Fine-tuned marine prediction method based on frequency domain enhancement neural network

Abstract

The application provides a fine ocean prediction method based on a frequency domain enhanced neural network, and relates to the technical field of ocean prediction, and specifically comprises the following steps: cleaning and standardizing an original data set; arranging time-space sequence data of multiple-source ocean environment variables into a unified format to form an input feature set; fusing time-space features and frequency domain features by using a constructed deep learning model to predict the ocean environment variables; performing inverse standardization processing on the prediction result to obtain a final prediction value; combining the predicted future n-day ocean environment variables with the past n-day historical data to form new input features, and repeating the prediction steps to realize the prediction of the future ocean environment variables. The technical scheme of the application overcomes the problems in the prior art that frequency domain and spatial domain information cannot be efficiently fused under limited computing resources, and the mask characteristics and multivariate correlation characteristics of ocean data cannot be fully considered.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of marine prediction, and in particular to a fine marine prediction method based on a frequency domain enhanced neural network. BACKGROUND

[0002] Fine marine prediction is an important research direction of modern marine science and engineering, and has important application value in disaster prevention and reduction, shipping safety, fishery resource management, offshore energy development and other fields. In recent years, with the rapid development of artificial intelligence and high-performance computing technology, data-driven intelligent marine prediction has become a new research hotspot. Deep learning methods can automatically extract high-dimensional features of multi-source data, learn the nonlinear laws of marine dynamics without relying on complex equation solving, and achieve fast and continuous prediction. Various representative model systems have been proposed in existing research, gradually forming a development trend centered on multi-scale feature extraction, attention mechanism modeling and mask constraint optimization. For example, the hierarchical Transformer architecture realizes the joint modeling of global and local information through block representation and encoding-decoding structure, and uses sea-land mask mechanism to reduce the interference of invalid areas; the model with autoregressive time series as the core describes the time evolution characteristics of different variables through time mixing and multi-step rolling; some frameworks introduce multi-source information of ocean and atmosphere in a unified network to realize cross-element coupled prediction; in addition, some researches also use multi-grid structure and cycle strategy to balance large-scale background and small-scale structure in long-term prediction, which improves the stability of medium and long-term prediction. These methods have made significant progress in small-scale structure reconstruction, model bias correction and fast rolling prediction, and have shown the application potential of artificial intelligence technology in the marine field.

[0003] However, as a whole, the current intelligent marine prediction model still has some limitations in the regional application of high resolution, three-dimensional multi-variable. On the one hand, although frequency domain modeling can effectively enhance the perception of global dynamic process, it is easy to appear details smoothing and energy loss in complex terrain and local structure; on the other hand, some multi-scale networks have information attenuation in the up-sampling and feature reconstruction stage, which is difficult to accurately restore the small-scale features while maintaining the consistency of large-scale. In addition, the internal coupling relationship between different physical variables is often not fully utilized, resulting in the incoordination of the phase and amplitude between multi-variable fields, which affects the expression of overall dynamic balance; at the same time, some model structures are large and complex, with high training and inference cost, which is difficult to meet the demand of high frequency and low delay prediction of business system.

[0004] Therefore, there is an urgent need for a method for fine-grained ocean prediction that can efficiently fuse frequency domain and spatial domain information under limited computing resources, has global scale perception ability and maintains local structure detail expression, can use spectral information to improve the description of the overall dynamics of the ocean system, and can realize continuous reconstruction of complex spatial fields through multi-scale feature interaction. At the same time, the model should fully consider the mask characteristics and multivariate correlation characteristics of ocean data, have good physical consistency and generalizability, and provide a new technical approach for realizing high-precision, fast response and engineering deployable intelligent ocean prediction. SUMMARY

[0005] The main purpose of the present application is to provide a fine-grained ocean prediction method based on a frequency domain enhanced neural network, to solve the problem that the prior art cannot efficiently fuse frequency domain and spatial domain information under limited computing resources, and cannot fully consider the mask characteristics and multivariate correlation characteristics of ocean data.

[0006] To achieve the above purpose, the present application provides a fine-grained ocean prediction method based on a frequency domain enhanced neural network, which specifically comprises the following steps:

[0007] S1, cleaning and standardizing the original data set.

[0008] S2, arranging the time and space sequence data of the multi-source ocean environmental variables into a unified format to form an input feature set, the multi-source ocean environmental variables including temperature, salinity, flow rate and sea surface height.

[0009] S3, using the constructed deep learning model to fuse the time and space features and the frequency domain features to predict the ocean environmental variables.

[0010] S4, performing inverse standardization processing on the prediction result to obtain the final prediction value.

[0011] S5, combining the predicted future n-day ocean environmental variables with the past n-day historical data to form new input features, and repeating steps S1-S4 to realize prediction of future ocean environmental variables.

[0012] Further, step S1 specifically comprises the following steps:

[0013] S1.1, data cleaning of the original data set, identifying and deleting outliers, and interpolating and filling invalid values on land.

[0014] S1.2, assuming is the original value of the variable after step S1 processing, and are the maximum and minimum values in the data set, and the maximum and minimum value standardization method is used to normalize the variable value to the interval [0, 1]:

[0015] ;

[0016] wherein, is the standardized variable value.

[0017] Further, the step S2 specifically comprises:

[0018] Suppose the spatial shape of the three-dimensional marine environment variable is wherein, is the length, is the width, is the depth; the temperature, salinity, flow field and sea surface temperature of the past n days are stacked along the channel dimension to construct a four-dimensional tensor ; wherein the channel dimension includes: variable type information and time series information of each variable.

[0019] Further, the step S3 specifically comprises the following steps:

[0020] S3.1, constructing a deep learning model comprising: an input layer, a spatial down-sampling layer, a spatio-temporal feature encoding layer, a frequency domain feature processing layer, an encoded feature fusion layer and a decoder connected in turn.

[0021] S3.2, the spatial down-sampling layer utilizes 3D convolution operation to realize spatial down-sampling, and gradually reduces the spatial size by controlling the convolution step length, and then uses LayerNorm layer for normalization; suppose the four-dimensional tensor input in the input layer is , the process of 2 times down-sampling by 3D convolution is represented as:

[0022] ;

[0023] wherein, and represent the weight and bias, and represent the size and convolution step length of the convolution kernel, is the 3D convolution operation.

[0024] S3.3, the spatio-temporal feature encoding layer encodes the feature tensor after the size reduction, and extracts local features at different scales; the spatio-temporal feature encoding layer includes: 3D deep separable convolution module, LayerNorm normalization, multilayer perception module and residual connection; the spatio-temporal feature tensor encoding process is represented as:

[0025] ;

[0026] wherein, and respectively represent the multi-layer perceptron module and the depthwise separable convolution module, to the spatio-temporal feature.

[0027] S3.4, in the frequency domain feature processing layer, the adaptive Fourier neural operator firstly maps the spatio-temporal feature to the frequency domain; then, a learnable weight matrix is introduced in the frequency domain; finally, the processed frequency domain feature is mapped back to the spatio-temporal domain through inverse Fourier transform, completing the feature reconstruction:

[0028] ;

[0029] wherein, and represent the Fourier transform and the inverse Fourier transform, respectively.

[0030] S3.5, repeat steps S3.2 to S3.4 multiple times to gradually extract spatio-temporal features and frequency domain features at multiple stages.

[0031] Further, step S3 further comprises the following steps:

[0032] S3.6, in the spatial upsampling layer, firstly stack the encoded features in the channel dimension as the input of the decoder; then, use tri-linear interpolation to gradually perform upsampling; after completing the upsampling, introduce 3D convolution operation to reconstruct the feature; assuming that the input four-dimensional tensor in the first layer of the deep learning model is , the upsampling process is represented as:

[0033] ;

[0034] wherein, is an activation function, is tri-linear interpolation.

[0035] S3.7, input the upsampled feature into the decoder for decoding, and the decoding process is realized using the same depthwise separable convolution module as the encoding process:

[0036] ;

[0037] wherein, the decoder output feature, is the input of the decoder, and represent the multi-layer perceptron module and the depthwise separable convolution module in the decoder, respectively.

[0038] S3.8, repeat steps S3.6 to S3.7 multiple times to gradually extract multi-scale spatio-temporal features and frequency domain features of the marine environmental variables, and gradually restore the spatial resolution of the features in the decoding process.​

[0039] Further, step S4 is specifically:

[0040] Suppose the variable value output through step S3 is , the deep learning model output result is inversely normalized according to formula (6) to obtain the final prediction value :

[0041] .

[0042] The present application has the following beneficial effects:

[0043] (1) Coding-decoding architecture design: the present application constructs a U-shaped coding-decoding network based on 3D depth separable convolution, and realizes multi-scale feature extraction through phased down-sampling and up-sampling. This architecture can directly model the marine environmental variables in three-dimensional space, effectively integrate time series information through channel dimension while capturing spatial dependence, so as to realize sufficient modeling and efficient learning of spatio-temporal features.

[0044] (2) Spatio-temporal and frequency domain feature fusion mechanism: in the proposed model, feature fusion and utilization are the core innovation points. While extracting multi-scale local spatio-temporal features, an adaptive Fourier neural operator is introduced to map the features to the frequency domain for noise reduction and global modeling, thereby enhancing the learning ability of frequency domain features. In the decoding stage, spatio-temporal and frequency domain features are fully fused, and a progressive decoding strategy is adopted, effectively improving the stability and prediction accuracy of the model.

[0045] The "frequency domain enhanced multi-scale coding-decoding neural network for fine marine prediction" proposed by the present application realizes the global and local integrated modeling of marine dynamic processes by introducing a frequency domain enhancement mechanism in the multi-scale coding-decoding structure. The frequency domain module can effectively capture large-scale circulation and long-range correlation features, while maintaining low computational cost and strengthening the expression of the overall dynamic structure; multi-scale convolution and up-sampling process focus on detail recovery and local structure reconstruction, so that the model can reproduce basin-scale background fields and small-scale changes in complex sea areas at the same time, improving spatial continuity and physical authenticity. The model provided by the present application has high precision and efficiency, and realizes fast prediction of complex marine three-dimensional dynamic field while maintaining strong computational controllability. Compared with existing methods, the present application can more accurately reconstruct multi-scale marine element fields under limited computing resources, has the comprehensive advantages of stable boundary, clear structure, good physical consistency, high training stability and flexible deployment, and provides an efficient and generalizable intelligent technical approach for high-resolution regional marine prediction. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the specific embodiments or prior art of the present application, the drawings required to be used in the description of the specific embodiments or prior art will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0047] Figure 1 A flow chart of a fine ocean prediction method based on a frequency domain enhanced neural network is shown.

[0048] Figure 2 The prediction results of east-west sea current flow velocity using the model in the prior art are shown.

[0049] Figure 3 The prediction results of east-west sea current flow velocity using the method provided by the present application are shown.

[0050] Figure 4 The time series of east-west flow velocity in a certain sea area predicted by the method provided by the present application is shown.

[0051] Figure 5 The east-west flow velocity power spectrum of Figure 4 is shown.

[0052] Figure 6 The time series of north-south flow velocity in a certain sea area predicted by the method provided by the present application is shown.

[0053] Figure 7 The north-south flow velocity power spectrum of Figure 6 is shown. DETAILED DESCRIPTION

[0054] The technical solutions of the present application will be described in detail below with reference to the drawings. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0055] As shown in Figure 1 a fine ocean prediction method based on a frequency domain enhanced neural network, specifically comprising the following steps:

[0056] S1, cleaning and standardizing the original data set.

[0057] S2, arranging the time-space sequence data of multi-source ocean environmental variables into a unified format to form an input feature set, the multi-source ocean environmental variables including temperature, salinity, flow velocity and sea surface height.

[0058] S3, fusing the space-time feature and the frequency domain feature by using the constructed deep learning model to predict the marine environment variable.

[0059] S4, performing inverse standardization processing on the prediction result to obtain a final prediction value.

[0060] S5, combining the predicted future n-day marine environment variable with the past n-day historical data to form a new input feature, and repeating steps S1-S4 to realize the prediction of the future marine environment variable.

[0061] The application proposes a U-shaped architecture deep learning model, which can be directly used for predicting marine environment three-dimensional space fields (such as temperature, salinity, flow field and marine surface height). The application only needs to input 4-day historical data to obtain a prediction result of 14 days in the future (the time resolution is 3 hours), and the process adopts a rolling prediction mode, which significantly improves the timeliness and applicability of the prediction.

[0062] Specifically, step S1 specifically includes the following steps:

[0063] S1.1, data cleaning is performed on the original data set, abnormal values are identified and deleted, and invalid values of the land part are interpolated and filled to ensure the stability of the model training.

[0064] S1.2, assuming is the original value of the variable after step S1 processing, and are the maximum and minimum values in the data set, in order to make the data distribution more concentrated and the gradient update of the training process more stable, the maximum and minimum value standardization method is used to normalize the variable value to the interval [0, 1]:

[0065] ;

[0066] wherein, is the variable value after standardization.

[0067] The original data is cleaned and standardized. Before model training, the original data set is denoised and abnormal values are identified and removed to ensure the continuity and reliability of the data. The data is standardized by using the maximum and minimum value normalization method, so that the distribution of each variable is more concentrated, and the stability of the model training is improved.

[0068] Specifically, step S2 is specifically:

[0069] Assuming that the spatial shape of the three-dimensional marine environment variable is wherein, is the length, is the width, depth; the temperature, salinity, flow field and sea surface temperature of the past n days along the channel dimension stacking, constructing a four-dimensional tensor ; wherein the channel dimension comprises variable type information and time series information of each variable, used to represent the input features of the model. The spatiotemporal sequence data of multi-source marine environmental variables (including temperature, salinity, flow rate and sea surface height, etc.) are sorted into a unified format to form an input feature set for model training, ensuring that the input data can reflect the characteristics of marine changes at different time and spatial scales.

[0070] Specifically, the constructed four-dimensional tensor is input into the model, and after multiple up-sampling and down-sampling, features of different scales are extracted. In addition, the convolutional network is organically combined with the adaptive Fourier neural operator, considering the information in the spatial and frequency domains, which can capture local details and global modeling, realizing accurate prediction of marine environmental variables.

[0071] Step S3 specifically includes the following steps:

[0072] S3.1, constructing a deep learning model includes: an input layer, a spatial down-sampling layer, a spatiotemporal feature encoding layer, a frequency domain feature processing layer, an encoding feature fusion layer and a decoder connected in turn.

[0073] S3.2, in the spatial down-sampling layer, 3D convolution operation is used to realize spatial down-sampling, and by controlling the convolution step length, the spatial size is gradually reduced, and then LayerNorm layer is used for normalization; assuming that the four-dimensional tensor (including channel) input in the input layer is , the process of 2 times down-sampling by 3D convolution is represented as:

[0074] ;

[0075] wherein, and represent the weight and bias respectively, and represent the size and convolution step length of the convolution kernel respectively, is a 3D convolution operation. The three-dimensional convolution down-sampling structure is used inside the model, and by controlling the sampling interval of the convolution kernel, the spatial resolution is gradually reduced, and a normalization operation is added after each layer to keep the features stable. Through the stacking of multiple convolution blocks, multi-scale spatial and frequency domain features are extracted, providing a basis for subsequent fusion and reconstruction.

[0076] S3.3, in the spatiotemporal feature encoding layer, the feature tensor after size reduction The local features at different scales are encoded and extracted; the spatio-temporal feature encoding layer includes: a 3D depth separable convolution module, a LayerNorm normalization, a multi-layer perception module and a residual connection; the spatio-temporal feature tensor encoding process is represented as:

[0077] ;

[0078] wherein, and represent the multi-layer perception module and the depth separable convolution module respectively, is the spatio-temporal feature.

[0079] The model adopts the separable convolution and the multi-layer perception module to reduce the parameter quantity and improve the non-linear expression ability, and adaptively filters the features of different spatial frequencies in the feature extraction process, enhances the sensitivity of the model to the key area, and realizes the collaborative representation of the frequency domain features and the spatial features.

[0080] S3.4, in the frequency domain feature processing layer, the adaptive Fourier neural operator first maps the spatio-temporal feature to the frequency domain, filters part of the high-frequency components in the frequency domain to weaken the noise interference and realize sparse modeling, so as to be able to adaptively capture the global dependence relationship. Subsequently, a learnable weight matrix is introduced in the frequency domain to realize the information interaction and fusion between channels. Finally, the processed frequency domain features are mapped back to the spatio-temporal domain through inverse Fourier transform to complete feature reconstruction:

[0081] ;

[0082] wherein, and represent the Fourier transform and the inverse Fourier transform respectively.

[0083] In the frequency domain enhancement process, the interaction and fusion of different frequency band features are realized through feature mapping and inverse transform. The low-frequency features are used to maintain the overall dynamic structure, and the high-frequency features are used to restore local details, realizing the coordination and unity of global and local information.

[0084] S3.5, steps S3.2 to S3.4 are repeated multiple times to gradually extract spatio-temporal features and frequency domain features at multiple stages; the extracted features are used for decoding of the deep learning model. The specific number of repetitions is four, and spatio-temporal features and frequency domain features at four stages are gradually extracted, which will be used for subsequent decoding of the model.

[0085] Specifically, step S3 further includes the following steps:

[0086] S3.6, the spatially sampled layer, first stacks the encoded features in the channel dimension as the input of the decoder; then, it gradually upsamples using tri-linear interpolation, where the magnification of each stage is consistent with the corresponding down-sampling stage. After completing the up-sampling, a 3D convolution operation is introduced to reconstruct the features, thus making up for the information loss caused by simple interpolation and improving the stability of the decoding process. Assuming that the input four-dimensional tensor in the deep learning model is , the up-sampling process is represented as:

[0087] ;

[0088] wherein, is the activation function, is the tri-linear interpolation.

[0089] The decoding stage uses a combination of three-dimensional deconvolution and tri-linear interpolation to gradually up-sample and gradually restore the spatial resolution. After each level of up-sampling, normalization and activation operations are performed to ensure the smoothness and stability of the output results.

[0090] S3.7, the up-sampled features are input into the decoder for decoding, and the decoding process uses the same depth separable convolution module as the encoding process:

[0091] ;

[0092] wherein, is the output feature of the decoder, is the input of the decoder, and represent the multi-layer perceptron module and the depth separable convolution module in the decoder, respectively, which are independent of and in the encoder.

[0093] The decoding process fully integrates local features and frequency domain features of different scales, ensuring that the network not only retains local details but also completely captures global patterns.

[0094] S3.8, steps S3.6 to S3.7 are repeated multiple times to gradually extract multi-scale spatio-temporal features and frequency domain features of the marine environmental variables and gradually restore the spatial resolution of the features in the decoding process. The specific number of repetitions is four. This gradual feature extraction and modeling mechanism not only effectively enhances the fusion of global and local information, but also significantly improves the stability and generalization ability of the model. After sufficient learning of the input data, the model uses a depth separable convolution module as the output layer to generate a 2-day prediction of the marine environmental variables.

[0095] ​The decoding features and the encoding features are fused at the same scale, so that the model not only retains global structure information, but also strengthens local detail performance, thereby realizing high-precision three-dimensional marine element prediction output.

[0096] After multiple rounds of iterative training, the hyperparameters and weights of the deep learning model are determined, the optimal model parameters are saved, and the model training is completed. In the inference stage, only the past 4 days of marine environment variables need to be input to quickly complete the prediction of the next 2 days. The model training adopts a supervised learning method, which is optimized through continuous time samples, so that the model has stable time extrapolation capability, and finally a deep neural network model capable of multi-step rolling prediction is obtained.

[0097] Specifically, step S4 is specifically:

[0098] Suppose the variable value output by step S3 is The deep learning model output result is inversely normalized according to formula (6) to obtain the final prediction value , and the prediction results of different marine environment variables are separated for subsequent analysis.

[0099] .

[0100] In the prediction stage, the model input is continuous historical marine environment variable data, and the output is the prediction result of the three-dimensional marine environment in the future period, including temperature, salinity, flow rate and sea surface height and other elements.

[0101] Specifically, in step S5, the predicted future 2-day marine environment variables are combined with the past 2-day historical data to form new input features, and steps 1-4 are repeated to realize the prediction of the future 4th to 6th day marine environment variables. Continue to perform the rolling prediction operation until the prediction result of the future 14-day marine environment variables is obtained. Through multi-step rolling prediction, the model can use the existing historical data to continuously generate the marine state result for several days in the future, significantly improving the accuracy and stability of the marine environment prediction.

[0102] Figure 2 and Figure 3 The spatial distribution of the surface east-west current velocity is shown. The left figure is the prediction result of the Transformer derived architecture commonly used in large models in the prior art, and the right figure is the prediction result of the method of the present application. The color bar represents the abnormal value of the current velocity (unit: m / s), and the red color represents the eastward flow enhancement, and the blue color represents the westward flow enhancement. In comparison, the result of the method of the present application is more continuous in the flow field structure and more clear in detail description, and can more accurately reflect the characteristics of small and medium scale flow.

[0103] Figures 4-7The time series and corresponding power spectrum distribution of the east-west and north-south components of the surface (about 2 m deep) current in a representative sea area are shown. From the power spectrum results, it can be seen that the predicted flow rate has significant energy peaks at about 0.5-day and 1-day periods, indicating that the method provided by the present application can effectively capture the main periodic signals such as tidal currents, and has high precision and physical consistency in the characterization of time series variation characteristics.

[0104] Of course, the above description is not a limitation on the present application, and the present application is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the spirit and scope of the present application should also be within the protection scope of the present application.

Claims

1. A fine-tuned ocean prediction method based on a frequency domain enhanced neural network, characterized in that, Specifically comprising the following steps: S1, cleaning and standardizing the original data set; S2, arranging the spatio-temporal sequence data of multi-source marine environment variables into a unified format to form an input feature set, the multi-source marine environment variables including temperature, salinity, flow rate and sea surface height; S3, using the constructed deep learning model to fuse the spatio-temporal features and frequency domain features to predict the marine environment variables; S4, performing inverse standardization processing on the prediction result to obtain the final prediction value; S5, combining the predicted future n-day marine environment variables with the past n-day historical data to form a new input feature, and repeating steps S1-S4 to realize the prediction of future marine environment variables; Step S3 specifically comprises the following steps: S3.1, constructing a deep learning model including an input layer, a spatial down-sampling layer, a spatio-temporal feature coding layer, a frequency domain feature processing layer, an encoded feature fusion layer and a decoder connected in turn; S3.2, the spatial down-sampling layer utilizes 3D convolution operation to realize spatial down-sampling, gradually reduces the spatial size by controlling the convolution step, and then uses the LayerNorm layer for normalization; assuming that the four-dimensional tensor input in the input layer is The process of 2 times down-sampling by 3D convolution is expressed as: ； wherein, and represent weights and biases, respectively, and represent the size and the convolution stride of the convolution kernel, respectively, is a 3D convolution operation; S3.3, the spatiotemporal feature encoding layer encodes the down-scaled feature tensor encoding, extracting local features at different scales; the spatiotemporal feature encoding layer includes: a 3D depth separable convolution module, a LayerNorm normalization, a multi-layer perception module and a residual connection; the spatiotemporal feature tensor encoding process is represented as: ； wherein, and represent a multi-layer perceptron module and a depthwise separable convolution module, respectively, is a spatio-temporal feature; S3.4, in the frequency domain feature processing layer, the adaptive Fourier neural operator first maps the spatio-temporal features to the frequency domain; then, a learnable weight matrix is introduced in the frequency domain; finally, the processed frequency domain features are mapped back to the spatio-temporal domain through inverse Fourier transform to complete feature reconstruction: ； wherein and represent the Fourier transform and the inverse Fourier transform, respectively; S3.5, repeating steps S3.2 to S3.4 multiple times to gradually extract spatio-temporal features and frequency domain features of multiple stages; S3.6, the layer is spatially up-sampled, firstly, the encoded features are stacked in the channel dimension as the input of the decoder; then, the up-sampling is performed step by step using tri-linear interpolation; after the up-sampling is completed, a 3D convolution operation is introduced to reconstruct the features; it is assumed that the input four-dimensional tensor in the deep learning model is layer is The process of up-sampling is represented as: ； wherein is an activation function, is a trilinear interpolation; S3.7, input the up-sampled features into the decoder for decoding, and the decoding process is realized by using the same deep separable convolution module as the encoding process: ； in, Decoder output features For the input of the decoder, and These represent the multilayer perceptron module and the depthwise separable convolutional module in the decoder, respectively. S3.8, repeating steps S3.6 to S3.7 multiple times to gradually extract multi-scale spatio-temporal features and frequency domain features of the marine environment variables, and gradually restore the spatial resolution of the features in the decoding process.

2. The method of claim 1, wherein the method is based on a frequency domain enhanced neural network. Step S1 specifically comprises the following steps: S1.1, cleaning the original data set, identifying and deleting outliers, and interpolating invalid values on land; S1.2, assuming for the variable raw values after processing by step S1, and are the maximum and minimum values in the dataset, the variable values are normalized into the interval [0, 1] using the max-min normalization method: ； wherein is the normalized variable value.

3. The method of claim 1, wherein the method is based on a frequency domain enhanced neural network. Step S2 is specifically: Assume the spatial shape of a three-dimensional marine environmental variable is where, is length, is width, is depth; the temperature, salinity, flow field and sea surface temperature of the past n days are stacked along the channel dimension to construct a four-dimensional tensor ; where the channel dimension includes both variable type information and time series information of each variable.

4. The method of claim 1, wherein the method is based on a frequency domain enhanced neural network. Step S4 is specifically: Assume that the variable value output through step S3 is The final prediction value is obtained by inversely normalizing the deep learning model output result according to formula (6) : 。

Images