Intelligent fusion method and system of three-dimensional ocean environment field based on deep learning

Abstract

The present invention discloses a method and system for intelligent fusion of three-dimensional ocean environment fields based on deep learning. The method includes the following steps: S1, obtaining multi-source ocean data, wherein the multi-source ocean data includes one-dimensional observation data, two-dimensional ocean satellite observation data, and three-dimensional gridded ocean models and reanalysis data; S2, performing data processing on the multi-source ocean data; S3, producing a sample set; S4, building a deep learning network model, and training it to construct an ocean data intelligent fusion model; S5, inputting the multi-source ocean data to be fused into the ocean data intelligent fusion model, and outputting an intelligent fusion result. The present invention comprehensively considers the spatiotemporal continuity and consistency of ocean environmental elements and the physical correlation between multiple elements, and realizes the high-precision, high-resolution, and spatiotemporal continuous intelligent fusion of multi-dimensional, multi-source heterogeneous ocean data by constructing an ocean data intelligent fusion model, providing high-quality data resource support for scientific research and application in the ocean field.

Description

Technical Field

[0001] The present invention belongs to the fields of marine science and machine learning, and in particular relates to an intelligent fusion method and system for three-dimensional marine environment fields based on deep learning. Background Art

[0002] High-quality oceanographic data is a crucial prerequisite and essential condition for marine science research, global change studies, and efficient offshore production operations. To further understand and predict climate change and the evolution of the marine environment, it is essential to obtain long-term, high-resolution, and reliable marine environmental data, and to conduct scientific research and numerical simulations based on this data. This has become a consensus in the field.

[0003] Currently, ocean observation methods are becoming increasingly diverse, and the amount of available ocean data is rapidly increasing. However, different ocean data sources have their own strengths and weaknesses. Conventional ocean observations, such as those from ocean stations, buoys, submersibles, and underway navigation, can detect information about the ocean's interior and offer advantages such as high accuracy. However, due to their limited observation range, they are globally sparse and have poor spatiotemporal continuity. Ocean satellite remote sensing data, on the other hand, offers advantages such as a wide detection range and good spatiotemporal continuity, but are limited in depth and cannot detect information about the ocean's interior. The largest existing model simulation, forecast, or reanalysis data is three-dimensional gridded data, which offers the best spatiotemporal continuity but is far less accurate than observational data. Therefore, it is necessary to combine the strengths of multiple ocean data sources to obtain a high-precision, high-resolution, and spatiotemporally continuous three-dimensional fusion field of ocean elements.

[0004] With the increasing variety and volume of observational data, the integration of multi-source ocean data poses significant challenges to computing power, numerical methods, and assimilation techniques. The development of traditional methods for ocean statistical analysis and numerical prediction of the marine environment has encountered certain bottlenecks: due to the oversimplification of traditional linear models, an incomplete understanding of ocean processes, the uncertainty of numerical models, and the limitations of model resolution, these methods are unable to meet the demands of increasingly sophisticated and accurate research and production.

[0005] In recent years, artificial intelligence technology with deep learning as its core has been introduced into the field of Earth system science and has achieved good results, providing new ideas and methods for marine science research. AI technology with machine learning as its core will become a new way to maximize the value of observation data and improve the level of marine environmental protection. Summary of the Invention

[0006] In response to the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide an intelligent fusion method and system for three-dimensional ocean environment fields based on deep learning, so as to solve the technical problems that the existing ocean observation data have poor spatiotemporal continuity, limited observation layer depth, low precision of model and reanalysis data, large amount of calculation, and low generation efficiency, resulting in the inability to quickly obtain high-precision, high-resolution, spatiotemporally continuous three-dimensional ocean element fusion fields.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] The first aspect of the present invention provides an intelligent fusion method for a three-dimensional ocean environment field based on deep learning, comprising the following steps:

[0009] S1. Acquire multi-source ocean data, wherein the multi-source ocean data includes one-dimensional observation data, two-dimensional ocean satellite observation data, and three-dimensional gridded ocean model and reanalysis data;

[0010] S2. Processing the acquired multi-source ocean data, including quality control, spatiotemporal interpolation, and normalization.

[0011] S3. Create a sample set and divide it into training set, test set and validation set;

[0012] S4. Build a deep learning network model and use the training set to train the deep learning network model to build an intelligent fusion model of ocean data;

[0013] S5. Input the multi-source ocean data to be fused into the ocean data intelligent fusion model and output the intelligent fusion result.

[0014] Preferably, it also includes S6, inputting the test set into the ocean data intelligent fusion model, outputting the model test results, and analyzing the performance of the model according to the speed and parameter quantity of the model during the test; at the same time, according to the model test results and the test sample labels, calculating and visualizing the error index, and evaluating the accuracy of the verification model and the data fusion quality.

[0015] Preferably, the one-dimensional observation data include Argo, and sea temperature profile data observed by buoys, submersible buoys and shipborne equipment, and the one-dimensional observation data are derived from the EN4-profiles data set of the UK Met Office Hadley Centre; the two-dimensional ocean satellite observation data include sea surface temperature data, sea surface height anomaly data and sea surface wind field data; wherein, the sea surface temperature data are derived from the ERSST data set, the GHRSST data set and the HadiSST data set; the sea surface height anomaly data are derived from the AVISO SLA data set; the sea surface wind field data are derived from the CCMP data set; the three-dimensional gridded ocean model and reanalysis data include HYCOM data, SODA ocean reanalysis data and EN4-analysis data.

[0016] Preferably, in step S2, the data processing includes sequentially performing quality control processing, spatiotemporal interpolation processing, and normalization processing on the obtained multi-source ocean data.

[0017] Preferably, the quality control processing of the multi-source ocean data is specifically: querying, eliminating, correcting and filling in abnormal values ​​and missing values ​​of one-dimensional observation data in the multi-source ocean data.

[0018] Preferably, the spatiotemporal interpolation processing of multi-source ocean data specifically includes the following steps:

[0019] A21. Use linear interpolation to interpolate the one-dimensional observation data to the vertical standard layer {z1, z2, …, zk};

[0020] A22. Use the spatiotemporal weighted interpolation method to interpolate the one-dimensional observation data with discrete distribution in the horizontal direction to the corresponding grid points of the horizontal grid {x1,x2,…,xm}×{y1,y2,…,yn}.

[0021] A23. Use bilinear interpolation to interpolate the two-dimensional ocean satellite observation data to the horizontal grid {x1,x2,…,xm}×{y1,y2,…,yn};

[0022] A24. Interpolate the three-dimensional gridded ocean model and reanalysis data to the three-dimensional grid {x1,x2,…,xm}×{y1,y2,…,yn}×{z1,z2,…,zk} using bilinear interpolation.

[0023] A25. Interpolate the spatially interpolated multi-source ocean data to the same temporal resolution.

[0024] Preferably, the normalization processing of the multi-source ocean data is specifically: normalizing the multi-source ocean data and the spatiotemporal information data after spatiotemporal interpolation, wherein the spatiotemporal information data includes the year, month, longitude and latitude.

[0025] Preferably, in step S3, the preparing of the sample set specifically includes the following steps:

[0026] S31. The one-dimensional observation data interpolated to the vertical standard layer is used as the sample label. The number of vertical standard layers is set to k, and the dimension of the sample label array is recorded as (1, k).

[0027] S32, using the EN4-analysis data, HYCOM data, SODA reanalysis data, the difference between the EN4-analysis data and the HYCOM data, the difference between the HYCOM data and the SODA reanalysis data, and the difference between the EN4-analysis data and the SODA reanalysis data at the horizontal grid positions corresponding to the one-dimensional observation data as sample feature data to form six feature channels;

[0028] S33. The spatiotemporal information data of the sample label and the two-dimensional ocean satellite observation data of the grid corresponding to the sample label are used as sample feature data to form a feature channel, which is made into a sample set, and the sample feature array dimension is recorded as (7, k).

[0029] Preferably, a global attention mechanism and an adaptive parameterized ReLU activation function are added to the deep learning network model, wherein the global attention mechanism adopts a serial method of first performing channel attention and then performing spatial attention.

[0030] Preferably, in step S4, the deep learning network model constructed is trained using a training set, and the specific steps are: the deep learning network model constructed is trained using a training set, and the output result is compared with the sample label data, and the difference between the network model output and the sample label data is minimized, and the parameters are optimized to construct an intelligent fusion model of ocean data.

[0031] A second aspect of the present invention provides an intelligent system for a three-dimensional ocean environment field based on deep learning, comprising:

[0032] Data acquisition and processing module, used to acquire multi-source ocean data and perform data processing;

[0033] A sample set production module is used to analyze multi-source ocean data and produce them into sample sets;

[0034] Ocean data intelligent fusion model building module, used to build and train deep learning network models;

[0035] An intelligent fusion result output module is used to input the multi-source ocean data to be fused into the ocean data intelligent fusion model and output the intelligent fusion result;

[0036] The result verification module is used to evaluate and verify the performance, accuracy and data fusion quality of the constructed ocean data intelligent fusion model.

[0037] The present invention has the following beneficial effects:

[0038] The intelligent fusion method of three-dimensional ocean environment field based on deep learning in the present invention can intelligently fuse many ocean element fields including temperature, salinity, flow field, sea level, waves, and sea surface wind field. Based on multi-source ocean environment data such as field observation data, satellite remote sensing data, ocean reanalysis data and model data, and taking into account the spatiotemporal continuity and consistency of ocean environment elements and the physical correlation between multiple elements, the intelligent fusion model of ocean data is constructed to realize high-precision, high-resolution, and spatiotemporal continuity of multi-dimensional, multi-source heterogeneous ocean data, namely three-dimensional ocean environment (sea temperature) field, providing high-quality data resource support for scientific research and application in the ocean field. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 This is a flow chart of the intelligent fusion method of the three-dimensional ocean environment field based on deep learning of the present invention;

[0041] Figure 2 Schematic diagram of sample labels;

[0042] Figure 3 Schematic diagram of the arrangement of sample feature data;

[0043] Figure 4 Schematic diagram of the ODF-Net network model structure;

[0044] Figure 5 Schematic diagram of the structure of the global attention mechanism;

[0045] Figure 6 Schematic diagram of the structure of the channel attention submodule;

[0046] Figure 7 Schematic diagram of the structure of the spatial attention submodule;

[0047] Figure 8 Schematic diagram of the basic principle of APReLU activation function;

[0048] Figure 9-aFigure a is the result of intelligent fusion of global ocean standard layer sea temperature data in Example 3;

[0049] Figure 9-b Figure b is the result of intelligent fusion of global ocean standard layer sea temperature data in Example 3;

[0050] Figure 9-c Figure c is the intelligent fusion result of the global ocean standard layer sea temperature data in Example 3. DETAILED DESCRIPTION

[0051] In the following description, specific details such as particular system structures and techniques are provided for purposes of illustration, not limitation, to facilitate a thorough understanding of the embodiments of the present invention. However, it will be apparent to those skilled in the art that the present invention may be practiced in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present invention with unnecessary detail.

[0052] Example 1

[0053] Reference Figure 1 , an embodiment of the present invention provides an intelligent fusion method of a three-dimensional ocean environment field based on deep learning, comprising the following steps:

[0054] S1. Acquire multi-source ocean data, wherein the multi-source ocean data includes one-dimensional observation data, two-dimensional ocean satellite observation data, and three-dimensional gridded ocean model and reanalysis data;

[0055] Specifically, the one-dimensional observation data includes Argo data and sea temperature profile data observed by buoys, submersible buoys, and shipborne equipment. The one-dimensional observation data comes from the EN4-profiles (sea temperature profile) dataset of the UK Met Office Hadley Center;

[0056] Two-dimensional ocean satellite observation data include sea surface temperature (SST), sea surface height anomaly (SHA), and sea surface wind data. The SST data are derived from the ERSST (Extended Reconstructed Sea Surface Temperature) dataset, the GHRSST (Global High-Resolution Sea Surface Temperature Experiment) dataset, and the Hadley Centre's Global Sea Surface Temperature Analysis (HadiSST) dataset. The SST data are derived from the AVISOSLA (Satellite Oceanography Archive Data Center, CNES), and the CCMP (NASA Earth Science Enterprise Multi-Platform Cross-Calibration Data).

[0057] 3D gridded ocean model and reanalysis data include HYCOM data (U.S. Navy Hybrid Coordinate Ocean Model), SODA reanalysis data (University of Maryland Simple Ocean Data Assimilation), and temperature and salinity analysis data (quality-controlled by the UK Meteorological Hadley Centre).

[0058] EN4-analysis data, etc.

[0059] S2. Processing the acquired multi-source ocean data;

[0060] Specifically, the obtained multi-source ocean data are sequentially subjected to quality control processing, spatiotemporal interpolation processing, and normalization processing;

[0061] Among them, (1) the quality control processing of the obtained multi-source ocean data is mainly to query, eliminate, correct and fill the outliers and missing values ​​of the one-dimensional observation data in the multi-source ocean data; specifically, all incorrect and uncertain profiles are eliminated according to the flag quality identifier in the EN4-profiles dataset, and then the deepest depth of the EN4-profiles dataset is selected to 1000m, and then the abnormal profiles with temperature values ​​outside the range of -5 to 40℃ are screened and eliminated;

[0062] (2) Perform spatiotemporal interpolation processing on multi-source ocean data;

[0063] That is, first, the multi-source (heterogeneous) ocean data (information) are interpolated to the same horizontal grid points, and then the spatially interpolated multi-source ocean data (information) are interpolated to the same temporal resolution; specifically, the following steps are included:

[0064] A21. Use linear interpolation to interpolate one-dimensional observation data to the vertical standard layer {z1, z2, ..., z k}, in this embodiment k = 23, the vertical depth of z1 is 0m, z 23 The vertical depth is 1000m;

[0065] A22. Use the spatiotemporal weighted interpolation method to interpolate the one-dimensional observation data with discrete distribution in the horizontal direction to the horizontal grid {x1, x2, ..., x m}×{y1,y2,…,y n} corresponding grid points; in this embodiment, the resolution of the horizontal grid is 0.25°, and the value corresponding to each horizontal grid to be interpolated is calculated according to the following formulas (I) and (II):

[0066]

[0067]

[0068] Among them, u estimateis the interpolated ocean feature, w k is the interpolation weight, x k ,y k , t k are the longitude, latitude and time of the kth point in the interpolation neighborhood, respectively, x o ,y o , t o are the latitude and longitude values ​​and time of the target interpolation point, respectively; D is the spatial horizontal interpolation radius; T is the temporal dimension interpolation radius; and N is the number of all observation data contained in the spatiotemporal neighborhood with the spatial radius D and the temporal dimension radius T centered on the target grid; in this embodiment, D = 0.5°, T = 15d;

[0069] A23. Use bilinear interpolation to interpolate the two-dimensional ocean satellite observation data to the horizontal grid {x1, x2, ..., x m}×{y1,y2,…,y n};

[0070] A24. Use bilinear interpolation to interpolate the three-dimensional grid ocean model and reanalysis data (including HYCOM data, SODA data, and EN4-analysis data) to the three-dimensional grid {x1, x2, ..., x m}×{y1,y2,…,y n}×{z1,z2,…,z k};

[0071] A25. Interpolate the spatially interpolated multi-source ocean data to the same temporal resolution.

[0072] (3) Normalization of multi-source ocean data: Specifically, normalization of multi-source ocean data and spatiotemporal information data after spatiotemporal interpolation to unify the dimensions. The spatiotemporal information data includes the year, month, longitude, and latitude. Due to the particularity of longitude (180°E and 180°W coincide), the longitude is normalized according to the following formula (III):

[0073]

[0074]

[0075] The remaining spatiotemporal information data (including year, month, and latitude) are normalized to their maximum and minimum values ​​according to the following formula (IV):

[0076]

[0077] Among them, x norm represents the value obtained after normalization of a certain spatiotemporal information data (feature), x represents the original value of the spatiotemporal information data, Xmax and X min Respectively represent the maximum and minimum values ​​of the spatiotemporal information data in all samples;

[0078] S3. Prepare a sample set and divide it into training set, test set and validation set;

[0079] By analyzing the relationship between spatiotemporal information data, sea surface information data and multi-source observation data, key elements are extracted to create a sample set, which specifically includes the following steps:

[0080] S31, the one-dimensional observation data (from the EN4-profiles dataset) interpolated to the vertical standard layer is used as the sample label;

[0081] Since the number of vertical standard layers in this embodiment is 23, the dimension of the label in a sample is recorded as 1×23, as shown in Figure 2 As shown, Li represents the observation value in the EN4-profiles data of the i-th layer;

[0082] S32. The EN4-analysis data, HYCOM data, SODA reanalysis data, the difference between EN4-analysis data and HYCOM data, the difference between HYCOM data and SODA reanalysis data, the difference between EN4-analysis data and SODA reanalysis data, the spatiotemporal information data where the sample label is located, and the two-dimensional ocean satellite observation data of the grid corresponding to the sample label are used as sample feature data, wherein the EN4-analysis data, HYCOM data, SODA reanalysis data, the difference between EN4-analysis data and HYCOM data, the difference between HYCOM data and SODA reanalysis data, and the difference between EN4-analysis data and SODA reanalysis data constitute six feature channels, the spatiotemporal information data where the sample label is located and the two-dimensional ocean satellite observation data (i.e., sea surface satellite observation data) of the grid corresponding to the sample label constitute one feature channel, and a sample set is made to obtain the arrangement of the above-mentioned sample feature data as follows: Figure 2As shown in , Ei, Hi, and Si represent the analysis data of the i-th layer in the same latitude and longitude grid as the label, (EH)i represents the difference between the EN4-analysis and HYCOM data of the i-th layer, (ES)i represents the difference between the EN4-analysis data and the SODA reanalysis data of the i-th layer, and (HS)i represents the difference between the HYCOM data and the SODA reanalysis data of the i-th layer. Hydro contains the spatiotemporal information of the sample label and the two-dimensional ocean satellite observation data of the corresponding grid. lat is the latitude of the grid corresponding to the sample label, lon1 and lon2 represent the longitude of the grid corresponding to the sample label, year is the year of the sample label, month is the month of the sample label, sla is the sea surface height anomaly at the location corresponding to the sample label, ccmp_u and ccmp_v are the latitudinal and longitudinal components of the sea surface wind field at the location corresponding to the sample label, ersst, ghrsst, and hadisst are three different sets of sea surface temperature data at the location corresponding to the sample label. Finally, Hydro is padded with 0 to a length of 23, so the dimension of the sample feature array can be recorded as 7×23.

[0083] S4. Build a deep learning network model and use the training set to train the deep learning network model. Compare the output results with the labeled data. By minimizing the difference between the network model output and the labeled data, optimize the parameters and build an intelligent fusion model for ocean data.

[0084] The present invention independently builds a deep learning network model and names the built deep learning network model ODF-Net. The ODF-Net network model structure is as follows Figure 2 As shown, it contains 4 basic blocks with residual structures (block1, block2, block3 and block4); among them, Input represents the input layer; GAM represents the global attention mechanism; Conv1d represents the 1D convolution layer with a convolution kernel size of 3×1; BN (Batch Normalization) represents the batch normalization layer; APReLU represents the adaptive parameterized ReLU activation function; the Dropout layer temporarily discards neurons from the network with a probability of p=0.5 to prevent the network from overfitting; Max-Pool represents the maximum pooling layer; the Linear layer is used to process the data into one dimension; FC represents the fully connected layer. The block indicated by the dotted rectangular box represents the basic block with a residual structure. Block1 has one less BN layer, one APReLU and one Dropout layer than other blocks. Experiments have confirmed that the four blocks (basic blocks) constructed by the present invention can bring the best fusion effect.

[0085] Among them, the Global Attention Mechanism (GAM) is introduced to amplify the cross-dimensional interactions so as to capture important features in all three dimensions. The structure of the GAM attention mechanism is as follows Figure 5 As shown in the figure, Channel Attention represents the channel attention submodule, Spatial Attention represents the spatial attention submodule, and GAM adopts a serial method of first performing channel attention and then performing spatial attention. Given an input mapping F1, F∈R C×H×W , the intermediate state F2 and output F3 are defined as:

[0086]

[0087]

[0088] Among them, M C and M S They are channel attention map and spatial attention map respectively; Represents element-wise multiplication operation.

[0089] The structure of the channel attention submodule is as follows Figure 6 As shown in the figure, the channel attention submodule uses a three-dimensional permutation to preserve information in three dimensions; it then uses a two-layer MLP (Multi-Layer Perceptron) to amplify cross-dimensional channel spatial dependencies. The MLP is an encoder-decoder structure with a compression ratio of r. Specifically, the input data of dimensions C×W×H is first transformed to W×H×C. It then passes through a simple MLP structure and is converted back to the original dimensions of C×W×H. Finally, a sigmoid function is used to calculate the weight value of each channel.

[0090] The structure of the spatial attention submodule is as follows Figure 7 As shown in the figure, in the spatial attention submodule, two convolutional layers are used to fuse spatial information to focus on spatial information. Specifically, input data of dimension C × H × W first passes through a convolutional layer with a 7 × 7 kernel. After passing through the convolutional layer, the data dimension becomes C / r × H × W. Then it passes through another convolutional layer with a 7 × 7 kernel. After passing through this convolutional layer, the data dimension returns to the original C × H × W. Finally, the sigmoid function is used to calculate the weight of each eigenvalue.

[0091] APReLU (Adaptively Parametric Rectifier Linear Unit, APReLU for short), or the adaptive parameterized ReLU activation function, is an optimization based on the PReLU (Parametric ReLU) activation function. PReLU sets the coefficient of features less than zero to a trainable parameter, and trains it along with other parameters using gradient descent during neural network training. APReLU, on the other hand, passes each sample through a small, fully connected network to obtain its own corresponding weight, and then uses this set of weights as the coefficient of the original eigenvalue when the eigenvalue is less than zero, i.e., the weight of the negative part of the PReLU function.

[0092] The basic principle of APReLU is as follows Figure 8 As shown in the figure, ReLU represents the ReLU activation function; min(x,0) represents the selection of the smaller value between x and 0; max(x,0) represents the selection of the larger value between x and 0; GAP represents global average pooling; Concat represents channel concatenation; FC represents the fully connected layer; BN represents batch normalization operation; Sigmoid represents the Sigmoid function, which is used to obtain a set of weight values;

[0093] S5. Input the multi-source ocean data to be fused into the ocean data intelligent fusion model and output the intelligent fusion result.

[0094] S6. Input the test set into the ocean data intelligent fusion model obtained in step S4 and output the model test results. Analyze the performance of the network model based on the speed and number of parameters of the model during the test. Calculate and visualize the error indicators based on the model test results and the test sample labels to evaluate the accuracy of the verification model and the quality of data fusion.

[0095] Example 2

[0096] This embodiment provides an intelligent system for a three-dimensional ocean environment field based on deep learning, including:

[0097] Data acquisition and processing module, used to acquire multi-source ocean data and perform data processing;

[0098] A sample set production module is used to analyze multi-source ocean data and produce them into sample sets;

[0099] Ocean data intelligent fusion model building module, used to build and train deep learning network models;

[0100] An intelligent fusion result output module is used to input the multi-source ocean data to be fused into the ocean data intelligent fusion model and output the intelligent fusion result;

[0101] The result verification module is used to evaluate the performance and accuracy of the constructed ocean data intelligent fusion model and the data fusion quality.

[0102] Example 3

[0103] This embodiment uses the intelligent fusion method of the three-dimensional ocean environment field based on deep learning in Example 1 to intelligently fuse the global ocean standard layer sea temperature data to obtain a three-dimensional fused sea temperature field. The specific results are shown in Tables 1, 2, Figure 9-a 、 Figure 9-b and Figure 9-c shown.

[0104] As shown in Table 1, the prediction error of the ODF-Net model constructed by the present invention is significantly smaller than that of the model without APReLU and GAM, indicating that APReLU and GAM have a positive effect on the prediction of the intelligent fusion model of ocean data.

[0105] From Table 2, Figure 9-a 、 Figure 9-b and Figure 9-c The results show that compared with En4_Anal, HYCOM and SODA data, the fused data obtained by the method of the present invention is closer to the labeled observation data.

[0106] Table 1

[0107] APReLU GAM MAE RMSE × × 0.384 0.687 √ × 0.369 0.662 √ √ 0.349 0.628

[0108] Note: √ indicates joining; × indicates not joining.

[0109] Table 2

[0110] Data Name MAE RMSE En4 Anal 0.468 0.799 HYCOM 0.576 0.966 SODA 0.476 0.870 Global Ocean Standard Layer Sea Temperature Fusion Data 0.349 0.628

[0111] The present invention is not limited to the above-mentioned specific implementation methods. Various changes made by ordinary technicians in this field based on the above-mentioned concept without creative work are all within the scope of protection of the present invention.

Claims

1. An intelligent fusion method for three-dimensional ocean environment fields based on deep learning, characterized by: The following steps are involved: S1. Acquire multi-source ocean data, wherein the multi-source ocean data includes one-dimensional observation data, two-dimensional ocean satellite observation data, and three-dimensional gridded ocean model and reanalysis data; The one-dimensional observation data includes Argo, as well as sea temperature profile data observed by buoys, submerged buoys and shipborne equipment; the two-dimensional ocean satellite observation data includes sea surface temperature data, sea surface height anomaly data and sea surface wind field data; the three-dimensional gridded ocean model and reanalysis data include HYCOM data, SODA ocean reanalysis data and EN4-analysis data; S2. Processing the acquired multi-source ocean data, including quality control, spatiotemporal interpolation, and normalization. The quality control processing of the multi-source ocean data is specifically: querying, removing, correcting and filling out abnormal values ​​and missing values ​​of one-dimensional observation data in the multi-source ocean data; The spatiotemporal interpolation processing of multi-source ocean data specifically includes the following steps: A21. Use linear interpolation to interpolate the one-dimensional observation data to the vertical standard layer {z1, z2, …, zk}; A22. Use the spatiotemporal weighted interpolation method to interpolate the one-dimensional observation data with discrete distribution in the horizontal direction to the corresponding grid points of the horizontal grid {x1, x2, …, xm}×{y1, y2, …, yn}. A23. Use bilinear interpolation to interpolate the two-dimensional ocean satellite observation data to the horizontal grid {x1, x2, …, xm}×{y1, y2, …, yn}; A24. Interpolate the three-dimensional gridded ocean model and reanalysis data to a three-dimensional grid {x1, x2, …, xm}×{y1, y2, …, yn}×{z1, z2, …, zk} using bilinear interpolation. A25. Interpolate the spatially interpolated multi-source ocean data to the same temporal resolution; The multi-source ocean data is normalized, specifically comprising the following steps: normalizing the multi-source ocean data after spatiotemporal interpolation and the spatiotemporal information data, wherein the spatiotemporal information data includes the year, month, longitude, and latitude; S3. Create a sample set and divide it into training set, test set and validation set; The preparation of the sample set specifically includes the following steps: S31. The one-dimensional observation data interpolated to the vertical standard layer is used as the sample label. The number of vertical standard layers is set to k, and the array dimension of the sample label is recorded as (1, k). S32, using the EN4-analysis data, HYCOM data, SODA reanalysis data, the difference between the EN4-analysis data and the HYCOM data, the difference between the HYCOM data and the SODA reanalysis data, and the difference between the EN4-analysis data and the SODA reanalysis data at the horizontal grid positions corresponding to the one-dimensional observation data as sample feature data to form six feature channels; S33. The spatiotemporal information data of the sample label and the two-dimensional ocean satellite observation data of the grid corresponding to the sample label are used as sample feature data to form a feature channel, and a sample set is prepared. The dimension of the sample feature array is recorded as (7, k); S4. Build a deep learning network model and use the training set to train the deep learning network model to build an intelligent fusion model for ocean data; A global attention mechanism and an adaptive parameterized ReLU activation function are added to the deep learning network model, wherein the global attention mechanism adopts a serial approach of first performing channel attention and then performing spatial attention; The deep learning network model constructed by using the training set is trained, and the specific steps are: using the training set to train the constructed deep learning network model, and comparing the output results with the sample label data, by minimizing the difference between the network model output and the sample label data, optimizing the parameters, and constructing the ocean data intelligent fusion model; S5. Input the multi-source ocean data to be fused into the ocean data intelligent fusion model and output the intelligent fusion result.

2. The intelligent fusion method of three-dimensional ocean environment field based on deep learning according to claim 1 is characterized in that: It also includes S6, inputting the test set into the ocean data intelligent fusion model, outputting the model test results, and analyzing the model performance based on the speed and number of parameters of the model during the test; at the same time, based on the model test results and test sample labels, calculating and visualizing the error indicators to evaluate the accuracy of the verification model and the quality of data fusion.

3. The intelligent fusion system of three-dimensional ocean environment field based on deep learning is characterized by: include: Data acquisition and processing module, used to acquire multi-source ocean data and perform data processing; A sample set production module is used to analyze multi-source ocean data and produce them into sample sets; Ocean data intelligent fusion model building module, used to build and train deep learning network models; An intelligent fusion result output module is used to input the multi-source ocean data to be fused into the ocean data intelligent fusion model and output the intelligent fusion result; The result verification module is used to evaluate and verify the performance and accuracy of the constructed ocean data intelligent fusion model and the data fusion quality.

Images