Intelligent inversion method and device for marine three-dimensional temperature and salinity field based on deep learning

Abstract

The present application relates to a kind of ocean three-dimensional temperature and salinity field intelligent inversion method and equipment based on deep learning, belong to ocean three-dimensional temperature and salinity field information and deep learning technical field.The present application carries out the time-space pretreatment to the abnormal value and derived variable of multi-source sea surface observation data, extracts spatial feature by inputting spatial data into spatial dimension branch network, extracts time feature by inputting time data into time dimension branch network, extracts spatial feature and time feature and inputs into space-time feature fusion module fusion, obtains inversion result after fusion by a fully connected layer.Through double branch architecture, respectively, the time and spatial dimension features of sea surface data are deeply mined, and the ability of representing complex ocean process is enhanced by improved transformer block, and the present application can significantly improve the inversion precision and stability of ocean three-dimensional temperature and salinity field, and provide high-quality data support for aquatic ecological protection.

Description

TECHNICAL FIELD

[0001] The present application relates to a deep learning-based intelligent inversion method and device for marine three-dimensional temperature and salinity field, and belongs to the technical field of marine three-dimensional temperature and salinity field information and deep learning. BACKGROUND

[0002] Seawater temperature and salinity are key parameters for studying internal oceanic processes, maintaining the health and stability of the marine ecosystem. About 30% of the heat penetrates into the seawater below 700 meters every year, and even a small change in seawater temperature will affect the ocean-atmosphere circulation. At the same time, seawater salinity plays a key role in global water cycle regulation by regulating the density structure of the ocean and the movement pattern of the water body. The change in seawater temperature and salinity not only affects the evaporation and precipitation process, but also further affects the wind direction, wind speed and other climate elements, thereby having an important influence on global climate change and hydrological cycle. In addition, the spatial gradient of seawater temperature and salinity is an important driving factor for the formation of ocean currents, and the movement of ocean currents helps to maintain the balance of the marine ecosystem. Therefore, accurately detecting the temperature and salinity structure inside the ocean is of great importance to marine science and climate change research.

[0003] At present, due to environmental complexity and vast sea area, the spatial distribution of buoys that can obtain subsurface temperature and salinity data is not uniform enough, especially in high latitude areas, and the observation interval is discontinuous, which limits the in-depth study of the ocean. Satellite remote sensing can provide a large amount of long-term continuous and high-resolution data for the ocean, however, satellite remote sensing can only obtain surface ocean data.

[0004] Since the change in the characteristics of the ocean surface is mainly due to the dynamic process inside the ocean, there is a significant correlation between the two, and a nonlinear mapping relationship between the sea surface parameters and the subsurface temperature and salinity field can be established to achieve high-precision inversion from two-dimensional sea surface information to three-dimensional ocean structure.

[0005] In the existing intelligent inversion method of three-dimensional temperature and salinity field, common deep learning models include fully connected neural network (FCNN), multi-scale convolutional neural network (CNN), long short-term memory neural network (LSTM) and Transformer. However, these methods have certain limitations in spatiotemporal feature extraction: FCNN can effectively model the nonlinear relationship between data, but cannot capture the spatiotemporal correlation of input data; multi-scale convolutional neural network can fully extract local spatial features, but ignores time information; LSTM and Transformer are good at modeling time series dependency, but have relatively insufficient spatial feature extraction capability. In addition, common methods usually ignore the vertical relationship between data, and build a model for each depth separately, resulting in a large number of models and slow training speed. SUMMARY

[0006] The purpose of the present application is to overcome the problems of insufficient utilization of spatial and temporal information and low training efficiency of the existing three-dimensional temperature and salinity field intelligent inversion method, and to propose a marine three-dimensional temperature and salinity field intelligent inversion method based on deep learning, which can fully utilize spatial and temporal information, and simultaneously invert temperature anomalies or salinity anomalies at different depths, effectively capture vertical correlation, and significantly speed up the training speed.

[0007] The technical scheme adopted by the present application is:

[0008] The marine three-dimensional temperature and salinity field intelligent inversion method based on deep learning comprises the following steps:

[0009] S1. Spatial and temporal data collection: collect satellite remote sensing data and marine three-dimensional temperature and salinity data, obtain multi-source sea surface observation data anomaly values including sea surface height, sea surface salinity, sea surface temperature, sea surface wind speed east component, sea surface wind speed north component, and three-dimensional temperature and salinity data anomaly values by subtracting climatic characteristics; calculate the sea surface temperature anomaly difference DSSTA and the sea surface salinity anomaly difference DSSSA as derived variables according to the existing data;

[0010] S2. Spatial and temporal data preprocessing:

[0011] The multi-source sea surface observation data anomaly values and the derived variables are subjected to spatial and temporal preprocessing, and the data is subjected to sub-regional sliding block processing by using the interlaced and intercolumnar method to obtain spatial data, and the data is stacked in the time dimension to obtain time data in the time sequence;

[0012] S3. Input the spatial data into the spatial dimension branch network to extract the spatial features, and input the time data into the time dimension branch network to extract the time features: in the spatial dimension branch, the data is first subjected to the processing of the spatial feature enhancement block, and then input into the spatial Transformer block, in the spatial Transformer block, the data is sequentially subjected to the channel multi-head attention module and the spatial multi-head attention module to comprehensively extract the spatial features, and then subjected to the convolution feedforward network to further enhance the output spatial features; in the time dimension branch, the data is first subjected to the preprocessing of the time feature enhancement block, and then transmitted into the Transformer block to extract the time features;

[0013] S4. The extracted spatial features and time features are input into the spatiotemporal feature fusion module for fusion, and the fusion result is obtained through a fully connected layer;

[0014] S5. Take the three-dimensional temperature and salinity anomaly values as the true observation values, use the inversion result as the predicted value to calculate the loss function, and train and optimize the model to obtain the final inversion model for inversion.

[0015] In the above method, the calculation formula of the sea surface temperature anomaly difference DSSTA in step S1 is:

[0016] ,

[0017] ,

[0018] ,

[0019] where Y represents year, M represents month, N represents years, D represents depth layer, SST represents sea surface temperature, ST represents subsurface temperature and climatology represents climatic characteristics; the multi-year average of sea surface temperature and sea water temperature at different depths are calculated respectively by month, and the difference is obtained by subtracting the latter from the former different M,D , (M = 1, 2, 3, …, 12, D = 1, 2, 3, …, 33). The sea surface temperature of the specific M month is subtracted from different M,D , the sea surface temperature difference of depth layer D is obtained, and the sea surface temperature difference is subtracted from the climatic characteristics to obtain DSSTA.

[0020] The calculation formula of sea surface salinity anomaly difference DSSSA is:

[0021] ,

[0022] ,

[0023] ,

[0024] where Y represents year, M represents month, N represents years, D represents depth layer, SSS represents sea surface salinity, SS represents subsurface salinity and climatology represents climatic characteristics; the multi-year average of sea surface salinity and sea water salinity at different depths are calculated respectively by month, and the difference is obtained by subtracting the latter from the former different' M,D , (M = 1, 2, 3, …, 12, D = 1, 2, 3, …, 33). The sea surface salinity of the specific M month is subtracted from different' M,D , the sea surface salinity difference of depth layer D is obtained, and the sea surface salinity difference is subtracted from the climatic characteristics to obtain DSSSA.

[0025] The sliding block processing in step S2 is to crop the entire research area according to the set window size (2I-1) x (2I-1) x C. The sliding block starts from the top left corner and moves pixel by pixel in the horizontal and vertical directions. Each time the sub-area with a size of I x I x C is extracted from the window by the interlaced column extraction method, until the entire research area is traversed.

[0026] The spatial feature enhancement block in step S3 uses two parallel convolution layers to capture horizontal and vertical features respectively, and integrates the outputs through a residual connection, and then uses a depth separable convolution layer to extract higher-level spatial features. The extracted features are connected with the original input processed by a 1x1 convolution layer through a residual connection to obtain the enhanced output.

[0027] In the temporal feature enhancement block in step S3, the input features are first added after being processed by causal convolution and causal dilated convolution respectively, then the long-distance dependency information that the causal dilated convolution fails to capture is further extracted by a bidirectional long short-term memory network (BiLSTM), and finally the current output is connected with the original input processed by a 1x1 convolution layer through a residual connection to obtain the final feature representation. This design helps the subsequent Transformer module to better understand the temporal dynamics in the sequence to capture interactions in a larger time range.

[0028] In the channel multi-head attention in step S3, the input feature map X E R I×I×C is mapped to H different subspaces by a 1x1 convolution layer, and the maximum feature and the average feature of each channel are calculated by a global maximum pooling layer (GMP) and an average pooling layer (GAP) to generate two vectors M 1 ，M 2 ∈R C / H×1 are concatenated and input into a shared multi-layer perceptron to learn the attention weights, adaptively identify the key channels, and add the output features to obtain the final attention weight vector M 1 ，M 2 . M E R C / H×1 The Sigmoid activation function is used to ensure that the weights are between 0 and 1, and the M shape is 1x1x C / H multiplied by the original feature map to obtain the weighted channel feature map. The outputs of multiple attention heads are spliced and the channel dimension is transformed by a 1x1 convolution layer to produce the final output O E R I×I×C .

[0029] In the spatial multi-head attention in step S3, the input feature map X E R I×I×C is mapped to H subspaces by a 1x1 convolution layer, and maximum pooling (MP) and average pooling (AP) are performed along the channel dimension to obtain two spatial information vectors S 1 ，S 2 ∈R I×I , the S 1 and S 2 Splicing along the channel dimension obtains S 3 ∈R I×I×2 , S 3 Contains different pooling information of each spatial position, uses a convolutional layer to S 3 Feature fusion and mapping are carried out to generate spatial attention weights S E R I×I×1 The generated spatial attention weights are applied to the original feature map to weight each position, and a weighted feature map is obtained, the outputs of multiple attention heads are spliced and transformed by a convolutional layer to output.

[0030] The transformer block for extracting time features in step S3 includes multi-head self-attention, feedforward neural network, residual connection and layer normalization processing.

[0031] The convolutional feedforward network includes a 1x1 convolutional layer and a 3x3 depth separable convolutional layer.

[0032] In the space-time feature fusion module in step S4, the spatial features are expanded in channels by a convolutional layer, and then merged into a comprehensive sea surface vector by a global maximum pooling operation, the time features are flattened and adjusted in channel number by a fully connected layer, and weighted by a channel attention layer, the two branch data are spliced in channel dimension to form a new variable, and the new variable is fused by a fully connected layer and a channel attention module to realize the fusion of time and space information.

[0033] The deep learning-based marine three-dimensional temperature and salinity field intelligent inversion electronic device comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, and the processor implements the deep learning-based marine three-dimensional temperature and salinity field intelligent inversion method as described above when executing the program.

[0034] The deep learning-based marine three-dimensional temperature and salinity field intelligent inversion storage device has a computer program stored thereon, and the program is executed to implement the steps of the deep learning-based marine three-dimensional temperature and salinity field intelligent inversion method as described above.

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

[0036] (1) The application takes abnormal value as input data, focuses on non-seasonal signals that are difficult to capture, introduces derivative variables (sea surface temperature anomaly difference / sea surface salinity anomaly difference) as model inputs to reveal the vertical correlation of temperature and salinity structure at different depths, improves the problem of insufficient response of traditional methods to subsurface signals, can simultaneously invert temperature anomalies or salinity anomalies at different depths, effectively captures the vertical relationship, significantly improves the model training speed, and improves the inversion accuracy of the model.

[0037] (2) The application deeply excavates the time and space dimension features of sea surface data through a double-branch architecture. In the spatial dimension branch, a spatial Transformer block (STB) is designed, a spatial-channel multi-head attention mechanism and a convolution feedforward network are designed, the ability of the transformer to process two-dimensional spatial data is expanded, the spatial correlation of seawater temperature and salinity is fully considered, and the ability to represent complex ocean processes is enhanced. At the same time, the time and space feature enhancement module (TFEBs / SFEBs) is proposed, which further enhances the inversion ability of the model to subsurface seawater temperature and salinity by learning multi-scale features.

[0038] (3) The application adopts an interlaced spatial data extraction method, which retains key information and expands the receptive field without changing the input size. The spatio-temporal feature fusion module is composed of convolution layers, fully connected layers and channel attention mechanisms, which is used to process and fully fuse the features output by the double-branch.

[0039] (4) The application can accurately invert the subsurface temperature or salinity profile at a depth of 5-2000m using only sea surface remote sensing data. By simultaneously inverting the temperature and salinity fields at 5-2000m, the model can effectively introduce and utilize vertical information. The model effectively learns the complex nonlinear relationship between sea surface remote sensing data and subsurface data, accurately inverts the subsurface seawater temperature and salinity anomalies, and effectively fills the gap in subsurface observation data. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 The method flowchart of the application;

[0041] Figure 2 The model architecture diagram of the application;

[0042] Figure 3 The structure diagram of the spatial feature enhancement block;

[0043] Figure 4 The structure diagram of the channel multi-head attention module;

[0044] Figure 5 The structure diagram of the spatial multi-head attention module;

[0045] Figure 6 is a convolutional feedforward network structure diagram;

[0046] Figure 7 is a time feature enhancement block structure diagram;

[0047] Figure 8 is a feature fusion module structure diagram;

[0048] Figure 9 is a channel attention module structure diagram;

[0049] Figure 10 is a spatial data sampling method;

[0050] Figure 11 is a spatial distribution of the inverted STA (a), the observed STA (b) and the difference (c);

[0051] Figure 12 is a spatial distribution of the inverted SSA (a), the observed SSA (b) and the difference (c). DETAILED DESCRIPTION

[0052] The application will be further described below with reference to specific embodiments.

[0053] Embodiment 1: A deep learning-based intelligent inversion method for marine three-dimensional temperature and salinity field, comprising the following steps (such as Figure 1 ):

[0054] S1. Spatio-temporal data collection:

[0055] Collect satellite remote sensing data (sea surface height, sea surface salinity, sea surface temperature, sea surface wind speed east component, sea surface wind speed north component) and marine three-dimensional temperature and salinity data, and obtain the anomaly values of the sea surface variables and three-dimensional temperature and salinity data by subtracting the climatic characteristics, focusing on the non-seasonal signals that are difficult to capture. The temperature and salinity can be obtained by adding the climatic characteristics to the inverted temperature and salinity anomaly values. Using these anomaly data as the training basis can improve the inversion accuracy of the model.

[0056] Delete grid points with missing values. If a grid point is null in any month, delete all data for that grid point. The data used in the present application are all processed into monthly average data with a resolution of 0.25°x0.25°. The monthly average values can be calculated from daily data. The data are unified to a 0.25°x0.25° grid using the bilinear difference method.

[0057] Standardize the calculated anomaly values to ensure that the mean of the data set is 0 and the variance is 1, in order to eliminate the magnitude differences between different data.

[0058] To make full use of the vertical correlation between the sea surface observation data and the subsurface temperature and salinity, the application introduces derivative variables: sea surface temperature anomaly difference / sea surface salinity anomaly difference (DSSTA / DSSSA), which are obtained only from existing data without additional data collection. They are used as inputs together with sea surface anomaly data to improve the accuracy of subsurface sea water temperature and salinity anomaly inversion. The calculation formula of the sea surface temperature anomaly difference DSSTA is:

[0059] ,

[0060] ,

[0061] ,

[0062] Where: Y represents the year, M represents the month, N represents the number of years, D represents the depth layer, SST represents the sea surface temperature, ST represents the subsurface temperature, and climatology represents the climatological characteristics; the multi-year average values of the sea surface temperature and the sea water temperature at different depths are calculated respectively by month, and the difference between the former and the latter is obtained different M,D , (M=1,2,3,….,12, D=1,2,3,…,33). The sea surface temperature of the specific M month is subtracted from different M,D , that is, the sea surface temperature difference DSST of the depth layer D is obtained, and the sea surface temperature difference is subtracted from the climatological characteristics to obtain DSSTA.

[0063] The calculation formula of the sea surface salinity anomaly difference DSSSA is:

[0064] ,

[0065] ,

[0066] ,

[0067] Where: Y represents the year, M represents the month, N represents the number of years, D represents the depth layer, SSS represents the sea surface salinity, SS represents the subsurface salinity, and climatology represents the climatological characteristics; the multi-year average values of the sea surface salinity and the sea water salinity at different depths are calculated respectively by month, and the difference between the former and the latter is obtained different' M,D , (M=1,2,3,….,12, D=1,2,3,…,33). The sea surface salinity of the specific M month is subtracted from different' M,D , that is, the sea surface salinity difference DSSS of the depth layer D is obtained, and the sea surface salinity difference is subtracted from the climatological characteristics to obtain DSSSA.

[0068] S2. Spatiotemporal data preprocessing:

[0069] Three-dimensional temperature and salinity anomalies were selected as the actual observed values, and anomalies in multi-source sea surface observation data (sea surface height anomalies, sea surface salinity anomalies, sea surface temperature anomalies, eastern component anomalies of sea surface wind speed anomalies, and northern component anomalies of sea surface wind speed anomalies) were used as the model inputs. These data were then processed as necessary to meet the model input requirements.

[0070] For each local region of (2I-1)×(2I-1)×C centered on the true value, an I×I×C sub-region is extracted using an alternating row and column method as the input for the model space branch, such as... Figure 10 This extraction method preserves key information and expands the receptive field without changing the input size; sub-regions with more than half of the invalid values ​​are removed, and sub-regions with less than half of the missing values ​​are filled with 0.

[0071] The sea surface observation anomaly data from the previous T months are stacked to form time series data, which is used as input to the time series branch of the model.

[0072] S3. Input spatial data into the spatial dimension branch network to extract spatial features, and input temporal data into the temporal dimension branch network to extract temporal features:

[0073] like Figure 2 The overall network consists of two main branches: the spatial dimension branch focuses on extracting the spatial dimension features of sea surface data, and the temporal dimension branch focuses on extracting the temporal dimension features of sea surface data.

[0074] In the spatial dimension branch, data is first processed by the spatial feature enhancement block before being input into the spatial Transformer block. Residual connections are used to enhance the model's generalization ability. Figure 3 This is a diagram of the spatial feature enhancement block structure. The spatial feature enhancement block adopts a branching structure, containing convolutional layers with different kernel sizes and depthwise separable convolutional layers. The model uses 1×5 and 5×1 convolutional layers to capture lateral and vertical features respectively, and integrates the output through residual connections. Then, a 3×3 depthwise separable convolutional layer is used to extract higher-level spatial features. To fuse finer-grained information, the current output is residually connected to the original input after 1×1 convolution processing to obtain the final output. Furthermore, batch normalization layers and the ReLU activation function are used to optimize performance.

[0075] In the spatial Transformer block, the input data first passes through the channel multi-head attention module and the spatial multi-head attention module sequentially to comprehensively extract spatial features. These features are then further enhanced by a convolutional feedforward network. Figure 4 This is a structural diagram of a channel multi-head attention module. In the channel multi-head attention module, the input feature map... X E R I×I×CMapped to via a 1×1 convolutional layer H In three distinct subspaces, a Global Max Pooling (GMP) layer and a Global Average Pooling (GAP) layer are used to compute the maximum and average features for each channel, generating two vectors. M 1 ，M 2 ∈R C / H×1 .Will M 1 ，M 2 The inputs are fed into a shared multilayer perceptron to learn attention weights and adaptively identify key channels. The two output features are then summed to obtain the final attention weight vector. M ∈R C / H×1 Using the Sigmoid activation function ensures the weights are between 0 and 1. Adjust... M The shape is 1×1× C / H The weighted channel feature map is obtained by multiplying it with the original feature map. Then, the outputs of multiple attention heads are concatenated and the channel dimensions are transformed through a 1×1 convolutional layer to produce the final output. O E R I×I×C This process can be represented as:

[0076] ,

[0077] ,

[0078] in: X h and Head h Representing the h Input and output of size. represent Sigmoid Activation function. * represents a weighted summation operation. This represents the convolution operation. W and b This represents the weights and biases of the last convolutional layer.

[0079] Figure 5 This is a structural diagram of the spatial multi-head attention module. The spatial multi-head attention is applied to the input feature map. X E R I ×I×C Mapped to a 1×1 convolutional layer H Each subspace. Max pooling (MP) and average pooling (AP) are performed along the channel dimension to obtain two spatial information vectors. S 1 ，S 2 ∈R I×I. will be S 1 and S 2 along the channel dimension S 3 ∈R I×I×2 , S 3 contains different pooling information for each spatial position. A convolutional layer is used to S 3 perform feature fusion and mapping to generate attention weights S E R I×I×1 . The generated spatial attention weights are applied to the original feature map to weight each position, highlighting important regions. The process for each head can be represented as:

[0080] ,

[0081] where: f 7×7 represents a two-dimensional convolutional layer with a kernel size of 7 and an output channel of 1.

[0082] Figure 6 is a convolutional feedforward network structure diagram. The convolutional feedforward network includes a 1x1 convolutional layer and a 3x3 depthwise separable convolutional layer.

[0083] In the time dimension branch, the input variable is first preprocessed by the time feature enhancement block, and then transmitted into the transformer block. Figure 7 is a time feature enhancement block structure diagram, which combines causal dilated convolution and bidirectional long short-term memory network (BiLSTM) in the time feature enhancement block; the input features are first processed by causal convolution and causal dilated convolution respectively and then added, and then the long-distance dependence information that the causal dilated convolution cannot capture is further extracted by bidirectional long short-term memory network (BiLSTM). Finally, the current output is connected in residual with the original input processed by 1x1 convolution to obtain the final feature representation. This design helps the subsequent Transformer module to better understand the temporal dynamics in the sequence, so as to capture interactions in a larger time range. Causal dilated convolution is a special form of convolution that expands the receptive field by using a dilation factor to increase the spacing between convolution kernel elements, and can ensure that the output only depends on the historical input and is not affected by future data. For input sequence X = {X1, X2, …, X N} and convolution kernel K = {k1, k2, …, k C}, each element y i in the output sequence Y = {y1, y2, …, y N ' operation can be represented as:

[0084] ,

[0085] where: C represents the convolution time step, d represents the dilation factor, which makes the model can track longer time series data.

[0086] The causal dilated convolution layer with dilation factors of 1 and 2 respectively is used to capture the local dependence in the input sequence. BiLSTM is used to process the long distance dependence that the causal dilated convolution fails to capture.

[0087] The Transformer block has efficient parallel processing capability, and can capture long-term and short-term information in time series data through self-attention mechanism, and focus on key parts in the sequence for weight allocation. The Transformer block includes multi-head self-attention, feedforward neural network, residual connection and layer normalization.

[0088] S4. The extracted spatial and temporal features are input into the space-time feature fusion module for fusion, and the fusion result is obtained through a fully connected layer:

[0089] Figure 8 is a feature fusion module, after the spatial feature is expanded in channel through two convolution layers, it is combined into a comprehensive sea surface vector through global maximum pooling operation. The time feature is flattened and adjusted in channel number through a fully connected layer, and is weighted in importance by a channel attention layer. The data processed by the two branches is spliced in the channel dimension to form a new variable. The variable is fused through two fully connected layers and a channel attention module to realize the fusion of time and space information to produce the final output. Figure 9 is a channel attention module structure diagram, the channel attention score in the channel attention module is calculated by two fully connected layers with different output channel numbers, the attention score is multiplied by the input feature to obtain the weighted channel feature map, then the weighted feature is processed through a fully connected layer to output the final inversion result.

[0090] S5. Take the three-dimensional temperature and salinity anomaly value as the true observation value, use the inversion result as the prediction value to calculate the loss function, and train and optimize the model to get the final inversion model for inversion:

[0091] (1) Data division, the preprocessed data is randomly divided into training set, validation set and test set according to the ratio of 7:2:1;

[0092] (2) Model training, the training set is used to train the network model, and the validation set is used to monitor and verify the performance of the training process;

[0093] (3) Loss calculation, the root mean square error is used as the loss function to calculate the error between the inversion result and the true value;

[0094] (4) Optimize the performance of the model, adopt the early stopping strategy, select the best inversion model on the validation dataset, and use the Adam optimizer to adjust the learning process of the ocean subsurface temperature and salinity field inversion model;

[0095] (5) Model testing, input the sea surface remote sensing observation data of the test dataset into the trained spatiotemporal fusion model, and detect the generalization performance of the trained network;

[0096] (6) Inversion result visualization.

[0097] Figure 11 and Figure 12 respectively show the spatial distribution of the model inversion results, the true observation values and the differences thereof at different depths in October 2021. Figure 11 The comparative results of seawater temperature anomalies at different depths are given. In the range of 100m to 1000m, the determination coefficient R² value of the model is between 0.62 and 0.82, and the average absolute error is less than 0.5℃, which can effectively reflect the temperature anomaly distribution characteristics of the equatorial and mid-latitude regions and accurately depict the spatial structure of the cold and warm water masses. With the increase of depth, the subsurface anomaly signal of seawater gradually weakens. At the depth of 2000m, the anomaly signal basically disappears, and the temperature field tends to be stable. Figure 12 The spatial distribution of seawater salinity anomalies at different depths is shown. At the depth of 100m, there is a low salinity anomaly area in the northeast Pacific, and the SSA value is-0.4 psu; the northwest Pacific is affected by the Kuroshio, and a high salinity anomaly of 0.4 psu appears. At this depth, the R² value of the model inversion result is 0.54, and the average absolute difference is 0.07 psu. At the depth of 300m, affected by the extension area of the Kuroshio, the area near 20°N-40°N and 130°E-170°E shows obvious salinity anomaly. At the depths of 600m and 1000m, the spatial distribution of salinity anomaly is similar to that at 300m. The average absolute difference of SSA at these depths is 0.027 psu, 0.014 psu and 0.009 psu respectively, and the R² value is 0.53, 0.32 and 0.40 respectively. At the depth of 2000m, the anomaly signal basically disappears, and the absolute average difference is 0.003 psu and the R² value is 0.33.

[0098] Overall, the model proposed in the present application exhibits good performance and can accurately capture the spatial characteristics of subsurface seawater temperature and salinity anomalies. The model not only has excellent inversion ability in shallow water areas, but also maintains high accuracy in deep sea areas. Therefore, the model exhibits wide applicability and reliability.

[0099] Embodiment 2: A deep learning-based intelligent inversion electronic device for ocean three-dimensional temperature and salinity field, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the deep learning-based intelligent inversion method for ocean three-dimensional temperature and salinity field as described in Embodiment 1.

[0100] A deep learning-based intelligent inversion storage device for ocean three-dimensional temperature and salinity field, having a computer program stored thereon, wherein execution of the program implements the steps of the deep learning-based intelligent inversion method for ocean three-dimensional temperature and salinity field as described in Embodiment 1.

[0101] The above is a further description of the application in combination with the embodiments, and the protection scope of the application is not limited thereto.

Claims

1. A deep learning-based intelligent inversion method for three-dimensional marine temperature and salinity fields, characterized by: The steps include the following: S1. Spatiotemporal data collection: Collect satellite remote sensing data of the sea surface and three-dimensional temperature and salinity data of the ocean. By subtracting climatological features, obtain outliers of multi-source sea surface observation data, including sea surface height, sea surface salinity, sea surface temperature, eastern component of sea surface wind speed, and northern component of sea surface wind speed, as well as outliers of three-dimensional temperature and salinity data. Based on existing data, sea surface temperature anomaly difference DSSTA and sea surface salinity anomaly difference DSSSA were calculated as derived variables. S2. Spatiotemporal data preprocessing: Outliers and derived variables from multi-source sea surface observation data are preprocessed in a spatiotemporal manner. Spatial data are obtained by extracting sub-regions from these data in a way that alternates rows and columns in local areas. These data are then stacked in the time dimension to obtain time series data. S3. Spatial data is input into the spatial dimension branch network to extract spatial features, and temporal data is input into the temporal dimension branch network to extract temporal features: In the spatial dimension branch, the data is first processed by the spatial feature enhancement block, and then input into the spatial Transformer block. In the spatial Transformer block, the data passes through the channel multi-head attention module and the spatial multi-head attention module in sequence to comprehensively extract spatial features, and then passes through the convolutional feedforward network to further enhance the output spatial features; In the temporal dimension branch, the data is first preprocessed by the temporal feature enhancement block, and then input into the Transformer block to extract temporal features. S4. Input the extracted spatial and temporal features into the spatiotemporal feature fusion module for fusion. After fusion, the inversion result is obtained by passing through a fully connected layer. S5. Take the three-dimensional temperature and salinity anomalies as the true observation values, use the inversion results as the predicted values ​​to calculate the loss function, and train and optimize the model to obtain the final inversion model for inversion.

2. The intelligent inversion method for three-dimensional marine temperature and salinity fields based on deep learning according to claim 1, characterized in that, The formula for calculating the sea surface temperature anomaly difference (DSSTA) mentioned in step S1 is as follows: ， ， ， Where: Y represents the year, M represents the month, N represents the number of years, D represents the depth layer, SST represents sea surface temperature, ST represents subsurface temperature, and climatology represents climate characteristics; the multi-year average of sea surface temperature and seawater temperature at different depths is calculated separately for each month, and the difference is obtained by subtracting the latter from the former. different M,D M=1,2,3,….,12, D=1,2,3,…,33, subtract the sea surface temperature of the specific month M from the given value. different M,D That is, the sea surface temperature difference DSST at depth layer D is obtained, and the DSSTA is obtained by subtracting the climate characteristics from the sea surface temperature difference. The formula for calculating the sea surface salinity anomaly difference (DSSSA) is: ， ， ， Where: Y represents the year, M represents the month, N represents the number of years, D represents the depth layer, SSS represents sea surface salinity, SS represents subsurface salinity, and climatology represents climate characteristics; the multi-year average of sea surface salinity and seawater salinity at different depths is calculated separately for each month, and the difference is obtained by subtracting the latter from the former. different' M,D M=1,2,3,….,12, D=1,2,3,…,33, subtract the sea surface salinity from the specific month M. different' M,D That is, the sea surface salinity difference (DSSS) at depth D is obtained, and the climatic characteristics are subtracted from the sea surface salinity difference to obtain the DSSSA.

3. The intelligent inversion method for three-dimensional marine temperature and salinity fields based on deep learning according to claim 1, characterized in that, The slider processing described in step S2 involves cropping the entire study area according to a set window size of (2I-1)×(2I-1)×C. The slider starts from the upper left corner and moves pixel by pixel along the horizontal and vertical directions. Each move extracts a sub-region of size I×I×C from the window by means of alternating rows and columns, until the entire study area is traversed.

4. The intelligent inversion method for three-dimensional marine temperature and salinity fields based on deep learning according to claim 1, characterized in that, The spatial feature enhancement block described in step S3 uses two parallel convolutional layers to capture horizontal and vertical features respectively, and integrates the output through residual connections. Then, a depthwise separable convolutional layer is used to extract higher-level spatial features. The extracted features are then residually connected with the original input after 1×1 convolution processing to obtain the enhanced output.

5. The intelligent inversion method for three-dimensional marine temperature and salinity fields based on deep learning according to claim 1, characterized in that, In the temporal feature enhancement block described in step S3, the input features are first processed by causal convolution and causal dilation convolution respectively and then added together. Subsequently, the long-distance dependency information that the causal dilation convolution failed to capture is further extracted by a bidirectional long short-term memory network. Finally, the current output is residually connected with the original input after 1×1 convolution to obtain the final feature representation.

6. The intelligent inversion method for three-dimensional marine temperature and salinity fields based on deep learning according to claim 1, characterized in that, In step S3, the channel multi-head attention method involves inputting the feature map. X∈R I×I×C Mapped to via a 1×1 convolutional layer H In each different subspace, the maximum and average features of each channel are calculated using a global max pooling layer and an average pooling layer, generating two vectors. M 1 ，M 2 ∈R C / H×1 ,Will M 1 ，M 2 The input is fed into a shared multilayer perceptron to learn attention weights, adaptively identify key channels, and the output features are summed to obtain the final attention weight vector. M∈R C / H×1 Using the Sigmoid activation function ensures that the weights are between 0 and 1, and adjusts them accordingly. M The shape is 1×1× C / H Multiplying the original feature map yields a weighted channel feature map. The outputs of multiple attention heads are concatenated and the channel dimensions are transformed through a 1×1 convolutional layer to produce the final output. O∈R I×I×C .

7. The intelligent inversion method for three-dimensional marine temperature and salinity fields based on deep learning according to claim 1, characterized in that, In the spatial multi-head attention described in step S3, the input feature map X∈R I×I×C Mapped to a 1×1 convolutional layer H In each subspace, max pooling and average pooling are performed along the channel dimension to obtain two spatial information vectors. S 1 ，S 2 ∈R I×I ,Will S 1 and S 2 By splicing along the channel dimension S 3 ∈R I×I×2 , S 3 It contains different pooling information for each spatial location, and uses a convolutional layer to... S 3 Feature fusion and mapping are performed to generate spatial attention weights. S∈R I×I×1 The generated spatial attention weights are applied to the original feature map to weight each position, resulting in a weighted feature map. The outputs of multiple attention heads are then concatenated and transformed through a convolutional layer.

8. The intelligent inversion method for three-dimensional marine temperature and salinity fields based on deep learning according to claim 1, characterized in that, In step S4, the spatial features in the spatiotemporal feature fusion module are expanded by the convolutional layer and then merged into a comprehensive sea surface vector through global max pooling. The temporal features are flattened and then the number of channels is adjusted by the fully connected layer and weighted by the channel attention layer. The two branches of data are concatenated in the channel dimension to form a new variable. The new variable is fused with temporal and spatial information through the fully connected layer and the channel attention module.

9. A deep learning-based intelligent inversion electronic device for three-dimensional marine temperature and salinity fields, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the program, it implements the deep learning-based intelligent inversion method for three-dimensional marine temperature and salinity fields as described in any one of claims 1-8.

10. A deep learning-based intelligent inversion and storage device for three-dimensional marine temperature and salinity fields, which stores computer programs, characterized in that... When the program is executed, the steps of the deep learning-based intelligent inversion method for three-dimensional marine temperature and salinity fields as described in any one of claims 1-8 are implemented.

Images