A method for reconstructing three-dimensional ocean temperature, salinity, density field
Patent Information
- Application Number
- CN202611429516.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-15
- Publication Date
- 2026-10-20
AI Technical Summary
[0005]本申请针对现有海洋环境场重建及观测数据质量控制方法在复杂海域条件下存在多源数据融合不足、近岸及岛礁边界重建失真、垂向结构不连续以及温盐密度物理一致性较差等问题，提出一种面向海洋水文气象观测设备与系统的三维海洋温度、盐度、密度场重建方法
[0024]1、通过分别编码温盐类海表数据、动力类海表数据及海底地形、有效海域和岸岛边界等地理数据，并利用动力门控和几何门控进行协同融合，增强了近岸及岛礁区域的空间特征表达，减少了重建结果的边界错位。
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of marine science and artificial intelligence, and specifically relates to a method for reconstructing three-dimensional ocean temperature, salinity and density fields using dynamic geometrically coupled residual networks. Background Technology
[0002] Marine environmental field reconstruction and observation data quality control are crucial data processing technologies for marine hydrological and meteorological observation equipment and systems. By processing sea surface and profile data acquired from buoys, moorings, shipborne, shore-based, and satellite observation equipment, continuous three-dimensional temperature, salinity, and density fields are formed, providing fundamental environmental information for environmental perception of marine engineering equipment, navigation operation support, offshore engineering construction, and marine observation data quality control. However, actual marine observations are easily affected by factors such as sparse equipment deployment, sensor drift or failure, communication interruptions, complex sea conditions, and land-sea boundary obstruction, leading to problems such as inconsistent spatiotemporal resolution, local missing values, outliers, and discontinuous depth layers in the observation data. Furthermore, marine engineering equipment is typically limited by computing power, energy, and communication bandwidth, and requires rapid acquisition of reliable environmental information from nearshore areas, islands, and complex sea areas. Therefore, proposing a method applicable to marine hydrological and meteorological observation equipment and systems that balances environmental field reconstruction and observation data quality control is of significant engineering importance.
[0003] As key technologies for ocean observation data processing, existing methods mainly include threshold- and climatological quality control, spatial interpolation, empirical orthogonal decomposition, data assimilation, and deep learning reconstruction methods. Traditional threshold testing and neighboring observation comparison methods can identify some obvious anomalous data, but they struggle to distinguish between real strong changes in complex sea conditions and sensor anomalies. Spatial interpolation and empirical orthogonal decomposition methods are insufficiently adaptable to large-scale continuous missing data and complex coastlines and island boundaries, easily leading to overly smoothed local structures. While deep learning methods possess nonlinear representation capabilities, they still suffer from the following problems: 1. Existing methods typically fuse multi-source sea surface observation data directly, lacking the coordinated utilization of temperature and salinity information, dynamic information, and geographical constraints such as seabed topography, effective sea area, and shore-island boundaries, resulting in boundary misalignment in nearshore and island / reef area reconstructions; 2. Existing methods often treat depth as a common spatial dimension, failing to adequately model the vertical transfer relationship between adjacent depth layers, easily leading to discontinuities in temperature and salinity profiles; 3. Existing methods do not adequately consider the state equation relationship between temperature, salinity, and density, static stability, and strong anomaly regions, making it difficult to balance the accuracy of environmental field reconstruction with the requirements of observation data quality control, which is detrimental to their application in resource-constrained marine observation equipment and engineering equipment.
[0004] Therefore, this application proposes a method for reconstructing three-dimensional ocean temperature, salinity, and density fields for marine hydrological and meteorological observation equipment and systems. This method first performs spatiotemporal registration, normalization, and climatological state construction on multi-source sea surface data and geographic data; secondly, it extracts marine environmental field characteristics using the geometric perception surface representation module and the depth perception vertical propagation module in a dynamic geometrically coupled residual network; finally, through a climatologically guided multi-head prediction module, it outputs three-dimensional temperature, salinity, and density fields, providing continuous and coordinated environmental information for environmental perception and observation data quality control of marine engineering equipment. Summary of the Invention
[0005] This application addresses the problems of insufficient multi-source data fusion, distortion in nearshore and island / reef boundary reconstruction, discontinuous vertical structure, and poor physical consistency of temperature, salinity, and density in existing marine environmental field reconstruction and observation data quality control methods under complex marine conditions. It proposes a three-dimensional marine temperature, salinity, and density field reconstruction method for marine hydrological and meteorological observation equipment and systems. The method is characterized by: firstly, spatiotemporal registration and climatological construction of sea surface temperature, salinity, dynamics, and geographic data; secondly, constructing a dynamic geometrically coupled residual network, and extracting three-dimensional marine environmental features using gated fusion, geographic constraint modulation, and vertical propagation; and finally, predicting temperature, salinity, and density residuals against the background of temperature, salinity, climatological conditions, and state equation reference density, and completing network training by combining physical consistency constraints, outputting three-dimensional temperature, salinity, and density fields for environmental perception and observation data quality control of marine engineering equipment.
[0006] To achieve the above objectives, the present invention adopts the following technical solution.
[0007] A method for reconstructing three-dimensional ocean temperature, salinity, and density fields for environmental perception of marine engineering equipment, characterized by the following steps:
[0008] Step 1: Collect sea surface input data and geographic data for the target sea area. The sea surface input data is divided into temperature and salinity data and dynamic data. The temperature and salinity data includes sea surface temperature and sea surface salinity, while the dynamic data includes sea surface height anomaly data and 10-meter height latitude and longitude wind data. The geographic data includes seabed topography, effective sea area mask, coastline boundary mask, and island boundary mask encoding.
[0009] Step 2: Perform spatiotemporal registration on the sea surface input data and geographic data according to a unified timestamp, horizontal coordinates, and depth layer, registering them as continuous spatiotemporal variables; perform interpolation registration on the continuous variables of temperature, salinity, and dynamics in the sea surface input data, and use nearest neighbor registration on the geographic data; divide the processed data into training set and validation set according to time order, and then calculate the temperature and salinity climatology of the training set data by month and depth layer;
[0010] Step 3: Construct a dynamic geometrically coupled residual network; the dynamic geometrically coupled residual network is divided into three modules: a geometrically sensing surface representation module, a depth-sensing vertical propagation module, and a climate-guided multi-head prediction module;
[0011] Step 4: Input the preprocessed data into the geometric perception surface representation module. Encode the temperature-salinity sea surface input data and the dynamic sea surface input data into temperature-salinity features and dynamic features, respectively. Fuse the temperature-salinity features and dynamic features to obtain the surface fusion features. Encode the geographic data into geographic features. Input the two types of encoded data into the parallel residual branch that has been geometrically gated and spatially gated and weighted, and output the two-dimensional surface feature data.
[0012] Step 5: Input the two-dimensional surface feature data output from Step 4 into the depth-aware vertical propagation module, expand the input data to a preset depth layer, superimpose learnable depth embedding data to form three-dimensional feature data, and then rearrange it into a depth sequence at each horizontal position. Using a vertical propagation layer with shared horizontal position parameters, the interaction between adjacent depth layers is realized through depth local convolution and gated residual update, and the three-dimensional feature volume is output.
[0013] Step Six: Input the 3D feature volume output from Step Five into the climatology-guided multi-head prediction module. Using temperature and salinity climatologies as background, predict the temperature residual and the regionally constrained salinity residual respectively, and superimpose them with the corresponding climatologies to obtain the reconstructed temperature field and reconstructed salinity field. Based on the reconstructed temperature and salinity fields, calculate the reference density using the seawater state equation. Aggregate the 3D feature volume obtained from vertical propagation along the depth direction and predict the density residual together with the reference density. Superimpose the density residual with the reference density element by element to obtain the reconstructed density field. Further, the specific implementation steps of data preprocessing are as follows:
[0014] Both types of data were preprocessed and uniformly transformed to the same coordinate system. Spatiotemporal registration was performed by establishing a correspondence between preset timestamps, horizontal grid coordinates, and depth layers. For data lacking corresponding timestamps, linear interpolation between adjacent time points was used for time registration, ultimately resulting in continuous spatiotemporal variables. Specifically, for continuous variables of temperature, salinity, and dynamics, bilinear weighted interpolation was performed based on the latitude and longitude positions of the target grid points in the original grid cells, using four adjacent grid points. For geographic data, nearest neighbor assignment was performed based on the nearest original grid point corresponding to the center of the target grid to maintain the geographic category and boundary information. Normalization was performed using the mean and standard deviation of the training set, and the training set and validation set were divided into the first 80% and the last 20% of the time series, respectively. Temperature and salinity climatologies were taken from the mean of the target field along the time dimension in the same month and at the same depth layer of the training set, respectively. Missing positions in the preprocessed data were filled with the corresponding mean, and the validation set did not participate in the climatology construction.
[0015] Furthermore, the specific details of the dynamic geometrically coupled residual network are as follows:
[0016] The dynamic geometrically coupled residual network comprises a geometrically perceptive surface representation module, a depth-perceptive vertical propagation module, and a climate-guided multi-head prediction module. The geometrically perceptive surface representation module includes surface feature encoding, temperature, salinity, and dynamic feature fusion, geographic feature encoding, geometrically gated modulation, and parallel residual branching. It is used to fuse sea surface temperature, salinity, and dynamic information and incorporate seabed topography and marine boundary constraints to extract two-dimensional surface features. The depth-perceptive vertical propagation module includes depth extension, learnable depth embedding, and a vertical propagation layer. It is used to achieve feature interaction between adjacent depth layers and generate a three-dimensional feature volume. The climate-guided multi-head prediction module includes a temperature prediction head, a salinity prediction head, a density prediction head, and a state equation calculation unit. It is used to reconstruct the temperature and salinity fields against the background of temperature and salinity climate states, and then reconstruct the salinity field using the seawater state equation.
[0017] Furthermore, the specific implementation steps of the geometric perception surface representation module are as follows:
[0018] Preprocessed sea surface temperature, salinity, and dynamic data are input into the temperature, salinity, and dynamic data encoders, respectively. Corresponding features are extracted through downsampling and bilinear upsampling. Dynamic gating weights are generated based on the temperature, salinity, and dynamic features. The dynamic features are modulated element-wise and then concatenated with the temperature, salinity, and dynamic features to obtain fused surface dynamic features. Geographic data is convolved, normalized, and activated before being concatenated to obtain geographic features. Geometric gating weights are generated by concatenating geographic features with fused surface dynamic features, and element-wise geometric modulation is applied to the fused surface dynamic features. The modulation results are input into four parallel residual branches. The branch weights are determined by normalizing the gating scores at each horizontal grid position. The weighted sum of the outputs of each branch is then obtained to obtain the two-dimensional surface features.
[0019] Furthermore, the specific implementation steps of the depth perception vertical propagation module are as follows:
[0020] Two-dimensional surface features are copied along the depth direction and superimposed with depth embedding data corresponding to each depth layer to form an initial three-dimensional feature volume. The initial three-dimensional feature volume is rearranged according to the dimensions of batch, horizontal grid position, channel, and depth, forming a depth sequence arranged along the depth direction at each horizontal grid position. Each depth sequence is then input into the vertical propagation block in the vertical propagation layer, which shares parameters at each horizontal position. One-dimensional convolution along the depth direction is used to extract information from adjacent depth layers, and channel mixing is performed through pointwise convolution. Candidate updates are generated through normalization, activation, discarding, and feature gating. Then, gated residual updates are performed using restricted learnable residual coefficients. After vertical propagation processing, the depth sequences are rearranged back into a three-dimensional arrangement to obtain the three-dimensional feature volume.
[0021] Furthermore, the specific implementation steps of the climate-guided multi-prediction module are as follows:
[0022] The 3D feature volume is input into the climate-guided multi-head prediction module. Temperature residuals are predicted using 3D convolution against a temperature climate state background. Salinity residuals are then incorporated, integrating their vertical gradient and geographical features. The resulting salinity residuals, constrained by a hyperbolic tangent function, are output via a 3D encoder and decoder. These residuals are then scaled using different coefficients for inland waters, coastlines, and island boundaries to obtain the predicted salinity residuals. The temperature and salinity residuals are then superimposed with their corresponding climate states to obtain the reconstructed temperature and salinity fields. Finally, a reference density is calculated using the linear state equation formula, combining the reconstructed temperature and salinity fields. The mean value of the 3D feature volume along the depth direction is then calculated, and the density residuals are obtained through one-dimensional pointwise convolution. The magnitude of these residuals is limited and added to the reference density to obtain the reconstructed density field.
[0023] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:
[0024] 1. By separately encoding sea surface data of temperature and salinity, sea surface data of dynamics, and geographic data such as seabed topography, effective sea area and shore-island boundaries, and by using dynamic gating and geometric gating for collaborative fusion, the spatial feature expression of nearshore and island areas is enhanced, and boundary misalignment of reconstruction results is reduced.
[0025] 2. By using the depth-sensing vertical propagation module, convolutional interaction and gated residual updates are performed on adjacent depth layers, which enhances the vertical continuity of the three-dimensional temperature, salinity, and density fields and improves the problems of discontinuity in temperature-salinity profiles and insufficient transmission of deep information.
[0026] 3. By introducing constraints such as seawater equation of state consistency, static stability, vertical gradient, thermocline sensitivity, and regional sensitivity, the reconstructed temperature, salinity, and density fields have better physical consistency and the ability to preserve anomalous regions. At the same time, the adoption of parameter sharing and lightweight prediction structure reduces the computational and storage overhead of the model, making it suitable for environmental perception and observation data quality control of resource-constrained marine observation equipment and engineering equipment. Attached Figure Description
[0027] Figure 1 This is a flowchart illustrating the overall process of a three-dimensional ocean temperature, salinity, and density field reconstruction method according to the present invention.
[0028] Figure 2 This is a diagram of the overall framework of the dynamic geometrically coupled residual network in this invention.
[0029] Figure 3The figures show a comparison of the quantitative results of temperature, salinity, and diagnostic density of the present invention and various comparative models on the validation set; wherein, (a) is a comparison of the root mean square error of temperature of the present invention and various comparative models on the validation set, (b) is a comparison of the root mean square error of salinity of the present invention and various comparative models on the validation set, and (c) is a comparison of the root mean square error of density of the present invention and various comparative models on the validation set. Detailed Implementation
[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings:
[0031] Example 1: Figure 1 The overall flow of a three-dimensional ocean temperature, salinity, and density field reconstruction method of the present invention is given; Figure 2 A general framework diagram of the dynamic geometrically coupled residual network is presented. The method for reconstructing the three-dimensional ocean temperature, salinity, and density fields using the dynamic geometrically coupled residual network includes the following steps:
[0032] Step 1: Collect sea surface input data and geographic data for the target sea area. The sea surface input data is divided into temperature and salinity data and dynamic data. The temperature and salinity data includes sea surface temperature and sea surface salinity, while the dynamic data includes sea surface height anomaly data and 10-meter height latitude and longitude wind data. The geographic data includes seabed topography, effective sea area mask, coastline boundary mask, and island boundary mask encoding.
[0033] Step 2: Perform spatiotemporal registration on the sea surface input data and geographic data according to a unified timestamp, horizontal coordinates, and depth layer, registering them as continuous spatiotemporal variables; perform interpolation registration on the continuous variables of temperature, salinity, and dynamics in the sea surface input data, and use nearest neighbor registration on the geographic data; divide the processed data into training set and validation set according to time order, and then calculate the temperature and salinity climatology of the training set data by month and depth layer;
[0034] Step 3: Construct a dynamic geometrically coupled residual network; the dynamic geometrically coupled residual network is divided into three modules: a geometrically sensing surface representation module, a depth-sensing vertical propagation module, and a climate-guided multi-head prediction module;
[0035] Step 4: Input the preprocessed data into the geometric perception surface representation module. Encode the temperature-salinity sea surface input data and the dynamic sea surface input data into temperature-salinity features and dynamic features, respectively. Fuse the temperature-salinity features and dynamic features to obtain the surface fusion features. Encode the geographic data into geographic features. Input the two types of encoded data into the parallel residual branch that has been geometrically gated and spatially gated and weighted, and output the two-dimensional surface feature data.
[0036] Step 5: Input the two-dimensional surface feature data output from Step 4 into the depth-aware vertical propagation module, expand the input data to a preset depth layer, superimpose learnable depth embedding data to form three-dimensional feature data, and then rearrange it into a depth sequence at each horizontal position. Using a vertical propagation layer with shared horizontal position parameters, the interaction between adjacent depth layers is realized through depth local convolution and gated residual update, and the three-dimensional feature volume is output.
[0037] Step Six: Input the 3D feature volume output from Step Five into the climate-guided multi-head prediction module. Using temperature and salinity climate states as background, predict the temperature residual and the regionally constrained salinity residual respectively, and superimpose them with the corresponding climate states to obtain the reconstructed temperature field and the reconstructed salinity field. Based on the reconstructed temperature field and salinity field, calculate the reference density using the seawater state equation. Aggregate the 3D feature volume obtained from vertical propagation along the depth direction and predict the density residual together with the reference density. Superimpose the density residual with the reference density element by element to obtain the reconstructed density field.
[0038] The specific steps for data preprocessing are as follows:
[0039] This case study covers the region from 107.06°E to 119.69°E and from 8.06°N to 20.69°N. Ocean temperature, salinity, and sea level height data are from global ocean physical products published by the Copernicus Marine Service, covering the period from January 1, 2018 to December 31, 2023, with a daily temporal resolution and a total of 2191 samples. After spatial cropping and grid alignment, the horizontal grid is 102×102, with an approximate spatial resolution of 0.125°, and the vertical range extends from the sea surface to 1000 meters, encompassing 36 depth layers. Sea surface temperature and salinity are taken from the uppermost layers of the three-dimensional temperature and salinity fields, respectively, while sea level height anomalies are derived from the two-dimensional sea level height field. The zonal and meridional winds at 10-meter height are from ERA5 single-layer reanalysis data, with the wind fields converted to daily averages and interpolated to the same temporal and spatial grid as the ocean data. The inputs also include monthly sine and cosine codes, normalized latitude and longitude, seabed topography, effective sea area mask, coastline boundary mask, island boundary mask, sea surface height anomaly gradient, wind field derivation, and smooth and Laplace fields constructed from sea surface data. The number of input channels in the formal network configuration is 49.
[0040] Both types of data were preprocessed and uniformly transformed to the same coordinate system. Spatiotemporal registration was performed by establishing a correspondence between preset timestamps, horizontal grid coordinates, and depth layers. For data lacking corresponding timestamps, linear interpolation between adjacent time points was used for time registration, ultimately resulting in continuous spatiotemporal variables. Specifically, for continuous variables of temperature, salinity, and dynamics, bilinear weighted interpolation was performed based on the latitude and longitude positions of the target grid points in the original grid cells, using four adjacent grid points. For geographic data, nearest neighbor assignment was performed based on the nearest original grid point corresponding to the center of the target grid to maintain geographic category and boundary information. Normalization was performed using the mean and standard deviation of the training set, and the training set and validation set were divided into the first 80% and the last 20% of the time series, respectively. Temperature and salinity climatologies were taken from the mean of the target field along the time dimension in the same month and at the same depth layer of the training set, respectively. Missing positions in the preprocessed data were filled with the corresponding mean, and the validation set did not participate in the climatology construction.
[0041] The specific details of the dynamic geometrically coupled residual network are as follows:
[0042] The dynamic geometrically coupled residual network comprises a geometrically perceptive surface representation module, a depth-perceptive vertical propagation module, and a climate-guided multi-head prediction module. The geometrically perceptive surface representation module includes surface feature encoding, temperature, salinity, and dynamic feature fusion, geographic feature encoding, geometrically gated modulation, and parallel residual branching. It is used to fuse sea surface temperature, salinity, and dynamic information and incorporate seabed topography and marine boundary constraints to extract two-dimensional surface features. The depth-perceptive vertical propagation module includes depth extension, learnable depth embedding, and a vertical propagation layer. It is used to achieve feature interaction between adjacent depth layers and generate a three-dimensional feature volume. The climate-guided multi-head prediction module includes a temperature prediction head, a salinity prediction head, a density prediction head, and a state equation calculation unit. It is used to reconstruct the temperature and salinity fields against the background of temperature and salinity climate states, and then reconstruct the salinity field using a linear state equation.
[0043] Furthermore, the specific implementation steps of the geometric perception surface representation module are as follows:
[0044] The preprocessed sea surface temperature, salinity, and dynamic data are input into the temperature, salinity, and dynamic encoders, respectively. Corresponding features are extracted through downsampling and bilinear upsampling. The process is as follows: ,in, and The input data are for thermo-salinity and dynamics, respectively. and These are a temperature-salt encoder and a power encoder, respectively. and These are temperature-salinity characteristics and dynamic characteristics, respectively; dynamic gating weights are generated based on these characteristics. ,in, For power gating mapping, It is the Sigmoid activation function. The dynamic gating weights are used to modulate the dynamic features element-wise, and then concatenated with the temperature and salinity features before being encoded by a fusion encoder into a 48-channel fused surface feature. ,in, Indicates channel splicing. This indicates element-wise multiplication. For fusion encoders, To integrate surface features, the seabed topography, effective sea area mask, coastline boundary mask, and island boundary mask are stitched together and processed through two 3×3 convolutions, batch normalization, and SiLU activation units to obtain 48-channel geographic features: ,in, It is an underwater topography. , and These are respectively effective sea area masks, coastline boundary masks, and island boundary masks. For geographic encoder, For geographic features, the geographic features are concatenated with the fused surface features, and then geometrically gated weights with spatial variations are generated through a geometrical gating mapping consisting of two 1×1 convolutions and an intermediate activation function. And residual geometric modulation of geographic features is performed using a coefficient of 0.25: ,in, For geometrically gated mapping, For geometrically gated weights, The geometric modulation features are used as inputs to four parallel residual experts, each consisting of two layers of convolutional residual blocks. A gating network generates gating scores for each horizontal grid position based on the geometric modulation features, and the scores are normalized using the spatial Softmax function to obtain the weights of the four residual experts. ,in, It is a gated network consisting of two 1×1 convolutions and an intermediate SiLU activation. For the first The weights of the four residual experts at each horizontal grid position are determined. The outputs of the four residual experts are then weighted and summed to obtain the two-dimensional surface features. ,in, For the first A residuals expert, This is the output two-dimensional surface feature.
[0045] Furthermore, the specific implementation steps of the depth perception vertical propagation module are as follows:
[0046] The two-dimensional surface features are copied to 36 depth layers, and the depth embedding parameters automatically learned during network training are added to each depth layer, ultimately forming an initial three-dimensional feature body with a vertical sequence of length 36: ,in, It is a two-dimensional surface feature. For depth layers, This is a learnable deep embedding; the deep embedding parameter data is for 36 preset depth layers, with a learnable channel feature vector set for each depth layer, and the initial value is a small random value; the initial three-dimensional feature volume is rearranged according to the dimensions of batch, horizontal grid position, channel, and depth, forming a depth sequence arranged along the depth direction at each horizontal grid position; then each depth sequence is input into the vertical propagation block in the vertical propagation layer, which shares parameters at each horizontal position, and one-dimensional convolution along the depth direction is used to extract information from adjacent depth layers, channel mixing is performed by pointwise convolution, and candidate updates are generated through normalization, activation, discarding, and feature gating, and then gated residual updates are performed using restricted learnable residual coefficients: ,in, and The first Depth sequences before and after vertical propagation block update, For candidate updates, For feature gating, For learnable residual parameters, This represents element-wise multiplication; after vertical propagation, the depth sequence is rearranged back into a three-dimensional arrangement to obtain a three-dimensional feature volume.
[0047] Furthermore, the specific implementation steps of the climate-guided multi-prediction module are as follows:
[0048] The 3D feature volume is input into the climatology-guided multi-head prediction module. Temperature residuals are predicted using a series of operations: 3×3×3 convolution, batch normalization, SiLU activation, and 1×1×1 convolution, with temperature as the background. Salinity residuals are then incorporated into the salinity prediction branch, which employs a 3D encoder-decoder with two-stage skip connections. During the encoding stage, downsampling is performed only along the two horizontal dimensions, maintaining 36 depth layers. During the decoding stage, trilinear interpolation is used to restore the horizontal resolution, and finally, the salinity residuals are output after being constrained by the hyperbolic tangent function. The salinity residuals are then scaled according to region type; in this example, scaling factors of 1, 0.8, and 0.6 are used for inland waters, coastlines, and island regions, respectively, to obtain the predicted salinity residuals. The temperature and salinity residuals are then superimposed with the corresponding climatologies to obtain the reconstructed temperature and salinity fields. Finally, the linear state equation formula is applied... ,in, Let T be the reference density of the state equation, S be the reconstructed temperature field, and S be the reconstructed salinity field. By combining the reconstructed temperature field and salinity field, the reference density is calculated. Then, the mean value of the three-dimensional feature volume along the depth direction is calculated. The density residual is obtained by one-dimensional pointwise convolution. Its amplitude is restricted within a certain limit and added to the reference density. In this example, the reconstructed density field is obtained by using a 0.1 times hyperbolic tangent function to restrict the correction amplitude.
[0049] Second embodiment: Figure 2 This is a comparison chart of the quantitative results of temperature, salinity, and diagnostic density of the present invention and various comparative models on the validation set. The method of the present invention is compared with climatological filling methods, 3D U-Net, SwinIR-type image restoration networks, MDRF-type multi-scale residual fusion networks, vertical sequence propagation networks, and Pix2Pix conditional generation networks. The MDRF-type multi-scale residual fusion network uses shared multi-scale residual feature extraction and task-specific output branches. The vertical sequence propagation network arranges two-dimensional surface features into a sequence along the depth direction and propagates the features layer by layer along the depth direction. Using the same experimental conditions, data partitioning, preprocessing methods, training parameters, and evaluation metrics as in Example 1, the first 80% of the time series samples were used for training, and the last 20% were used for validation. The root mean square error (RMSE) was calculated within the effective sea area. The RMSE in the figure is calculated separately for variables and depth layers. On the validation set, the predicted values of all effective sea area grid points at the same depth layer are compared with the target values, and the square root of the average of the squared errors is taken.
[0050]
[0051] in, Indicates temperature, salinity, or seawater density. Indicates depth layer, For reconstruction value, For the target value, For effective sea area masking, To validate samples or time points in the set, Number the depth layers. For the latitudinal or row direction index of the horizontal grid, For the longitudinal or column direction index of the horizontal grid, The root mean square error (RMSE) is used; a RMSE value is obtained for each depth layer, and the curves formed along the depth direction are the curves shown in the figure. The units for salinity, salt content, and seawater density are ℃, psu, and kg·m³, respectively. -3 ;Depend on Figure 2It can be seen that the root mean square error (RMSE) of the method of this invention is the lowest among the comparative methods for the three reconstruction tasks of temperature, salinity, and diagnostic density, which are 0.259 degrees Celsius, 0.043 psu, and 0.066 kg per cubic meter, respectively. Compared with the comparative method that performs best among the variables, the RMSE of the method of this invention is reduced by 11.6%, 44.2%, and 29.0%, respectively. This shows that the three-dimensional ocean temperature, salinity, and density field reconstruction method of this invention, which utilizes a dynamic geometrically coupled residual network, can simultaneously improve the reconstruction accuracy of the three-dimensional temperature, salinity, and diagnostic density fields, providing reliable three-dimensional ocean environmental information for environmental perception of marine engineering equipment and quality control of ocean observation data.
[0052] Although some embodiments of the present invention have been shown and described, those skilled in the art will recognize that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present invention, and all such solutions are within the scope of protection of the present invention.
Claims
1. A method for reconstructing three-dimensional ocean temperature, salinity, and density fields, characterized in that, Includes the following steps: Step 1: Collect sea surface input data and geographic data for the target sea area. The sea surface input data is divided into temperature and salinity data and dynamic data. The temperature and salinity data includes sea surface temperature and sea surface salinity, while the dynamic data includes sea surface height anomaly data and 10-meter height latitude and longitude wind data. The geographic data includes seabed topography, effective sea area mask, coastline boundary mask, and island boundary mask encoding. Step 2: Perform spatiotemporal registration on the sea surface input data and geographic data according to a unified timestamp, horizontal coordinates, and depth layer, registering them as continuous spatiotemporal variables; perform interpolation registration on the continuous variables of temperature, salinity, and dynamics in the sea surface input data, and use nearest neighbor registration on the geographic data; divide the processed data into training set and validation set according to time order, and then calculate the temperature and salinity climatology of the training set data by month and depth layer; Step 3: Construct a dynamic geometrically coupled residual network; the dynamic geometrically coupled residual network is divided into three modules: a geometrically sensing surface representation module, a depth-sensing vertical propagation module, and a climate-guided multi-head prediction module; Step 4: Input the preprocessed data into the geometric perception surface representation module. Encode the temperature-salinity sea surface input data and the dynamic sea surface input data into temperature-salinity features and dynamic features, respectively. Fuse the temperature-salinity features and dynamic features to obtain the surface fusion features. Encode the geographic data into geographic features. Input the two types of encoded data into the parallel residual branch that has been geometrically gated and spatially gated and weighted, and output the two-dimensional surface feature data. Step 5: Input the two-dimensional surface feature data output from Step 4 into the depth-aware vertical propagation module, expand the input data to a preset depth layer, superimpose learnable depth embedding data to form three-dimensional feature data, and then rearrange it into a depth sequence at each horizontal position. Using a vertical propagation layer with shared horizontal position parameters, the interaction between adjacent depth layers is realized through depth local convolution and gated residual update, and the three-dimensional feature volume is output. Step Six: Input the 3D feature volume output from Step Five into the climate-guided multi-head prediction module. Using temperature and salinity climate states as background, predict the temperature residual and the regionally constrained salinity residual respectively, and superimpose them with the corresponding climate states to obtain the reconstructed temperature field and the reconstructed salinity field. Based on the reconstructed temperature field and salinity field, calculate the reference density using the seawater state equation. Aggregate the 3D feature volume obtained from vertical propagation along the depth direction and predict the density residual together with the reference density. Superimpose the density residual with the reference density element by element to obtain the reconstructed density field.
2. The method for reconstructing a three-dimensional ocean temperature, salinity, and density field as described in claim 1, characterized in that, The specific implementation steps of step two are as follows: Both types of data were preprocessed and uniformly transformed to the same coordinate system. Spatiotemporal registration was performed by establishing a correspondence between preset timestamps, horizontal grid coordinates, and depth layers. For data lacking corresponding timestamps, linear interpolation between adjacent time points was used for time registration, ultimately resulting in continuous spatiotemporal variables. Specifically, for continuous variables of temperature, salinity, and dynamics, bilinear weighted interpolation was performed based on the latitude and longitude positions of the target grid points in the original grid cells, using four adjacent grid points. For geographic data, nearest neighbor assignment was performed based on the nearest original grid point corresponding to the center of the target grid to maintain the geographic category and boundary information. Normalization was performed using the mean and standard deviation of the training set, and the training set and validation set were divided into the first 80% and the last 20% of the time series, respectively. Temperature and salinity climatologies were taken from the mean of the target field along the time dimension in the same month and at the same depth layer of the training set, respectively. Missing positions in the preprocessed data were filled with the corresponding mean, and the validation set did not participate in the climatology construction.
3. The method for reconstructing a three-dimensional ocean temperature, salinity, and density field as described in claim 1, characterized in that, The specific details of step three are as follows: The dynamic geometrically coupled residual network comprises a geometrically perceptive surface representation module, a depth-perceptive vertical propagation module, and a climate-guided multi-head prediction module. The geometrically perceptive surface representation module includes surface feature encoding, temperature, salinity, and dynamic feature fusion, geographic feature encoding, geometrically gated modulation, and parallel residual branching. It is used to fuse sea surface temperature, salinity, and dynamic information and incorporate seabed topography and marine boundary constraints to extract two-dimensional surface features. The depth-perceptive vertical propagation module includes depth extension, learnable depth embedding, and a vertical propagation layer. It is used to achieve feature interaction between adjacent depth layers and generate a three-dimensional feature volume. The climate-guided multi-head prediction module includes a temperature prediction head, a salinity prediction head, a density prediction head, and a state equation calculation unit. It is used to reconstruct the temperature and salinity fields against the background of temperature and salinity climate states, and then reconstruct the salinity field using the seawater state equation.
4. The method for reconstructing a three-dimensional ocean temperature, salinity, and density field as described in claim 1, characterized in that, The specific implementation steps for step four are as follows: The preprocessed sea surface temperature, salinity and dynamic data are input into the temperature, salinity and dynamic encoders, respectively. Corresponding features are extracted through downsampling and bilinear upsampling. Dynamic gating weights are generated based on the temperature, salinity and dynamic features. The dynamic features are modulated element by element and then concatenated with the temperature, salinity and dynamic features to obtain the fused surface dynamic features. The geographic data are convolved, normalized and activated before being concatenated to obtain the geographic features. Geometric gating weights are generated by concatenating geographic features with fused surface dynamic features, and element-wise geometric modulation is applied to the fused surface dynamic features. The modulation result is input into four parallel residual branches. The branch weights are determined by normalizing the gating scores at each horizontal grid position. The outputs of each branch are weighted and summed to obtain the two-dimensional surface features.
5. The method for reconstructing a three-dimensional ocean temperature, salinity, and density field as described in claim 1, characterized in that, The specific implementation steps of step five are as follows: Two-dimensional surface features are copied along the depth direction and superimposed with depth embedding data corresponding to each depth layer to form an initial three-dimensional feature volume. The initial three-dimensional feature volume is rearranged according to the dimensions of batch, horizontal grid position, channel, and depth, forming a depth sequence arranged along the depth direction at each horizontal grid position. Each depth sequence is then input into the vertical propagation block in the vertical propagation layer, which shares parameters at each horizontal position. One-dimensional convolution along the depth direction is used to extract information from adjacent depth layers, and channel mixing is performed through pointwise convolution. Candidate updates are generated through normalization, activation, discarding, and feature gating. Then, gated residual updates are performed using restricted learnable residual coefficients. After vertical propagation processing, the depth sequences are rearranged back into a three-dimensional arrangement to obtain the three-dimensional feature volume.
6. The method for reconstructing a three-dimensional ocean temperature, salinity, and density field as described in claim 1, characterized in that, The specific implementation steps for step six are as follows: The 3D feature volume is input into the climate-guided multi-head prediction module. Temperature residuals are predicted using 3D convolution against a temperature climate state background. Salinity residuals are then incorporated, integrating their vertical gradient and geographical features. The resulting salinity residuals, constrained by a hyperbolic tangent function, are output via a 3D encoder and decoder. These residuals are then scaled using different coefficients for inland waters, coastlines, and island boundaries to obtain the predicted salinity residuals. The temperature and salinity residuals are then superimposed with their corresponding climate states to obtain the reconstructed temperature and salinity fields. Finally, a reference density is calculated using the linear state equation formula, combining the reconstructed temperature and salinity fields. The mean value of the 3D feature volume along the depth direction is then calculated, and the density residuals are obtained through one-dimensional pointwise convolution. The magnitude of these residuals is limited and added to the reference density to obtain the reconstructed density field.