Double-stage ocean subsurface temperature super-resolution reconstruction method and system

Abstract

The invention provides a double-stage ocean subsurface temperature super-resolution reconstruction method and system, and relates to the technical field of ocean science, and the method comprises the steps: obtaining multi-source seawater surface remote sensing observation data; performing quality control and normalization processing on the seawater surface remote sensing observation data to obtain a to-be-measured data set; training the initial model based on a preset historical data set to obtain a trained super-resolution reconstruction model of the underwater three-dimensional temperature structure; the working process of the super-resolution reconstruction model comprises an inversion stage and a super-resolution stage; and inputting a to-be-measured data set into the super-resolution reconstruction model to obtain seawater subsurface high-resolution temperature data. According to the invention, accurate reconstruction of the high-resolution temperature data of the seawater subsurface layer is realized, and the capabilities of marine environment monitoring and climate research are enhanced.

Description

Technical Field

[0001] The present invention relates to the field of marine science and technology, and particularly to a two-stage method and system for super-resolution reconstruction of ocean subsurface temperature. Background Art

[0002] The ocean has long been regarded as one of the most important natural resources for humanity, and temperature is one of its key characteristics. Many phenomena in the ocean are related to the temperature structure in the ocean, such as the formation of underwater sound shadow zones in the ocean. High-resolution ocean temperature data is crucial for understanding mesoscale processes in the ocean, protecting the marine ecosystem, enhancing the potential of ocean numerical forecasting, evaluating the acoustic properties of seawater, and more accurately understanding the global ocean climate. Therefore, obtaining high-resolution ocean temperature data is crucial for promoting marine scientific research.

[0003] Currently, traditional observation methods for the subsurface temperature structure of seawater include direct observation methods such as Argo floats, expendable bathythermographs, or on-site detection by exploration vessels. However, due to the vast area of the ocean, it is difficult to obtain high-resolution subsurface seawater temperature data through this method; through indirect observation methods such as satellite remote sensing, large-scale high-resolution long-term continuous observation hydrological information can be obtained, but it can only obtain hydrological parameters on the seawater surface. And satellite observation methods are divided into two methods: microwave and infrared wave. The microwave has strong penetration, but low resolution; the infrared wave has weak penetration and is easily blocked by clouds, making it difficult to form continuous observations, but the observation resolution is high. Therefore, under certain climate conditions, the resolution of the surface hydrological data obtained by satellite remote sensing may be low. The hydrological parameters on the water surface are closely related to the temperature structure underwater, and many physical phenomena inside the ocean can be manifested through the data on the ocean surface. Using satellite observation data to invert the subsurface temperature structure of the ocean and further study the fine structure inside the ocean, such as mesoscale eddies, ocean fronts, and internal ocean waves, is of great significance for in-depth understanding of ocean phenomena. However, the current seawater subsurface temperature reconstruction technology still faces some challenges, such as large errors and insufficient resolution.

[0004] Chinese Patent Application CN117934274B discloses a method for super-resolution reconstruction of ocean salinity based on a deep learning SROSRN model. This method uses multi-source sea surface remote sensing observation data, combines with in-situ measured data inside the ocean, and adopts a deep learning method combining convolutional neural network and self-attention mechanism to establish a remote sensing super-resolution reconstruction model based on multi-source satellite observations to perform super-resolution reconstruction on the salinity inside the ocean, and obtains a good reconstruction effect. However, this model requires the input of high-resolution sea surface temperature, and the ordinary self-attention mechanism does not have the ability to perceive features of long sequences, which is not sufficient to balance the extraction of local and global spatial information features. This patent is restricted by the resolution of the sea surface temperature parameters when reconstructing the salinity structure of the subsurface seawater.

[0005] Most of the current models for reconstructing the subsurface temperature of seawater rely on variational assimilation-based methods or deep learning-based methods. Traditional variational or assimilation-based methods have problems such as large computational resource overheads. With the rise of deep learning in recent years, researchers have found that data-driven deep learning methods can also achieve good reconstruction results. However, there are few studies on super-resolution reconstruction of ocean subsurface temperature. Existing studies and solutions basically reconstruct the ocean subsurface temperature on the premise of the same spatial resolution as satellite observation data. This will limit the spatial resolution of the reconstructed subsurface temperature by the spatial resolution of sea surface satellite remote sensing data, resulting in the inability to reconstruct high-spatial-resolution subsurface temperature data in some cases where the satellite can only obtain low-resolution sea surface data.

[0006] Among the few existing studies on super-resolution reconstruction methods for underwater data, the output does not generate all the data on the entire two-dimensional depth profile at once, but sequentially outputs the one-dimensional vertical profile data of each point in the reconstruction area. By using the sea surface hydrological parameters in a specific area as features, the vertical profile data of more points in each local area compared to the original data is predicted, thus achieving super-resolution reconstruction. However, there is a problem with this method: due to the differences in input parameters and features in different local areas, the model cannot learn the transitional relationships between different local areas, resulting in transitional discontinuities between these areas, or unnatural transitions. This discontinuity may not be significant on a large scale, but within a local range, it may manifest as insufficient smoothness in the transition between different small areas, with obvious faults or unnatural transition phenomena, leading to poor reconstruction results. Summary of the Invention

[0007] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a two-stage super-resolution reconstruction method and system for ocean subsurface temperature, which realizes the accurate reconstruction of high-resolution temperature data of seawater subsurface, enhances the ability of ocean environmental monitoring and climate research, and provides more reliable data support for ocean scientific research.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] A two-stage super-resolution reconstruction method for ocean subsurface temperature, comprising:

[0010] Obtain multi-source remote sensing observation data of the sea surface;

[0011] Perform quality control and normalization processing on the remote sensing observation data of the sea surface to obtain a dataset to be measured;

[0012] Train the initial model based on a preset historical dataset to obtain a trained super-resolution reconstruction model for the underwater three-dimensional temperature structure; the working process of the super-resolution reconstruction model includes: an inversion stage and a super-resolution stage; the inversion stage is used to invert the subsurface temperature of different depth profiles using satellite data to obtain an inversion result; the super-resolution stage is used to perform super-resolution processing based on the inversion result to generate subsurface temperature data; the historical dataset includes a sea surface input dataset for a historical period of time and temperature data of the seawater subsurface as labels.

[0013] Input the dataset to be measured into the super-resolution reconstruction model to obtain high-resolution temperature data of the seawater subsurface.

[0014] Preferably, the seawater surface remote sensing observation data includes sea surface height data, sea surface temperature data, sea surface salinity data, sea surface temperature anomaly data, and sea surface height anomaly data.

[0015] Preferably, the data preprocessing process of the historical dataset includes:

[0016] Unify the time resolution of the preset sample seawater surface remote sensing observation data and the sample seawater subsurface temperature reanalysis data to one month. For the sample seawater surface remote sensing observation data and the sample seawater subsurface temperature reanalysis data with multiple data values within one month, average the data within the month; the formula for average calculation is where data i represents the average data for the i-th month, data i,j represents the j-th data for the i-th month, and M i represents the total number of data for the i-th month;

[0017] Unify the spatial resolution of the sample seawater surface remote sensing observation data to 1 / 4°×1 / 4°, and unify the spatial resolution of the sample seawater subsurface temperature reanalysis data to 1 / 12°×1 / 12°. Interpolate the sample seawater surface remote sensing observation data and the sample seawater subsurface temperature reanalysis data onto the same grid coordinates to obtain an interpolated dataset.

[0018] Normalize the interpolated dataset and other relevant features to obtain a sea surface input dataset for model training and temperature data of the seawater subsurface respectively; the formula for normalization is: where X is the normalized sea surface input dataset and temperature data of the seawater subsurface, x is the interpolated dataset and other relevant features, and μ and σ are the mean and variance of the interpolated dataset and other relevant features; the other relevant features include longitude, latitude, and month.

[0019] Preferably, in both the inversion stage and the super-resolution stage, a 3×3 convolutional layer is used for shallow feature extraction; the output of the convolutional layer is fed into a stacked Transformer-based encoder for deep feature extraction, and then added to the output of the previous convolutional layer through a residual connection.

[0020] Preferably, in the inversion stage, the output after the residual connection is used as the input of the next convolutional layer, and after being processed by the convolutional layer, the output will be used as the input of the super-resolution stage.

[0021] Preferably, in the super-resolution stage, the output after the residual connection first passes through a convolutional layer, then undergoes a pixel rearrangement operation, and then passes through another convolutional layer to obtain the final subsurface super-resolution temperature data; the super-resolution temperature data is concatenated with the output of the inversion stage to generate the final output for input into the loss function for iterative optimization.

[0022] Preferably, the Transformer-based encoder also incorporates a multi-scale self-attention mechanism and a coordinate attention mechanism; the multi-scale self-attention mechanism is used to extract spatial information features of different scales by dividing windows of different sizes; the coordinate attention mechanism is used to enhance the learning of global spatial information features and channel information by embedding spatial position information.

[0023] Preferably, the training process of the super-resolution reconstruction model is carried out using the Adam optimizer with a learning rate of 0.0002 for backpropagation of errors.

[0024] Preferably, the other relevant features include the longitude and latitude of the data in the sea surface input dataset and the month to which the data belongs.

[0025] A two-stage subsurface ocean temperature super-resolution reconstruction system, comprising:

[0026] A data acquisition unit for acquiring multi-source remotely sensed sea surface observation data;

[0027] A data processing unit for performing quality control and normalization processing on the remotely sensed sea surface observation data to obtain a dataset to be measured;

[0028] A model training unit for training an initial model based on a preset historical dataset to obtain a trained super-resolution reconstruction model for underwater three-dimensional temperature structure; the working process of the super-resolution reconstruction model includes: an inversion stage and a super-resolution stage; the inversion stage is used to invert the subsurface temperature of different depth profiles using satellite data to obtain an inversion result; the super-resolution stage is used to perform super-resolution processing based on the inversion result to generate subsurface temperature data; the historical dataset includes a sea surface input dataset for a historical period of time and temperature data of the seawater subsurface as labels.

[0029] A data reconstruction unit for inputting the dataset to be measured into the super-resolution reconstruction model to obtain high-resolution temperature data of the seawater subsurface.

[0030] According to the specific embodiments provided by the present invention, the following technical effects are disclosed by the present invention:

[0031] The present invention provides a two-stage super-resolution reconstruction method and system for ocean subsurface temperature. By integrating multi-source remote sensing observation data and performing quality control and normalization processing, the accuracy and reliability of the data can be improved. Secondly, the super-resolution reconstruction model trained using the historical dataset can effectively invert the subsurface temperature of different depth profiles, and then generate high-resolution temperature data. This method not only improves the spatial resolution of the temperature data, but also enhances the monitoring ability of ocean environmental changes, providing more accurate basic data support for ocean scientific research, climate change assessment and ocean resource management. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0033] Figure 1 It is a flowchart of the method provided by the embodiment of the present invention;

[0034] Figure 2 It is a schematic diagram of the technical route provided by the embodiment of the present invention;

[0035] Figure 3 It is a schematic diagram of the overall structure of the model provided by the embodiment of the present invention;

[0036] Figure 4 It is a schematic diagram of the Transformer encoder structure provided by the embodiment of the present invention;

[0037] Figure 5Schematic diagram for comparing reconstruction metrics of different models provided by embodiments of the present invention; among them Figure 5 (a) is RMSE; Figure 5 (b) is SSIM; Figure 5 (c) is PSNR. Detailed implementation manners

[0038] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0039] The purpose of the present invention is to provide a two-stage ocean subsurface temperature super-resolution reconstruction method and system, which realizes the accurate reconstruction of high-resolution temperature data of the seawater subsurface, enhances the ability of ocean environment monitoring and climate research, and provides more reliable data support for ocean scientific research.

[0040] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0041] Figure 1 Flowchart of the method provided by embodiments of the present invention, as Figure 1 shown, the present invention provides a two-stage ocean subsurface temperature super-resolution reconstruction method, including:

[0042] Step 100: Obtain multi-source remote sensing observation data of the sea surface;

[0043] Step 200: Perform quality control and normalization processing on the remote sensing observation data of the sea surface to obtain a dataset to be measured;

[0044] Step 300: Train an initial model based on a preset historical dataset to obtain a trained super-resolution reconstruction model for the underwater three-dimensional temperature structure; the working process of the super-resolution reconstruction model includes: an inversion stage and a super-resolution stage; the inversion stage is used to invert the subsurface temperature of different depth profiles using satellite data to obtain an inversion result; the super-resolution stage is used to perform super-resolution processing based on the inversion result to generate subsurface temperature data; the historical dataset includes a sea surface input dataset for a historical period of time and the temperature data of the seawater subsurface as a label;

[0045] Step 400: Input the dataset to be measured into the super-resolution reconstruction model to obtain high-resolution temperature data of the seawater subsurface.

[0046] As Figure 2As shown in the figure, this embodiment proposes a two-stage ocean subsurface temperature super-resolution reconstruction method based on Transformer, including the following steps:

[0047] (1) Data acquisition

[0048] Obtain multi-source remote sensing observation data of the sea surface and reanalysis data of the sea subsurface temperature. The gridded data used in this embodiment includes multi-source remote sensing observation data of the sea surface and reanalysis data of the sea subsurface. The remote sensing observation data of the sea surface includes sea surface height data SSH (Sea Surface Height), sea surface temperature data SST (Sea Surface Temperature), sea surface salinity data SSS (Sea Surface Salt), sea surface temperature anomaly SSTA (Sea Surface Temperature Anomaly), and sea level anomaly data SLA (Sea Level Anomaly); the sea level anomaly data is provided by the Copernicus Climate Change Service (CCCS), and the sea surface temperature anomaly data is provided by the National Oceanic and Atmospheric Administration (NOAA). The temperature data of the sea subsurface is the reanalysis data provided by the Copernicus Marine Environment Monitoring Service (CMEMS).

[0049] (2) Data preprocessing

[0050] Due to the existence of islands in the ocean and other situations, complete gridded data cannot be obtained at some locations, and data quality control is required. For the data at these missing locations, this embodiment uses the method of setting the data value at its location to 0 for processing.

[0051] This embodiment involves multiple data sets, but different data sets have different time resolutions and spatial resolutions. To avoid errors caused by inconsistent data set formats, it is necessary to unify the time resolution and spatial resolution of the multiple data sets involved in this article.

[0052] In this embodiment, the time resolution of the data set is unified to one month. For a data set with multiple data values within a month, the data within the month is averaged. The calculation formula is as follows:

[0053]

[0054] Where datai Represents the average data for the i-th month, data i,j Represents the j-th data for the i-th month, M i Represents the total number of data for the i-th month.

[0055] In this embodiment, the spatial resolution of the sea surface dataset is unified to 1 / 4°×1 / 4°, and the spatial resolution of the ocean subsurface temperature dataset is unified to 1 / 12°×1 / 12°, and they are interpolated onto the same grid coordinates. In this embodiment, the xarray library of python is used to adjust the spatial resolution.

[0056] (3) Data normalization

[0057] To improve the model convergence speed, the input and output data are normalized. Normalized data can make the influence weights of each feature dimension on the objective function consistent, and improve the convergence speed of iterative solution. The processing formula is as follows

[0058]

[0059] Where X is the normalized data, x is the original data, and μ and σ are the mean and variance of the original data.

[0060] (4) Establishment of super-resolution reconstruction model

[0061] The input of the super-resolution reconstruction model is the satellite hydrological observation dataset of the sea surface, and the output of the inversion model is the temperature data of the sea subsurface. The specific steps are as follows:

[0062] (4.1) Model input and output

[0063] Model input: What is input into the model are the historical sea surface hydrological parameters, longitude and latitude, and the month to which the data belongs for a period of time. Five measurements (sea surface temperature, sea surface salinity, sea surface height, sea surface temperature anomaly, sea surface height anomaly) are produced monthly on the sea surface. Each measurement is gridded data with spatial correlation of size H×W. The eight measurements can be represented as 8-channel data of the corresponding size. The final tensor shape input into the model is: (B,8,H,W).

[0064] Model output: What the model outputs is the temperature data of the sea subsurface after super-resolution reconstruction. Similar to the analysis in the model input, the output of the model is super-resolution gridded data with spatial correlation of size 3H×3W. The final tensor shape output by the model is: (B,,H,W,10).

[0065] (4.2) Overall model structure

[0066] As Figure 3 and Figure 4As shown in the figure, the present invention proposes a two-stage ocean subsurface temperature super-resolution reconstruction model based on Transformer. In both stages, a 3×3 convolutional layer is used for shallow feature extraction. The output of the convolutional layer is then fed into a stacked Transformer-based encoder for deep feature extraction, and then added to the output of the previous convolutional layer through a residual connection. The difference is that in the first stage (inversion stage), the output after the residual connection is used as the input of the next convolutional layer. After being processed by the convolutional layer, this output will be used as the input of the second stage (super-resolution stage). In the second stage, the output after the residual connection first passes through a convolutional layer, then through a pixel rearrangement operation, and then through another convolutional layer to obtain the final subsurface super-resolution temperature data. This data will be concatenated with the output of the inversion stage to generate the final output, which is used as the input to the loss function for iterative optimization.

[0067] (4.3) Multi-scale self-attention mechanism

[0068] To improve the model's ability to extract and learn local spatial feature information, a multi-scale self-attention mechanism is introduced into the Transformer encoder. Given an input of shape B×C×H×W, the shape of the input must be reshaped to B×(H×W)×C to meet the requirements of the multi-head attention mechanism. However, during the reconstruction process of ocean subsurface temperature, (H×W) is usually too large, and the original multi-head attention mechanism cannot effectively process such a sequence length, exceeding its feature extraction ability, thus reducing the model's prediction performance. In addition, a larger sequence length will result in higher computational costs and longer computation times. To calculate spatial feature information at different scales more accurately and efficiently, in this embodiment, the input feature X is first divided into N groups, and then the window size W n is used to calculate the self-attention for the Nth group of features. In this way, this embodiment can flexibly control the computational cost by setting different window sizes. For example, the sequence length of the self-attention for the Nth group can be reduced to thereby reducing the computational cost and extracting features more accurately. In addition, by controlling the window size, this embodiment can also extract corresponding features at different scales. After calculating the self-attention for different groups, these results are concatenated and fused through a 1×1 convolutional operation.

[0069] (4.4) Coordinate attention mechanism

[0070] Since the multi-scale self-attention mechanism focuses more on capturing local spatial information, in this embodiment, the coordinate attention mechanism is used to enhance the learning of global spatial information features and channel information. By embedding spatial position information in the channel attention, the model can capture cross-channel information and position-sensitive details simultaneously. First, the input feature map is globally averaged along the height and width directions respectively to obtain two feature vectors, which represent the global information in the height dimension and the width dimension respectively; then the two feature vectors are concatenated, and a non-linear mapping is performed on the original spatial information to learn the complex relationships between different positions in the input features, and separated into two different attention vectors; finally, the generated attention vectors are recombined onto the original feature map.

[0071] (5) Model training

[0072] (5.1) The obtained sea surface hydrological parameters are composed into a four-dimensional tensor with a shape of (16, 96, 64, 8), and then the tensor is rearranged into a four-dimensional tensor with a shape of (16, 8, 96, 64) and input into the model for processing to obtain the underwater temperature data in the form of a four-dimensional tensor of (16, 96, 64, 10).

[0073] (5.2) In the present invention, the model prediction result is compared with the true value, the MSE loss value is calculated, and the Adam optimizer is used to backpropagate the error and train the model at a learning rate of 0.0002 to update the model parameters.

[0074] As Figure 5 shown, the present invention uses a two-stage deep learning model based on Transformer for super-resolution reconstruction of the underwater temperature structure, and makes a comparison with other parallel models. The comparison models are the super-resolution model based on convolutional neural network (Super-Resolution Convolutional Neural Network, SRCNN), the very deep super-resolution model (Very Deep Super-Resolution, VDSR), and the U-net based on attention mechanism (Attention U-net). The experimental result evaluation indexes for different models are as follows: The experimental results show that the effect of using the reconstruction method proposed in the present invention is the best, followed by Attention U-net. The experimental results can prove the super-resolution reconstruction performance of the reconstruction method proposed in the present invention.

[0075] Corresponding to the above method, the present invention also provides a two-stage ocean subsurface temperature super-resolution reconstruction system, including:

[0076] A data acquisition unit for acquiring multi-source remote sensing observation data of the sea surface;

[0077] A data processing unit for performing quality control and normalization processing on the remotely sensed observation data of the sea surface to obtain a dataset to be measured;

[0078] A model training unit for training an initial model based on a preset historical dataset to obtain a super-resolution reconstruction model of the underwater three-dimensional temperature structure; the working process of the super-resolution reconstruction model includes: an inversion stage and a super-resolution stage; the inversion stage is used to invert the subsurface temperature of different depth profiles using satellite data to obtain an inversion result; the super-resolution stage is used to perform super-resolution processing based on the inversion result to generate subsurface temperature data; the historical dataset includes a sea surface input dataset for a historical period of time and the temperature data of the seawater subsurface as labels;

[0079] A data reconstruction unit for inputting the dataset to be measured into the super-resolution reconstruction model to obtain high-resolution temperature data of the seawater subsurface.

[0080] The beneficial effects of the present invention are as follows:

[0081] (1) The present invention uses a large amount of historical data to learn the physical pattern between sea surface hydrological parameters and seawater subsurface temperature, avoiding complex physical formula operations and the problems of low efficiency and large memory caused by fine-grained construction. While meeting the data accuracy, it reduces memory and improves the algorithm efficiency.

[0082] (2) The present invention uses a two-stage model framework to decouple the super-resolution reconstruction task and reduce the difficulty of model optimization.

[0083] (3) The present invention uses a feature encoder that combines a multi-scale self-attention mechanism and a coordinate attention mechanism to extract local spatial information and global spatial information of input features.

[0084] (4) In practical applications, the present invention can predict the temperature structure profile of the seawater subsurface by using the trained super-resolution reconstruction model and a small amount of sea surface data.

[0085] Each embodiment in this specification is described in a progressive manner. The key points of each embodiment are the differences from other embodiments. The same or similar parts among the embodiments can be referred to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0086] In this text, specific examples are used to elaborate on the principles and implementation manners of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A two-stage ocean subsurface temperature super-resolution reconstruction method, characterized in that: include: Obtain sea surface remote sensing observation data from multiple sources; Performing quality control and normalization processing on the sea surface remote sensing observation data to obtain a data set to be tested; The initial model is trained based on a preset historical data set to obtain a trained super-resolution reconstruction model of underwater three-dimensional temperature structure; The working process of the super-resolution reconstruction model includes: an inversion stage and a super-resolution stage; the inversion stage is used to use satellite data to invert the subsurface temperature of different depth profiles to obtain inversion results; the super-resolution stage is used to perform super-resolution processing based on the inversion results to generate subsurface temperature data; the historical data set includes a sea surface input data set for a period of history and the temperature data of the subsurface of seawater as a label; The data set to be tested is input into a super-resolution reconstruction model to obtain high-resolution temperature data of the subsurface layer of seawater.

2. The two-stage ocean subsurface temperature super-resolution reconstruction method according to claim 1 is characterized in that: The sea surface remote sensing observation data includes sea surface height data, sea surface temperature data, sea surface salinity data, sea surface temperature anomaly data and sea surface height anomaly data.

3. The two-stage ocean subsurface temperature super-resolution reconstruction method according to claim 1 is characterized in that: The data preprocessing process of the historical data set includes: The time resolution of the preset sample seawater surface remote sensing observation data and the sample seawater subsurface temperature reanalysis data is unified to one month. For the sample seawater surface remote sensing observation data and the sample seawater subsurface temperature reanalysis data with multiple data values ​​within one month, the data within the month are averaged; the average calculation formula is: Where data i Represents the average data of the i-th month, data i,j represents the jth data item in the i-th month, M i Represents the total number of data in month i; The spatial resolution of the sample seawater surface remote sensing observation data is unified to 1 / 4°×1 / 4°, and the spatial resolution of the sample seawater subsurface temperature reanalysis data is unified to 1 / 12°×1 / 12°, and the sample seawater surface remote sensing observation data and the sample seawater subsurface temperature reanalysis data are interpolated to the same grid coordinates to obtain an interpolated data set; The interpolated data set and other related features are normalized to obtain the sea surface input data set and the seawater subsurface temperature data for model training. The normalization formula is: Where X is the normalized sea surface input data set and the temperature data of the subsurface layer of the seawater, x is the interpolated data set and other related features, and μ and σ are the mean and variance of the interpolated data set and other related features.

4. The two-stage ocean subsurface temperature super-resolution reconstruction method according to claim 1 is characterized in that: In both the inversion stage and the super-resolution stage, a 3×3 convolutional layer is used for shallow feature extraction; the output of the convolutional layer is fed into a stacked Transformer-based encoder for deep feature extraction, and then added to the output of the previous convolutional layer through a residual connection.

5. The two-stage ocean subsurface temperature super-resolution reconstruction method according to claim 4 is characterized in that: In the inversion stage, the output after the residual connection is used as the input of the next convolutional layer, and after being processed by the convolutional layer, the output will be used as the input of the super-resolution stage.

6. The two-stage ocean subsurface temperature super-resolution reconstruction method according to claim 5 is characterized in that: In the super-resolution stage, the output after the residual connection first passes through a convolution layer, then undergoes a pixel rearrangement operation, and then passes through another convolution layer to obtain the final sub-surface super-resolution temperature data; the super-resolution temperature data is spliced ​​with the output of the inversion stage to generate the final output, which is used to input the loss function for iterative optimization.

7. The two-stage ocean subsurface temperature super-resolution reconstruction method according to claim 4 is characterized in that: The encoder based on the Transformer structure also has a multi-scale self-attention mechanism and a coordinate attention mechanism embedded therein; the multi-scale self-attention mechanism is used to extract spatial information features of different scales by dividing windows of different sizes; The coordinate attention mechanism is used to enhance the learning of global spatial information features and channel information by embedding spatial position information.

8. The two-stage ocean subsurface temperature super-resolution reconstruction method according to claim 4 is characterized in that: The training process of the super-resolution reconstruction model is performed using the Adam optimizer with a learning rate of 0.0002 to back-propagate the error.

9. The two-stage ocean subsurface temperature super-resolution reconstruction method according to claim 3 is characterized in that: The other relevant features include the latitude and longitude of the data in the sea surface input data set and the month to which the data belongs.

10. A two-stage ocean subsurface temperature super-resolution reconstruction system, characterized in that: include: A data acquisition unit, used to acquire multi-source sea surface remote sensing observation data; A data processing unit, used for performing quality control and normalization processing on the sea surface remote sensing observation data to obtain a data set to be measured; A model training unit, used to train the initial model based on a preset historical data set to obtain a trained super-resolution reconstruction model of the underwater three-dimensional temperature structure; The working process of the super-resolution reconstruction model includes: an inversion stage and a super-resolution stage; the inversion stage is used to use satellite data to invert the subsurface temperature of different depth profiles to obtain inversion results; the super-resolution stage is used to perform super-resolution processing based on the inversion results to generate subsurface temperature data; the historical data set includes a sea surface input data set for a period of history and the temperature data of the subsurface of seawater as a label; The data reconstruction unit is used to input the data set to be tested into the super-resolution reconstruction model to obtain high-resolution temperature data of the subsurface layer of seawater.

Images