Three-dimensional sea surface temperature long-term intelligent prediction method considering multi-element influence

Abstract

The application proposes a three-dimensional sea temperature long-term intelligent prediction method considering the influence of multiple factors, belonging to the field of intelligent prediction of marine environment. In view of the defects of the existing three-dimensional sea temperature prediction model, such as insufficient consideration of the coupling effect of multiple factors, high calculation complexity, strong dependence on computing resources and insufficient stability of long-term prediction, the application designs a deep learning algorithm by fusing multi-scale convolution feature extraction, space-time decoupling Transformer structure, multiple factor interaction attention mechanism and deep perception attention mechanism, and introduces a marine physical constraint loss function in the model training process, which significantly reduces the complexity and parameter size of the model calculation while improving the long-term prediction accuracy, meets the demand of three-dimensional sea temperature long-term intelligent prediction under the condition of limited computing power, and finally generates the three-dimensional sea temperature long-term prediction product of the target sea area in the future 30 days based on the optimal performance model by inputting the historical 30-day multi-source sea surface data.

Description

Technical Field

[0001] This application relates to the field of intelligent marine environmental forecasting, specifically proposing a three-dimensional intelligent forecasting method for long-term sea surface temperature that considers the influence of multiple factors. Background Technology

[0002] The three-dimensional ocean temperature field is an important physical quantity characterizing the ocean's thermal structure and its evolution, and it plays a significant role in ocean dynamics research, climate change analysis, underwater platform environmental protection, and marine engineering applications. Achieving long-term, continuous forecasting of the three-dimensional ocean temperature field is of vital practical importance for improving our understanding of the marine environment and ensuring the safety of marine activities.

[0003] Currently, three-dimensional sea surface temperature (SST) forecasting primarily relies on numerical modeling methods, which calculate and predict the SST by solving ocean dynamic and thermodynamic equations. These methods are highly dependent on initial and boundary conditions, and the computational process is complex, typically requiring high-performance computing resources. This makes it difficult to run stably and for extended periods on platforms or devices with limited computing power. Furthermore, numerical models are susceptible to the accumulation of model and initial value errors during long-term extrapolation, limiting their practicality in certain application scenarios.

[0004] With the development of artificial intelligence technology, data-driven marine environmental forecasting methods have gradually attracted attention. In recent years, attention mechanisms and Transformer architectures have been gradually introduced into the fields of meteorological and marine environmental forecasting due to their advantages in time series modeling. For example, in the field of wave forecasting, patent application No. 202511812161.5 proposes a high-precision intelligent wave forecasting method based on Transformer, which models historical wave sequences by constructing a model containing a multi-head attention mechanism; in the field of hydrological forecasting, patent application No. 202511775030.4 proposes a hydrological forecasting method based on multi-feature combination and Transformer model to capture long-term dependencies in time series.

[0005] However, when applying existing deep learning architectures and related patented technologies directly to long-term forecasting of three-dimensional sea surface temperature fields, the following significant defects and shortcomings still exist:

[0006] High-dimensional spatiotemporal data leads to a sharp increase in computational complexity: Existing general-purpose Transformer architectures typically require global self-attention computation when processing data. When applied to high-dimensional grid data such as 3D sea surface temperature fields, the computational complexity increases quadratically with the spatial scale, resulting in huge computational overhead and massive parameter scale, making it difficult to meet the long-term forecasting needs under computing-constrained environments.

[0007] The lack of effective mapping capabilities from two-dimensional surface information to three-dimensional vertical structures is a significant weakness: most existing multi-element fusion prediction technologies are limited to two-dimensional space or single-dimensional expansion. For example, in the field of precipitation prediction, although the patent application with application number 202610057758.1 proposes a multi-scale deep learning method that integrates ocean, land, and atmosphere signals through a multi-level attention mechanism, its essence remains primarily two-dimensional spatial modeling. Existing technologies generally lack effective "depth perception" mechanisms, making it difficult to adaptively learn the interlayer coupling relationship of sea surface temperature in the vertical direction, and thus achieve high-precision reconstruction of three-dimensional structures, relying solely on a limited number of sea surface elements.

[0008] The lack of explicit physical constraints on the three-dimensional thermodynamic structure of the ocean is a significant issue. Existing data-driven models typically use only mathematical error metrics as optimization targets during training. Although some models can capture statistical correlations in the data, the lack of constraints on the physical consistency of the three-dimensional temperature field of the ocean can easily lead to non-physical predictions that violate the stability and spatial continuity of ocean stratification during long-term extrapolation, thus significantly reducing the stability and reliability of long-term forecasts.

[0009] Therefore, how to reduce the computational complexity of the model while ensuring prediction accuracy and making full use of multi-element information of the sea surface, and construct a three-dimensional long-term sea surface temperature prediction model that can simultaneously characterize the multi-scale spatial structure features, long-term time dependence and vertical structural coupling relationship of the ocean, and improve the physical consistency of the prediction results by introducing ocean physical constraints, so as to realize a three-dimensional intelligent sea surface temperature prediction method suitable for computing power-constrained environments, has become a key technical problem that urgently needs to be solved in the field of intelligent marine environmental forecasting. Summary of the Invention

[0010] The three-dimensional sea surface temperature (SST) long-term intelligent forecasting method considering multiple factors described in this application aims to address the problems in existing three-dimensional SST forecasting technologies, such as insufficient consideration of the coupling effects of multiple sea surface elements, high model computational complexity, strong dependence on computing resources, and insufficient long-term forecast stability. This application integrates multi-scale convolutional feature extraction, spatiotemporal decoupling Transformer structure, multi-factor interactive attention mechanism, and depth-aware attention mechanism, and introduces a marine physical constraint loss function during model training. This approach reduces the computational complexity and parameter scale of the model while ensuring long-term prediction accuracy, meeting the needs of long-term intelligent three-dimensional SST forecasting under computationally limited conditions.

[0011] To achieve the above-mentioned objectives, the three-dimensional sea surface temperature long-term intelligent forecasting method considering the influence of multiple factors first collects multi-factor sea surface data and three-dimensional sea surface temperature reanalysis data of the target sea area, constructs a model training dataset, including sea surface temperature, sea surface height anomaly, sea surface wind field and three-dimensional ocean reanalysis data, determines the model input and prediction labels, and preprocesses the multi-factor sea surface data and three-dimensional sea surface temperature reanalysis data.

[0012] The training, validation, and test sets are divided chronologically. Based on this, an improved lightweight Transformer prediction model is constructed, integrating a multi-scale convolutional feature extraction module, a spatiotemporal decoupling Transformer encoding module, a multi-element interactive attention module, and a depth-sensing attention module. During model training, a marine physical constraint mechanism is introduced to ensure the physical consistency of the 3D sea surface temperature prediction results through spatial smoothness and vertical stability constraints. Specifically, the multi-scale convolutional feature extraction module extracts multi-scale spatial features, the spatiotemporal decoupling Transformer encoding module models time-series features, the multi-element interactive attention module dynamically fuses multiple elements, and the depth-sensing attention module maps 2D sea surface features to the 3D sea surface temperature vertical structure.

[0013] A three-dimensional long-term sea surface temperature (SST) prediction model is obtained through model training, and the prediction performance of the model is evaluated using a test set. Finally, based on the model with the best performance, a three-dimensional long-term SST forecast product for the target sea area in the future is generated by inputting historical sea surface multi-element data sequences.

[0014] Furthermore, it includes the following steps:

[0015] Step 1) Collect multi-element sea surface data and three-dimensional sea surface temperature reanalysis data of the target sea area, construct the model training dataset, and determine the model training input and training labels;

[0016] Step 2), Data Preprocessing;

[0017] Determine if there are any default values ​​in the model training dataset. If so, replace them with preset anomaly marker values ​​and introduce the corresponding missing mask into the model input. Normalize the model training dataset.

[0018] Step 3) Divide the dataset into training, validation, and test sets;

[0019] The normalized model training dataset is divided into a training set, a validation set, and a test set;

[0020] Step 4) Establish a lightweight, multi-element, three-dimensional sea surface temperature forecasting model;

[0021] The lightweight multi-element three-dimensional sea surface temperature forecasting model includes: a multi-scale convolutional feature extraction module, a spatiotemporal decoupling Transformer encoding module, a multi-element interactive attention module, and a depth perception attention module;

[0022] The data flow of the model is as follows: the multi-element input of the sea surface first enters the multi-scale convolutional feature extraction module for spatial feature encoding, then enters the spatiotemporal decoupling Transformer encoding module for time series modeling, then enters the multi-element interactive attention module to complete the dynamic fusion of different sea surface elements, and finally enters the depth perception attention module to realize the mapping to the three-dimensional sea surface temperature vertical structure and output the prediction results.

[0023] Its network model input is a two-dimensional data sequence of multiple sea surface elements over a continuous period of time in history;

[0024] The multi-element two-dimensional data of the sea surface spatially constitutes a two-dimensional matrix of the number of longitude grids × the number of latitude grids in the sea area, and in the element dimension constitutes a multi-channel feature tensor. It is input into the model in the form of time series to predict the three-dimensional sea temperature distribution in future periods.

[0025] Step 5) Model training and optimization;

[0026] The model is trained using a training set. During the training process, a marine physical constraint loss function is introduced to constrain the physical rationality of the predicted sea surface temperature field. The model parameters are then optimized using a validation set.

[0027] Step 6) Evaluate the performance of the three-dimensional sea surface temperature long-term forecast model using a test set;

[0028] The evaluation metric uses mean squared error, calculated using the following formula:

[0029] ;

[0030] in, Indicates the first Time sequence Long-term forecast results for each grid, Indicates the first Time sequence The actual values ​​of each grid are total. Each time period One grid;

[0031] Step 7) Forecast product generation;

[0032] After the lightweight multi-element three-dimensional sea surface temperature forecast model is trained, the corresponding multi-element sea surface data sequence is used as input to generate a long-term three-dimensional sea surface temperature forecast product for future periods.

[0033] Step 1) The model training dataset includes historical ocean reanalysis data and multi-element sea surface data. The model training input and training labels are constructed using a time series approach, with the multi-element sea surface data sequence of historical continuous periods as the model input and the three-dimensional sea surface temperature data sequence of future periods as the model training labels. The multi-element two-dimensional sea surface data includes, but is not limited to, sea surface temperature, sea surface height anomaly, sea surface wind field, etc. Each element constitutes a two-dimensional spatial field on the same latitude and longitude grid and is input into the model in the form of a multi-channel tensor.

[0034] Step 2) involves normalizing data including, but not limited to, sea surface temperature, sea surface height anomalies, and sea surface wind field variables. Each variable is then subjected to maximum and minimum normalization using the following formula to unify the numerical range of each variable to between 0 and 1:

[0035] ;

[0036] in: Represents a specific value in the original data; , These represent the minimum and maximum values ​​in the dataset, respectively. This is the normalized result, with a value between 0 and 1.

[0037] Step 3) involves dividing the preprocessed model training dataset into a training set, a validation set, and a test set in chronological order. Each of the training set, validation set, and test set consists of multi-factor sea surface input samples and three-dimensional sea surface temperature label samples for the corresponding time period. The training set is used for model parameter learning, the validation set is used for model parameter adjustment and training process optimization, and the test set is used for independent evaluation of the model's long-term forecast performance.

[0038] Step 4) describes a multi-scale convolutional feature extraction module for extracting multi-scale spatial features from sea surface multi-element two-dimensional spatial data. It captures local small-scale marine structural information and large-scale ocean circulation features using convolutional kernels of different scales. This module extracts multi-scale features from sea surface multi-element two-dimensional spatial data by setting convolutional layers with different scale kernels. The convolutional kernel sizes include 3×3, 5×5, and 7×7, used to simultaneously extract local small-scale ocean structural information and large-scale ocean circulation structural information. The spatiotemporal decoupling Transformer encoding module is used to model the time series of sea surface multi-element data. This module adopts a structure that separates spatial feature encoding from temporal feature encoding. Spatial features are extracted by the convolutional feature extraction module, while the Transformer encoding structure only performs self-attention calculations in the time dimension to model the temporal evolution of sea surface multi-element data. Residual connections and layer normalization structures are introduced to improve the training stability of deep networks and enhance the model's ability to express long-term sea surface temperature (SST) changes. The multi-factor interactive attention module receives multi-factor temporal features output by the Transformer encoding module, establishes feature branches for different ocean variables, and constructs a fusion branch. Through interactive attention calculation between the feature branches and the fusion branch, information exchange and coupling modeling between different ocean variable features are achieved, resulting in fused multi-factor features. The depth-sensing attention module receives the fused multi-factor features and incorporates ocean depth information into the 3D SST prediction process. By setting depth weight parameters for different depth layers, the model can learn the inter-layer coupling relationship of SST in the vertical direction and map the 2D sea surface multi-factor fusion features to the 3D SST vertical structure, generating 3D SST prediction results corresponding to different depth layers.

[0039] Step 4) includes,

[0040] Step 4.1) Extract feature information at different spatial scales;

[0041] The multi-scale convolutional feature extraction module extracts feature information at different spatial scales from a continuous multi-day two-dimensional spatiotemporal data sequence of sea surface elements within a historical period, using convolutional kernels of different scales, as follows:

[0042] ;

[0043] ;

[0044] ;

[0045] The final fusion yields multi-scale spatial features:

[0046] ;

[0047] in, Concat Indicates feature concatenation operation;

[0048] After this step, the model input is transformed from raw two-dimensional multi-element sea surface data into a feature representation containing multi-scale spatial structure information. This feature not only preserves the spatial distribution information of multiple elements on the original sea surface, but also enhances the model's ability to represent local disturbance structures and large-scale ocean circulation background, providing more effective input features for subsequent time series modeling;

[0049] Step 4.2), Temporal modeling stage;

[0050] The multi-scale spatial features output from step 4.1) are input into the spatiotemporally decoupled Transformer encoding module. Since the spatial feature extraction task has been completed by the front-end convolutional module, the Transformer structure only performs self-attention computation in the time dimension. Let the temporal features output by the multi-scale convolutional feature extraction module be:

[0051] ;

[0052] in, The spatial features at time t are represented by ; T represents the length of the input time series; C represents the number of feature channels;

[0053] During the Transformer encoding phase, the query, key, and value matrix is ​​first obtained through linear mapping as follows:

[0054] ;

[0055] ;

[0056] ;

[0057] in, , , The weight matrix is ​​trainable.

[0058] Subsequently, self-attention weights are calculated along the time dimension to measure the correlation between different time steps, and the value matrix is ​​weighted and summed to obtain the temporally encoded feature representation:

[0059] ;

[0060] Where d is the feature dimension;

[0061] After this step, the input multi-scale spatial feature sequence is transformed from a simple time-by-time spatial representation into a temporal feature representation containing long-term temporal dependencies, thus modeling the temporal evolution of multiple sea surface elements.

[0062] Step 4.3) Introduce a multi-head attention mechanism;

[0063] To enhance the model's ability to represent complex temporal change patterns, a multi-head attention structure is introduced into the Transformer encoding module. Specifically, the input features are mapped to multiple different feature subspaces, attention calculations are performed in each subspace, and the outputs of the multiple attention heads are concatenated and linearly mapped, as shown below:

[0064] ;

[0065] in, h represents the number of attention heads;

[0066] After obtaining the multi-head attention output, it is residually connected to the module input features and then subjected to layer normalization to preserve the original temporal feature information, alleviate the gradient decay problem during deep network training, and improve training stability. After residual connection and layer normalization, enhanced temporal coding features are obtained and used as input to the subsequent multi-factor interactive attention module.

[0067] The multi-head attention mechanism enables the model to extract time-dependent features from multiple different representation subspaces simultaneously, thereby more fully capturing the changing patterns of multiple sea surface elements at different time scales.

[0068] Step 4.4) Construct a multi-factor interactive attention module;

[0069] The time-series encoded features output in step 4.3) are split according to different ocean variables to form a multi-factor feature set: ;

[0070] in, Represents the characteristic of the i-th ocean variable;

[0071] To enhance the expressive power of coupling relationships among different ocean variables, feature branches are established for each ocean variable, and a fusion branch is constructed by concatenating all feature branches, as shown below:

[0072] ;

[0073] in, Indicates the fusion branch features, Indicates a splicing operation;

[0074] For the The interaction attention weights between the ocean variable feature branches and the fusion branch are defined as follows:

[0075] ;

[0076] in, Indicates the first Attention weights of each ocean variable feature branch relative to the fusion branch. and For a trainable parameter matrix, Indicates the feature dimension;

[0077] After obtaining the interaction attention weights, for the first... Each ocean variable feature branch is updated to obtain its fused feature representation:

[0078] ;

[0079] in, For a trainable parameter matrix, Indicates the first Feature representation after fusion of feature branches of ocean variables;

[0080] All updated feature branch features are merged, and the merged multi-feature feature is represented as follows:

[0081] ;

[0082] in, This represents the multi-element characteristics after fusion. This indicates a feature fusion operation.

[0083] After this step, the model input is transformed from multi-head time-series encoded features that only contain time dependencies into multi-factor fusion features that comprehensively consider the coupling relationships between different ocean variables.

[0084] Step 4.5) Construct the depth-sensing attention module;

[0085] The fused multi-element features obtained in step 4.4) Input the depth-sensing attention module to achieve the mapping from two-dimensional sea surface multi-element features to three-dimensional sea surface vertical structure. The predicted depth layer set is as follows: ;

[0086] in, This represents the k-th depth layer;

[0087] First, depth weights are generated through linear mapping: ;

[0088] in, This is the depth weight matrix;

[0089] Then, the depth weights are used to perform vertical mapping on the fused multi-element features to obtain the three-dimensional sea surface temperature prediction results corresponding to different depth layers, the expression of which is: ;

[0090] in, This indicates a predicted sea surface temperature. Representation of spatiotemporal characteristics;

[0091] Through this step, the model further maps the previously obtained two-dimensional sea surface multi-element fusion features into temperature feature representations at different depth layers, thereby completing the transformation from two-dimensional sea surface information to three-dimensional sea surface temperature vertical structure.

[0092] In summary, the data flow transformation process of the model is as follows: First, historical sea surface multi-element two-dimensional spatiotemporal data are input into a multi-scale convolutional feature extraction module to obtain multi-scale spatial features; then, the multi-scale spatial features are input into a spatiotemporal decoupling Transformer encoding module to obtain temporal features containing time dependencies; next, the temporal features are input into a multi-element interactive attention module, and through interactive modeling between feature branches and fusion branches, fused multi-element features are obtained; finally, the fused multi-element features are input into a depth-sensing attention module to generate predicted sea surface temperature values ​​corresponding to different depth layers, and output three-dimensional sea surface temperature prediction results for future periods. The output results are a four-dimensional sea surface temperature prediction sequence with spatial dimensions of time × depth × longitude × latitude, where the time dimension is the prediction time step, and each prediction time corresponds to a three-dimensional sea surface temperature field, which is used to generate long-term sea surface temperature forecast products.

[0093] Step 5) includes, during model training, constructing a total loss function consisting of data fitting error loss and physical constraint loss, the total loss function being defined as:

[0094] ;

[0095] in, The total loss function for model training; This represents the loss due to data fitting error. Loss due to physical constraints; This is the physical constraint weighting coefficient, used to adjust the trade-off between data error and physical constraints;

[0096] Data error loss is defined using mean squared error:

[0097] ;

[0098] in, This represents the sea surface temperature value predicted by the model at the i-th sample point; This represents the actual sea surface temperature at the corresponding location; This represents the total number of samples. This loss function measures the overall error between the model's predictions and the actual observed data.

[0099] Let the three-dimensional sea surface temperature field be: ;in, Indicates spatial latitude and longitude location; Indicates depth layer; Indicates time;

[0100] Define the vertical temperature difference between adjacent depth layers: ;

[0101] When the deep layer temperature is higher than the shallow layer temperature, that is This indicates the presence of a physically unstable stratification.

[0102] Then, the vertical stability constraint loss is defined as:

[0103] ;

[0104] In addition, a spatial smoothing constraint loss is introduced; first, the spatial gradient of the sea surface temperature field is defined: ;

[0105] The spatial gradient magnitude is: ;

[0106] Spatial smoothing loss is defined as: ;

[0107] By penalizing the square of the spatial gradient, unreasonable and drastic spatial variations in the predicted sea surface temperature field can be suppressed, making the prediction results smoother and more consistent with the actual ocean structure.

[0108] Combining vertical stability constraints and spatial smoothness constraints, the final physical constraint loss function is obtained as follows:

[0109] ;

[0110] in, The weights are for vertical stability constraints; The weights are used for spatial smoothing constraints.

[0111] In summary, this application has the following advantages and beneficial effects compared to the prior art:

[0112] This application significantly reduces the computational complexity and parameter scale of the model by integrating multi-scale convolutional feature extraction with a spatiotemporal decoupling Transformer structure, thus achieving the lightweight characteristics of a three-dimensional sea surface temperature prediction model.

[0113] This application fully explores the coupling relationship between multiple factors such as sea surface temperature, sea surface height anomaly, salinity and wind field through a multi-factor interactive attention mechanism, thereby effectively improving the model's ability to characterize ocean dynamic processes.

[0114] This application introduces a depth-sensing attention mechanism into the vertical structure information of the ocean, enabling the model to effectively learn the temperature coupling relationship between different depth layers, thereby significantly improving the accuracy of three-dimensional sea surface temperature prediction.

[0115] This application significantly improves the stability of long-term predictions by introducing a marine physical constraint loss function during model training to enhance the physical consistency of prediction results.

[0116] By applying this application, continuous and stable three-dimensional long-term sea surface temperature forecast products can be generated, which can efficiently and effectively meet the needs of underwater platform mission support and marine engineering environmental forecasting. Attached Figure Description

[0117] Figure 1 This is a flowchart of the three-dimensional intelligent sea surface temperature forecasting method with multi-factor influence proposed in this application.

[0118] Figure 2 It is an improved network structure diagram of the Transformer model.

[0119] Figure 3 This is a flowchart for creating a three-dimensional long-term sea surface temperature forecast product. Detailed Implementation

[0120] To better understand the above-mentioned objectives, features, and advantages of this application, the application will be further described below in conjunction with the accompanying drawings and embodiments. Many specific details are set forth in the following description to provide a thorough understanding of this application; however, this application may be implemented in other ways than those described herein, and therefore, this application is not limited to the specific embodiments disclosed below.

[0121] Example 1, such as Figures 1 to 3 As shown, this application proposes a three-dimensional intelligent sea surface temperature forecasting method that considers the influence of multiple factors.

[0122] This method first collects multi-element sea surface data and three-dimensional sea surface temperature reanalysis data of the target sea area, constructs a model training dataset, determines the model input and prediction labels, and preprocesses the multi-element sea surface data and three-dimensional sea surface temperature reanalysis data.

[0123] The training, validation, and test sets are divided according to time series. Based on this, an improved lightweight Transformer prediction model is constructed, which integrates a multi-scale convolutional feature extraction module, a spatiotemporal decoupling Transformer encoding module, a multi-element interactive attention module, and a depth-sensing attention module. During the model training process, a marine physical constraint mechanism is introduced to ensure the physical consistency of the three-dimensional sea surface temperature prediction results through spatial smoothness constraints and vertical stability constraints. Among them, the spatiotemporal decoupling Transformer encoding module is used for time series feature modeling, the multi-element interactive attention module is used for multi-element dynamic fusion, and the depth-sensing attention module is used to realize the mapping from two-dimensional sea surface features to three-dimensional sea surface temperature vertical structure.

[0124] A three-dimensional long-term sea surface temperature (SST) prediction model is obtained through model training, and the prediction performance of the model is evaluated using a test set. Finally, based on the model with the best performance, a three-dimensional long-term SST forecast product for the target sea area is generated from the input multi-element sea surface data sequence.

[0125] Specifically, it includes the following steps:

[0126] Step 1) Construct the model training dataset and determine the model training input and training labels;

[0127] Collect multi-element sea surface data and three-dimensional sea surface temperature reanalysis data of the target sea area within a preset time range, and construct a model training dataset based on the collected data. The multi-element sea surface data is used as the model input, and the three-dimensional sea surface temperature reanalysis data is used as the label data for model training and evaluation.

[0128] The model training input and training labels are constructed using a time series approach, including using the sea surface multi-element two-dimensional data sequence from T-30 to T days as the model training input to supervise the model's learning of the mapping relationship from sea surface state to future three-dimensional sea surface temperature structure, and using three-dimensional ocean reanalysis data from T+1 to T+30 days as the training labels.

[0129] The aforementioned multi-element two-dimensional sea surface data includes, but is not limited to, sea surface temperature, sea surface height anomaly, sea surface wind field and other multi-element data. Each element constitutes a two-dimensional spatial field on the same latitude and longitude grid and is input into the model in the form of a multi-channel tensor.

[0130] Step 2), Data Preprocessing;

[0131] Determine if there are any default values ​​in the model training dataset. If so, fill in the default values. Specifically, since the training model cannot recognize NAN values, all default values ​​need to be replaced by a preset anomaly marker value, and the corresponding missing mask is introduced into the model input.

[0132] Normalize the model training dataset;

[0133] Variables including, but not limited to, sea surface temperature, sea surface height anomaly, and sea surface wind field are subjected to maximum-min normalization according to the following formula to unify the numerical range of each variable to between 0 and 1:

[0134] ;

[0135] in: Represents a specific value in the original data; , These represent the minimum and maximum values ​​in the dataset, respectively. This is the normalized result, with a value between 0 and 1.

[0136] Step 3) Divide the dataset into training, validation, and test sets;

[0137] The normalized model training dataset is divided into a training set, a validation set, and a test set;

[0138] The training set, validation set, and test set are all composed of multi-factor sea surface input samples and three-dimensional sea surface temperature label samples for the corresponding time period. The training set is used for model parameter learning, the validation set is used for model parameter adjustment and training process optimization, and the test set is used for independent evaluation of the model's long-term forecast performance.

[0139] The training set, validation set, and test set can account for 60%, 20%, and 20% of the total sample size of the model training dataset, respectively, or other time-order division ratios can be adopted according to actual application needs.

[0140] Step 4) Establish a lightweight, multi-element, three-dimensional sea surface temperature forecasting model;

[0141] The lightweight multi-element three-dimensional sea surface temperature forecasting model includes: a multi-scale convolutional feature extraction module, a spatiotemporal decoupling Transformer encoding module, a multi-element interactive attention module, and a depth perception attention module;

[0142] The network model input is a series of two-dimensional multi-element sea surface data for multiple consecutive days within a historical period. The two-dimensional multi-element sea surface data forms a two-dimensional matrix of longitude grid number × latitude grid number in the spatial dimension and a multi-channel feature tensor in the element dimension. It is input into the model in the form of a time series to predict the three-dimensional sea surface temperature distribution in the future period.

[0143] Specifically, the multi-scale convolutional feature extraction module is used to extract multi-scale spatial features from multi-element two-dimensional spatial data of the sea surface. It captures local small-scale structural information and large-scale ocean circulation features through convolutional kernels of different scales. This module extracts multi-scale features from multi-element two-dimensional spatial data of the sea surface by setting convolutional layers with convolutional kernels of different scales. The convolutional kernel sizes include 3×3, 5×5 and 7×7, which are used to extract local small-scale ocean structural information and large-scale ocean circulation structural information simultaneously.

[0144] The aforementioned spatiotemporal decoupling Transformer encoding module is used to model the time series of multiple sea surface elements. This module adopts a structure that separates spatial feature encoding and temporal feature encoding. The spatial features are extracted by the convolutional feature extraction module, and the Transformer encoding structure only performs self-attention calculation in the time dimension to model the temporal evolution relationship of the multiple sea surface elements data. Residual connections and layer normalization structures are introduced in the encoding process to improve the training stability of deep networks and enhance the model's ability to express long-term sea surface temperature changes.

[0145] The multi-element interactive attention module is used to model the coupling relationship between different ocean variables. It establishes feature branches for different ocean variables and constructs a fusion branch. Through interactive attention calculation between each feature branch and the fusion branch, it realizes information exchange and coupling modeling between the features of different ocean variables, and obtains the fused multi-element features.

[0146] The aforementioned depth perception attention module is used to introduce ocean depth information in the three-dimensional sea surface temperature prediction process. By setting depth weight parameters for different depth layers, the model can learn the interlayer coupling relationship of sea surface temperature in the vertical direction and realize the mapping of two-dimensional sea surface multi-element information to three-dimensional sea surface temperature vertical structure.

[0147] Includes the following steps:

[0148] Step 4.1) Extract feature information at different spatial scales;

[0149] The lightweight multi-element three-dimensional sea surface temperature forecasting model described above uses multi-element two-dimensional sea surface data as input. The input data includes, but is not limited to, observable sea surface variables such as sea surface temperature (SST), sea surface height anomaly (SLA), sea surface salinity, and wind field.

[0150] Each sea surface element constitutes a two-dimensional spatial field on the same latitude and longitude grid, and is superimposed in the form of multi-channel features in the element dimension to form a two-dimensional feature field of sea surface multi-element that changes with time. The two-dimensional feature fields of multiple consecutive moments are arranged in chronological order to form a two-dimensional data sequence of sea surface multi-element, which is used as the model input.

[0151] To effectively extract local spatial features from multi-element sea surface data and reduce the computational complexity of the Transformer model in high-dimensional scenarios, a multi-scale convolutional feature extraction module is introduced at the input end of the model. This module extracts feature information at different spatial scales through convolutional kernels of different scales, as follows:

[0152] ;

[0153] ;

[0154] ;

[0155] The feature maps obtained by convolution at different scales are concatenated along the channel dimension and finally fused to obtain multi-scale spatial features:

[0156] ;

[0157] in, Concat Indicates feature concatenation operation;

[0158] Multi-scale convolutional structures can simultaneously capture small-scale ocean disturbances, meso-scale vortex structures, and large-scale ocean thermal structures, improving the model's ability to perceive changes in spatial structure.

[0159] Step 4.2), Temporal modeling stage;

[0160] After extracting spatial features at each time point, the model needs to further learn the dynamic patterns of the evolution of multiple sea surface elements over time. Existing technologies like Transformer, when processing high-dimensional spatial data, require calculating global self-attention across all spatial locations, resulting in a computational complexity of:

[0161] ;

[0162] Where H is the number of latitude grids and W is the number of longitude grids;

[0163] When the ocean grid resolution is high, the spatial dimension scale The large size of the self-attention computation overhead leads to a quadratic increase in computational overhead, which significantly increases the computational burden on the model and is not conducive to long-term ocean forecasting in environments with limited computing power.

[0164] Therefore, this application adopts the following spatiotemporal decoupling strategy:

[0165] First, the front-end convolutional module completes the spatial structure extraction, and then the Transformer only models the feature sequence in the time dimension. The Transformer structure only needs to perform self-attention calculation in the time dimension, thereby significantly reducing the overall computational complexity of the model.

[0166] Let the temporal features output by the multi-scale convolutional feature extraction module be:

[0167] ;

[0168] in, The spatial features at time t are represented by ; T represents the length of the input time series; C represents the number of feature channels;

[0169] During the Transformer encoding phase, the query, key, and value matrix is ​​first obtained through linear mapping as follows:

[0170] ;

[0171] ;

[0172] ;

[0173] in, , , The weight matrix is ​​trainable.

[0174] Based on this, the model calculates self-attention only along the time dimension to measure the degree of correlation between different time points:

[0175] ;

[0176] Where d is the feature dimension;

[0177] Through this time attention mechanism, the model can weight and aggregate information from different time steps based on the correlation between different moments in the historical sequence, better capture the evolution of multi-factor variables of the sea surface over long time scales, and establish long-distance time dependencies.

[0178] Step 4.3) Introduce a multi-head attention mechanism;

[0179] To further improve the model's expressive power, this application employs a multi-head attention structure in the Transformer encoding module, mapping temporal features to multiple different subspaces and performing parallel attention computations on each. The multi-head attention output is as follows:

[0180] ;

[0181] in, h represents the number of attention heads;

[0182] After multi-head attention, the model extracts temporal dependency patterns from multiple different feature subspaces simultaneously;

[0183] After obtaining the multi-head attention output, it is residually connected with the module input features and then subjected to layer normalization to preserve the original temporal feature information, alleviate the gradient decay problem during deep network training, and improve training stability.

[0184] Step 4.4) Construct a multi-factor interactive attention module;

[0185] Based on time feature modeling, this application introduces a multi-factor interactive attention mechanism to enhance the ability to express the coupling relationship between different marine environmental elements.

[0186] Let the input multi-feature set be: ;

[0187] in, Represents the characteristic of the i-th ocean variable;

[0188] First, the features of each ocean variable are concatenated along the channel dimension to construct a fused branch feature representation:

[0189] ;

[0190] in, Indicates the fusion branch features, This indicates a feature concatenation operation. This fusion branch retains the joint information of all features and can be regarded as a global multi-feature context representation.

[0191] For the The interaction attention weights between the ocean variable feature branches and the fusion branch are defined as follows:

[0192] ;

[0193] in, Indicates the first Attention weights of each ocean variable feature branch relative to the fusion branch. and For a trainable parameter matrix, Indicates the feature dimension;

[0194] No. The feature representation of each ocean variable feature branch after interactive attention update is as follows:

[0195] ;

[0196] in, For a trainable parameter matrix, Indicates the first Feature representation after fusion of feature branches of ocean variables;

[0197] The fused multi-factor features are represented as follows:

[0198] ;

[0199] in, This represents the multi-element characteristics after fusion. Indicates feature fusion operation;

[0200] Through this interactive attention module, the lightweight multi-factor 3D sea surface temperature forecasting model can realize information exchange between the feature branches and fusion branches of various marine variables, and adaptively learn the dynamic coupling relationship between different marine environmental elements, thereby improving the model's ability to characterize the multi-factor collaborative driving process of sea surface temperature change.

[0201] Step 4.5) Construct the depth-sensing attention module;

[0202] After obtaining the fusion features containing time-dependent relationships and multi-element coupling relationships, this application further constructs a deep perception attention module to realize the mapping from two-dimensional multi-element information of the sea surface to three-dimensional sea surface temperature vertical structure.

[0203] Let the set of prediction depth layers be: ;in, This represents the k-th depth layer;

[0204] First, depth weights are generated through linear mapping: ;

[0205] in, This is the depth weight matrix;

[0206] Three-dimensional sea surface temperature prediction is expressed as: ;in, This indicates a predicted sea surface temperature. Representation of spatiotemporal characteristics;

[0207] Through this depth-sensing attention mechanism, the model can automatically learn the temperature correlation between different depth layers, enabling the two-dimensional sea surface multi-element information to be effectively mapped to the three-dimensional ocean temperature structure.

[0208] In summary, the data flow transformation process of the model is as follows: First, historical sea surface multi-element two-dimensional spatiotemporal data are input into the multi-scale convolutional feature extraction module to obtain multi-scale spatial features; then, the multi-scale spatial features are input into the spatiotemporal decoupling Transformer encoding module to obtain temporal features containing time dependencies; next, the temporal features are input into the multi-element interactive attention module, and through interactive modeling between the feature branches and the fusion branches, the fused multi-element features are obtained; finally, the fused multi-element features are input into the depth perception attention module to generate predicted sea surface temperature values ​​corresponding to different depth layers, and output the three-dimensional sea surface temperature prediction results for future periods. The output results are four-dimensional sea surface temperature sequences with spatial dimensions of time × depth × longitude × latitude, where the time dimension is the prediction time step, and each prediction time corresponds to a three-dimensional sea surface temperature field, which can be directly used to generate long-term sea surface temperature forecast products.

[0209] Under the aforementioned lightweight design, the lightweight multi-element three-dimensional sea surface temperature forecasting model can significantly reduce the parameter scale and computational complexity. Attention calculation only applies to the time dimension, enabling the model to achieve stable long-term three-dimensional sea surface temperature extrapolation prediction on platforms or devices with limited computing power. It can still maintain high prediction accuracy and physical consistency without relying on high-resolution three-dimensional driving field data.

[0210] Step 5) Model training and optimization;

[0211] The model is trained using a training set. During the training process, a marine physical constraint loss function is introduced to constrain the physical rationality of the predicted sea surface temperature field. The model parameters are then optimized using a validation set.

[0212] The aforementioned marine physics constraint loss function, by adding a marine physics constraint term to the existing mean square error loss, can effectively constrain the prediction results to meet the basic physical laws of the ocean temperature field, including the stable stratification characteristics and spatial continuity characteristics of seawater temperature in the vertical direction. This ensures that the predicted three-dimensional sea temperature field not only closely approximates the actual observation data in numerical terms, but also satisfies the basic marine physics laws.

[0213] Specifically, during model training, the total loss function is defined as:

[0214] ;

[0215] in, The total loss function for model training; This represents the loss due to data fitting error. Loss due to physical constraints; This is the physical constraint weighting coefficient, used to adjust the trade-off between data error and physical constraints;

[0216] Data error loss is defined using mean squared error:

[0217] ;

[0218] in, This represents the sea surface temperature value predicted by the model at the i-th sample point; This represents the actual sea surface temperature at the corresponding location; This represents the total number of samples. This loss function measures the overall error between the model's predictions and the actual observed data.

[0219] To ensure that the model prediction results satisfy the basic vertical stability law, this application introduces a vertical stability constraint loss;

[0220] Let the three-dimensional sea surface temperature field be: ;in, Indicates spatial latitude and longitude location; Indicates depth layer; Indicates time;

[0221] Define the vertical temperature difference between adjacent depth layers: ;

[0222] When the deep layer temperature is higher than the shallow layer temperature, that is This indicates the presence of a physically unstable stratification.

[0223] Then, the vertical stability constraint loss is defined as:

[0224] ;

[0225] The loss function described above only penalizes deep temperatures when they are higher than shallow temperatures, thus guiding the model to learn predictions that better reflect the actual vertical temperature structure of the ocean.

[0226] To avoid drastic, non-physical spatial changes in the predicted field, this application introduces a spatial smoothing constraint loss;

[0227] First, define the spatial gradient of the sea surface temperature field: ;

[0228] The spatial gradient magnitude is: ;

[0229] Spatial smoothing loss is defined as: ;

[0230] By penalizing the square of the spatial gradient, unreasonable and drastic spatial variations in the predicted sea surface temperature field can be suppressed, making the prediction results smoother and more consistent with the actual ocean structure.

[0231] Combining vertical stability constraints and spatial smoothness constraints, the final physical constraint loss function is obtained as follows:

[0232] ;

[0233] in, The weights are for vertical stability constraints; Weights for spatial smoothing constraints. These are determined by adjusting parameters. and This allows for control over the degree of influence of different physical constraints on model training;

[0234] By introducing the ocean physical constraint loss function as described above, this application considers both data fitting error and ocean physical law constraints during model training, so that the prediction results not only have high numerical accuracy, but also maintain reasonable ocean temperature distribution characteristics in three-dimensional spatial structure, thereby improving the stability, reliability and physical consistency of long-term three-dimensional sea surface temperature forecasts.

[0235] Step 6) Evaluate the performance of the three-dimensional sea surface temperature long-term forecast model using a test set;

[0236] Since the test set was not used in model training, it can be used to evaluate the prediction performance of the prediction model. The mean squared error is used as the evaluation metric.

[0237] The calculation formula is as follows:

[0238] ;

[0239] in, Indicates the first Time sequence Long-term forecast results for each grid, Indicates the first Time sequence The actual values ​​of each grid are total. Each time period One grid;

[0240] Step 7) Forecast product generation;

[0241] After the lightweight multi-element three-dimensional sea surface temperature forecast model is trained, the corresponding sea surface multi-element data sequence is used as input to generate a long-term three-dimensional sea surface temperature forecast product for future periods.

[0242] Specifically, it includes the following steps:

[0243] Step 7.1) Data preprocessing: fill in the missing values ​​for the multi-element data series of the sea surface and perform normalization.

[0244] Step 7.2) The normalized data above are sequentially fed into the lightweight multi-element three-dimensional sea surface temperature forecasting model based on the improved Transformer to obtain the three-dimensional intelligent long-term sea surface temperature forecasting sequence.

[0245] Step 7.3) Perform inverse normalization on the three-dimensional intelligent long-term sea surface temperature forecast sequence to complete the production of the three-dimensional intelligent long-term sea surface temperature forecast product.

[0246] The embodiments described above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. It should be noted that, for those skilled in the art, any modifications and improvements made within the scope of the technology disclosed in this application, based on the technical solution and inventive concept of this application, are all within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the appended claims.

Claims

1. A three-dimensional intelligent sea surface temperature forecasting method considering the influence of multiple factors, characterized by: Multi-source sea surface feature data and three-dimensional sea surface temperature reanalysis data were collected for the target sea area. The time range of the collected data is a long-term series covering a continuous period of many years of history, and the spatial range covers a unified latitude and longitude grid range of the target sea area. Among them, the multi-source sea surface feature data is used to characterize the dynamic and thermal state of the ocean surface, and the three-dimensional sea surface temperature reanalysis data is used as the real value label for model training and evaluation. Based on the acquired data, a training dataset for a multi-element sea surface model was constructed, the model input and prediction labels were defined, and the multi-source sea surface element data and reanalysis data were preprocessed. The training, validation, and test sets were divided according to time series. An improved lightweight Transformer prediction model was constructed, integrating a multi-scale convolutional feature extraction module, a spatiotemporal decoupling Transformer encoding module, a multi-factor interactive attention module, and a depth-sensing attention module. A marine physical constraint mechanism was introduced during training to ensure the physical consistency of the 3D sea surface temperature prediction results. The specific functions of each module are as follows. The multi-scale convolution feature extraction module is used to extract multi-scale spatial features from multi-element two-dimensional spatial data of the sea surface, and capture local small-scale structural information and large-scale ocean circulation features through convolution kernels of different scales. The aforementioned spatiotemporal decoupling Transformer encoding module is used to model the time series of multiple sea surface elements. This module adopts a structure that separates spatial feature encoding and temporal feature encoding. It models the temporal dimension of the spatial features output by the multi-scale convolutional feature extraction module to characterize the temporal evolution relationship of the multiple sea surface elements. It also introduces residual connections and layer normalization structures in the encoding to improve the training stability of deep networks and enhance the model's ability to express long-term sea surface temperature changes. The multi-element interactive attention module establishes feature branches for different ocean variables and constructs a fusion branch. Through interactive attention calculation between each feature branch and the fusion branch, it realizes information exchange and coupling modeling between the features of different ocean variables, and obtains multi-element fusion features. The aforementioned depth perception attention module is used to introduce ocean depth information in the process of three-dimensional sea surface temperature prediction. By constructing weight representations of different depth layers, the model can learn the interlayer coupling relationship of sea surface temperature in the vertical direction and realize the mapping of two-dimensional sea surface multi-element information to three-dimensional sea surface temperature vertical structure. The features fused by the multi-element interactive attention module are further input into the depth perception attention module to generate three-dimensional sea surface temperature prediction results corresponding to different depth layers, thus completing the mapping from two-dimensional sea surface multi-element information to three-dimensional sea surface temperature field. A three-dimensional long-term sea surface temperature (SST) prediction model is obtained through model training, and the prediction performance of the model is evaluated using a test set. Finally, based on the model with the best performance, a three-dimensional long-term SST forecast product for the target sea area for the next 30 days is generated by inputting multi-source sea surface multi-element data from the past 30 days. During model training, the total loss function is defined as: in, The total loss function for model training; This represents the loss due to data fitting error. Loss due to physical constraints; This is the physical constraint weighting coefficient, used to adjust the trade-off between data error and physical constraints; Data error loss is defined using mean squared error: in, This represents the sea surface temperature value predicted by the model at the i-th sample point; This represents the actual sea surface temperature at the corresponding location; This represents the total number of samples; the loss function is used to measure the overall error between the model's predictions and the actual observed data. Let the three-dimensional sea surface temperature field be: ;in, Indicates spatial latitude and longitude location; Indicates depth layer; Indicates time; Define the vertical temperature difference between adjacent depth layers: ; When the deep layer temperature is higher than the shallow layer temperature, that is This indicates the presence of a physically unstable stratification. Then, the vertical stability constraint loss is defined as: Introducing a spatial smoothing constraint loss; first, defining the spatial gradient of the sea surface temperature field: ; The spatial gradient magnitude is: ; Spatial smoothing loss is defined as: ; By penalizing the square of the spatial gradient, unreasonable and drastic spatial variations in the predicted sea surface temperature field can be suppressed, making the prediction results smoother and more consistent with the actual ocean structure. Combining vertical stability constraints and spatial smoothness constraints, the final physical constraint loss function is obtained as follows: ； in, The weights are for vertical stability constraints; The weights are used for spatial smoothing constraints.

2. The three-dimensional intelligent sea surface temperature forecasting method considering the influence of multiple factors according to claim 1, characterized in that: Includes the following steps, Step 1) Collect multi-source sea surface feature data and three-dimensional sea surface temperature reanalysis data of the target sea area, construct the model training dataset, and determine the model training input and training labels; Step 2), Data Preprocessing; Determine if there are any missing values ​​in the model training dataset, replace them with preset anomaly marker values, and introduce the corresponding missing mask into the model input; normalize the model training dataset to unify the numerical range of each variable. Step 3) Divide the dataset into training, validation, and test sets; The normalized model training dataset is divided into a training set, a validation set, and a test set; Step 4) Establish a lightweight, multi-element, three-dimensional sea surface temperature forecasting model; The lightweight multi-element three-dimensional sea surface temperature forecasting model includes: a multi-scale convolutional feature extraction module, a spatiotemporal decoupling Transformer encoding module, a multi-element interactive attention module, and a depth-sensing attention module; its network model input is multi-element spatiotemporal data of sea surface over several consecutive days in historical time periods; The multi-element two-dimensional data of the sea surface spatially constitutes a two-dimensional matrix of the number of longitude grids × the number of latitude grids in the sea area, and in the element dimension constitutes a multi-channel feature tensor. It is then input into the model in the form of a time series to predict the three-dimensional sea surface temperature distribution in future periods. Step 5) Model training and optimization; The model is trained using a training set, and a marine physics constraint loss function is introduced during the training process to constrain the physical rationality of the predicted sea surface temperature field; the model parameters are evaluated and optimized using a validation set. Step 6) Evaluate the performance of the three-dimensional sea surface temperature long-term forecast model using a test set; mean square error is used as the evaluation metric. Step 7) Forecast product generation; After the lightweight multi-element three-dimensional sea surface forecast model is trained, a long-term three-dimensional sea surface temperature forecast product for future periods is generated by using historical multi-source sea surface element data sequences as input.

3. The three-dimensional intelligent sea surface temperature forecasting method considering the influence of multiple factors according to claim 2, characterized in that: Step 1), the model training dataset includes ocean history reanalysis data and sea surface multi-element data; The model training input and training labels are constructed using a time series approach, with historical continuous sea surface multi-element data sequences as model inputs and future three-dimensional sea surface temperature data sequences as model training labels. The aforementioned multi-element two-dimensional sea surface data includes sea surface temperature, sea surface height anomaly, sea surface wind field, and heat flux. Each element constitutes a two-dimensional spatial field on the same latitude and longitude grid and is input into the model in the form of a multi-channel tensor.

4. The three-dimensional intelligent sea surface temperature forecasting method considering the influence of multiple factors according to claim 3, characterized in that: Step 2) involves normalizing data including sea surface temperature, sea surface height anomaly, sea surface wind field, and heat flux. Each variable is normalized to its maximum and minimum values ​​according to the following formula, so that the values ​​of each variable are uniformly between 0 and 1.

5. The three-dimensional intelligent sea surface temperature forecasting method considering the influence of multiple factors according to claim 4, characterized in that: Step 3) involves dividing the preprocessed model training dataset into a training set, a validation set, and a test set in chronological order. Each of the training set, validation set, and test set consists of multi-factor sea surface input samples and three-dimensional sea surface temperature label samples for the corresponding time period. The training set is used for learning model parameters, the validation set is used for adjusting model parameters and optimizing the training process, and the test set is used for independent evaluation of the model's long-term forecast performance.

6. The three-dimensional intelligent sea surface temperature forecasting method considering the influence of multiple factors according to claim 5, characterized in that: Step 4) specifically includes the following sub-steps: Step 4.1) Extract feature information at different spatial scales; The input spatiotemporal data of multiple sea surface elements is fed into the multi-scale convolutional feature extraction module. Feature information of different spatial scales is extracted through convolutional kernels of different scales and fused to obtain multi-scale spatial features. Step 4.2), Temporal modeling stage; The multi-scale spatial features are input into the spatiotemporally decoupled Transformer encoding module, where the Transformer structure performs self-attention computation only in the time dimension; time series modeling is performed on the multi-element time series of the sea surface. Step 4.3) Introduce a multi-head attention mechanism; A multi-head attention structure is adopted in the Transformer encoding module to improve the model's ability to capture temporal features. Residual connections and layer normalization are introduced after the multi-head attention output to improve the training stability of deep networks. Step 4.4) Construct a multi-factor interactive attention module; After time-series modeling, the multi-element features are established into feature branches and a fusion branch is constructed. Through the interaction attention calculation between each feature branch and the fusion branch, the multi-element information exchange and coupled modeling are realized to obtain the fused multi-element features. Step 4.5) Construct the depth-sensing attention module; The fused multi-element features are input into the depth perception attention module, which introduces ocean depth information and generates depth weights corresponding to different depth layers, thereby realizing the mapping of two-dimensional sea surface multi-element information to three-dimensional ocean temperature structure. The input multi-element spatiotemporal data of the sea surface is processed by a multi-scale convolutional feature extraction module to obtain multi-scale spatial features. The multi-scale spatial features are then used by a spatiotemporal decoupling Transformer encoding module to complete temporal modeling. A multi-element interactive attention module is then used to realize the interaction between the branches of each element and the fusion branch to generate fused multi-element features. Finally, a depth perception attention module is used to map the results to the three-dimensional sea surface temperature prediction results for future time periods. The output results are in spatial dimensions of time × depth × longitude × latitude, where the time dimension is the prediction time step, and are used to generate long-term sea surface temperature forecast products.

Images