A method for predicting sea surface height changes based on multi-factor input deep convolutional networks

Abstract

This invention belongs to the field of marine element prediction technology and discloses a method for predicting sea surface height changes based on a multi-element input deep convolutional network. The method includes the following steps: S1, selecting several years of sea level anomaly (SLA) data, sea surface temperature (SST) data, and sea surface wind field data within a region to establish a historical mixed dataset; S2, using the sea level anomaly (SLA) data, sea surface temperature (SST) data, and sea surface wind field data to establish a sea surface environmental field, combining an attention mechanism to weightedly filter spatiotemporal information, and constructing a multi-element input sea surface height change prediction network (M-A-DNN); S3, training the sea surface height change prediction model using the historical mixed dataset, and inputting the data to be predicted into the model to obtain the prediction result. This invention organically combines the physical constraints of marine dynamic processes with deep learning methods, improving the accuracy of short-term sea level anomaly prediction and providing important reference information for sea state prediction and climate forecasting.

Description

Technical Field

[0001] This invention belongs to the field of marine element prediction technology, and specifically relates to a method for predicting sea surface height changes based on a multi-element input deep convolutional network. Background Technology

[0002] Changes in sea level height are a comprehensive reflection of various ocean dynamic processes. Therefore, based on changes in sea level height, ocean dynamic processes can be analyzed more intuitively. High-precision predictions of sea level height changes can also provide necessary data for forecasting the formation and dissipation of mesoscale eddies, thus supporting the study of ocean dynamic processes.

[0003] There are two main methods commonly used for predicting sea level height: modeling methods based on ocean dynamic processes (i.e., ocean numerical models) and statistical methods based on mathematical principles (i.e., data-driven models). Modeling methods based on ocean dynamic processes primarily establish a series of equations to describe the overall dynamic changes within an ocean region. This method typically analyzes nonlinear phenomena through multiple equations and simplifies complex equations using boundary conditions, thereby obtaining an ocean dynamic model for the entire region and achieving the goal of predicting sea level height. However, due to the large number of parameters affecting the prediction results and the highly complex interactions between these elements, existing theories cannot account for all nonlinear changes, making it difficult to construct highly accurate numerical models. Furthermore, the ocean dynamic environment varies significantly between different ocean regions, and the selection of conditions during modeling is often highly correlated with regional ocean conditions. Therefore, this modeling method exhibits strong regional characteristics and is difficult to apply to the global ocean. Statistical methods based on mathematical principles, on the other hand, mainly utilize mathematical and statistical principles to conduct data-level predictions by exploring the patterns of change in the data itself. The advantage of this method is faster modeling, making it suitable for operational needs. In addition, since its principle is based on the data itself, it does not need to consider the complexity of the marine environment in the region, thus enabling the establishment of a larger-scale sea surface height change prediction model.

[0004] This invention attempts to introduce correlations between sea surface elements based on deep learning algorithms, organically combining the two methods mentioned above to further improve the prediction accuracy of sea surface height changes (represented by sea level anomalies). This provides important reference information for sea state prediction and climate forecasting, and further supports fields such as ship navigation, fisheries resource prediction, marine engineering, and industry. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a sea surface height change prediction method based on a multi-factor input deep convolutional network. The method constructs a multi-factor input sea surface height change prediction network, introduces an attention mechanism and a feature preservation layer, and comprehensively considers the temporal and spatial relationships of the historical sea surface environment field, effectively improving the prediction accuracy of short-term sea level anomalies in the region.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A method for predicting sea surface height changes based on a multi-factor input deep convolutional network includes the following steps:

[0008] S1. Select several years of sea level anomaly (SLA) data, sea surface temperature (SST) data, and sea surface wind field data within the region to establish a historical mixed dataset;

[0009] S2. A sea surface environmental field is established using sea level anomaly SLA data, sea surface temperature SST data, and sea surface wind field data. The spatiotemporal information is weighted and filtered using an attention mechanism to construct a multi-factor input sea surface height change prediction network MA-DNN.

[0010] S3. Train the sea level height change prediction model using the historical mixed dataset; then input the data to be predicted into the model to obtain the prediction results.

[0011] Furthermore, the process of creating a historical mixed dataset in S1 includes the following steps:

[0012] S101. Unify the temporal and spatial resolutions of sea level anomaly SLA data, sea surface temperature SST data, and sea surface wind field data;

[0013] S102. Extract the model training region from the data;

[0014] S103. Mark land areas or areas with missing values ​​using -9999;

[0015] S104. The preprocessed regional data are iteratively connected according to the input time series, so that the data is stored in the form of a four-dimensional matrix. The first dimension is the number of samples, the second dimension is the time series, that is, the number of days of input data when the model is predicting, and the third and fourth dimensions are the latitude and longitude span of the selected data area.

[0016] Furthermore, the multi-factor input sea surface height change prediction network MA-DNN constructed in step S2 is divided into an internal network and an external network. The external network includes: environmental field data fusion layer, hybrid layer, spatial attention layer, temporal attention layer, SLA feature preservation layer, fusion layer and regression layer. The internal network includes three layers of MA-DNN units.

[0017] The input to the environmental field data fusion layer consists of time-series meridional wind speed, zonal wind speed, and sea surface temperature. The environmental field data fusion layer uses meridional and zonal wind speeds to calculate the sea surface wind stress curl, and superimposes it with the sea surface temperature field to obtain mixed sea surface environmental field data. The mixed sea surface environmental feature E0 is extracted by convolution, and environmental features E1, E2, and E3 at different scales are obtained by first-time pooling, second-time pooling, and third-time pooling, respectively.

[0018] The mixed environmental features E0 of the sea surface are spatially weighted and fused with the sea level height anomaly SLA data to obtain mixed field data D. Mixed features at different scales of the regional sea surface are obtained through training with three layers of MA-DNN units. Each MA-DNN unit includes a pooling layer, a spatial attention layer, a temporal attention layer, a convolutional layer, an SLA feature preservation layer, and an information fusion layer. In each MA-DNN unit, the receptive field is first increased through a pooling layer. Then, spatial attention and mixed attention are combined to perform spatiotemporal weighted filtering of the input sea surface mixed information. The attention-filtered sea surface mixed information is then processed through multiple convolutional layers to extract features, resulting in a spatiotemporal feature map of sea surface elements. Finally, the spatial features of the last day's SLA data (without attention weighting) are preserved and combined with the spatiotemporal feature map of sea surface elements to form a residual structure. The specific details of obtaining mixed features at different scales of the regional sea surface through the training of the three-layer MA-DNN units are as follows:

[0019] After passing through the first layer of MA-DNN units, the first-scale spatiotemporal hybrid feature F1 of the sea surface is obtained. F1 is then passed to the second layer of MA-DNN units to obtain the second-scale spatiotemporal hybrid feature F2 of the sea surface. F2 is then passed to the third layer of MA-DNN units to obtain the third-scale spatiotemporal hybrid feature F3 of the sea surface. The third-scale spatiotemporal hybrid feature F3 of the sea surface is then convolved and differencing with the environmental feature E3 of the same scale in the information fusion layer to obtain the preliminary predicted SLA feature distribution H3. After upsampling to reduce the receptive field, this H3 is passed to the information fusion layer of the previous layer of MA-DNN units. For The second-layer MA-DNN unit's information fusion layer connects the obtained spatial mixture feature F2 with the preliminary predicted SLA feature distribution H3 after the same scale of the lower-layer unit, and then performs convolution and difference operations with the second-scale environmental feature E2 to obtain the predicted SLA feature distribution H2. Similarly, H2 after the same scale is fed into the information fusion layer of the first-layer MA-DNN unit, connects the obtained spatial mixture feature F1 with H2, performs convolution and difference operations with the environmental feature E1 to obtain the predicted SLA feature distribution H1, and then upsamples it before feeding it into the fusion layer of the external network.

[0020] For the external network, in the hybrid layer, the mixed data information D is weighted and filtered using spatial attention and temporal attention, and then fed into the fusion layer. At the same time, the spatial details of the original SLA data are preserved and also fed into the fusion layer. Both are connected with the predicted SLA feature distribution H1 and fed into the regression layer. The loss is calculated with the spatial distribution of the SLA to be predicted, and the model parameters are updated.

[0021] Furthermore, in the MA-DNN unit, in the spatial attention layer, the mean and maximum values ​​of the data time series are superimposed and weighted to train the influence of the long-term average sea level and peak sea level in different regions on the spatial attention of the prediction results. The calculation process is shown in formula (1).

[0022] The temporal attention layer superimposes the mean and maximum values ​​of the data in the entire space at the same time, thereby training the weight of the influence of each time on the result. The calculation process is shown in formula (2).

[0023] output1 = Sigmoid(i1*(max) t (input)+mean t (input))) (1) output=Sigmoid((Relu(i2*max s (output1))+Relu(i3*mean s (output1)))) (2)

[0024] In the formula, i1, i2, i3 represent the weight matrices obtained during training, input represents the input to the attention layer, output and output1 represent the outputs of the spatial attention and temporal attention layers, respectively, and mean t and mean s The averaging is performed over the time and space dimensions, respectively, and the max is calculated. t and max s They are used to maximize the time and space dimensions respectively, with Sigmoid and ReLU being activation functions used to train the non-linearity of the weights.

[0025] Furthermore, in S3, training the model using the network follows these steps:

[0026] 1) During model training, 75% of the training set is used for training, and 25% is used as a validation set to validate the parameters of the iteration. The model with the smallest error in the validation set is saved.

[0027] 2) Use a custom loss function to train the model to reduce the impact of missing land values ​​on regional feature training. The specific formula for the loss function is as follows:

[0028]

[0029]

[0030] In the formula, F is the weighted variance function of points in the spatial region, and y i For the true value, is the model prediction value, and m is the total number of sample points in the region; the missing values ​​on land are marked with the minimum value -9999. Before calculating the loss value of each point in the two-dimensional space, a judgment is made. If the target value of the point is -9999, the impact of its loss value on the model weight update is ignored.

[0031] Furthermore, in step S3, when using the trained prediction model, the sea level anomaly for the predicted date is obtained by following these steps:

[0032] 1) First, mark the missing values ​​in the grid data of the prediction area, and then splice them according to the time series to obtain a four-dimensional matrix, so that the first dimension is the number of samples, the second dimension is the input duration, and the third and fourth dimensions are the latitude and longitude span of the selected data area;

[0033] 2) Input the data into the trained sea surface height change prediction model to obtain the prediction results.

[0034] Compared with the prior art, the advantages of the present invention are as follows:

[0035] This invention presents a sea surface height change prediction method based on a multi-factor input deep convolutional network. It achieves the fusion of multiple environmental factors and comprehensively considers the temporal and spatial correlation between historical sea surface environmental factors and sea surface height changes. By introducing an attention mechanism and feature preservation, it improves the prediction accuracy of short-term sea level anomalies, provides important reference information for sea state prediction and climate forecasting, and further provides data support for ship navigation, fishery resource prediction, marine engineering and industry. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0038] Figure 2 This is a diagram of the multi-factor input sea surface height change prediction network structure of the present invention.

[0039] Figure 3 This is a structural diagram of the MA-DNN unit of the present invention. Detailed Implementation

[0040] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0041] Example 1

[0042] This embodiment provides a method for predicting sea surface height changes based on a multi-factor input deep convolutional network, such as... Figure 1 As shown, it includes the following steps:

[0043] S1. Select several years of sea level anomaly (SLA) data, sea surface temperature (SST) data, and sea surface wind field data within the region to establish a historical mixed dataset.

[0044] Creating a historical mixed dataset involves the following steps:

[0045] S101. Unify the temporal and spatial resolutions of sea level anomaly SLA data, sea surface temperature SST data, and sea surface wind field data;

[0046] S102. Extract the model training region from the data;

[0047] S103. Mark land areas or areas with missing values ​​using -9999;

[0048] S104. The preprocessed regional data are iteratively connected according to the input time series, so that the data is stored in the form of a four-dimensional matrix. The first dimension is the number of samples, the second dimension is the time series, that is, the number of days of input data when the model is predicting, and the third and fourth dimensions are the latitude and longitude span of the selected data area.

[0049] S2. A sea surface environmental field is established using sea level anomaly SLA data, sea surface temperature SST data, and sea surface wind field data. The sea surface wind stress curl is calculated using the sea surface wind field data. The spatiotemporal information is weighted and filtered using an attention mechanism to construct a multi-factor input sea surface height change prediction network MA-DNN (Multi-Factor-InputAttention-mechanism Deep Neural Network).

[0050] S3. Train the sea level height change prediction model using the historical mixed dataset; then input the data to be predicted into the model to obtain the prediction results.

[0051] The following section details the multi-factor input sea surface height change prediction network MA-DNN constructed in this invention. Figure 2 This is the structure diagram of the multi-factor input deep convolutional network constructed in this invention. The network extracts features at different spatial scales through multiple pooling operations. Therefore, the resulting mixed environment field also needs to extract environmental features at different scales through pooling. Furthermore, to prevent information redundancy caused by the input environmental field data from leading to partial forgetting of historical sea level anomaly features by the model, an SLA feature preservation layer is added to retain spatial details of sea level height variation features. To maximize the preservation of mixed sea surface information, the attention mechanism in the hybrid network combines the mixed sea surface features of the sea surface environment field. Simultaneously, to prevent the model from losing sea level height variation features, attention weights for sea level anomaly features are trained separately before being fed into the network. The main body of the network obtains high-order features of regional sea level height variation through training with three layers of MA-DNN units. Figure 3 This is a structural diagram of the MA-DNN unit, a basic building block in the multi-factor input deep convolutional network constructed in this invention. In the unit, the input sea surface height change features are first pooled to obtain the spatial feature map of the next scale. The spatial and temporal weights are assigned through an attention mechanism. Features are extracted through convolutional layers and then fused and connected to the upper and lower layers respectively.

[0052] Combination Figure 2 and Figure 3As shown, the MA-DNN network for predicting sea surface height change based on multi-factor input consists of an internal network and an external network. The external network includes: an environmental field data fusion layer, a hybrid layer, a spatial attention layer, a temporal attention layer, an SLA feature preservation layer, a fusion layer, and a regression layer. The internal network consists of three MA-DNN units. The specific regression process is as follows:

[0053] 1) The input to the environmental field data fusion layer is divided into time series meridional wind speed, zonal wind speed and sea surface temperature. The environmental field data fusion layer uses meridional wind speed and zonal wind speed to calculate sea surface wind stress curl, and superimposes it with the sea surface temperature field to obtain mixed sea surface environmental field data. The mixed sea surface environmental feature E0 is extracted by convolution, and environmental features E1, E2 and E3 at different scales are obtained by first pooling, second pooling and third pooling respectively.

[0054] 2) The mixed environment features E0 of the sea surface and the SLA data of sea level anomalies are spatially weighted and fused to obtain the mixed field data D. The mixed features of the sea surface at different scales in the region are obtained by training a three-layer MA-DNN unit. Each MA-DNN unit includes a pooling layer, a spatial attention layer, a temporal attention layer, a convolutional layer, an SLA feature preservation layer, and an information fusion layer. In each MA-DNN unit, the receptive field is first increased by the pooling layer. Then, the input sea surface mixed information is spatiotemporally weighted and filtered by combining spatial attention and mixed attention to retain the effective information to the greatest extent. In the spatial attention layer, the mean and maximum values ​​of the data time series are superimposed and weighted to train the influence of the long-term average sea level and peak sea level in different regions on the spatial attention of the prediction results. The calculation process is shown in formula (1).

[0055] The temporal attention layer superimposes the mean and maximum values ​​of the data in the entire space at the same time, thereby training the weight of the influence of each time on the result. The calculation process is shown in formula (2).

[0056] output1 = Sigmoid(i1*(max) t (input)+mean t (input))) (1) output=Sigmoid((Relu(i2*max s (output1))+Relu(i3*mean s (output1)))) (2)

[0057] In the formula, i1, i2, i3 represent the weight matrices obtained during training, input represents the input to the attention layer, output and output1 represent the outputs of the spatial attention and temporal attention layers, respectively, and mean t and means The averaging is performed over the time and space dimensions, respectively, and the max is calculated. t and max s These are the maximum values ​​taken over the time and space dimensions, respectively. Sigmoid and ReLU are activation functions used to non-linearize the training weights. Relu(x) = max(0,x).

[0058] Then, the attention-filtered mixed sea surface information is processed through multiple convolutional layers to extract features, resulting in a spatiotemporal feature map of sea surface elements. To prevent information distortion caused by multiple convolutions, spatial details need to be preserved. Since the spatial distribution of the sea surface changes continuously over time, the spatial features of the last day's SLA data in the input data have the highest correlation with the SLA to be predicted. Therefore, the spatial features of the last day's SLA data without attention weighting are preserved and combined with the spatiotemporal feature map of sea surface elements to form a residual structure, thus preserving the spatial details of the original data to the maximum extent.

[0059] The three-layer MA-DNN unit training obtains mixed features of different scales of the regional sea surface as follows:

[0060] First, for the first layer MA-DNN unit, the first-scale sea surface spatiotemporal mixture feature F1 can be obtained by the above steps. F1 is then passed to the second layer MA-DNN unit, and the second-scale sea surface spatiotemporal mixture feature F2 is obtained through the same training steps. Similarly, F2 is passed to the third layer MA-DNN unit to obtain the third-scale sea surface spatiotemporal mixture feature F3. The spatial-temporal hybrid feature F3 of the third-scale sea surface and the environmental feature E3 of the same scale are convolved and differencing in the information fusion layer to obtain the preliminary predicted SLA feature distribution H3. After upsampling to reduce the receptive field, it is passed to the information fusion layer of the next layer MA-DNN unit. For the second layer MA-DNN unit, its information fusion layer connects the obtained spatial hybrid feature F2 with the preliminary predicted SLA feature distribution H3 of the next layer unit at the same scale, and then convolves and differencing it with the second-scale environmental feature E2 to obtain the predicted SLA feature distribution H2. Similarly, H2 at the same scale is passed to the information fusion layer of the first layer MA-DNN unit, and the obtained spatial hybrid feature F1 is connected with H2. Then, it is convolved and differencing it with the environmental feature E1 to obtain the predicted SLA feature distribution H1. After upsampling, it is passed to the fusion layer of the external network.

[0061] 3) For the external network, in the hybrid layer, the hybrid data information D is weighted and filtered using spatial attention and temporal attention, and then fed into the fusion layer. At the same time, the spatial details of the original SLA data are preserved and also fed into the fusion layer. The two are connected with the predicted SLA feature distribution H1 and fed into the regression layer. The loss is calculated with the spatial distribution of the SLA to be predicted, and the model parameters are updated.

[0062] It should be noted that in step S3 of this invention, training the model using the network follows these steps:

[0063] 1) During model training, 75% of the training set is used for training, and 25% is used as a validation set to validate the parameters of the iteration. The model with the smallest error in the validation set is saved.

[0064] 2) Use a custom loss function to train the model to reduce the impact of missing land values ​​on regional feature training. The specific formula for the loss function is as follows:

[0065]

[0066]

[0067] In the formula, F is the weighted variance function of points in the spatial region, and y i For the true value (target value), is the model prediction value, and m is the total number of sample points in the region; the land missing value is marked with the minimum value -9999. Before calculating the loss value of each point in the two-dimensional space, a judgment is made. If the target value of the point is -9999, the impact of its loss value on the model weight update is ignored.

[0068] In S3, when using the trained prediction model, the sea level anomaly for the predicted date is obtained by following these steps:

[0069] 1) First, mark the missing values ​​in the grid data of the prediction area, and then splice them according to the time series to obtain a four-dimensional matrix, so that the first dimension is the number of samples, the second dimension is the input duration, and the third and fourth dimensions are the latitude and longitude span of the selected data area;

[0070] 2) Input the data into the trained sea surface height change prediction model to obtain the prediction results.

[0071] In practical application, this invention selects historical data of the area to be predicted to establish a hybrid dataset, uses a multi-factor input sea surface height change prediction network to adjust parameters and train a sea surface height change prediction model, and uses historical sea level anomaly, sea surface temperature, and sea surface wind field data as inputs to the prediction model to obtain the sea surface height change prediction results.

[0072] Example 2

[0073] In this embodiment, the model training area is the South China Sea, ranging from 109.875°E to 119.625°E and 4.375°N to 20.125°N. Data from January 1, 1998 to December 31, 2018 (20 years) was collected to build the training set. Models were constructed to predict sea level anomalies for the next 1, 3, 5, and 7 days using 10 days of historical sea level data. This embodiment uses data from the same area from January 1, 2019 to June 3, 2020 to build a test set to evaluate the accuracy of the trained model. The test set format is consistent with the training set. In this embodiment, the average prediction accuracy (ACC), average linear correlation coefficient, and root mean square error (RMSE) of all samples in the test set are used to evaluate the model's prediction accuracy. The calculation methods are as follows:

[0074]

[0075]

[0076]

[0077] In the formula: m represents the total number of predicted days, and n represents the total number of points where the space ocean is located. This represents the predicted sea level anomaly value at point i in space on day t of the test set. This represents the true (target) value of the sea level anomaly at point i in space on day t of the test set. and These represent the average values ​​of the predicted and true values ​​over the ocean region up to day t, respectively.

[0078] This embodiment compares the accuracy of the multi-factor input sea surface height change prediction network MA-DNN constructed in this invention with the traditional multiple linear regression (MLR) method in sea surface height change prediction. The results are shown in Table 1:

[0079] Table 1 Comparison of Model Prediction Accuracy

[0080]

[0081] As shown in the table above, compared with the traditional MLR method, the multi-feature input depthwise convolutional network prediction method of the present invention has higher accuracy. Moreover, the accuracy improvement of the multi-feature input depthwise convolutional network prediction method of the present invention is even greater with the increase of prediction time.

[0082] In summary, by utilizing the technical solution described above, the correlation between sea surface elements can be introduced into the deep learning model, organically combining the physical constraints of ocean dynamic processes with the deep learning model to achieve higher-precision short-term sea surface height change prediction. This provides important reference information for sea state prediction and climate forecasting, and further provides data support for ship navigation, fisheries resource prediction, marine engineering, and industry.

[0083] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should be protected by the present invention.

Claims

1. A method for sea surface height change prediction based on multi-element input deep convolutional network, characterized in that, The method comprises the following steps: S1, selecting sea level anomaly SLA data, sea surface temperature SST data and sea surface wind field data in a region for a plurality of years to establish a historical mixed data set; S2, using the sea level anomaly SLA data, the sea surface temperature SST data and the sea surface wind field data to establish a sea surface environment field, combining an attention mechanism to weight and filter spatio-temporal information, and constructing a sea surface height change prediction network M-A-DNN with multiple element inputs; The multi-element input sea surface height change prediction network M-A-DNN constructed in step S2 is divided into an internal network and an external network, the external network comprises an environmental field data fusion layer, a mixed layer, a spatial attention layer, a temporal attention layer, an SLA feature reservation layer, a fusion layer and a regression layer, and the internal network comprises three M-A-DNN units; The environmental field data fusion layer inputs are time series zonal wind speed, meridional wind speed and sea surface temperature, the zonal wind speed and the meridional wind speed are used to calculate sea surface wind stress curl in the environmental field data fusion layer, and are superimposed with the sea surface temperature field to obtain sea surface mixed environmental field data, sea surface mixed environmental features E0 are extracted through convolution, and different scale environmental features E1, E2 and E3 are obtained through one-time pooling, two-time pooling and three-time pooling; The sea surface mixed environmental features E0 are spatially weighted and fused with the sea level anomaly SLA data to obtain mixed field data D, and the mixed features of different scales of the regional sea surface are obtained through training of three M-A-DNN units, wherein each M-A-DNN unit comprises a pooling layer, a spatial attention layer, a temporal attention layer, a convolution layer, an SLA feature reservation layer and an information fusion layer; the sea surface mixed information is first weighted and filtered in space and time by combining the spatial attention and the mixed attention through the pooling layer in each M-A-DNN unit to increase the receptive field, the sea surface mixed information filtered by the attention is then extracted through a plurality of convolution layers to obtain sea surface element spatio-temporal feature maps, and finally the spatial features of the last day SLA data without attention weighting are reserved, combined with the sea surface element spatio-temporal feature maps, and a residual structure is formed; the mixed features of different scales of the regional sea surface obtained through training of the three M-A-DNN units are as follows: After the first layer M-A-DNN unit, the first scale sea surface spatio-temporal mixed feature F1 is obtained, F1 is transmitted into the second layer M-A-DNN unit, the second scale sea surface spatio-temporal mixed feature F2 is obtained, F2 is transmitted into the third layer M-A-DNN unit, the third scale sea surface spatio-temporal mixed feature F3 is obtained, the third scale sea surface spatio-temporal mixed feature F3 and the environmental feature E3 of the same scale are subjected to convolution operation and difference in the information fusion layer, and the preliminary predicted SLA feature distribution H3 is obtained, which is transmitted into the information fusion layer of the upper layer M-A-DNN unit after being reduced in the receptive field by upsampling; for the second layer M-A-DNN unit, the obtained spatial mixed feature F2 is connected with the preliminary predicted SLA feature distribution H3 of the same scale after the lower unit, and then subjected to convolution operation and difference with the second scale environmental feature E2, and the predicted SLA feature distribution H2 is obtained; similarly, h2 is transmitted into the information fusion layer of the first layer M-A-DNN unit after being of the same scale, the obtained spatial mixed feature F1 is connected with H2, and subjected to convolution operation and difference with the environmental feature E1, and the predicted SLA feature distribution h1 is obtained, which is transmitted into the fusion layer of the external network after upsampling; For the external network, in the mixing layer, the mixed field data D is weighted and screened by using spatial attention and time attention, and is transmitted into the fusion layer, and the spatial detail information of the original SLA data is also transmitted into the fusion layer for feature reservation, and the two are connected with the predicted SLA feature distribution H1, and are transmitted into the regression layer, and the loss calculation is performed with the SLA spatial distribution to be predicted, and the model parameters are updated; S3, a sea surface height change prediction model is trained using a historical mixed data set; and the prediction result is obtained by inputting the to-be-predicted data into the model. 2.The sea surface height change prediction method based on multi-element input deep convolutional network according to claim 1, wherein, The historical mixed data set in S1 comprises the following steps: S101, unify the time resolution and the spatial resolution of the sea level anomaly SLA data, the sea surface temperature SST data and the sea surface wind field data; S102, cut the data in the model training area; S103, mark the land area or the missing value area using -9999; S104, iteratively connect the region data obtained by preprocessing respectively according to the input time sequence, so that the data is stored in the form of a four-dimensional matrix, the first dimension is the sample number, the second dimension is the time sequence, that is, the input data days when the model is predicted, and the third dimension and the fourth dimension are the latitude and longitude span of the selected data region. 3.The sea surface height change prediction method based on multi-element input deep convolutional network according to claim 1, characterized in that, In the M-A-DNN unit, in the spatial attention layer, the mean value and the maximum value of the data time sequence are superimposed and weighted, so as to train the influence of the average sea surface and the peak sea surface in different regions on the spatial attention of the prediction result, and the calculation process is shown in formula (1); The time attention layer superimposes the mean value and the maximum value of the data in the whole space at the same time, so as to train the influence weight of each time on the result, and the calculation process is shown in formula (2); output1 = Sigmoid(i1 * (max t (input) + mean t (input)))(1) output = Sigmoid((Relu(i2 * max s (output1)) + Relu(i3 * mean s (output1))))(2) where i1, i2, i3 represent the trained weight matrix, input represents the input of the attention layer, output and output1 represent the output of the spatial attention and temporal attention layer respectively, mean t and mean s are the average of the time dimension and the space dimension respectively, max t and max s are the maximum of the time dimension and the space dimension respectively, Sigmoid and Relu are activation functions for the non-linearization of the trained weights.

4. The method of claim 1-3, wherein, In the S3, the network is used to train the model, which follows the following steps: 1) In the model training process, 75% of the training set is used for training, and 25% is used as a validation set to verify the parameters of iterations, and the model with the minimum validation set error is saved; 2) Use a custom loss function to train the model to reduce the impact of land missing values on regional feature training. The loss function formula is as follows: In the formula, F is a weighted variance function of the spatial region points, y i is a true value, is a model predicted value, m is the total number of region sample points; the minimum value -9999 is used to mark the land missing value, and before calculating the loss value of each point in the two-dimensional space, if the target value of the point is -9999, the influence of the loss value on the model weight update is ignored.

5. The method of claim 1-3, wherein, In S3, when using the trained prediction model, the following steps are used to obtain the sea level anomaly of the date to be predicted: 1) Mark the missing values of the prediction area grid data first, and splice according to the time sequence to obtain a four-dimensional matrix, so that the first dimension is the sample number, the second dimension is the input time length, and the third and fourth dimensions are the latitude and longitude span of the selected data area; 2) Input the data into the trained sea surface height change prediction model to obtain the prediction result.

Images