and the radiation brightness temperature TB in pairs as the data set of the CNN network, and the label is the sea-land mark on each brightness temperature image pixel;

[0033] a network training unit configured to take the observed brightness temperature TA and the radiative brightness temperature TB as inputs of the CNN network, train the CNN network and adjust parameters of the CNN network reversely to make loss converge, aiming at minimizing feature difference loss between a predicted brightness temperature TB0 output by the CNN network and the radiative brightness temperature TB.

[0034] The application module is configured to input the observed brightness temperature image of the spaceborne synthetic aperture microwave radiometer into the trained CNN network, and output a land pollution error corrected brightness temperature image.

[0035] According to another aspect of the present application, a computer readable storage medium is provided, which stores a computer program, and the program is executed by a processor to implement the method according to any one of the first aspect.

[0036] In general, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0037] (1) The method of the present application uses a CNN network to identify and classify ocean and land features, learn the sea-land pollution error features in the synthetic aperture salinity remote sensing brightness temperature image, and correct the sea-land pollution error by remapping the observed brightness temperature image using the CNN network. Therefore, the observed brightness temperature TA of the spaceborne synthetic aperture microwave radiometer affected by sea-land pollution and the earth surface radiative brightness temperature TB with clear sea-land boundaries are required. However, since the surface radiative brightness temperature (ideal brightness temperature) corresponding to the true SMOS satellite brightness temperature image (polluted brightness temperature image) cannot be obtained, the present application constructs the earth surface radiative brightness temperature TB with clear sea-land boundaries as a simulated ideal brightness temperature, and constructs the observed brightness temperature TA of the spaceborne synthetic aperture microwave radiometer affected by sea-land pollution based on the radiative brightness temperature TB. In this way, the sea-land pollution error in the spaceborne synthetic aperture microwave radiometer ocean salinity remote sensing can be trained by the neural network, and finally the brightness temperature image with clear sea-land boundaries is obtained to realize efficient and real-time correction of the sea-land pollution error of the spaceborne synthetic aperture microwave remote sensing brightness temperature image. At the same time, the CNN is used to extract features for error correction, which does not have the defects of the existing brightness temperature compensation method, windowing method and node sampling method, etc. mainstream error correction methods, can effectively reduce the computational complexity of sea-land pollution error correction, and can improve the accuracy of error correction.

[0038] (2) Further, in the sea-land pollution error correction using the CNN, since the integrated aperture salinity remote sensing brightness temperature image is one-dimensional non-uniform hexagonal grid data, which does not match the two-dimensional non-uniform data of the convolutional neural network, therefore, the application designs a convolutional neural network connected through full connection layers at the beginning and the end, to complete the mapping of the one-dimensional non-uniform hexagonal grid brightness temperature data and the two-dimensional deep convolution space, so as to adapt to the feature extraction of the intermediate layer two-dimensional convolutional neural network.

[0039] (3) Further, a specific sea-land surface radiation brightness temperature TB construction method and a method for obtaining a satellite-borne integrated aperture microwave radiometer observation brightness temperature TA based on the sea-land surface radiation brightness temperature TB are provided, and a pair of training data sets are constructed based on the radiation brightness temperature TB and the observation brightness temperature TA.

[0040] (4) As a preferred, FTT flat target transformation is performed on the spatial spectrum sampling, so as to reduce the system error of the satellite-borne integrated aperture microwave radiometer.

[0041] (5) As a preferred, since the number of one-dimensional non-uniform data of each snapshot unit is different, the data of each snapshot is truncated to the same length by truncation, so as to realize the matching of the two-dimensional neural network data.

[0042] (6) Further, the CNN network constructed by the application is used to correct the pollution of the SMOS satellite observation brightness temperature in the sea-land boundary area, therefore, the SMOS satellite observation brightness temperature is used as a test set to verify the correction effect on the SMOS satellite observation brightness temperature, and the reliability of the network is improved.

[0043] In summary, the application obtains the sea-land pollution characteristics in the satellite-borne integrated aperture microwave radiometer ocean salinity remote sensing through the CNN supervised learning, and the sea-land pollution error correction process is completed by the deep convolutional neural network, so as to solve the problems of high complexity and unsatisfactory effect of the current method, and the sea-land pollution error correction in the satellite-borne integrated aperture microwave radiometer ocean salinity remote sensing can be realized more efficiently and in real time. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 It is a schematic diagram of the sea-land pollution error correction method of the satellite-borne integrated aperture microwave radiometer of the application.

[0045] Figure 2 It is a training flowchart in the embodiment of the application.

[0046] Figure 3 It is a simulated brightness temperature graph of Snapshot ID 565902844 in the method embodiment of the application.

[0047] Figure 4The observed brightness temperature map for the Snapshot ID 565902844 in the embodiment of the method of the application.

[0048] Figure 5 The test brightness temperature map for the Snapshot ID 565902844 in the embodiment of the method of the application.

[0049] Figure 6 The brightness temperature image remapped by the trained convolutional neural network CNN in the embodiment of the method of the application. DETAILED DESCRIPTION

[0050] In order to make the objects, technical solutions and advantages of the application clearer, the application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application and not to limit the application. In addition, the technical features involved in each embodiment of the application described below can be combined with each other as long as they do not conflict with each other.

[0051] In the application, the terms "first", "second", etc. in the application and the drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0052] As shown in Figure 1 , Figure 2 The spaceborne synthetic aperture microwave radiometer sea-land pollution error correction method of the application mainly includes:

[0053] Training phase:

[0054] Obtain the sea-land surface radiation brightness temperature TB, and take the radiation brightness temperature TB as the input of the spaceborne synthetic aperture radiometer observation model to calculate the spaceborne synthetic aperture microwave radiometer observed brightness temperature TA;

[0055] Normalize and match the radiation brightness temperature TB and the spaceborne synthetic aperture microwave radiometer observed brightness temperature TA into pairs as the data set of the CNN network, and the label is the sea-land label on each pixel of the brightness temperature image;

[0056] Take the spaceborne synthetic aperture microwave radiometer observed brightness temperature TA and the radiation brightness temperature TB as the input of the CNN network, take the feature difference loss between the predicted brightness temperature TB0 output by the CNN network and the radiation brightness temperature TB as the target, train the CNN network and adjust the parameters of the CNN network in the reverse direction to make the loss converge;

[0057] Application phase:

[0058] Input the spaceborne synthetic aperture microwave radiometer observed brightness temperature image into the trained CNN network, and output the brightness temperature image after sea-land pollution error correction, that is, the brightness temperature image with clear sea-land boundary.

[0059] Specifically, in the training stage, obtaining the sea-land surface radiance brightness temperature TB includes the following steps:

[0060] S11, obtaining the physical parameters of the land surface and the sea surface, in the embodiment of the present application, the physical parameters of the land surface and the sea surface include sea surface temperature, wind speed, wind direction, sea-land mask and the like; specifically, in S11, taking the ERA-5 historical reanalysis data as the initial value, driving the numerical weather prediction model WRF to obtain the physical data of the land surface and the sea surface;

[0061] S12, interpolating the obtained physical parameters of the land surface and the sea surface to the observation pixel position of the spaceborne synthetic aperture microwave radiometer, and dividing the interpolated data into marine and land area data;

[0062] S13, inputting the marine area data into the FASTEM-5 model and inputting the land area data into the CMEM model, respectively obtaining the corresponding marine surface radiance brightness temperature and land surface radiance brightness temperature, that is, the sea-land surface radiance brightness temperature TB.

[0063] Specifically, calculating the spaceborne synthetic aperture microwave radiometer observation brightness temperature TA includes the following steps:

[0064] S21, based on the observation index of the spaceborne synthetic aperture microwave radiometer, converting the sea-land surface radiance brightness temperature TB coordinates on the latitude and longitude into the antenna coordinate grid points on the spatial cosine plane; wherein the observation index of the spaceborne synthetic aperture microwave radiometer includes orbit parameters, satellite position, satellite orientation, scene coordinate information and the like.

[0065] S22, calculating the corrected brightness temperature with the coordinate-converted radiance brightness temperature TB, and performing DFT calculation on the corrected brightness temperature to obtain the spatial frequency spectrum sampling of the brightness temperature distribution, that is, the visibility function;

[0066] S23, performing inverse discrete Fourier transform IDFT on the spatial frequency spectrum sampling of the brightness temperature distribution to obtain the spaceborne synthetic aperture microwave radiometer observation brightness temperature TA.

[0067] As a preferred, before performing the inverse discrete Fourier transform IDFT on the spatial frequency spectrum sampling of the brightness temperature distribution, it further includes performing FTT flat target transform on the spatial frequency spectrum sampling of the brightness temperature distribution to correct the system error of the spaceborne synthetic aperture microwave radiometer.

[0068] Specifically, normalizing and matching the radiance brightness temperature TB and the spaceborne synthetic aperture microwave radiometer observation brightness temperature TA in pairs, including:

[0069] Normalizing the radiance brightness temperature TB and the observation brightness temperature TA and mapping them in the interval of 0-1;

[0070] The normalized radiation brightness temperature TB, the observed brightness temperature TA and the sea-land mark of the corresponding pixel points under the same polarization are arranged into groups in a snapshot unit;

[0071] The data of each snapshot is cut to the same length, and the length is empirically statistically determined according to the data set, so that the data set can be divided into a uniform length and the data loss is minimized.

[0072] Specifically, the CNN network in the embodiment of the application includes a first fully connected layer, a first deconvolution layer, a convolution layer, a second deconvolution layer and a second fully connected layer connected in sequence, denoted as SLPCN convolutional neural network.

[0073] The first fully connected layer is used to map the one-dimensional non-uniform hexagonal grid data observed brightness temperature TA and the radiation brightness temperature TB to a two-dimensional deep convolution space, so as to adapt to the feature extraction of the CNN network.

[0074] After the spatial mapping of the fully connected layer, the spaceborne synthetic aperture microwave radiometer observed brightness temperature TA is sequentially subjected to feature extraction through the first deconvolution layer, the convolution layer and the second deconvolution layer; wherein the sea and land are distinguished by adding a sea-land mask in the first deconvolution layer.

[0075] The second fully connected layer is used to map the extracted two-dimensional uniform data features to one-dimensional non-uniform hexagonal grid data, to obtain the predicted brightness temperature TB0 output by the CNN network.

[0076] The intermediate layer uses the TensorFlow framework to design the structure of the network, and includes the first deconvolution layer, the convolution layer and the second deconvolution layer connected in sequence; the neural network is trained by randomly extracting n samples in each iteration, and the loss function is:

[0077]

[0078] Wherein, M represents the total number of the training set, n represents the number of samples in each training set, y i represents the predicted brightness temperature output by the neural network in the current iteration; x i represents the corresponding radiation brightness temperature in the sample in the current iteration.

[0079] In the training process, the hyperparameters of the deep convolutional neural network are set, and the loss function is optimized by using the stochastic gradient descent method to make the network converge, to obtain the trained deep convolutional neural network. The hyperparameters of the deep convolutional neural network include: the mapping dimensions of the two fully connected layers, the learning rate of the network, the number of network iterations, the running mode of the network and the location of the network parameter saving.

[0080] The method further comprises: inputting historical real observation data of the SMOS satellite into the trained CNN network to obtain a land-sea pollution error corrected brightness temperature image.

[0081] The root mean square error (RMSE) between the land-sea pollution error corrected brightness temperature image and the radiative brightness temperature TB is calculated to test the performance of the trained CNN network, and the smaller the root mean square error RMSE, the better the performance of the CNN network.

[0082] In the embodiment of the application, the microwave remote sensing brightness temperature image with Snapshot ID of 565902844 of the SMOS satellite is taken as an example to further illustrate the method of the application.

[0083] The L2 level product data of the SMOS satellite is downloaded to obtain physical information such as sea surface temperature, wind speed, wind direction and land mask. According to a numerical weather prediction model, ERA-5 historical reanalysis data is taken as initial values to drive the WRF model to predict the physical data of the ground and sea surface.

[0084] Based on the FATEST-5 and CMEM models, the microwave remote sensing simulation brightness temperature TB is generated according to the above-mentioned physical parameters, and the simulation brightness temperature TB is input into the satellite-borne synthetic aperture radiometer observation model to calculate the satellite-borne synthetic aperture microwave remote sensing observation brightness temperature image TA. Finally, 500 groups of TA, TB and land-sea label data are generated.

[0085] The one-dimensional observation brightness temperature TA and the land-sea label data of each Snapshot are stacked together in the second dimension, and the simulation brightness temperature TB is saved as a mat format file. Finally, 500 groups of integrated brightness temperature and land-sea labels are generated. 450 groups of data are randomly selected as a training set, and 50 groups of data are selected as a test set.

[0086] The CNN network constructed above is used to set the learning rate, running mode and network parameter storage position of the network in the code, and the network is trained.

[0087] The trained network is tested by using the test set, and the performance of the network is tested by using the brightness temperature image directly read from the historical real observation data of the SMOS satellite as the test set.

[0088] As shown in Figure 3 , it is the simulation brightness temperature image of Snapshot ID 565902844 in the embodiment of the method of the application, that is, the radiative brightness temperature TB; Figure 4 , it is the observation brightness temperature image TA calculated based on the simulation brightness temperature image in Figure 3 ; and Figure 5 , it is the test brightness temperature image of Snapshot ID 565902844 in the embodiment of the method of the application, which is directly read from the historical real observation data of the SMOS satellite.Figure 6 For the remapped brightness temperature image in the embodiment of the method of the application by the trained convolutional neural network CNN, the remapping is performed by taking the test brightness temperature image in the image as input. Figure 5

[0089] Table 1 is the root mean square error (RMSE) of the test brightness temperature of each region before and after remapping with the simulated brightness temperature, in K.

[0090] Table 1 is the root mean square error (RMSE) of the test brightness temperature of each region before and after remapping with the simulated brightness temperature

[0091]

[0092] Comparison Figure 5 and Figure 6 From the visual comparison, it can be clearly seen that the remapped brightness temperature obtained by the trained network of the application has a good similarity with the original brightness temperature, and has a clearer sea-land boundary than the original brightness temperature. From Table 1, it can be seen from the numerical values that after remapping, the root mean square error of the observed brightness temperature and the simulated brightness temperature in the sea-land junction region and the near-sea region is obviously reduced. Therefore, the sea-land pollution error correction of the synthetic aperture microwave remote sensing brightness temperature image can be effectively performed.

[0093] In the embodiment of the application, the sea-land pollution feature in the ocean salinity remote sensing of the spaceborne synthetic aperture microwave radiometer is obtained by using supervised learning, and the sea-land pollution error correction process is completed by the deep convolutional neural network. Thus, the problem of high complexity and unsatisfactory effect of the current method is solved, and the sea-land pollution error correction in the ocean salinity remote sensing of the spaceborne synthetic aperture microwave radiometer can be more efficiently and timely realized.

[0094] According to another aspect of the application, a sea-land pollution error correction system for a spaceborne synthetic aperture microwave radiometer is provided, which is used to perform the steps corresponding to the sea-land pollution error correction method for the spaceborne synthetic aperture microwave radiometer, and specifically includes a training module and an application module.

[0095] The training module includes:

[0096] The brightness temperature data construction unit is configured to obtain a sea-land surface radiation brightness temperature TB, and take the radiation brightness temperature TB as an input of a spaceborne synthetic aperture radiometer observation model to calculate a spaceborne synthetic aperture microwave radiometer observation brightness temperature TA.

[0097] The data set construction unit is configured to normalize and match the observation brightness temperature TA and the radiation brightness temperature TB into pairs as a data set of the CNN network, and label the sea-land marks on each pixel of the brightness temperature image.

[0098] ​a network training unit configured to take the observed brightness temperature TA and the radiative brightness temperature TB as inputs of a CNN network, train the CNN network and adjust parameters of the CNN network reversely to minimize a feature difference loss between a predicted brightness temperature TB0 output by the CNN network and the radiative brightness temperature TB, and make the loss converge;

[0099] The application module is configured to input the observed brightness temperature image of the space-borne synthetic aperture microwave radiometer into the trained CNN network, and output a land pollution error corrected brightness temperature image.

[0100] According to another aspect of the present application, a computer readable storage medium is provided, which stores a computer program, and the program is executed by a processor to implement steps corresponding to the space-borne synthetic aperture microwave radiometer land pollution error correction method as described above.

[0101] The method of the present application uses the CNN network to identify and classify the sea and land features, learns the land pollution error features in the synthetic aperture salinity remote sensing brightness temperature image, and uses the CNN network to remap the observed brightness temperature image to correct the land pollution error. Therefore, the space-borne synthetic aperture microwave radiometer observed brightness temperature TA affected by the land pollution and the earth surface radiative brightness temperature TB with clear land and sea boundaries are required. However, since the surface radiative brightness temperature (ideal brightness temperature) corresponding to the true SMOS satellite brightness temperature image (polluted brightness temperature image) cannot be obtained, the present application constructs the earth surface radiative brightness temperature TB with clear land and sea boundaries as a simulated ideal brightness temperature, and constructs the space-borne synthetic aperture microwave radiometer observed brightness temperature TA affected by the land pollution based on the radiative brightness temperature TB. In this way, the land pollution error in the space-borne synthetic aperture microwave radiometer sea salinity remote sensing can be trained by the neural network, and finally the brightness temperature image with clear land and sea boundaries is obtained to realize efficient and real-time correction of the land pollution error of the space-borne synthetic aperture microwave remote sensing brightness temperature image. At the same time, the CNN is used to extract features for error correction, and there is no defect corresponding to the mainstream error correction methods such as the existing brightness temperature compensation method, the windowing method and the node sampling method, which can effectively reduce the calculation complexity of the land pollution error correction and improve the accuracy of the error correction. Moreover, the experimental results further prove the effectiveness of the method of the present application.

[0102] When the CNN is used to correct the land pollution error, the synthetic aperture salinity remote sensing brightness temperature image is one-dimensional non-uniform hexagonal grid data, which does not match the two-dimensional non-uniform data of the convolutional neural network. Therefore, the SLPCN convolutional neural network is designed to connect the beginning and the end through the full connection layer mapping, to complete the mapping of the one-dimensional non-uniform hexagonal grid brightness temperature data and the two-dimensional deep convolution space to adapt to the feature extraction of the intermediate layer two-dimensional convolutional neural network.

[0103] The sea-land pollution error correction method for the satellite-borne synthetic aperture microwave remote sensing brightness temperature image of the application is based on the SMOS satellite, and the brightness temperature image with clear sea-land boundary is reconstructed from the input synthetic aperture microwave remote sensing brightness temperature image under the same polarization and the sea-land label, so as to solve the pollution problem of the synthetic aperture microwave remote sensing brightness temperature observation image at the sea-land junction. Different from the conventional convolutional neural network CNN which is usually used for regular two-dimensional image feature extraction, the network model in the application can adapt to the irregular two-dimensional image of the SMOS satellite data through special design. The training data set of the convolutional neural network CNN in the application is composed of pairs of synthetic aperture microwave remote sensing radiation brightness temperature TB data and observation brightness temperature TA data, and is supplemented by the sea-land label data of the corresponding region. The method can efficiently correct the sea-land pollution error in the satellite-borne synthetic aperture microwave radiometer ocean salinity remote sensing, effectively improve the measurement brightness temperature precision of the satellite-borne synthetic aperture microwave radiometer in the near sea area, and reduce the calculation complexity of the sea-land pollution error correction. It is a new type of sea-land pollution error correction method in the satellite-borne synthetic aperture microwave radiometer ocean salinity remote sensing.

[0104] Those skilled in the art will readily understand that the above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for correcting land and sea pollution errors in a spaceborne integrated aperture microwave radiometer, characterized in that, include: Training phase: The surface radiation brightness temperature TB of land and sea is obtained, and the radiation brightness temperature TB is used as the input of the observation model of the spaceborne synthetic aperture radiometer to calculate the observation brightness temperature TA of the spaceborne synthetic aperture microwave radiometer. The observed brightness temperature TA and the radiation brightness temperature TB are normalized and matched into pairs, which are used as the dataset for the CNN network. The labels are the land and sea markers on each brightness temperature image pixel. Using the observed brightness temperature TA and radiation brightness temperature TB as inputs to the CNN network, the goal is to minimize the feature difference loss between the predicted brightness temperature TB0 and the radiation brightness temperature TB output by the CNN network. The CNN network is trained and its parameters are adjusted in reverse to converge the loss. The CNN network includes a first fully connected layer, a first deconvolution layer, a convolution layer, a second deconvolution layer, and a second fully connected layer connected in sequence. The first fully connected layer maps the one-dimensional non-uniform hexagonal grid data of observed brightness temperature TA and radiation brightness temperature TB to a two-dimensional deep convolutional space to adapt to the feature extraction of the CNN network. After spatial mapping by the first fully connected layer, the observed brightness temperature TA is then processed sequentially through the first deconvolution layer, the convolution layer, and the second deconvolution layer to extract features, resulting in two-dimensional uniform data features. A land-sea mask is added in the first deconvolution layer to distinguish between ocean and land. The second fully connected layer maps the extracted two-dimensional uniform data features to one-dimensional non-uniform hexagonal grid data to obtain the predicted brightness temperature TB0 output by the CNN network. Application phase: The brightness temperature images observed by the spaceborne integrated aperture microwave radiometer are input into a trained CNN network, which outputs brightness temperature images corrected for land and sea pollution errors. Obtaining the surface radiative brightness temperature (TB) of the land and sea includes: S11. Obtain physical parameters of the land surface and sea surface; S12. Interpolate the obtained surface and sea surface physical parameters to the observation pixel positions of the spaceborne integrated aperture microwave radiometer, and divide the differenced data into ocean and land area data; S13. Input the marine area data into the FASTEM-5 model and the land area data into the CMEM model to obtain the surface radiation brightness temperature TB of the sea and land. The calculation of the brightness temperature (TA) observed by the spaceborne synthetic aperture microwave radiometer includes: S21. Based on the observation indicators of the spaceborne integrated aperture microwave radiometer, the radiation brightness temperature TB coordinates in latitude and longitude are converted into antenna coordinate grid points in the spatial cosine plane. S22. Calculate the corrected brightness temperature using the coordinate-transformed radiative brightness temperature TB, and perform DFT calculation on the corrected brightness temperature to obtain the spatial spectrum sampling of the brightness temperature distribution. S23. Perform Inverse Discrete Fourier Transform (IDFT) on the spatial spectrum sampling to obtain the brightness temperature (TA) observed by the spaceborne integrated aperture microwave radiometer.

2. The method according to claim 1, characterized in that, In S11, the ERA-5 historical reanalysis data are used as initial values ​​to drive the WRF numerical weather prediction model to forecast the physical parameters of the land surface and sea surface.

3. The method according to claim 1, characterized in that, In S23, before performing the Inverse Discrete Fourier Transform (IDFT) on the spatial spectrum samples, the method further includes performing the Flat Target Transform (FTT) on the spatial spectrum samples.

4. The method according to claim 1, characterized in that, Normalizing and pairing the observed brightness temperature TA with the radiative brightness temperature TB includes: The normalized radiation brightness temperature TB, observed brightness temperature TA, and corresponding land and sea markers of the corresponding pixels under the same polarization are grouped into snapshots. The data from each snapshot is truncated to the same length.

5. The method according to claim 1, characterized in that, The training phase also includes the following steps: Historical real observation data from the SMOS satellite is input into the trained CNN network to obtain the predicted brightness temperature image; The root mean square error between the predicted brightness temperature image and the radiation brightness temperature TB is calculated to test the performance of the trained CNN network.

6. A spaceborne integrated aperture microwave radiometer sea-land pollution error correction system, characterized in that, The method for performing any one of claims 1-5 includes a training module and an application module; The training module includes: The brightness temperature data construction unit is used to obtain the radiation brightness temperature TB of the land and sea surface, and use the radiation brightness temperature TB as the input of the observation model of the spaceborne synthetic aperture radiometer to calculate the brightness temperature TA observed by the spaceborne synthetic aperture microwave radiometer. The dataset construction unit is used to normalize and match the observed brightness temperature TA and the radiation brightness temperature TB into pairs, which are used as the dataset for the CNN network. The labels are the land and sea markers on each brightness temperature image pixel. The network training unit is used to train the CNN network with the observed brightness temperature TA and the radiation brightness temperature TB as inputs, aiming to minimize the feature difference loss between the predicted brightness temperature TB0 and the radiation brightness temperature TB of the CNN network output, and to adjust the parameters of the CNN network in reverse so that the loss converges. The application module is used to input the brightness temperature images observed by the spaceborne integrated aperture microwave radiometer into the trained CNN network and output the brightness temperature images after correction for marine and land pollution errors.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-5.

Images