A method for retrieving the three-dimensional ocean temperature and salinity field using satellite remote sensing data

Abstract

The present invention discloses a method for inverting the three-dimensional ocean temperature and salinity field using satellite remote sensing data, which belongs to the field of ocean observation. Specifically, first, for the ocean area to be processed, historical data is obtained to construct a training set; then, a deep learning network model based on the U-Net neural network is constructed; by establishing jump connections in different convolutional layers, all information in different convolutional layers can be retained during the upsampling process. Next, the deep learning network is trained using the training set to obtain an intelligent inversion model. Finally, the satellite remote sensing data at the time to be collected is input, and the corresponding three-dimensional ocean temperature and salinity field is obtained using the intelligent inversion model. Due to the addition of the time variable, the present invention enables the inversion model to have the ability to learn the characteristics of time laws. While ensuring practical feasibility, the inversion accuracy, the utilization value and efficiency of satellite remote sensing data are improved, and at the same time, it has good versatility.

Description

Technical Field

[0001] The invention belongs to the technical field of ocean observation, and in particular relates to a method for inverting three-dimensional ocean temperature and salinity fields using satellite remote sensing data. Background Art

[0002] Satellites can provide a wide range of ocean remote sensing information, but their limitation is that satellite remote sensing can only observe the ocean surface. In order to obtain richer ocean parameters below the sea surface, it is necessary to use satellite remote sensing inversion technology, combined with the parameter characteristics of the ocean surface, such as sea surface temperature (SST), sea surface salinity (SSS) and sea surface height (SSH), to invert and calculate the parameter information below the sea surface, thereby improving the utilization value and efficiency of remote sensing data.

[0003] Three types of neural networks are commonly used in existing satellite remote sensing intelligent inversion algorithms: fully connected neural networks (ANNs), convolutional neural networks (CNNs), and long short-term memory networks (LSTMs). Among them, ANNs can capture sufficient information, but their computational complexity is relatively large; CNNs can capture spatial features while reducing computational complexity, but they suffer from information loss problems; and LSTMs can better capture temporal regularity features, but they are lacking in capturing spatial features.

[0004] At present, there is no intelligent inversion algorithm that can make up for the above shortcomings while ensuring feasibility. Summary of the Invention

[0005] In response to the above problems, the present invention proposes a method for inverting the three-dimensional ocean temperature and salinity field using satellite remote sensing data, which can effectively make up for the defects in the existing intelligent inversion algorithm.

[0006] The specific steps are as follows:

[0007] Step 1: For the ocean area to be processed, obtain historical satellite remote sensing data and ocean historical reanalysis data, and construct a training sample dataset and a training label dataset respectively;

[0008] Satellite remote sensing data include SST, SSS and SSH, each of which is a two-dimensional matrix;

[0009] Each satellite remote sensing data is marked with a corresponding time tag according to its own time to form a time two-dimensional matrix;

[0010] Then, the two-dimensional matrix of the three satellite remote sensing data and the two-dimensional time matrix are combined into a four-layer three-dimensional matrix, that is, the training sample data set S = {S1, S2, ..., S n};

[0011] At the same time, the temperature and salinity fields at the corresponding time in the historical reanalysis data are extracted to construct a training label dataset L = {L1, L2, ..., L n}.

[0012] Step 2: Build a deep learning network model based on the U-Net neural network;

[0013] The deep learning network model includes: convolutional layers, skip connections, downsampling and upsampling.

[0014] Among them, the convolution layer includes convolution operation, batch normalization operation and activation function, downsampling is average pooling operation, upsampling is bilinear interpolation; the convolution kernel size is 3×3;

[0015] The first half of the U-Net neural network is used for feature extraction, and the second half is used for upsampling. By establishing skip connections between different convolutional layers, all information in different convolutional layers can be retained during the upsampling process, avoiding information loss.

[0016] Step 3: Use the training sample dataset and the training label dataset to train the deep learning network to obtain an intelligent inversion model.

[0017] The specific steps are:

[0018] First, the hyperparameters of the deep learning network are set according to actual needs: learning rate and number of training times, etc.

[0019] The learning rate uses an adaptive variable learning rate to adapt to different application requirements; the number of training times is mainly set directly based on the grid size.

[0020] Then, the training sample data set S={S1,S2,...,S n}, input into the deep learning network in the form of 4 layers of images;

[0021] The input data is 4*512*512; after two convolutions, it becomes a 16*512*512 three-dimensional feature matrix S1; this matrix S1 will flow into two different paths:

[0022] One path is downsampling, where the number of horizontal grids is reduced by 2 times through pooling operation to 16*256*256, and then converted to 32*256*256 matrix feature S2 through 2 convolutions;

[0023] The matrix feature S2 continues to flow into two different paths: one is downsampling, where the number of horizontal grids is reduced by 2 times through the pooling operation to 32*128*128, and then becomes the matrix feature S3 of 64*128*128 through two convolutions;

[0024] The matrix feature S3 continues to flow into two different paths: one is downsampling, where the number of horizontal grids is reduced by 2 times through the pooling operation to become 64*64*64, and then becomes the 128*64*64 matrix feature S4 through two convolutions;

[0025] The matrix feature S4 continues to flow into two different paths: one is downsampling, where the number of horizontal grids is reduced by 2 times through the pooling operation, becoming a 128*32*32 matrix feature S5, which is output through two convolutions; the other is a skip connection, where the matrix feature S4 is merged with the upsampled feature matrix S5 to form a 64*64*64 feature matrix S6, which is output through two convolutions;

[0026] The other feature matrix S3 is a skip connection, which is merged with the upsampled feature matrix S6 to form a 128*128*128 feature matrix S7, which is output through two convolutions.

[0027] The other feature matrix S2 is a skip connection, which is merged with the upsampled feature matrix S7 to form a 64*256*256 feature matrix S8, which is output through two convolutions.

[0028] Another path of the matrix feature S1 is the jump connection. This matrix will be merged with the up-sampled feature matrix S8 to form a 32*512*512 feature matrix and output through 2 convolutions.

[0029] The U-Net neural network uses the Adam gradient descent method as the optimizer, the MSEloss loss function as the cost function, and the ReLU activation function; the parameters in the U-Net neural network are gradually trained using the adaptive learning rate method, and the optimal parameters are obtained after multiple iterative cycles. The deep learning network with the optimal parameters is used as the intelligent inversion model;

[0030] Step 4: Input the satellite remote sensing data collected at the new time and use the intelligent inversion model to obtain the corresponding three-dimensional ocean temperature and salinity field;

[0031] The specific steps are:

[0032] First, the satellite remote sensing data SST, SSS, and SSH collected at the new time are marked with corresponding time tags according to the time to form a time matrix set;

[0033] Then, the satellite remote sensing data and the time matrix set are stacked to obtain the test sample data set T = {T1, T2, ..., T n};

[0034] Finally, the test sample data set T={T1,T2,...,T n}Directly input into the intelligent inversion model to obtain the temperature and salinity field below the sea surface.

[0035] The beneficial effects of the present invention are:

[0036] This invention discloses a method for inverting three-dimensional ocean temperature and salinity fields using satellite remote sensing data. Using a U-Net neural network, an intelligent inversion scheme with spatiotemporal learning characteristics is designed. This method effectively captures spatial characteristic patterns and mitigates the information loss problem of convolutional neural networks. Furthermore, the inclusion of a time variable enables the inversion model to learn temporal regularity. This method improves inversion accuracy while ensuring practical feasibility, enhances the value and efficiency of satellite remote sensing data, and exhibits good versatility. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flowchart of the steps of the intelligent algorithm for inverting the three-dimensional ocean temperature and salinity field using satellite remote sensing data;

[0038] Figure 2 This is a schematic diagram of the present invention inputting the training sample data set into the deep learning network in the form of 4 layers of images;

[0039] Figure 3 1 is a schematic diagram of the inversion model architecture in an embodiment of the present invention; DETAILED DESCRIPTION

[0040] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0041] The present invention provides a method for inverting the three-dimensional ocean temperature and salinity field using satellite remote sensing data. Figure 1 As shown, the following steps are included:

[0042] Step 1: For the ocean area to be processed, obtain historical satellite remote sensing data and ocean historical reanalysis data, and construct a training sample dataset and a training label dataset respectively;

[0043] Satellite remote sensing data include SST, SSS and SSH, each of which is a two-dimensional matrix;

[0044] Each satellite remote sensing data is labeled with a corresponding time label according to its own time to form a two-dimensional time matrix; the purpose of constructing the time matrix set is to add the time parameter to the neural network, thereby enabling the inversion model to have the ability to learn the regular characteristics of time.

[0045] Then, the two-dimensional matrix of the three satellite remote sensing data and the two-dimensional matrix of time are stacked and merged into a four-layer three-dimensional matrix, that is, the training sample data set S = {S1, S2, ..., S n};

[0046] At the same time, the temperature and salinity fields at the corresponding time in the historical reanalysis data are extracted to construct a training label dataset L = {L1, L2, ..., L n}.

[0047] Step 2: Build a deep learning network model based on the U-Net neural network;

[0048] The deep learning network model adopts the U-Net neural network model architecture, including: convolutional layers, skip connections, downsampling and upsampling.

[0049] Among them, the convolution layer includes convolution operation, batch normalization operation and activation function, downsampling is average pooling operation, upsampling is bilinear interpolation; the convolution kernel size is 3×3;

[0050] Feature extraction capability is the most important factor that determines the inversion effect; among the existing intelligent inversion methods, the convolutional neural network is the most effective in extracting sea surface parameter features. During the convolution process of the convolutional neural network, the neurons in the convolution layer are only connected to some neurons in the upper layer, which realizes local perception. Local perception can strengthen the feature relationship between close grid points and weaken the feature relationship between distant grid points. Global feature extraction can be achieved through the sliding of the convolution kernel and continuous convolution. The role of the pooling layer is to still be able to identify features through the average value when the parameters of the local grid points change, thereby improving the universality of the convolutional neural network. However, the convolutional neural network inevitably has the problem of information loss during the convolution process, especially the pooling process, and the lost detailed information will reduce the accuracy of the intelligent inversion. Therefore, based on the convolutional neural network, the present invention proposes an intelligent inversion method based on the U-Net neural network to solve the information loss problem existing in the traditional convolutional neural network.

[0051] The U-Net neural network is an encoder-decoder convolutional neural network. The first half of the neural network is used for feature extraction, while the second half is used for upsampling. The solution to information loss using the U-Net neural network is to establish skip connections between different convolutional layers, ensuring that all information in different convolutional layers is preserved during the upsampling process. This ensures feature extraction capabilities while preventing loss of detailed information.

[0052] Step 3: Use the training sample data set S = {S1, S2, ..., S n} and training label dataset L={L1,L2,...,L n}Train the deep learning network to obtain an intelligent inversion model.

[0053] The specific steps are:

[0054] First, the hyperparameters of the deep learning network are set according to actual needs: learning rate and number of training times, etc.

[0055] The learning rate uses an adaptive variable learning rate to adapt to different application requirements; the number of training times is mainly set directly based on the grid size.

[0056] Traditional inversion algorithms only utilize SST, SSS, and SSH. In fact, the ocean temperature and salinity field also has regular characteristics that change over time. This invention improves the inversion method by adding a time element, enabling the model to predict the changing patterns of time series.

[0057] The training sample data set S={S1,S2,...,S n}, input into the deep learning network in the form of 4 layers of pictures; Figure 2 shown.

[0058] like Figure 3 As shown, the input data is 4*512*512; after two convolutions, it becomes a 16*512*512 three-dimensional feature matrix S1; the matrix S1 will flow into two different paths:

[0059] One path is downsampling, where the number of horizontal grids is reduced by 2 times through pooling operation to 16*256*256, and then converted to 32*256*256 matrix feature S2 through 2 convolutions;

[0060] The matrix feature S2 continues to flow into two different paths: one is downsampling, where the number of horizontal grids is reduced by 2 times through the pooling operation to 32*128*128, and then becomes the matrix feature S3 of 64*128*128 through two convolutions;

[0061] The matrix feature S3 continues to flow into two different paths: one is downsampling, where the number of horizontal grids is reduced by 2 times through the pooling operation to become 64*64*64, and then becomes the 128*64*64 matrix feature S4 through two convolutions;

[0062] The matrix feature S4 continues to flow into two different paths: one is downsampling, where the number of horizontal grids is reduced by 2 times through the pooling operation, becoming a 128*32*32 matrix feature S5, which is output through two convolutions; the other is a skip connection, where the matrix feature S4 is merged with the upsampled feature matrix S5 to form a 64*64*64 feature matrix S6, which is output through two convolutions;

[0063] The other feature matrix S3 is a skip connection, which is merged with the upsampled feature matrix S6 to form a 128*128*128 feature matrix S7, which is output through two convolutions.

[0064] The other feature matrix S2 is a skip connection, which is merged with the upsampled feature matrix S7 to form a 64*256*256 feature matrix S8, which is output through two convolutions.

[0065] Another path of the matrix feature S1 is the jump connection. This matrix will be merged with the up-sampled feature matrix S8 to form a 32*512*512 feature matrix and output through 2 convolutions.

[0066] The jump connection contains both global feature information and original matrix information, thus solving the information loss problem in traditional convolutional neural networks. Repeating the above process and increasing the number of convolution layers can further improve the inversion effect.

[0067] The U-Net neural network uses the Adam gradient descent method as the optimizer, the MSEloss loss function as the cost function, and the ReLU activation function; the parameters in the U-Net neural network are gradually trained using the adaptive learning rate method, and the optimal parameters are obtained after multiple iterative cycles. The deep learning network with the optimal parameters is used as the intelligent inversion model;

[0068] Step 4: Input the satellite remote sensing data collected at the new time and use the intelligent inversion model to obtain the corresponding three-dimensional ocean temperature and salinity field;

[0069] The specific steps are:

[0070] First, the satellite remote sensing data SST, SSS, and SSH collected at the new time are marked with corresponding time tags according to the time to form a time matrix set;

[0071] Then, the SST, SSS, SSH data and the time matrix set of the newly collected time are stacked to obtain the test sample data set T = {T1, T2, ..., T n};

[0072] Finally, the test sample data set T={T1,T2,...,T n}Directly input into the intelligent inversion model to obtain the temperature and salinity field below the sea surface.

[0073] The present invention will be specifically described below with reference to a practical example:

[0074] Taking the remote sensing inversion of the temperature and salinity field of the South China Sea in 2016 as an example, the three-dimensional temperature and salinity field of the ocean is inverted using satellite remote sensing data. The satellite remote sensing data and ocean historical reanalysis data are from CMEMS. First, the monthly average satellite remote sensing data of the area to be processed from 1993 to 2015 are obtained, including SST, SSS and SSH. The corresponding time tags are marked according to the time to form a time matrix set. The SST, SSS, SSH and time matrix set are stacked to construct a training sample dataset S = {S1, S2, ..., S n}; extract the temperature and salinity fields from the historical reanalysis data at the corresponding time, and construct a training label dataset L = {L1, L2, ..., L n Then, a deep learning network model based on the U-Net neural network is constructed, including convolutional layers, skip connections, downsampling, and upsampling.

[0075] Using the training sample data set S={S1,S2,...,S n} and training label dataset L={L1,L2,...,L n}The deep learning network model is trained to obtain an intelligent inversion model; finally, the satellite remote sensing data at the time to be collected is input, and the three-dimensional ocean temperature and salinity field is obtained using the intelligent inversion model.

[0076] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for inverting the three-dimensional ocean temperature and salinity field using satellite remote sensing data, characterized in that: The specific steps include: Step 1: For the ocean area to be processed, historical satellite remote sensing data and ocean historical reanalysis data are obtained to construct a training sample dataset and a training label dataset respectively; The satellite remote sensing data include SST, SSS and SSH, each of which is a two-dimensional matrix. At the same time, each satellite remote sensing data is marked with a corresponding time tag according to its own time to form a time two-dimensional matrix. Then, the two-dimensional matrix of the three satellite remote sensing data and the time two-dimensional matrix are combined into a four-layer three-dimensional matrix, that is, a training sample data set S = {S1, S2, ..., S n }; At the same time, the temperature and salinity fields at the corresponding time in the historical reanalysis data are extracted to construct a training label dataset L = {L1, L2, ..., L n }; Step 2: Build a deep learning network model based on the U-Net neural network, including convolutional layers, skip connections, downsampling, and upsampling. By establishing skip connections in different convolutional layers, all information in different convolutional layers can be preserved during the upsampling process. Step 3: Use the training sample data set and the training label data set to train the deep learning network to obtain an intelligent inversion model; The specific steps are: First, the hyperparameters of the deep learning network are set manually according to actual needs: learning rate and number of training times; Then, the training sample data set S={S1,S2,...,S n }, input into the deep learning network in the form of 4 layers of images; The input data is 4*512*512; after two convolutions, it becomes a 16*512*512 three-dimensional feature matrix S1; this matrix S1 will flow into two different paths: One path is downsampling, where the number of horizontal grids is reduced by 2 times through pooling operation to 16*256*256, and then converted to 32*256*256 matrix feature S2 through 2 convolutions; The matrix feature S2 continues to flow into two different paths: one is downsampling, where the number of horizontal grids is reduced by 2 times through the pooling operation to 32*128*128, and then becomes the matrix feature S3 of 64*128*128 through two convolutions; The matrix feature S3 continues to flow into two different paths: one is downsampling, where the number of horizontal grids is reduced by 2 times through the pooling operation to become 64*64*64, and then becomes the 128*64*64 matrix feature S4 through two convolutions; The matrix feature S4 continues to flow into two different paths: one is downsampling, where the number of horizontal grids is reduced by 2 times through the pooling operation, becoming a 128*32*32 matrix feature S5, which is output through two convolutions; the other is a skip connection, where the matrix feature S4 is merged with the upsampled feature matrix S5 to form a 64*64*64 feature matrix S6, which is output through two convolutions; The other feature matrix S3 is a skip connection, which is merged with the upsampled feature matrix S6 to form a 128*128*128 feature matrix S7, which is output through two convolutions. The other feature matrix S2 is a skip connection, which is merged with the upsampled feature matrix S7 to form a 64*256*256 feature matrix S8, which is output through two convolutions. Another path of the matrix feature S1 is the jump connection. This matrix will be merged with the up-sampled feature matrix S8 to form a 32*512*512 feature matrix and output through 2 convolutions. Next, the Adam gradient descent method is used as the optimizer, the MSEloss loss function is used as the cost function, and the ReLU activation function is used. The adaptive learning rate method is used to gradually train the parameters of the U-Net neural network. After multiple iterative cycles, the optimal parameters are obtained, and the deep learning network with the optimal parameters is used as the intelligent inversion model. Step 4: Input the satellite remote sensing data collected at the new time and use the intelligent inversion model to obtain the corresponding three-dimensional ocean temperature and salinity field.

2. The method for inverting the three-dimensional ocean temperature and salinity field using satellite remote sensing data according to claim 1, characterized in that: The convolution layer includes a convolution operation, a batch normalization operation, and an activation function. Downsampling is an average pooling operation, and upsampling is a bilinear interpolation operation. The convolution kernel size is 3×3.

3. The method for inverting the three-dimensional ocean temperature and salinity field using satellite remote sensing data according to claim 1, characterized in that: The satellite remote sensing data collected at the new time is used to obtain the corresponding three-dimensional ocean temperature and salinity field using an intelligent inversion model; specifically: First, the satellite remote sensing data SST, SSS, and SSH collected at the new time are marked with corresponding time tags according to the time to form a time matrix set; Then, the satellite remote sensing data and the time matrix set are stacked to obtain the test sample data set T = {T1, T2, ..., T n }; Finally, the test sample data set T={T1,T2,...,T n Directly input into the intelligent inversion model to obtain the temperature and salinity field below the sea surface; Test sample data set T={T1,T2,...,T n }Input into the intelligent inversion model to obtain the temperature and salinity field below the sea surface.

Images