Method and system for inverting ocean three-dimensional temperature structure based on sea surface data

Abstract

The invention discloses a method and system for inverting an ocean three-dimensional temperature structure based on sea surface data, and relates to the technical field of remote sensing, and the method comprises the steps: obtaining the reanalysis data of a target sea area, and carrying out the data preprocessing; according to the processed reanalysis data, training the ocean three-dimensional temperature reconstruction model to obtain a trained ocean three-dimensional temperature reconstruction model; the ocean three-dimensional temperature reconstruction model is a reconstruction model of an ocean three-dimensional temperature field constructed by selecting a U-net network; in the training process of the ocean three-dimensional temperature reconstruction model, a loss function sensitivity test and a temperature field inversion model sensitivity test based on a U-net network are adopted to optimize the performance of the ocean three-dimensional temperature reconstruction model; and according to the real-time sea surface data of the target sea area, constructing an ocean three-dimensional temperature structure based on the trained ocean three-dimensional temperature reconstruction model. According to the invention, the reconstruction precision of the ocean three-dimensional temperature field can be improved, and rapid and real-time temperature field reconstruction can be realized.

Description

Technical Field

[0001] The present application relates to the field of remote sensing technology, and in particular to a method and system for inverting the three-dimensional temperature structure of the ocean based on sea surface data. Background Art

[0002] In recent years, with the development of ocean satellite technology, ocean surface observation data has also seen a huge increase, and data on various characteristics of the ocean surface have become increasingly abundant. Ocean big data can provide us with more comprehensive ocean information and provide more data for the analysis of ocean characteristics.

[0003] However, it is difficult for ocean remote sensing observations to penetrate the ocean surface to obtain changes in underwater temperature structures, and most ocean dynamic processes are not limited to the surface, but can usually extend to areas thousands of meters deep underwater, so ocean remote sensing data cannot provide us with comprehensive ocean information. The lack of continuous and high-temporal and spatial resolution underwater data makes it difficult for us to understand the subtle and rapid changes in underwater features. Summary of the invention

[0004] The purpose of this application is to provide a method and system for inverting the three-dimensional temperature structure of the ocean based on sea surface data, which can improve the reconstruction accuracy of the three-dimensional temperature field of the ocean and realize fast and real-time reconstruction of the three-dimensional temperature field of the ocean.

[0005] To achieve the above objectives, this application provides the following solutions:

[0006] In a first aspect, the present application provides a method for inverting a three-dimensional ocean temperature structure based on sea surface data, comprising:

[0007] Obtain reanalysis data for the target sea area.

[0008] The reanalysis data are preprocessed to obtain processed reanalysis data; the processed reanalysis data include a training set and a validation set divided according to a set ratio, and a test set.

[0009] According to the processed reanalysis data, the ocean three-dimensional temperature reconstruction model is trained to obtain a trained ocean three-dimensional temperature reconstruction model; the ocean three-dimensional temperature reconstruction model is a reconstruction model of the ocean three-dimensional temperature field constructed using the U-net network; during the training process of the ocean three-dimensional temperature reconstruction model, a loss function sensitivity test and a temperature field inversion model sensitivity test based on the U-net network are used to optimize the performance of the ocean three-dimensional temperature reconstruction model.

[0010] According to the real-time sea surface data of the target sea area, based on the trained ocean three-dimensional temperature reconstruction model, an ocean three-dimensional temperature structure is constructed; the ocean three-dimensional temperature structure is represented by an ocean three-dimensional temperature field.

[0011] In a second aspect, the present application provides a system for inverting the three-dimensional temperature structure of the ocean based on sea surface data, comprising:

[0012] The data acquisition module is used to obtain the reanalysis data of the target sea area.

[0013] The data preprocessing module is used to perform data preprocessing on the reanalysis data to obtain processed reanalysis data; the processed reanalysis data includes a training set and a validation set divided according to a set ratio, and a test set.

[0014] A model training module is used to train an ocean three-dimensional temperature reconstruction model according to the processed reanalysis data to obtain a trained ocean three-dimensional temperature reconstruction model; the ocean three-dimensional temperature reconstruction model is a reconstruction model of the ocean three-dimensional temperature field constructed using a U-net network; during the training process of the ocean three-dimensional temperature reconstruction model, a loss function sensitivity test and a temperature field inversion model sensitivity test based on a U-net network are used to optimize the performance of the ocean three-dimensional temperature reconstruction model.

[0015] The ocean three-dimensional temperature structure construction module is used to construct an ocean three-dimensional temperature structure according to the real-time sea surface data of the target sea area and based on the trained ocean three-dimensional temperature reconstruction model; the ocean three-dimensional temperature structure is represented by an ocean three-dimensional temperature field.

[0016] The present application provides a method and system for inverting the three-dimensional temperature structure of the ocean based on sea surface data, the method comprising: obtaining reanalysis data of the target sea area; performing data preprocessing on the obtained reanalysis data to generate a processed data set, the data set comprising a training set, a validation set and a test set divided according to a predetermined ratio; training the three-dimensional temperature reconstruction model of the ocean based on the processed data set to obtain a trained model; the three-dimensional temperature reconstruction model of the ocean is constructed using a U-net network architecture, and is intended to reconstruct the three-dimensional temperature field of the ocean; during the model training process, the model performance is optimized through loss function sensitivity tests and temperature field inversion model sensitivity tests based on the U-net network; using real-time sea surface data of the target sea area, based on the trained three-dimensional temperature reconstruction model of the ocean, the three-dimensional temperature structure of the ocean is constructed; the structure is expressed through the three-dimensional temperature field of the ocean. The present application aims to improve the reconstruction accuracy of the three-dimensional temperature field of the ocean and realize fast and real-time reconstruction of the three-dimensional temperature field of the ocean. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 A schematic flow chart of a method for inverting the three-dimensional temperature structure of the ocean based on sea surface data provided in one embodiment of the present application.

[0019] Figure 2 A flowchart of a method for inverting three-dimensional ocean structure from sea surface data based on a U-net deep learning model is provided in accordance with an embodiment of the present application.

[0020] Figure 3 A schematic diagram of the U-net deep learning framework structure provided in one embodiment of the present application.

[0021] Figure 4 A schematic diagram of a MAE loss function provided in an embodiment of the present application.

[0022] Figure 5 A schematic diagram of the MSE loss function provided in one embodiment of the present application.

[0023] Figure 6 This is a scatter plot of the numerical distribution of the reconstructed temperature field and the original temperature field provided in an embodiment of the present application.

[0024] Figure 7 This is a statistical graph of the RMSE percentage of the reconstructed ocean three-dimensional temperature field provided in an embodiment of the present application.

[0025] Figure 8 A graph showing the percentage of RMSE of the ocean three-dimensional temperature field reconstructed in different intervals provided in an embodiment of the present application.

[0026] Fig. 9 A typical warm vortex reconstructed temperature field and original data temperature field vortex latitudinal profile distribution diagram provided in an embodiment of the present application.

[0027] Fig.10 A typical cold vortex reconstructed temperature field and a vortex latitudinal profile distribution diagram of the original data temperature field provided in an embodiment of the present application. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0029] The vertical distribution of the three-dimensional temperature field determines the location of characteristic structures such as the ocean mixed layer, thermocline and barrier layer. The three-dimensional temperature field of the ocean significantly affects the air-sea interaction, ocean circulation and climate change. The ocean subsurface temperature is one of the important indicators of climate change. There are many dynamic processes of various temporal and spatial scales in the ocean, which have an important effect on the thermodynamic state parameters such as density, temperature, salinity of seawater, marine ecosystems and global climate change, which will significantly change the distribution of the three-dimensional temperature field of the ocean.

[0030] Accurately grasping the three-dimensional temperature structure of the ocean is crucial for the study of oceanography. At present, there is an urgent need for continuous and high-temporal and spatial resolution ocean three-dimensional data. Field observation is the most effective way to obtain three-dimensional ocean data. For the observation of ocean data, ocean observation data is obtained by scientific research ships, buoys and other observation methods, but the field observation method is not enough to achieve real-time description of the changes in the temperature structure of the ocean, and it is difficult to accurately reflect the changing laws of the temperature structure of the ocean. The numerical simulation method is to obtain ocean environmental data by establishing a regional model. However, this method is generally applicable to specific sea areas, has certain limitations, and has high requirements for data and computing resources, and is difficult to achieve business applications. In recent years, with the development of marine satellite technology, ocean surface observation data has also ushered in a huge increase, and data on various characteristics of the ocean surface are becoming increasingly abundant. Big data of the ocean can provide more comprehensive ocean information for this application and provide more information for the analysis of ocean characteristics. However, it is difficult for ocean remote sensing observations to penetrate the ocean surface to learn about the changes in the underwater temperature structure, and most ocean dynamic processes are not limited to the surface, but usually extend to areas with a depth of thousands of meters underwater, so ocean remote sensing data cannot provide comprehensive ocean information for this application. The lack of continuous and high-temporal and spatial resolution underwater data makes it difficult for this application to understand the fine and rapid changes in underwater features. Therefore, how to obtain ocean data that meets small and medium scales is still a problem that needs to be solved in oceanographic research. If the correlation between sea surface data and underwater data can be established, and the sea surface data is used to invert and reconstruct the underwater environment, and relevant underwater data is obtained, it will help this application to have a more comprehensive understanding of the ocean dynamic process.

[0031] The methods for reconstructing underwater temperature fields based on sea surface data mainly include statistical methods, dynamic models and artificial intelligence methods. The statistical method is to establish the correlation between sea surface variables and underwater parameters through mathematical statistics. This method is suitable for the study of ocean areas with relatively stable ocean dynamic changes, but for sea areas with more complex dynamic environments, the reconstruction results usually have large errors. The dynamic model is developed based on the theory of ocean dynamics. Through idealized dynamic equations, the dynamic relationship between ocean surface variables and underwater parameters is obtained. Although this method can reflect the actual dynamic background of the ocean to a certain extent, the constructed model is usually only applicable to specific sea areas, and it is difficult to accurately reconstruct the temperature field of all periods and sea areas. In recent years, the rise of artificial intelligence methods has provided new ideas for the reconstruction of ocean temperature fields. Artificial intelligence models can accurately reconstruct underwater temperature fields by learning the relevant characteristics of sea surface variables, and artificial intelligence methods have good applicability and can be applied to the reconstruction of temperature fields in different periods and regions. Therefore, by combining deep learning data with satellite data, giving full play to the advantages of satellite data and deep learning methods, and constructing the correlation between sea surface variables and underwater parameters, it may be possible to achieve the reconstruction of the temperature field of the sea area, providing important data support for this application to deeply understand the ocean dynamic phenomena.

[0032] The Northwest Pacific and China's coastal waters are part of the northern Pacific Ocean, including the seas in the range of 100°-180°E, 20°N-65°N. As one of the largest regions in the Pacific Ocean, the temperature structure of the Northwest Pacific and China's coastal waters is affected by a variety of cyclones and monsoons. The seasonal changes in the marine environment are obvious, and there is a vast ocean space, which provides favorable conditions for the occurrence and evolution of ocean dynamic processes such as mesoscale eddies. However, due to the lack of vertical data, the study of mesoscale eddies is still in the exploratory stage. Therefore, it is very necessary to reconstruct the three-dimensional temperature field of the ocean. The deep learning algorithm has a strong feature extraction and learning ability. It can fully learn the characteristics of the data under the condition of a large amount of data input, and can further improve the reconstruction accuracy of the model by adjusting the input data and the model, providing new ideas and methods for the study of ocean temperature structure. The Pacific Ocean is an important area for the exchange of heat and salt energy between the tropical and subtropical Pacific Oceans. The energy exchange in the Northwest Pacific Pacific Ocean will directly affect the marine environment of my country's coastal waters. Therefore, studying the changing laws of the dynamic and thermal processes in the Northwest Pacific and China's coastal waters is of great value to my country's climate, ecology, fisheries, shipping and other aspects.

[0033] By inverting the underwater temperature structure, we can better understand the formation and change laws of deep ocean circulation systems and vertical mixing processes. At the same time, the underwater temperature of the ocean is closely related to global climate change. By reconstructing the subsurface temperature of the ocean, we can better understand the characteristics and trends of climate change and provide a reference for predicting future climate change. The distribution of subsurface temperature in the ocean is closely related to ocean circulation, which is of great significance for studying the changing laws of the ocean circulation system and deeply understanding the mechanism of ocean movement. In addition, the three-dimensional temperature structure of the ocean is closely related to dynamic phenomena such as mesoscale eddies. Compared with large-scale ocean phenomena, the research on small and medium-scale processes such as mesoscale eddies and ocean internal waves often faces greater challenges. This is because the changes in small and medium-scale dynamic processes in space and time scales are often small, the frequency of changes is higher, and the structure is more refined. Therefore, compared with large-scale ocean phenomena, the research on small and medium-scale dynamic processes puts forward higher technical requirements. Small and medium-scale dynamic processes are very frequent, and the demand for three-dimensional temperature data with high temporal and spatial resolution in related research is very urgent. Therefore, the reconstruction of ocean temperature information will provide more help for this application to better understand the phenomena such as mesoscale eddies in the ocean.

[0034] Most of the existing ocean temperature field reconstruction models use a single sea surface variable as input, which makes the model only learn the characteristics of one sea surface during reconstruction, and cannot fully learn the characteristics of the ocean, which in turn leads to certain limitations in the accuracy of reconstruction. Under the existing data conditions, how to further improve the reconstruction accuracy of the ocean temperature field through data processing and input is a problem that needs to be solved urgently. Deep learning algorithms can be well applied to the inversion of the three-dimensional structure of underwater temperature in the ocean. In addition, different input data and experimental models will also affect the results of reconstruction. In current related research, it is difficult to achieve long-term and large-scale continuous observation of the ocean due to objective conditions. Ocean satellites provide comprehensive sea surface data for this application, but effective underwater observation information is still difficult to obtain. If high-precision underwater information can be obtained, this application will be able to have a better understanding of the multi-scale dynamic process of the ocean. With the explosive growth of satellite data, artificial intelligence provides new ideas and methods for the acquisition of ocean information. Using a deep learning model, learn the characteristics of sea surface data, and use underwater label data to establish a connection with underwater features, so that underwater parameters can be reconstructed using sea surface data. By adjusting the combination of model input variables and label data, high-precision reconstruction results are obtained.

[0035] The purpose of this application is to provide a method and system for inverting the three-dimensional temperature structure of the ocean based on sea surface data, which can improve the reconstruction accuracy of the three-dimensional temperature field of the ocean and realize fast and real-time reconstruction of the three-dimensional temperature field of the ocean, thereby dynamically updating the three-dimensional temperature field of the ocean.

[0036] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0037] Embodiment 1

[0038] like Figure 1 As shown, this embodiment provides a method for inverting the three-dimensional temperature structure of the ocean based on sea surface data, the method comprising:

[0039] Step 101: Obtain reanalysis data of the target sea area.

[0040] Step 102: performing data preprocessing on the reanalysis data to obtain processed reanalysis data; the processed reanalysis data includes a training set and a validation set divided according to a set ratio, and a test set.

[0041] Step 103: According to the processed reanalysis data, the ocean three-dimensional temperature reconstruction model is trained to obtain a trained ocean three-dimensional temperature reconstruction model; the ocean three-dimensional temperature reconstruction model is a reconstruction model of the ocean three-dimensional temperature field constructed by using a U-net network; during the training process of the ocean three-dimensional temperature reconstruction model, a loss function sensitivity test and a temperature field inversion model sensitivity test based on a U-net network are used to optimize the performance of the ocean three-dimensional temperature reconstruction model.

[0042] Step 104: constructing an ocean three-dimensional temperature structure according to the real-time sea surface data of the target sea area and based on the trained ocean three-dimensional temperature reconstruction model; the ocean three-dimensional temperature structure is represented by an ocean three-dimensional temperature field.

[0043] In some embodiments, when executing step 101, the specific steps may be as follows:

[0044] The acquired sea surface data of the target sea area includes sea surface temperature data, salinity data, mixed layer depth data and sea surface height data.

[0045] Specifically, the input data and label data used are GLORYS12V1 reanalysis data, and GLORYS12V1 products are reanalysis data of CMENS global ocean eddy resolution (spatial resolution is 1 / 12 °, time resolution is 1d, and vertical stratification is 50 layers), covering altimeter data (after 1993). Mainly based on real-time global CMENS system, observation data are assimilated by price reduction Kalman filtering. Including altimeter data, sea surface temperature data, sea ice data and vertical profile assimilation of temperature and salinity observed in the field. This product includes daily average data of sea surface temperature data, salinity data, mixed layer depth data and sea surface height data. The reanalysis data time range used in the present embodiment is from January 1, 2010 to December 31, 2018, a total of 3357 samples.

[0046] In some embodiments, when executing step 102, the specific steps may be as follows:

[0047] like Figure 2 As shown, the data preprocessing includes a data normalization processing method, a nearest interpolation method, a NaN value processing method, and a data set division method.

[0048] Specifically, before the sea surface data and label data are input into the model, all sample data need to be preprocessed. First, the spatial resolution of the sea surface input data (sea surface data) and the subsurface temperature data must be unified. Through the nearest interpolation method, the spatial resolution of the GLORYS reanalysis data is interpolated to a 0.25° grid, and the temporal resolution remains at 1d.

[0049] The data set of the temperature field reconstruction model is then divided. The data time range used in this application is from January 1, 2010 to December 31, 2018. Among them, 2992 samples from 2010 to 2017 are used to train the model, obtain the optimal reconstruction parameters, and are divided into training set and validation set in a ratio of 8:2; the temperature data in 2018 is used as a test set to evaluate the accuracy of the model's inversion in the reconstruction of the ocean subsurface temperature structure.

[0050] In order to explore the impact of different processing methods of sea surface data on the reconstruction of underwater temperature field, the input data of the model will be obtained through different processing methods. Among them, for the reconstruction of underwater temperature field, the reconstruction results of the original value and the abnormal value as the input variable are compared. The original value is the interpolated sea surface temperature (SST) and sea surface height (SSH) data in the GLORYS reanalysis data; the abnormal value is the sea surface temperature anomaly (SSTA) and sea surface height anomaly (SLA) obtained by subtracting the average value of the original value from the original value. The calculation formula is as follows:

[0051] SSTA=SST-SST mean(1).

[0052] SLA = SSH-SSH mean (2).

[0053] In addition, since different variable data have differences in magnitude, in order to avoid such differences affecting the accuracy of the reconstruction results, before the data is input into the model, it is necessary to perform unified normalization processing to make different variables in the same magnitude and dimension, so that all variables are distributed in the same value range. The normalization expression is:

[0054]

[0055] In the formula, x is the initial value of the sea surface variables in all data, μ and σ are their mean and standard deviation respectively, and X is the normalized sea surface variable. After normalization, all sea surface variables will be concentrated in the numerical range with a mean of 0 and a variance of 1.

[0056] After normalization, the invalid values ​​(NotANumber, NAN) corresponding to the terrain need to be further processed. Experiments have shown that setting the NaN value to 0 can minimize the interference of the NaN value on the reconstruction result. In this study, after comparing multiple experimental results, the NAN value caused by the terrain is set to 0.

[0057] In some embodiments, before executing step 103, the following may be performed:

[0058] Based on the U-net deep learning network, an ocean three-dimensional temperature reconstruction model is constructed; the U-net deep learning network adopts an encoder-decoder symmetrical structure; the encoder part consists of a convolution layer and a pooling layer, which is used to extract features of input data in the ocean three-dimensional temperature reconstruction model through downsampling and pooling; the decoder part consists of a convolution layer and a pooling layer, which is used to reconstruct the features provided by the encoder through upsampling and deconvolution to generate a high-resolution three-dimensional temperature field image.

[0059] Specifically, Figure 3As shown in the figure, the U-net network is used to build a reconstruction model of the ocean three-dimensional temperature field. The U-net network is developed from the CNN network, and its main structure is composed of the CNN network. Compared with CNN, the structure of the U-net network can retain more data details and can extract context information for feature fusion. U-net is an improved fully convolutional network that uses an image or a complete data matrix as input, and outputs a result of the same size as the input data through a symmetrical upsampling and downsampling process. U-net is mainly used for image segmentation tasks. Since 2015, the U-net network has been widely used in the field of biomedical image segmentation. U-net is an end-to-end semantic segmentation network, which was initially valued for its ability to better solve biomedical problems.

[0060] U-net is an encoder-decoder symmetrical structure. The whole model is U-shaped and is mainly divided into two parts: encoder and decoder. The structure on the left is the encoder, which consists of convolutional layers and pooling layers. Its main functions are downsampling and pooling. It is mainly used for feature extraction and is divided into multiple layers. The input data is downsampled through a series of convolutional layers and pooling layers to gradually reduce the resolution. The encoder is used to capture the high-level features of the input data. Therefore, through the convolution operation, the model can learn the local features in the data. The right part is the decoder, where the input data is gradually restored to the original spatial resolution through a series of upsampling and deconvolution operations. The role of the decoder is to generate prediction results based on the high-level features provided by the encoder. The decoder has a similar structure to the encoder, except that the decoder uses deconvolution operations to gradually increase the resolution of the data. In addition, there are fully connected layers in the U-net network, including intermediate connections (Bridge) and skip connections (Skip Connections). Among them, the intermediate connection is the channel connecting the encoder and the decoder. Its role is to retain high-level feature information so that it can be used in the decoder to more accurately restore details and shapes. The intermediate connection ensures that the decoder has sufficient contextual information when generating predictions, because the last few layers of the encoder usually contain information about the overall image structure; the skip connection is directly connected from a certain layer of the encoder to the corresponding layer of the decoder. Such a connection retains a higher resolution feature map and helps to refine the segmentation results. Through this connection, the model can integrate information at different scales to better capture local details and global context. The skip connection also helps to alleviate the problem of gradient disappearance because the gradient can be more easily passed back to the earlier layer through the skip connection. Both direct connections and skip connections are designed to solve the problem of information loss and gradient disappearance, which enables the U-net network to simultaneously utilize high-level abstract features and detail information.

[0061] The operation process of the U-net framework is as follows: data is input into the model, enters the convolution layer, and performs two convolution operations. Then, the convolutional data is downwardly pooled; the pooled data enters the next convolution layer, and the same process is repeated four times in total. This process is performed by the left part of the framework and is mainly used to obtain the feature information of the data; after that, the right part of the framework performs the upsampling process: first, the upsampling part is used to perform upsampling calculations on the downsampling results, and the results are spliced ​​with the corresponding downsampling results on the left to form data with twice the number of channels as the downsampling results. Then, the number of data channels is reduced to the same as the downsampling through convolution calculations, and the results are upsampled again. The same process is repeated four times. The calculation process is mainly used to increase the resolution of the data and achieve accurate positioning of the data.

[0062] In some embodiments, when executing step 103, the specific steps may be as follows:

[0063] The loss function sensitivity test design is as follows:

[0064] The loss function is a calculation function used to measure the difference between the model's predicted value and the true value. It is an important criterion for evaluating the quality of the output results. It is a non-negative real-valued function. In deep learning operations, the smaller the loss function, the better the model fit.

[0065] The loss function is mainly used in the model training stage. After each batch of training data is sent to the model, the predicted value is output through forward propagation, and then the loss function calculates the difference between the predicted value and the true value, that is, the loss value. After obtaining the loss value, the model updates each parameter through back propagation to reduce the loss between the true value and the predicted value, so that the predicted value generated by the model is closer to the true value, thereby achieving the purpose of learning.

[0066] The loss function of a deep learning model is an important parameter used to evaluate the quality of the output results during the model training process, helping to determine the accuracy of the prediction results. The sensitivity test uses the mean absolute error loss function (MAE) and the root mean square error loss function (MSE) to test the impact of different loss functions on the training results.

[0067] (1) MAE loss function

[0068] like Figure 4 As shown in the figure, the calculation principle of the mean absolute error (MeanAbsoluteError) loss function is to take the average of the absolute values ​​of the difference between the estimated value and the true value.

[0069]

[0070] In the formula, f(xi) represents the model output value, yi represents the true value, when the output value and the true value are the same, the MAE loss reaches the minimum, and the minimum value is 0. As the difference between the output value and the true value increases, the MAE loss value shows a linear growth trend.

[0071] Depend on Figure 4 It can be seen that although the MAE curve is continuous, it is not differentiable at f(tion)-yi=0, and in most cases, the MAE loss gradient is the same, which is not conducive to fast convergence.

[0072] (2) MSE loss function

[0073] The calculation principle of the mean square error (MSE) loss function is to take the average of the sum of squares of the difference between the estimated value and the true value.

[0074]

[0075] Depend on Figure 5 It can be seen that MSE is a smooth function and can be differentiated everywhere. Therefore, it is beneficial to calculate the error gradient when solving optimization problems. As the error decreases, the gradient also decreases. Therefore, using a fixed learning rate can also converge to the minimum value quickly. However, the distance between the estimated value and the true value is magnified through square calculation, which has a great negative impact on outliers, thereby reducing the prediction effect of normal values. In addition, when the error is large, the gradient is also large, which is unstable in the early stage of training and is prone to gradient explosion.

[0076] Among them, the sensitivity analysis of the temperature field inversion model based on the U-net network is as follows:

[0077] In order to explore the influence of different input factors on the inversion accuracy, four groups of experiments were set up. Among them, the combination of input variables as sea surface temperature anomaly SSTA, sea surface height anomaly SLA, and the normalized data set labeled as ocean subsurface temperature anomaly TA was set as experimental group one; the combination of input variables as sea surface temperature SST, sea surface height anomaly SLA, and the original data set labeled as ocean subsurface temperature anomaly TA was set as experimental group two; the combination of input variables as sea surface temperature SST, sea surface absolute dynamic height ADT, and the normalized data set labeled as subsurface temperature T was set as experimental group three; the combination of input variables as sea surface temperature SST, sea surface absolute dynamic height ADT, and the original data set labeled as subsurface temperature T was set as experimental group four.

[0078] Taking the GLORYS original data as the benchmark, the reconstruction accuracy of each set of experiments is evaluated by calculating the root mean square error and correlation coefficient of the temperature field of each water layer reconstructed in the test set. The root mean square error is calculated by comparing the reconstruction result with the original data, and the formula is as follows:

[0079]

[0080] Among them, X pre,i is the temperature value reconstructed by the model, X model,i is the original value of the GLORYS reanalysis data. When the RMS error is smaller, it means that the difference between the reconstructed data and the original data is smaller, and the accuracy of the model reconstruction is higher.

[0081] Table 1 Input variable combinations of four groups of reconstruction experiments of three-dimensional temperature field

[0082]

[0083] In some embodiments, when executing step 104, the specific steps may be as follows:

[0084] Verification of the temperature inversion field is evaluated.

[0085] Specifically, based on the above-constructed model, a long-term ocean three-dimensional temperature field is constructed, and the accuracy of the constructed ocean three-dimensional temperature is tested. For example, the overall accuracy of the reconstruction result is statistically analyzed through a scatter plot of the temperature value distribution. Figure 6 The scatter plots of the reconstructed temperature field data at different times and the original data distribution are shown. The distribution characteristics of the reconstruction results are similar to those of the original data, which indicates that the model has been well trained.

[0086] in, Figure 7 as well as Figure 8 (a), (b), (c) and (d) in the figure count the RMSE proportion of the reconstructed temperature field and its distribution in different seasons. It can be seen that in the four seasons of 2018, the smaller the RMSE value, the higher the proportion. As the RMSE value increases, its proportion gradually decreases. 0.1℃ accounts for the largest proportion, and values ​​below 0.5℃ account for the vast majority. This shows that in the reconstruction results, the RMSE is generally small and the accuracy of the reconstruction results is high.

[0087] in, Fig. 9 and Fig.10 They are respectively the vertical distribution diagrams of the reconstructed temperature field and the original temperature field of typical warm vortex and cold vortex, among which, Fig. 9 Reconstructed temperature field of typical warm vortex and vortex latitudinal profile distribution of original data temperature field, Fig. 9 (a), (b), (c), and (d) are the reconstructed vertical distributions of the vortex. Fig. 9(e), (f), (g), and (h) are the vertical distribution of vortices in the original data. It can be seen that the U-net network still has a high accuracy in reconstructing small-scale vortices. Fig.10 Reconstructed temperature field of typical cold vortex and vortex latitudinal profile distribution of original data temperature field. Fig.10 (a), (b), (c), and (d) are the reconstructed vertical distributions of the vortex. Fig.10 (e), (f), (g), and (h) are the vertical distribution of vortices of the original data.

[0088] Among them, the position of the warm vortex does not change significantly in different seasons, and the depth remains basically unchanged, which can extend to the area 600m underwater. In winter and autumn, the water surface temperature is high, and the warm vortex in the upper area will be more obvious. The cold vortex will shift in position in different seasons, which is also more obvious in winter and spring. Through the analysis of the three-dimensional structural characteristics of the reconstructed vortex, it can be seen that the reconstructed vortex is consistent with the vortex at the corresponding position in the original data, which further verifies that the vertical distribution of the temperature in the vortex area can well reflect the vertical temperature distribution of the vortex in the original data, indicating that the three-dimensional temperature field of the ocean reconstructed by the U-net network can well distinguish the three-dimensional structure of the ocean mesoscale vortex.

[0089] Embodiment 2

[0090] This embodiment provides a system for inverting a three-dimensional ocean temperature structure based on sea surface data, including:

[0091] The data acquisition module is used to obtain the reanalysis data of the target sea area.

[0092] The data preprocessing module is used to perform data preprocessing on the reanalysis data to obtain processed reanalysis data; the processed reanalysis data includes a training set and a validation set divided according to a set ratio, and a test set.

[0093] A model training module is used to train an ocean three-dimensional temperature reconstruction model according to the processed reanalysis data to obtain a trained ocean three-dimensional temperature reconstruction model; the ocean three-dimensional temperature reconstruction model is a reconstruction model of the ocean three-dimensional temperature field constructed using a U-net network; during the training process of the ocean three-dimensional temperature reconstruction model, a loss function sensitivity test and a temperature field inversion model sensitivity test based on a U-net network are used to optimize the performance of the ocean three-dimensional temperature reconstruction model.

[0094] The ocean three-dimensional temperature structure construction module is used to construct an ocean three-dimensional temperature structure according to the real-time sea surface data of the target sea area and based on the trained ocean three-dimensional temperature reconstruction model; the ocean three-dimensional temperature structure is represented by an ocean three-dimensional temperature field.

[0095] The system also includes: a model building module, which is used to build an ocean three-dimensional temperature reconstruction model based on a U-net deep learning network; the U-net deep learning network adopts an encoder-decoder symmetrical structure; the encoder part consists of a convolution layer and a pooling layer, which is used to extract features of input data in the ocean three-dimensional temperature reconstruction model through downsampling and pooling; the decoder part consists of a convolution layer and a pooling layer, which is used to reconstruct the features provided by the encoder through upsampling and deconvolution to generate a high-resolution three-dimensional temperature field image.

[0096] Specifically, the data acquisition module includes a sea surface temperature data acquisition submodule, a salinity data acquisition submodule, a mixed layer depth data acquisition submodule and a sea surface height data acquisition submodule.

[0097] In summary, this application has the following technical effects:

[0098] 1) This application can improve the accuracy of ocean three-dimensional temperature field reconstruction and provide a reliable data source for marine scientific research and environmental protection. In order to better understand and monitor the multi-scale variation characteristics of the ocean three-dimensional temperature field, a model of ocean three-dimensional temperature field reconstruction based on sea surface information was built using the U-net deep learning model. Based on this model, sea surface input field and model loss function sensitivity tests were carried out in the northwest Pacific and the coastal waters of China. The setting of sea surface input variables and loss functions can be optimized. This method can effectively extract multi-scale information in the ocean temperature field, improve the accuracy of three-dimensional temperature field reconstruction, and provide a reliable data source for marine scientific research and environmental protection, with broad application prospects.

[0099] 2) It can dynamically construct the ocean multi-scale temperature field and improve the ability to reconstruct the ocean's small and medium-scale. This method can quickly and in real time reconstruct the ocean's three-dimensional temperature field and dynamically update the ocean's three-dimensional temperature field. This method can more realistically reflect the spatiotemporal variation characteristics in the observed data, especially can better capture the three-dimensional structural characteristics of the mesoscale eddy, and can provide technical support for refined marine environmental protection.

[0100] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0101] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for inverting the three-dimensional temperature structure of the ocean based on sea surface data, characterized in that: The method comprises: Obtain reanalysis data of the target sea area; Performing data preprocessing on the reanalysis data to obtain processed reanalysis data; the processed reanalysis data includes a training set and a validation set divided according to a set ratio, and a test set; According to the processed reanalysis data, the ocean three-dimensional temperature reconstruction model is trained to obtain a trained ocean three-dimensional temperature reconstruction model; the ocean three-dimensional temperature reconstruction model is a reconstruction model of the ocean three-dimensional temperature field constructed by using a U-net network; during the training process of the ocean three-dimensional temperature reconstruction model, a loss function sensitivity test and a temperature field inversion model sensitivity test based on a U-net network are used to optimize the performance of the ocean three-dimensional temperature reconstruction model; According to the real-time sea surface data of the target sea area, based on the trained ocean three-dimensional temperature reconstruction model, an ocean three-dimensional temperature structure is constructed; the ocean three-dimensional temperature structure is represented by an ocean three-dimensional temperature field.

2. The method for inverting the three-dimensional temperature structure of the ocean based on sea surface data according to claim 1, characterized in that: Before training the ocean three-dimensional temperature reconstruction model according to the processed reanalysis data to obtain the trained ocean three-dimensional temperature reconstruction model, the method further includes: Based on the U-net deep learning network, an ocean three-dimensional temperature reconstruction model is constructed; the U-net deep learning network adopts an encoder-decoder symmetrical structure; the encoder part consists of a convolution layer and a pooling layer, which is used to extract features of input data in the ocean three-dimensional temperature reconstruction model through downsampling and pooling; the decoder part consists of a convolution layer and a pooling layer, which is used to reconstruct the features provided by the encoder through upsampling and deconvolution to generate a high-resolution three-dimensional temperature field image.

3. The method for inverting the three-dimensional temperature structure of the ocean based on sea surface data according to claim 2, characterized in that: The encoder part of the U-net deep learning network includes four convolutional layers. After performing two convolution operations, the convolutional layer performs maximum pooling on the convolved data downward and inputs the pooled data into the next convolutional layer.

4. The method for inverting the three-dimensional temperature structure of the ocean based on sea surface data according to claim 1, characterized in that: The loss function formula expression in the loss function sensitivity test is: Among them, fxi represents the model output value and yi represents the true value.

5. The method for inverting the three-dimensional temperature structure of the ocean based on sea surface data according to claim 1, characterized in that: The calculation formula for reconstruction accuracy in the sensitivity test of the temperature field inversion model based on the U-net network is: Among them, X pre,i is the temperature value reconstructed by the model, X model,i is the original value of the GLORYS reanalysis data, and RMSE is the root mean square error. The smaller the root mean square error, the higher the reconstruction accuracy of the ocean three-dimensional temperature reconstruction model.

6. The method for inverting the three-dimensional temperature structure of the ocean based on sea surface data according to claim 1, characterized in that: The sea surface data of the target sea area includes sea surface temperature data, salinity data, mixed layer depth data and sea surface height data.

7. The method for inverting the three-dimensional temperature structure of the ocean based on sea surface data according to claim 1, characterized in that: The data preprocessing includes a data normalization processing method, a nearest interpolation method, a NaN value processing method and a data set division method.

8. A system for inverting the three-dimensional temperature structure of the ocean based on sea surface data, characterized in that: include: Data acquisition module, used to obtain reanalysis data of the target sea area; A data preprocessing module, used for performing data preprocessing on the reanalysis data to obtain a processed reanalysis; The processed reanalysis data includes a training set and a validation set divided according to a set ratio, and a test set; A model training module is used to train the ocean three-dimensional temperature reconstruction model according to the processed reanalysis data to obtain a trained ocean three-dimensional temperature reconstruction model; the ocean three-dimensional temperature reconstruction model is a reconstruction model of the ocean three-dimensional temperature field constructed by using a U-net network; during the training process of the ocean three-dimensional temperature reconstruction model, a loss function sensitivity test and a temperature field inversion model sensitivity test based on a U-net network are used to optimize the performance of the ocean three-dimensional temperature reconstruction model; The ocean three-dimensional temperature structure construction module is used to construct an ocean three-dimensional temperature structure according to the real-time sea surface data of the target sea area and based on the trained ocean three-dimensional temperature reconstruction model; the ocean three-dimensional temperature structure is represented by an ocean three-dimensional temperature field.

9. The system for inverting the three-dimensional temperature structure of the ocean based on sea surface data according to claim 7, characterized in that: Also includes: A model building module is used to build a three-dimensional ocean temperature reconstruction model based on a U-net deep learning network; the U-net deep learning network adopts an encoder-decoder symmetric structure; the encoder part consists of a convolution layer and a pooling layer, which is used to extract features of input data in the three-dimensional ocean temperature reconstruction model through downsampling and pooling; The decoder part consists of convolutional layers and pooling layers, which are used to reconstruct the features provided by the encoder through upsampling and deconvolution to generate a high-resolution three-dimensional temperature field image.

10. The system for inverting the three-dimensional temperature structure of the ocean based on sea surface data according to claim 7, characterized in that: The data acquisition module includes a sea surface temperature data acquisition submodule, a salinity data acquisition submodule, a mixed layer depth data acquisition submodule and a sea surface height data acquisition submodule.

Images