Method and application of reconstructing subsurface feature field data based on deep network and scattered observation data

Abstract

The present invention discloses a method for reconstructing subsurface element field data based on a deep network and scattered observation data, belonging to the field of marine environment stereoscopic monitoring technology. Specifically, the method solves the problem in the background technology that traditional complex deep networks cannot use scattered field observation data sets as label data. By using the sea surface element framework observed by satellite as a factor and field observation profile data as a label, subsurface temperature and salinity data can be reconstructed. The main parts of the present invention include: (1) a deep neural network part, which predicts the subsurface element field data by inputting the sea surface element field; (2) a cost function part of a bilinear interpolation operator, which uses a bilinear interpolation operator to correspond the predicted subsurface element field to the observation data set labels at scattered locations, thereby achieving the purpose of comprehensively utilizing multi-source ocean data to carry out subsurface element field data reconstruction.

Description

Technical Field

[0001] The present invention belongs to the technical field of marine environment stereo monitoring, and in particular relates to a method for reconstructing subsurface element field data based on a deep network and scattered observation data. Background Art

[0002] Ocean data is essential for the development of modern oceanography and the protection of the marine environment. Ocean data observations are primarily divided into in situ observations and remote sensing observations. In situ observations provide vertical profiles of ocean temperature, salinity, and other factors, but existing in situ observations are still relatively sparse. Even though the development of the Argo observing network has enabled all-weather, large-scale observations of the ocean environment, it only provides one profile per 3° × 3° area every 10 days on average. Reflecting mesoscale ocean phenomena requires a temporal and spatial resolution of at least 0.25° × 0.25° per week. Therefore, with the exception of individual targeted marine surveys of localized small and medium-scale phenomena, in situ observations cannot effectively reflect the ubiquitous small and medium-scale phenomena in the ocean. Remote sensing observations provide ocean surface fields of factors such as sea surface temperature and height. Existing remote sensing observations generally achieve a temporal and spatial resolution of 0.25° per day, enabling the effective identification of mesoscale phenomena such as ocean fronts and mesoscale eddies. However, remote sensing observations are limited to the ocean surface and cannot effectively reflect information about the marine environment below the surface. Overall, observation data from any single source cannot reflect the three-dimensional structure of the ocean.

[0003] The technology of reconstructing high-precision three-dimensional ocean data by comprehensively utilizing multi-source data is called ocean three-dimensional data reconstruction (hereafter referred to as reconstruction). Reconstruction, mapping a small amount of ocean information onto the entire three-dimensional ocean space, is a typical inverse problem and inherently ill-posed. To address this problem, researchers have mostly adopted empirical modeling methods, with statistical regression being the most widely used approach. The core idea of ​​statistical regression is to establish a statistical regression relationship between surface ocean elements such as sea surface temperature and sea surface height observed by satellites and underwater in-situ observation profiles. Based on this approach, previous researchers used linear regression to establish a statistical relationship between sea surface temperature, sea surface height, and in-situ temperature and salinity profiles, achieving surface-to-underwater reconstruction of ocean temperature and salinity data. This approach was applied to the renowned Modular Ocean Data Assimilation System (MODAS) and has been widely used. Based on this approach, previous researchers have attempted to combine reconstruction with nonlinear regression algorithms, such as self-organizing maps (SOMs) and support vector machines (SVMs). After the advent of the machine learning craze, some more popular methods have been put into use, such as generalized regression neural network (GRNN), random forest (RF), extreme gradient boosting learning machine (XGBoost), lightweight gradient boosting learning machine (LightGBM), feedforward neural network (FFNN), etc.

[0004] In the era of ocean big data, deep neural networks have experienced rapid development and application. The most classic networks are convolutional neural networks (CNNs) and recurrent neural networks (RNNs). Neural network structures based on CNNs are primarily suitable for "surface" data such as images and feature fields, requiring the data to be a regular grid of continuous points in space. Neural network structures based on RNNs are primarily suitable for "line" data such as time series, requiring the data to be a temporally continuous sequence. Unfortunately, traditional reconstruction techniques cannot effectively apply deep neural networks. This is primarily due to the fact that ocean observations, such as those from Argo floats, are not fixed in point locations and are discontinuous in time and space, making them incapable of meeting the data format requirements for deep learning. Although gridded sea surface data is available as regression factors, since labeled data originates from scattered and sparse observation profiles, modeling can only be performed by interpolating the sea surface grid products to the corresponding observation profile locations. This means that reconstruction is based on a "point-to-point" modeling approach. However, this "point-to-point" modeling approach presents the following problems: First, the "point-to-point" modeling process interpolates satellite observations to spatially and temporally discontinuous observation locations, eliminating the spatiotemporal evolution of all-weather satellite observations. Second, this "point-to-point" modeling process fails to fully incorporate physical constraints into the cost function. For example, physical formulas such as geostrophic balance require spatial or temporal differencing, and only field data or time series can satisfy these constraints. Third, this "point-to-point" modeling process is not conducive to model iteration. Advances in deep learning have largely focused on images or time series. For example, large-scale modeling techniques, which have gained significant attention in recent years, cannot be directly migrated into existing reconstruction architectures as modules. Therefore, this "point-to-point" modeling approach has led to bottlenecks in reconstruction technology, preventing it from keeping pace with advanced deep learning methods.

[0005] Therefore, there is an urgent need to design a method to reconstruct subsurface feature field data based on deep networks and scattered observation data. This method breaks through the bottleneck that traditional reconstruction schemes cannot utilize deep networks, and can assimilate the grid feature fields of satellite observations and scattered profiles of field observations to reconstruct high-precision three-dimensional ocean feature fields. Summary of the Invention

[0006] In response to the problem in the above-mentioned prior art that traditional complex deep networks cannot use scattered field observation data sets as label data, the purpose of the present invention is to provide a method for reconstructing subsurface feature field data based on deep networks and scattered observation data. The method constitutes the architecture OI-Rec-Net (Observation-Informed Reconstruction Net), which uses the sea surface feature architecture of satellite observations as factors and field observation profile data as labels to reconstruct subsurface temperature and salinity data. The architecture includes: (1) a deep neural network part, which predicts the subsurface feature field data by inputting the sea surface feature field; (2) a cost function part of a bilinear interpolation operator, which uses a bilinear interpolation operator to interpolate the predicted subsurface feature field to the location of the observation dataset, thereby realizing deep network training with the observation dataset as label. The present invention aims to solve the problem that traditional complex deep networks cannot use scattered field observation datasets as label data, and achieve the purpose of comprehensively utilizing multi-source ocean data to reconstruct subsurface feature field data.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] The method for reconstructing subsurface feature field data based on a deep network and scattered observation data is characterized by comprising:

[0009] S1. Based on the input data sea surface feature field X and the output data underwater feature field Y, select the corresponding deep neural network F;

[0010] S2. Using the on-site observation profile information and the grid information of the subsurface 3D reconstruction element field output by the neural network F, construct a bilinear interpolation matrix H for multiple observation points;

[0011] S3.y based on the bilinear interpolation matrix H pred , with the profile observation value y obs For the label, construct a loss function;

[0012] S4. Use the constructed loss function to train the deep neural network F, adjust the network weights, and obtain the trained neural network F;

[0013] S5. The sea surface feature field X is input as input data to the trained neural network F, and the trained neural network F outputs the final subsurface three-dimensional reconstruction feature field Y1.

[0014] Preferably, in step S1, the input layer dimension of the neural network F is the same as the sea surface element field X, the output layer dimension is the same as the subsurface three-dimensional reconstruction element field Y, and the sea surface element field X, the underwater element field Y and the neural network F satisfy the mapping relationship Y=F(X).

[0015] Preferably, the deep neural network F is a convolutional neural network or a transformer network.

[0016] Preferably, in step S2, the process of constructing a bilinear interpolation matrix H of multiple observation points includes:

[0017] S201. Construct bilinear interpolation of multiple observation points;

[0018] S202. Obtain a bilinear interpolation matrix H based on bilinear interpolation of multiple observation points.

[0019] Preferably, the multiple observation points include an observation point P and an observation point Q, and the bilinear interpolation construction method of the observation point P and the observation point Q is:

[0020] Assume that the horizontal coordinate of the observation point P falling within the given grid is (x o ,y o ), there are four grid points around it, and the values ​​corresponding to the four grid points are Y 11 、Y 21 、Y 12 and Y 22 , the corresponding coordinates are (x1, y1), (x2, y1), (x1, y2) and (x2, y2);

[0021] Therefore, the bilinear interpolation expression at the observation point P can be expressed as:

[0022]

[0023] Among them, y P is the value interpolated to the observation point;

[0024] Y 11 The factor of Y 12 The factor of Y 21 The factor of Y 22 The factor of Formula (1) can be simplified as:

[0025]

[0026] Different from the observation point P, the observation point Q is assumed to fall on Y 33 、Y 34 、Y 43 and Y 44 Within the coordinates, the bilinear interpolation expression at the observation point Q can be abbreviated as:

[0027]

[0028] Among them, y Q is the value interpolated to the observation point.

[0029] Preferably, the process of obtaining the bilinear interpolation matrix H based on bilinear interpolation of multiple observation points includes: combining formula (2) and formula (3) and converting them into a matrix form to obtain the bilinear interpolation matrix H;

[0030] The expression of the bilinear interpolation matrix H is:

[0031]

[0032] in, Y=(Y 11 ,Y 21 ,Y 12 ,Y 22 ,Y 33 ,Y 43 ,Y 34 ,Y 44 ) T .

[0033] Preferably, in step S3,

[0034] The constructed loss function expression is: J = MSE(y pred ,y obs );

[0035] Among them, y obs is the element value of ocean observation, and MSE refers to the mean square error.

[0036] The second object of the present invention is to provide a system for reconstructing subsurface feature field data based on a deep network and scattered observation data, which is characterized by comprising:

[0037] The network selection module selects the corresponding deep neural network F based on the input data sea surface feature field X and the output data underwater feature field Y;

[0038] The interpolation matrix construction module uses the on-site field observation profile information and the grid information of the subsurface three-dimensional reconstruction element field output by the neural network F to construct a bilinear interpolation matrix H of multiple observation points;

[0039] The loss function construction module is based on the bilinear interpolation matrix H of y pred , with the profile observation value y obs For the label, construct a loss function;

[0040] The training module uses the constructed loss function to train the deep neural network F, adjust the network weights, and obtain the trained neural network F;

[0041] The output module takes the sea surface feature field X as input data and inputs it into the trained neural network F. The trained neural network F outputs the final subsurface 3D reconstruction feature field Y1.

[0042] Among them, the system is implemented based on a method of reconstructing subsurface feature field data using deep networks and scattered observation data.

[0043] The third object of the present invention is to provide a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable the computer to execute a method for reconstructing subsurface feature field data based on a deep network and scattered observation data.

[0044] The fourth object of the present invention is to provide an electronic device comprising: at least one processor, and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute a method for reconstructing subsurface feature field data based on a deep network and scattered observation data.

[0045] The beneficial effects of the present invention are as follows: the present invention discloses a method for reconstructing subsurface feature field data based on a deep network and scattered observation data. Compared with the prior art, the improvements of the present invention are:

[0046] The present invention designs a deep network architecture (OI-Rec-Net) that can assimilate field observation data. It uses scattered observation information as labels to train deep neural networks and realize deep neural network predictions based on pure observations. The OI-Rec-Net architecture consists of a deep network module + a bilinear interpolation module. It breaks through the bottleneck that traditional deep networks cannot assimilate scattered observation profiles, and allows neural networks that use pure observations as training data to have the following functions: 1. It takes into account the spatiotemporal evolution information of all-weather satellite observations; 2. It allows the addition of physical constraints with spatial or temporal differences; 3. It allows the deep network module to be iteratively updated according to the latest progress of neural networks. And when the method of the present invention is actually used, the effect of reconstructing subsurface temperature is better than the results of feedforward neural networks with traditional architectures, and can be comparable to reanalysis data. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flowchart of the present invention;

[0048] Figure 2 Schematic diagram of the OI-Rec-Net architecture of the present invention;

[0049] Figure 3 Schematic diagram of the bilinear interpolation principle of the present invention;

[0050] Figure 4 A schematic diagram of the experimental implementation of the present invention;

[0051] Figure 5 The temperature distribution of the reconstructed element field at a depth of 100 m for the present invention (the selected date is January 1, 2018);

[0052] Figure 6 is the RMSD of the different data of the present invention compared with the measured data;

[0053] Figure 7 Corr is the comparison between different data of the present invention and measured data. DETAILED DESCRIPTION

[0054] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0055] Example 1:

[0056] The network architecture OI-Rec-Net of the present invention has the following specific structure: it is composed of a deep network module and a bilinear interpolation module; Figure 1-2 As shown in the figure, the left half is the neural network reconstruction part used to reconstruct the 3D feature field. This part inputs the sea surface feature field X into the deep neural network F and obtains the output subsurface 3D reconstruction feature field Y. It should be noted that there is no fixed combination of the sea surface input field, the neural network, and the subsurface 3D reconstruction feature field Y. Figure 2 The examples shown here are for reference only; any model that satisfies the Y = F(X) mapping can be used. For example, X can be sea surface temperature, sea surface height, sea surface salinity, sea surface temperature gradient, sea surface height gradient, longitude and latitude grid, and so on; F can be a convolutional neural network (CNN) or a transformer network architecture; and Y can be temperature, salinity, current velocity, chlorophyll, and other oceanographic elements.

[0057] The steps of the OI-Rec-Net architecture to reconstruct subsurface feature field data based on deep networks and scattered observation data include:

[0058] S1. Based on the input data sea surface feature field X and the output data underwater feature field Y, select the corresponding deep neural network F;

[0059] Specifically, the selected deep neural network F needs to be able to meet the requirements when the input data is the sea surface feature field X and the output data is the subsurface feature field Y. For example, if the input sea surface feature field is selected as two types of sea surface temperature and sea surface height, and the number of horizontal grid points is 1440×720, then the dimension of X is 1440×720×2, where the number 2 represents two channels, corresponding to the two types of input data. The output underwater feature field is the ocean temperature at a depth of 200m, and the number of horizontal grid points is 1440×720, then the dimension of Y is 1440×720×1. Therefore, the selected neural network F should consist of an input layer dimension of 1440×720×2, an output layer dimension of 1440×720×1, and several hidden layers.

[0060] In this embodiment, the neural network F can be selected from a convolutional neural network, a transformer, or other network architectures. Note that there is no fixed combination of the sea surface input field X, the deep neural network F, and the subsurface 3D reconstruction element field Y in this embodiment. Figure 2 The examples shown here are for reference only; any model that satisfies the Y = F(X) mapping can be used. For example, X can be sea surface temperature, sea surface height, sea surface salinity, sea surface temperature gradient, sea surface height gradient, longitude and latitude grid, and so on; F can be a convolutional neural network or a network architecture such as the Transformer; and Y can be temperature, salinity, or other oceanographic elements such as current velocity and chlorophyll.

[0061] S2. Construct a bilinear interpolation matrix H for multiple observation points using the on-site observation profile information and the grid information of the subsurface 3D reconstruction feature field Y output by the deep neural network F;

[0062] Field observation profile information includes profile location information and profile observation value y obs ,In this step, only the profile position information is used;

[0063] S201. Construct bilinear interpolation of multiple observation points;

[0064] Assume that at a certain time, m observations can be collected in the corresponding area, and the subsurface 3D reconstruction feature field Y output by the deep neural network F has a total of n grid points, then the constructed bilinear interpolation matrix is ​​H(m×n); please note that the 3D interpolation operation is actually performed by the 2D data of each layer, that is, each layer uses the same matrix H.

[0065] The bilinear interpolation construction process of a single observation point is as follows;

[0066] like Figure 3 As shown, if the observation point P falls within the given grid, then the horizontal coordinate of the observation point P is (x o ,yo ), there are four grid points closest to it, and the values ​​corresponding to the four grid points are Y 11 、Y 21 、Y 12 and Y 22 , the corresponding coordinates are (x1, y1), (x2, y1), (x1, y2) and (x2, y2) respectively. Therefore, the bilinear interpolation expression at the observation point P can be expressed as:

[0067]

[0068] Among them, y P is the value interpolated to the observation point;

[0069] For simplicity, Y 11 The factor of Y 12 The factor of Y 21 The factor of Y 22 The factor of Among them, such as Figure 3 As shown, in In the formula (1), the superscript 12 indicates that the grid point index range is between 1 and 2, the subscript 11 indicates that the coordinates are (x1, y1), and the P in the brackets refers to the observation point P. Similarly, formula (1) can be abbreviated as:

[0070]

[0071] Among them, y P is the value interpolated to the observation point;

[0072] Consider that there is an observation point Q that is different from the observation point P. Assume that the observation point Q falls on Y 33 、Y 34 、Y 43 and Y 44 Within the coordinates, according to the above principle:

[0073]

[0074] S202. Obtain a bilinear interpolation matrix H based on bilinear interpolation of multiple observation points;

[0075] Specifically, formula (2) and formula (3) are combined and converted into a matrix form to obtain the bilinear interpolation matrix H; that is, formula (4):

[0076]

[0077] in, Y=(Y 11 ,Y21 ,Y 12 ,Y 22 ,Y 33 ,Y 43 ,Y 34 ,Y 44 ) T ; Thus, the interpolation from the grid field Y to multiple scattered observations can be realized in the form of matrix H, that is, y pred =HY, please note that Y here is obtained from the subsurface 3D reconstruction element field Y output by the deep neural network F in step S101;

[0078] In formula (4), the bilinear interpolation matrix H reflects the y obtained by interpolating the grid field Y to the two observation points. pred If there are multiple observations within the range of more grid points, then Formula (4) can be expanded by analogy. Please note that the tensor dimension output by deep neural networks represented by CNN is a fixed grid field, while the number of observations in the ocean every day is a scattered profile with non-fixed dimensions. Since the dimension of the output data of the deep neural network does not correspond to the observation data, traditional deep neural networks cannot be trained directly using the observation data as labels. Constructing the bilinear interpolation matrix H corresponds the fixed-dimensional grid field output by the neural network to the observation data of any dimension, thus realizing deep network training using the observation data as labels.

[0079] S3.y based on the bilinear interpolation matrix H pred , with the profile observation value y obs For labels, construct loss functions as needed.

[0080] y obs is the profile observation value of the field observation profile information, that is, y obs The element value of ocean observation, such as the ocean temperature value observed by Argo float, pred and the actual observed value y obs The position and dimension of y are exactly the same. For example, if there are 30 observations at a certain depth, then y obs The dimension of is 30×1, and the y interpolated to the observation point position is pred The dimension is also 30×1;

[0081] Based on y pred and y obs The loss function can be constructed according to the needs of the user. In this step, the loss function can be constructed as J = MSE (y pred ,y obs ); where MSE stands for mean squared error, this construction can be implemented directly in Python via torch.nn.MSEloss.

[0082] S4. Use the constructed loss function to train the deep neural network F, adjust the network weights, and obtain the trained deep neural network F.

[0083] Network weight training is performed in Python. Specifically, the optimizer, batch division, and loss backpropagation steps are selected in the Python torch toolbox. In this step, Adam (torch.optim.Adam) is selected as the optimizer, 64 samples are taken as a batch (batch_size=64), and loss backpropagation is implemented using the loss.backward() code. After setting the number of training epochs, training is terminated early according to the early stopping strategy. If the training results do not improve after three consecutive epochs, the model training is terminated early.

[0084] S5. The sea surface feature field X is input as input data to the trained deep neural network F, which outputs the final subsurface 3D reconstructed feature field Y1;

[0085] like Figure 2 As shown in the figure, the left half is the neural network reconstruction part used to reconstruct the 3D feature field. This part inputs the sea surface feature field X into the deep neural network F and outputs the subsurface 3D reconstructed feature field Y1.

[0086] The present invention also provides a system for reconstructing subsurface feature field data based on a deep network and scattered observation data, which is characterized by comprising:

[0087] The network selection module selects the corresponding deep neural network F based on the input data sea surface feature field X and the output data underwater feature field Y;

[0088] The interpolation matrix construction module uses the on-site field observation profile information and the grid information of the subsurface three-dimensional reconstruction element field output by the neural network F to construct a bilinear interpolation matrix H of multiple observation points;

[0089] The loss function construction module is based on the bilinear interpolation matrix H of y pred , with the profile observation value y obs For the label, construct a loss function;

[0090] The training module uses the constructed loss function to train the deep neural network F, adjust the network weights, and obtain the trained neural network F;

[0091] The output module takes the sea surface feature field X as input data and inputs it into the trained neural network F. The trained neural network F outputs the final subsurface 3D reconstruction feature field Y1.

[0092] The control system is implemented based on the method of Example 1.

[0093] The present invention also provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable the computer to execute the method of embodiment 1.

[0094] The fourth object of the present invention is to provide an electronic device comprising: at least one processor, and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method of Example 1.

[0095] Example 2:

[0096] Example 2 is to reconstruct the subsurface temperature of the northwest Pacific Ocean using the method of Example 1. Specifically, the sea surface temperature data and sea surface height data observed by satellite are used as input, and the subsurface sea temperature data is output.

[0097] The specific process is:

[0098] 1. Basic situation of the experiment and data preparation

[0099] The input data has a temporal and spatial resolution of 0.25° × 0.25° per day, and the output data has the same temporal and spatial resolution as the input data, with 59 vertical layers ranging from 10 m to 1000 m. The experimental area is the northwestern Pacific (130°E-160°E, 30°N-50°N), with training covering the period 2010-2017 and validation covering the period 2018-2019.

[0100] The sea surface temperature (SST) data used in this experiment were obtained from the Optimum Interpolation Sea Surface Temperature (OISST) | National Centers for Environmental Information (NCEI) (noaa.gov), with a temporal and spatial resolution of 0.25° × 0.25° / day. Sea surface height (SSH) data were obtained from the Copernicus Marine Data Store | Copernicus Marine Service, product number SEALEVEL_GLO_PHY_L4_MY_008_047, with a temporal and spatial resolution of 0.25° × 0.25° / day. The climatological data used were from the World Ocean Atlas 2023 (WOA23) World Ocean Atlas-ClimateFields Data Access (noaa.gov), which provides monthly climatological data with a spatial resolution of 0.25° × 0.25° and 57 vertical layers from 0 to 1500 m. The observational profile data used are from the Global Ocean Science Dataset (CODC-GOSD) OceanandClimate (iap.ac.cn) of the Ocean Data Center of the Chinese Academy of Sciences. This dataset collects multi-source ocean profile data from 1940 to 2023 and has undergone quality control. The Global Ensemble Reanalysis Dataset (GREP) was used as a comparison data. This data is also from the Copernicus Marine Data Store | Copernicus Marine Service, product number GLOBAL_MULTIYEAR_PHY_ENS_001_031, with a temporal and spatial resolution of 0.25° × 0.25° / day and 75 vertical layers ranging from 0.5 m to 5902.1 m.

[0101] (2) Data preprocessing

[0102] Horizontal preprocessing: SST, SSH, WOA23, and GREP data are horizontally interpolated onto the SST grid and the regional extent of the northwest Pacific is intercepted.

[0103] Vertical preprocessing: The CODC-GOSD observation profiles, WOA23 and GREP data were vertically interpolated to 59 layers using the Akima interpolation method.

[0104] Outlier extraction: SSTA is obtained by deducting the WOA23 climate state of the corresponding month from the SST. For example, the SST on January 6 is obtained by deducting the WOA23 climate state of January from the SST. The SSH is deducted from the multi-year monthly average SSH to obtain SSHA. The WOA23 climate state data is horizontally interpolated to the position of the CODC-GOSD profile to obtain the temperature climate state profile T corresponding to the temperature profile T. clim , then the temperature anomaly profile is T anom =TT clim .

[0105] (3) Constructing a reconstruction model

[0106] During the reconstruction process, in order to facilitate the specific implementation, a layer-by-layer reconstruction method is adopted, that is, the sea surface information and the first layer, the sea surface information and the second layer are modeled layer by layer, and 59 reconstruction models are constructed. The 59 reconstruction models are trained using the same network, and the reconstruction model of the nth layer is trained according to Figure 2 For detailed implementation of the reconstructed network, see Figure 4 Among them, the network for training the reconstruction model is the OI-Rec-Net model. The schematic diagram of the architecture of the OI-Rec-Net model is as follows: Figure 2 As shown,

[0107] like Figure 4 As shown in the figure, a relatively simple CNN model is used, forming a three-layer deep neural network. The input layer contains four channels: longitude (Lon), latitude (Lat), sea surface temperature anomaly (SSTA), and sea surface height anomaly (SSHA). Based on the selected range, there are a total of 80×120 spatial grid points. Note that the longitude and latitude here are grid points. The temperature field is reconstructed through the nth layer of the model output and matched to the in-situ field observations in the nth layer through a bilinear interpolation module, enabling network training using field measurements as labels.

[0108] (4) Reconstruction model verification

[0109] The validation of the model is mainly considered from three perspectives: the first is the horizontal distribution of the reconstructed factor field, the second is the root mean square error RMSD compared with the on-site observations, and the third is the Pearson correlation coefficient Corr compared with the on-site observations.

[0110] The calculation formula of RMSD is as follows:

[0111]

[0112] in, is the i-th (i=1, 2, ..., N) observation value at depth z, is the i-th (i=1, 2, ..., N) prediction value at depth z, which is obtained by bilinear interpolation of the reconstructed field of the corresponding layer.

[0113] The calculation formula for Corr is as follows:

[0114]

[0115] Figure 5 The main morphologies of the reconstructed temperature element fields are plotted in Figure 1. It can be seen that the temperature fields from GREP and OI-Rec-Net exhibit relatively good temperature morphology, clearly reflecting fronts and eddies. In contrast, the temperature from WOA23 is overly smooth, revealing only the basic pattern of warmer sea temperatures and colder northern temperatures. While the temperature field from FFNN appears to reflect eddies, it is readily apparent that the reconstructed temperature field exhibits numerous localized fine structures, which are unphysical and appear as noise. This means that while the traditional "point-to-point" reconstruction architecture using FFNN can reconstruct more detailed structure than WOA23, it is prone to subsurface horizontal gradient anomalies. This is primarily due to the inadequate consideration of the interrelationships between adjacent grid points in this "point-to-point" reconstruction. In contrast, the reconstructed temperature morphology from OI-Rec-Net exhibits a more reasonable eddy structure, such as the entrainment of cold water by the cyclonic vortex in the region (153°E-156°E, 30°N-35°E).

[0116] Figure 6 The vertical profiles of RMSD are plotted in Figure 1. As can be seen, WOA23 has the highest error, considering that WOA23 only contains monthly signals, while the other three datasets contain valid small- and medium-scale signals within a month. The SST error reconstructed by OI-Rec-Net is much lower than that of FFNN and WOA23, and even slightly outperforms GREP at 500 meters above the ocean. It is important to emphasize that GREP data are derived by combining the results of multiple models and undergoing a complex data assimilation process. Despite this, OI-Rec-Net achieves comparable results to reanalysis data with its lightweight computational cost and workflow.

[0117] Figure 7 The vertical profile of Corr is drawn in Figure 2. It can be seen that the correlation coefficient of OI-Rec-Net is optimal at a depth of 100m-500m. At a depth of 500m, GREP has the highest Corr. The correlation sparsity of OI-Rec-Net is superior to that of traditional FFNN at all depths.

[0118] As can be seen, despite the very simple CNN used in this experiment, its reconstruction performance far exceeds that of traditional FFNN architectures and is comparable to GREP reanalysis products. In the future, with further iterations of the network structure and rapid optimization of the cost function, the OI-Rec-Net architecture is expected to achieve even higher-precision reconstruction.

[0119] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for reconstructing subsurface feature field data based on a deep network and scattered observation data, characterized by: include: S1. Based on the input data sea surface feature field X and the output data underwater feature field Y, select the corresponding deep neural network F; S2. Using the on-site observation profile information and the grid information of the subsurface 3D reconstruction element field output by the neural network F, construct a bilinear interpolation matrix H for multiple observation points; S3.y based on the bilinear interpolation matrix H pred , with the profile observation value y obs For the label, construct a loss function; S4. Use the constructed loss function to train the deep neural network F, adjust the network weights, and obtain the trained neural network F; S5. The sea surface feature field X is input as input data to the trained neural network F, and the trained neural network F outputs the final subsurface 3D reconstructed feature field Y1; In step S2, the process of constructing a bilinear interpolation matrix H of multiple observation points includes: S201. Construct bilinear interpolation of multiple observation points; S202. Obtain a bilinear interpolation matrix H based on bilinear interpolation of multiple observation points; The multiple observation points include observation point P and observation point Q, and the bilinear interpolation construction method of observation point P and observation point Q is: Assume that the horizontal coordinate of the observation point P falling within the given grid is (x o ,y o ), there are four grid points around it, and the values ​​corresponding to the four grid points are Y 11 、Y 21 、Y 12 and Y 22 , the corresponding coordinates are (x1, y1), (x2, y1), (x1, y2) and (x2, y2); The bilinear interpolation expression at the observation point P is expressed as: Among them, y P is the value interpolated to the observation point; Y 11 The factor of Y 12 The factor of Y 21 The factor of Y 22 The factor of Formula (1) can be simplified as: Assume that the observation point Q falls on Y 33 、Y 34 、Y 43 and Y 44 Within the coordinates, the bilinear interpolation expression at the observation point Q is obtained as: Among them, y Q is the value interpolated to the observation point.

2. The method for reconstructing subsurface feature field data based on a deep network and scattered observation data according to claim 1, characterized in that: In step S1, the input layer dimension of the neural network F is the same as the sea surface feature field X, the output layer dimension is the same as the subsurface three-dimensional reconstruction feature field Y, and the sea surface feature field X, the underwater feature field Y and the neural network F satisfy the mapping relationship Y=F(X).

3. The method for reconstructing subsurface feature field data based on a deep network and scattered observation data according to claim 1, characterized in that: The deep neural network F is a convolutional neural network or a transformer network.

4. The method for reconstructing subsurface feature field data based on a deep network and scattered observation data according to claim 1, characterized in that: Based on the bilinear interpolation of multiple observation points, the process of obtaining the bilinear interpolation matrix H includes: combining formula (2) and formula (3) into a matrix form to obtain the bilinear interpolation matrix H; The expression of the bilinear interpolation matrix H is: Among them, Y = (Y 11 , Y 21 , Y 12 , Y 22 , Y 33 , Y 43 , Y 34 , Y 44 ) T .

5. The method for reconstructing subsurface feature field data based on a deep network and scattered observation data according to claim 1, characterized in that: In step S3, The constructed loss function expression is: J = MSE(y pred ,y obs ); Among them, y obs is the element value of ocean observation, and MSE refers to the mean square error.

6. A system for reconstructing subsurface feature field data based on deep networks and scattered observation data, characterized by: include: The network selection module selects the corresponding deep neural network F based on the input data sea surface feature field X and the output data underwater feature field Y; The interpolation matrix construction module uses the on-site field observation profile information and the grid information of the subsurface three-dimensional reconstruction element field output by the neural network F to construct a bilinear interpolation matrix H of multiple observation points; The process of constructing a bilinear interpolation matrix H of multiple observation points includes: S201. Construct bilinear interpolation of multiple observation points; S202. Obtain a bilinear interpolation matrix H based on bilinear interpolation of multiple observation points; The multiple observation points include observation point P and observation point Q, and the bilinear interpolation construction method of observation point P and observation point Q is: Assume that the horizontal coordinate of the observation point P falling within the given grid is (x o ,y o ), there are four grid points around it, and the values ​​corresponding to the four grid points are Y 11 、Y 21 、Y 12 and Y 22 , the corresponding coordinates are (x1, y1), (x2, y1), (x1, y2) and (x2, y2); The bilinear interpolation expression at the observation point P is expressed as: Among them, y P is the value interpolated to the observation point; Y 11 The factor of Y 12 The factor of Y 21 The factor of Y 22 The factor of Formula (1) can be simplified as: Assume that the observation point Q falls on Y 33 、Y 34 、Y 43 and Y 44 Within the coordinates, the bilinear interpolation expression at the observation point Q is obtained as: Among them, y Q is the value interpolated to the observation point; The loss function construction module is based on the bilinear interpolation matrix H of y pred , with the profile observation value y obs For the label, construct a loss function; The training module uses the constructed loss function to train the deep neural network F, adjust the network weights, and obtain the trained neural network F; The output module takes the sea surface feature field X as input data and inputs it into the trained neural network F. The trained neural network F outputs the final subsurface 3D reconstruction feature field Y1. The system is implemented based on the method according to any one of claims 1 to 5.

7. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 5.

8. An electronic device comprising: At least one processor, and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 5.

Images