A deep learning-based method for reconstructing a three-dimensional temperature field of the ocean

Abstract

The application belongs to the technical field of marine remote sensing and artificial intelligence, and particularly relates to a marine three-dimensional temperature field reconstruction method based on deep learning, which comprises the following steps: acquiring multi-source and multi-resolution satellite data and reanalysis data, and constructing a sample library after preprocessing; a three-branch MAUS Transformer model is constructed, a U-Net network is used as a framework, a Swin Transformer module is used to replace convolution blocks embedded in an encoding path and a decoding path, and a CBAM attention module is embedded to adaptively optimize feature expression, and the three-branch model can extract features and fuse them at resolutions of 1 / 100°, 1 / 8° and 1 / 4° respectively; meanwhile, an adaptive loss function with a fusion horizontal partition and a vertical layering is designed, and a real temperature gradient weighted constraint is introduced to reconstruct a thermocline layer. The application can output a marine three-dimensional temperature field with high spatio-temporal resolution, improve the accuracy, stability and robustness of the reconstruction, and has important significance for marine environmental protection, marine data development, marine engineering construction and marine environmental security.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of marine remote sensing technology and artificial intelligence, specifically relating to a method for reconstructing a three-dimensional temperature field of the ocean based on deep learning. Background Technology

[0002] Seawater temperature is a key parameter for describing ocean physical properties and understanding ocean physical processes. Its vertical structure is closely related to seawater stratification, marine ecosystems, and global climate change. Meanwhile, three-dimensional temperature data is also of great significance for marine environmental protection, marine resource development, marine engineering construction, and ensuring marine environmental safety. However, obtaining high spatiotemporal resolution three-dimensional temperature information is an extremely challenging task, especially in marginal sea areas with complex topography.

[0003] In-situ measurements are the most effective way to acquire ocean data, but they have limitations such as high equipment costs, low observation efficiency, and poor data continuity and spatial resolution. Satellite remote sensing provides continuous, large-scale, and high-frequency real-time monitoring of ocean phenomena, with continuous and high-resolution data, but it struggles to penetrate the surface to understand the underwater marine environment. The scarcity of continuous underwater temperature data with high spatiotemporal resolution makes it difficult to fully and clearly understand the rapid and subtle changes occurring within the ocean. Therefore, how to acquire underwater temperature data quickly and efficiently remains a pressing scientific challenge in oceanographic research.

[0004] In recent years, emerging deep learning methods have provided new insights for the reconstruction of three-dimensional ocean temperature fields. By learning the characteristics of historical data and deeply mining the relationship between ocean surface parameters and internal temperature fields, deep learning methods can quickly establish temperature reconstruction relationships and possess good generalization capabilities, enabling large-scale and efficient temperature profile inversion. Due to the combined influence of complex topography and ocean currents in certain sea areas, the marine dynamic environment is extremely active, and the internal temperature structure exhibits highly complex and rapidly changing characteristics. In particular, the thermocline region, with its dramatic vertical temperature gradient and active internal dynamic disturbances, presents significant challenges in reconstructing its three-dimensional structure. Existing methods often employ single-scale network structures, which are prone to problems such as over-smoothing of local features and loss of detail when characterizing the complex and detailed dynamic changes of the ocean, failing to meet the requirements for high-precision reconstruction. Therefore, constructing a multi-scale fusion artificial intelligence method to accurately acquire information on the three-dimensional temperature field, especially the thermocline, is of great scientific value for in-depth exploration of the fine hydrological structure and variation patterns of this sea area. Summary of the Invention

[0005] This invention overcomes the above-mentioned defects and provides a deep learning-based method for reconstructing the three-dimensional temperature field of the ocean. It solves the problems of local information smoothing and low accuracy of thermocline reconstruction caused by the single model scale in the prior art, and improves the reconstruction accuracy of the three-dimensional temperature field of the ocean.

[0006] To achieve the above objectives, this invention provides a deep learning-based method for reconstructing a three-dimensional ocean temperature field, comprising the following steps: S1. Acquire sea surface variable data and ocean reanalysis data from multi-source satellite remote sensing observations at different resolutions in the study area, and perform quality control and spatiotemporal matching; S2. Add the corresponding latitude and longitude information and time information to the dataset and preprocess it to form the sample database required for model training; S3. Construct a multi-branch MAUS Transformer reconstruction model, design an adaptive loss function that integrates horizontal partitioning and vertical layering, complete model training, and save the optimal model. S4. Input the sea surface data of the area to be reconstructed into the optimal model and output the final three-dimensional ocean temperature field reconstruction result.

[0007] Furthermore, the multi-source satellite remote sensing observation data at different resolutions are used as input features, including sea surface temperature at 1 / 100°, sea surface precipitation at 1 / 10°, sea surface salinity at 1 / 8°, sea level anomaly, and sea surface wind field and sea surface current field data at 1 / 4°. The ocean reanalysis data, used as ground truth labels, represents the temperature field at depths of 5m to 222m, spanning 18 layers, with a spatial resolution of 1 / 12°.

[0008] Furthermore, the input features also include wind stress curl data, formulated as follows: , Where τ is the sea surface wind stress, and These are the zonal and meridional components of sea surface wind stress, respectively. This is the drag coefficient. air density; u and v are the zonal and meridional wind speeds at a height of 10m above the sea surface, respectively; sea surface wind speed. .

[0009] Furthermore, the quality control and spatiotemporal matching are specifically as follows: The spatial resolution of the 1 / 10° sea surface precipitation data was resampled to 1 / 8°, and the temporal resolution of all data was resampled to daily data. Finally, input feature data with three different spatial resolutions of 1 / 100°, 1 / 8°, and 1 / 4° and daily temporal resolution, and ocean reanalysis data with a spatial resolution of 1 / 12° and daily temporal resolution were generated.

[0010] Furthermore, in step S3, the multi-branch MAUS Transformer model uses the U-Net network as a framework, replaces the convolutional blocks with the Swin Transformer module in the encoding and decoding paths, and embeds the CBAM attention module to adaptively optimize feature representation.

[0011] Furthermore, the multi-branch MAUS Transformer model has a three-branch structure, corresponding to input features at resolutions of 1 / 100°, 1 / 8°, and 1 / 4°, respectively.

[0012] Furthermore, the feature extraction branch with a resolution of 1 / 100° is specifically implemented as follows: S301, Encoding stage: Data undergoes preliminary feature mapping through patch embedding layer and preliminary feature extraction through two layers of Swing Transformer blocks; S302. The initially extracted features are downsampled twice, and the features after each downsampling are further processed by two layers of Swing Transformer blocks for deep feature extraction. After the features after the second downsampling are extracted by two layers of Swing Transformer blocks, they are then processed by the CBAM module for channel and spatial dual attention weighting to highlight key features. S303. Entering the decoding stage, the features output by the encoder are upsampled twice. After each upsampling, the features are skipped and connected with the features of the corresponding scale in the encoding path that have been enhanced by the CBAM module. The spliced ​​and fused features are integrated through two layers of Swing Transformer blocks. S304. Resample the integrated features and reshape the feature map dimension to match the subsurface temperature field to obtain the final output feature of this branch. The feature extraction branches for 1 / 8° and 1 / 4° resolutions perform downsampling only once during the encoding stage and upsampling once during the decoding stage. The remaining steps are consistent with those for the feature extraction branch for 1 / 100° resolution.

[0013] Furthermore, the adaptive loss function incorporates gradient loss to evaluate the difference in temperature loss values ​​of the thermocline. The adaptive loss function formula is as follows: , in, The Huber loss values ​​represent the mixed layer and deep layer in nearshore or offshore environments. The Huber loss value represents the thermocline in the nearshore or offshore area. This represents the gradient loss of the thermocline near the shore or offshore, where a, b, and c represent the weight parameters of each loss.

[0014] Furthermore, the gradient loss incorporates the absolute value of the true temperature gradient. To accurately fit key physical abrupt change regions, the gradient loss formula is as follows: , , , in, and Let these represent the target value and the reconstructed value of the temperature field in the i-th layer, respectively. This represents the temperature difference between the (i+1)th layer and the target temperature field value of the ith layer. The temperature difference between the reconstructed temperature field values ​​of layer i+1 and layer i This indicates the number of layers between the upper and lower boundaries of the ocean thermocline. This represents the absolute value of the actual temperature gradient.

[0015] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: (1) This invention can simultaneously output three-dimensional temperature field data of sea areas from 5m to 222m. By considering the physical correlation of water temperature between layers, physical constraints are introduced, which significantly improves the time efficiency of the model. (2) This invention innovatively constructs an adaptive loss function that integrates regional partitioning and thermocline gradient. Based on physical heterogeneity, the system divides the region into nearshore and offshore areas at a water depth of 200m at the horizontal level and divides the profile into thermocline, mixed layer and deep layer at the vertical level. It also introduces gradient loss with real temperature gradient weight to thermocline, and applies stronger constraints to the vertical temperature drastic change area, effectively reducing thermocline reconstruction error and greatly improving the vertical structure reconstruction accuracy. (3) The present invention innovatively constructs a multi-branch MAUS Transformer architecture, which can efficiently integrate the features of sea surface elements with high, medium and low spatial resolutions, avoid the problem of excessive smoothing in traditional feature extraction, effectively capture ocean dynamic features at different spatial scales, and achieve high-fidelity characterization of fine structures such as mesoscale eddies. Attached Figure Description

[0016] Figure 1 This is a flowchart of the ocean three-dimensional temperature field reconstruction method of the present invention; Figure 2 This is a network structure diagram of the marine three-dimensional temperature field reconstruction model of the present invention; Figure 3 This is a cross-sectional comparison of the adaptive loss function and the Huber loss function RMSE and R in an embodiment of the present invention; Figure 4 This is a comparison chart of the average RMSE and R values ​​of the adaptive loss function and the Huber loss function in each month according to an embodiment of the present invention. Figure 5 This is a schematic diagram comparing the RMSE and R values ​​of the three-dimensional temperature field reconstruction model of this invention with those of other reconstructed temperature fields. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, the deep learning-based method for reconstructing the three-dimensional temperature field of the ocean provided by this invention is as follows: I. Acquiring multi-source satellite remote sensing observation data and ocean reanalysis data required for the study area (1) Multi-source satellite remote sensing observation data The data obtained from multi-source satellite remote sensing observations of a certain sea area are used as input features, including sea surface temperature data at 1 / 100°, sea surface precipitation data at 1 / 10°, sea surface salinity data at 1 / 8°, sea level anomaly data at 1 / 8°, sea surface wind field data at 1 / 4°, and sea surface current field data at 1 / 4°. Considering that sea surface wind stress curl can influence the distribution of the thermocline through Ekman suction, wind stress curl is used as one of the model inputs in the reconstruction of the upper ocean temperature field to improve the reconstruction accuracy of the temperature field in the subsurface ocean, especially near the thermocline, from a dynamic perspective. Based on the zonal and meridional components of the sea surface wind field, the wind stress curl can be calculated, and its specific expression is as follows: , Where τ represents sea surface wind stress, x represents zonal stress, and y represents meridional stress. and These are the zonal and meridional components of sea surface wind stress, respectively. This represents the rate of change of the meridional component of wind stress in the x-direction; This represents the rate of change of the zonal component of wind stress in the y-direction; This is the drag coefficient. air density; u and v are the zonal and meridional wind speeds at a height of 10m above sea level, respectively. This represents the spatial rate of change of meridional wind speed in the x-direction. Represents the spatial rate of change of zonal wind speed in the y-direction; sea surface wind speed , This represents the rate of change of sea surface wind speed in the x-direction. This represents the rate of change of sea surface wind speed in the y-direction.

[0019] (2) Ocean reanalysis data The ocean reanalysis data used in this embodiment of the invention is GLORYS12V1, a CMEMS global ocean eddy resolution reanalysis data product with a spatial resolution of 1 / 12°, a temporal resolution of daily average, and 50 vertical layers, covering altimeter data (since 1993). Based on the real-time global CMEMS system, the observation data is assimilated using reduced-priority Kalman filtering, including altimeter data, sea surface temperature data, sea ice data, and vertical profiles of temperature and salinity observed in the field. This product provides daily average data including sea surface temperature (SST), salinity (SSS), mixed layer depth (MLD), and sea surface height (SSH). In this embodiment, the reanalysis data is used as the ground truth label, selecting 18 layers of temperature fields with a spatial resolution of 1 / 12° from 5m to 222m underwater, covering the time range from January 1, 2010 to December 31, 2020.

[0020] (3) Perform quality control and spatiotemporal matching on the data. The spatial resolution of the 1 / 10° sea surface precipitation data was resampled to 1 / 8°, and the temporal resolution of all data was resampled to daily data. Finally, input feature data with three different spatial resolutions (1 / 100°, 1 / 8°, and 1 / 4°) and daily temporal resolutions were generated, along with monthly ocean reanalysis data with a spatial resolution of 1 / 12° and daily temporal resolutions.

[0021] 2. Add the corresponding latitude, longitude, and time information to the dataset, and perform data preprocessing.

[0022] Longitude, latitude, and sin(2π×yearly days / 365.25) and cos(2π×yearly days / 365.25) are added to the current sea surface variables at 1 / 100°, 1 / 8°, and 1 / 4° respectively to highlight the spatial and temporal differences of the variables. The data are then preprocessed as follows: First, the quality of the matched sea surface data and ocean reanalysis data is checked to identify and handle possible outliers or missing values, and these are set to -999. Secondly, to address the differences in range and magnitude among sea surface data from different channels, and to ensure consistency in numerical distribution between input features and prediction targets, a unified Z-Score normalization method was applied to all sea surface data. The specific calculation formula is as follows: , in, This indicates that the raw data that needs to be standardized is a numerical value of a single sea surface data point; and These are the mean and standard deviation calculated for the corresponding data; The standardized data will approximate a standard normal distribution with a mean of 0 and a standard deviation of 1. Standardization can make the subsequent model training process more stable and converge faster.

[0023] Finally, the dataset is divided. In this embodiment of the invention, data from a certain sea area from 2010 to 2019 are randomly divided into a training set and a validation set at a ratio of 9:1, and the data from 2020 is used as a test set to evaluate the effectiveness of subsequent methods.

[0024] III. Constructing a Multi-Branch MAUS Transformer Reconstruction Model and Adaptive Loss Function like Figure 2 As shown, this invention constructs a multi-branch MAUS Transformer ocean three-dimensional temperature field reconstruction model. Using the U-Net network as the basic framework, it embeds the Swin Transformer and CBAM attention modules to design a three-branch deep learning model. Data with different spatial resolutions are used to extract features according to the corresponding branches. Then, the features finally extracted from different branches are fused to output the reconstructed temperature field.

[0025] According to the embodiments of the present invention, feature extraction modules with different complexities are designed based on data with three different spatial resolutions: 1 / 100°, 1 / 8°, and 1 / 4°. Each module extracts multi-level depth features through an encoder and restores spatial details through a decoder combined with skip connections, and finally outputs feature extraction results with the same size as the input.

[0026] Specifically, using the U-Net network as the basic framework, the convolutional blocks in the encoding and decoding paths are replaced with Swing Transformer blocks. A CBAM attention module is added, and feature representation is adaptively optimized through channel and spatial attention mechanisms to accurately focus on high-weight key features. The feature data extracted from each branch is resampled to 1 / 12° and concatenated in the feature fusion module, ultimately outputting a reconstructed three-dimensional temperature field, as follows: (1) Feature extraction branch with spatial resolution of 1 / 100° Taking advantage of the rich detail in 1 / 100° data, this branch employs a double-downsampling depth feature extraction architecture to fully extract the deep information from ocean surface data: In the encoding path, data undergoes initial feature mapping through a patch embedding layer. The mapped feature maps are then fed into a two-layer Swing Transformer block for processing. The Swing Transformer utilizes multi-head self-attention with local windows (W-MSA) and shift windows (SW-MSA) to simultaneously capture short-range and long-range spatial dependencies. The features processed by the two-layer Swing Transformer are then computed in parallel along two paths: the main path A and skip connections A. The main path A undergoes a first downsampling and deep feature extraction using two Swing Transformer layers. It is then further divided into two parallel paths: the main path B and skip connections B. The main path B also undergoes a second downsampling and two Swing Transformer layers. Afterward, it is processed by the CBAM module for channel and spatial dual attention weighting to highlight key features and suppress noise interference before entering the decoding path. Features in both skip connections A and B are enhanced by the CBAM module and then concatenated with features of the corresponding scale in the decoding path.

[0027] In the decoding path, the features gradually restore spatial resolution through two upsampling steps. The features after the first upsampling are concatenated with the features in the skip connection B path. The concatenated and fused features are then integrated by a two-layer Swing Transformer block before entering the second upsampling. The features after the second upsampling are concatenated with the features in the skip connection A path. The concatenated and fused features are then integrated and optimized again by a two-layer Swing Transformer block.

[0028] This invention combines upsampled features with skip connections to fuse shallow local information with deep global information, thus compensating for the spatial information loss caused by downsampling. Finally, after a resampling operation, the dimensions of the feature map are directly reshaped to the corresponding dimensions of the subsurface parameters, thereby generating the final output of this branch.

[0029] (2) Feature extraction branch with spatial resolution of 1 / 8°

[0030] Compared to the 1 / 100° fine-grained branch, the 1 / 8° feature extraction branch has a relatively simplified network structure: after the data undergoes initial feature mapping through the Patch Embedding layer, it is immediately fed into a two-layer Swing Transformer block for processing. To minimize the loss of detail in low-resolution remote sensing data and improve model convergence speed, this branch performs only one downsampling operation in the feature encoding path. The downsampled features are then extracted through the two-layer Swing Transformer block. Subsequently, in the decoding stage, the features undergo one upsampling and are then stitched together with the feature maps of the same size in the encoding path that have already passed through the CBAM attention module. The stitched features are then processed again by the two-layer Swing Transformer block and reshaped into the corresponding dimensions of the subsurface field through resampling, thereby generating the final output of this branch.

[0031] The feature extraction branch with a spatial resolution of 1 / 4° is consistent with the 1 / 8° branch, and it also adopts a simplified structure with single downsampling. Finally, it also outputs a feature map that is reshaped into the subsurface field dimension.

[0032] Finally, the high-dimensional feature maps generated by the three branches are spliced ​​together in the channel dimension to achieve feature fusion. After processing by a 1*1 convolutional layer and a linear activation function, the model finally generates the temperature field reconstruction result of the subsurface layer of a certain sea area.

[0033] This invention's model, based on the shift window design of the Swing Transformer module, overcomes the bottleneck of the local receptive field in traditional convolutional blocks, constructing extensive spatial correlations. While ensuring computational efficiency, this architecture enhances the model's ability to mine deep information from sea surface remote sensing observations by collaboratively representing local features and macroscopic background field information, thereby achieving high-fidelity reconstruction of the subsurface three-dimensional temperature field. Simultaneously, the CBAM modules embedded at each level utilize a dual channel and spatial attention mechanism to accurately focus on and extract high-weight key features, effectively suppressing background noise interference.

[0034] Furthermore, the multi-branch parallel feature extraction architecture of this invention effectively solves the spatial smoothing problem in multi-source data fusion. By processing heterogeneous data independently at their respective original resolutions, the model avoids the detail loss caused by traditional interpolation and preserves high-frequency dynamic signals of the sea surface, thereby significantly improving the inversion accuracy and generalization ability of the subsurface three-dimensional dynamic structure.

[0035] The loss function used in model training is a computational function used to measure the degree of difference between the model's predicted values ​​and the true values. It is an important standard for evaluating the quality of the output results. The smaller the loss function, the better the model fits. Therefore, in this embodiment of the invention, the region is divided into nearshore and offshore areas horizontally, with a water depth of 200m as the boundary, and vertically into a mixing layer, thermocline, and deep layer, based on physical heterogeneity. During the reconstruction process, the loss value is calculated separately for each region. Since the differences between the layers in the thermocline are large and the structure is relatively complex, gradient loss is added to the loss function to further evaluate the difference in the loss value of the thermocline temperature. The formula for the constructed loss function is as follows: , in, The Huber loss values ​​represent the mixed layer and deep layer in nearshore or offshore environments. The Huber loss value represents the thermocline in the nearshore or offshore area. This represents the gradient loss of the thermocline near the coast or offshore, where a, b, and c represent the weight parameters of each loss, which are automatically tuned in the model.

[0036] Because the vertical temperature gradient of the thermocline changes drastically and the reconstruction error is concentrated, this invention introduces the absolute value of the true temperature gradient into the gradient loss formula. A heavier error penalty is applied to the thermocline region where temperature changes drastically, thus forcing the model to focus on fitting key physical abrupt change regions. The Gradloss formula is as follows: , , , Where y and ˆy represent the target value and the reconstructed value, respectively. This represents the temperature difference between the (i+1)th layer and the target temperature field value of the ith layer. The temperature difference between the reconstructed temperature field values ​​of the (i+1)th layer and the ith layer, where D represents the number of layers between the upper and lower boundaries of the ocean thermocline.

[0037] This invention combines gradient loss and standard Huber loss to construct a comprehensive loss function to learn the patterns of 3D ocean structural changes and improve the overall accuracy of the model.

[0038] IV. Reconstructing the 3D Temperature Field Based on the Trained MAUS Transformer Model In the application stage of the three-dimensional temperature field reconstruction, the sea surface dataset is input into the trained MAUS Transformer underwater three-dimensional temperature field reconstruction model to obtain the final three-dimensional temperature field reconstruction result of 5-222m, which does not require normalization and inverse normalization.

[0039] like Figure 3 As shown, this invention compares the adaptive loss function of the hybrid layer, thermocline layer, and deep layer with the Huber loss function, using two evaluation metrics: RMSE and R. RMSE is the root mean square error, used to measure the degree of deviation between the reconstructed temperature and the true temperature value; the smaller the value, the higher the prediction accuracy. R is the correlation coefficient, used to measure the consistency between the reconstructed temperature and the true temperature in terms of spatial distribution and variation trend; the closer the value is to 1, the stronger the consistency. The profiles of RMSE and R for the two loss functions show that the loss function constructed in this invention is significantly better than the Huber loss function, especially in the thermocline layer region (40m-110m), where its RMSE increases from 0.976 to 0.939 and its R increases from 0.851 to 0.864.

[0040] Figure 4 The average values ​​of RMSE and R were calculated for the two loss functions in each month of the thermocline region (40m-110m) in 2020. It can be seen that the loss function constructed in this invention has better error and correlation than the Huber loss function in each month, especially in January and July.

[0041] Figure 5 The chart comparing the RMSE and R-values ​​of the MAUS Transformer model with other models in 2020 shows that the MAUS Transformer model exhibits a significant performance advantage in the upper ocean (0-200m). Specifically, at all assessment depths, the model maintains the lowest RMSE across the entire field, indicating the smallest absolute reconstruction error and the highest prediction accuracy. Simultaneously, its correlation coefficient R-value is the highest among all models, demonstrating an excellent match between the reconstructed temperature field and actual observations in terms of spatial distribution and evolution trends. Overall, the MAUS Transformer's reconstruction performance in the upper water body comprehensively surpasses other benchmark models.

[0042] The embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A deep learning-based method for reconstructing a three-dimensional ocean temperature field, characterized in that, Includes the following steps: S1. Acquire sea surface variable data and ocean reanalysis data from multi-source satellite remote sensing observations at different resolutions in the study area, and perform quality control and spatiotemporal matching; S2. Add the corresponding latitude and longitude information and time information to the dataset and preprocess it to form the sample database required for model training; S3. Construct a multi-branch MAUS Transformer reconstruction model, design an adaptive loss function that integrates horizontal partitioning and vertical layering, complete model training, and save the optimal model. The multi-branch MAUS Transformer model uses the U-Net network as a framework, replaces the convolutional blocks with the Swing Transformer module to embed in the encoding and decoding paths, and embeds the CBAM attention module to adaptively optimize feature representation. The multi-branch MAUS Transformer model has a three-branch structure, corresponding to input features at resolutions of 1 / 100°, 1 / 8°, and 1 / 4°, respectively. The specific implementation of the 1 / 100° resolution feature extraction branch is as follows: S301, Encoding stage: Data undergoes preliminary feature mapping through patch embedding layer and preliminary feature extraction through two layers of Swing Transformer blocks; S302. The initially extracted features are downsampled twice, and the features after each downsampling are further processed by two layers of Swing Transformer blocks for deep feature extraction. After the features after the second downsampling are extracted by two layers of Swing Transformer blocks, they are then processed by the CBAM module for channel and spatial dual attention weighting to highlight key features. S303. Entering the decoding stage, the features output by the encoder are upsampled twice. After each upsampling, the features are skipped and connected with the features of the corresponding scale in the encoding path that have been enhanced by the CBAM module. The spliced ​​and fused features are integrated through two layers of Swing Transformer blocks. S304. Resample the integrated features and reshape the feature map dimension to match the subsurface temperature field to obtain the final output feature of this branch. The feature extraction branches for 1 / 8° and 1 / 4° resolutions perform downsampling only once during the encoding stage and upsampling once during the decoding stage, with the remaining steps consistent with the feature extraction branch for 1 / 100° resolution. The adaptive loss function incorporates gradient loss to evaluate the difference in temperature loss values ​​of the thermocline. The formula for the adaptive loss function is as follows: ， in, The Huber loss values ​​represent the mixed layer and deep layer in nearshore or offshore environments. The Huber loss value represents the thermocline in the nearshore or offshore area. This represents the gradient loss of the thermocline near the coast or offshore, where a, b, and c represent the weight parameters of each loss. S4. Input the sea surface data of the area to be reconstructed into the optimal model and output the final three-dimensional ocean temperature field reconstruction result.

2. The method for reconstructing a three-dimensional ocean temperature field based on deep learning according to claim 1, characterized in that, The multi-source satellite remote sensing observation data at different resolutions are used as input features, including sea surface temperature at 1 / 100°, sea surface precipitation at 1 / 10°, sea surface salinity at 1 / 8°, sea level anomaly at 1 / 8°, sea surface wind field at 1 / 4°, and sea surface current field data at 1 / 4°. The ocean reanalysis data, used as ground truth labels, represents the temperature field at depths of 5m to 222m, spanning 18 layers, with a spatial resolution of 1 / 12°.

3. The method for reconstructing a three-dimensional ocean temperature field based on deep learning according to claim 2, characterized in that, The input features also include wind stress curl data, formulated as follows: ， in, For sea surface wind stress, and These are the zonal and meridional components of sea surface wind stress, respectively. This is the drag coefficient. air density; u and v are the zonal and meridional wind speeds at a height of 10m above the sea surface, respectively; sea surface wind speed. .

4. The deep learning-based method for reconstructing a three-dimensional ocean temperature field according to claim 3, characterized in that, The specific details of the quality control and spatiotemporal matching are as follows: The spatial resolution of the 1 / 10° sea surface precipitation data was resampled to 1 / 8°, and the temporal resolution of all data was resampled to daily data. Finally, input feature data with three different spatial resolutions of 1 / 100°, 1 / 8°, and 1 / 4° and daily temporal resolution, and ocean reanalysis data with a spatial resolution of 1 / 12° and daily temporal resolution were generated.

5. The method for reconstructing a three-dimensional ocean temperature field based on deep learning according to claim 1, characterized in that, The gradient loss introduces the absolute value of the true temperature gradient. To accurately fit key physical abrupt change regions, the gradient loss formula is as follows: ， ， ， in, and Let these represent the target value and the reconstructed value of the temperature field in the i-th layer, respectively. This represents the temperature difference between the (i+1)th layer and the target temperature field value of the ith layer. The temperature difference between the reconstructed temperature field values ​​of layer i+1 and layer i This indicates the number of layers between the upper and lower boundaries of the ocean thermocline. This represents the absolute value of the actual temperature gradient.

Images