A sea wave significant wave height super-resolution reconstruction method, device and electronic equipment

Abstract

The application provides a sea wave significant wave height super-resolution reconstruction method and device and electronic equipment. An initial sea wave significant wave height field reconstruction model is trained by using a real low-resolution multivariate sea wave field sequence of continuous T time points before a plurality of time points τ and a real high-resolution sea wave significant wave height field to construct a training data set. The sea wave significant wave height field reconstruction model learns the propagation law of waves along the wave direction according to the significant wave height, average wave period, and sine and cosine components of the wave direction in the sea wave field sequence of continuous T time points, and extracts the physical structure of multiple scales in the wave field sequence. The sea wave significant wave height field is reconstructed according to the learned propagation law and the extracted physical structure under different scales. Compared with the traditional reconstruction method which cannot accurately reflect the wave evolution law and cannot accurately reflect the detailed information of the sea wave field, the resolution of the sea wave significant wave height field constructed by the application can be significantly improved.

Description

Technical Field

[0001] This application relates to the field of marine environmental analysis technology, and in particular to a method, apparatus and electronic equipment for super-resolution reconstruction of the effective wave height of ocean waves. Background Technology

[0002] Significant wave height (SWH) is a key parameter describing ocean wave conditions and has important applications in ocean monitoring, marine navigation, and coastal engineering. Specifically, significant wave height refers to the average wave height of the first third of the largest waves in a wave observation sequence, ordered from highest to lowest wave height.

[0003] Traditional methods for obtaining significant wave height (SWH) involve satellite observations or numerical simulations. However, due to the spatial resolution limitations of these methods, acquiring high-resolution SWHs is extremely costly. To reduce the cost of obtaining high-resolution SWHs, traditional methods employ interpolation or deep learning-based image super-resolution methods. However, existing deep learning-based image super-resolution methods are primarily designed for single-moment images and still have significant limitations when applied to SWH reconstruction tasks. Summary of the Invention

[0004] In view of this, embodiments of this application provide a method, apparatus and electronic device for super-resolution reconstruction of ocean wave significant wave height, which utilizes historical time-series information to improve the spatial resolution of the ocean wave significant wave height field at the current moment, so as to achieve low-cost reconstruction of a high spatial resolution ocean wave significant wave height field.

[0005] In a first aspect, embodiments of this application provide a super-resolution reconstruction method for the significant wave height of ocean waves, wherein the method includes: Obtain a low-resolution multivariate wave field sequence for T consecutive time points from the current time t, wherein the low-resolution multivariate field sequence at each time point includes the following physical variables: significant wave height, mean wave period, wave direction sine component, and wave direction cosine component; The target wave significant wave height field reconstruction model is invoked to reconstruct the wave significant wave height field based on the low-resolution multivariate wave field sequence of the continuous T time periods, so as to obtain the target high-resolution wave significant wave height field at the current time t. The target wave significant wave height field reconstruction model is pre-trained using the following method: A training dataset is constructed using a sequence of real low-resolution multivariate ocean wave fields at several times τ and real high-resolution effective wave height fields at each time τ. The training dataset is input into the initial wave effective wave height field reconstruction model. The wave effective wave height field reconstruction model learns the wave propagation law along the wave direction and extracts the physical structure of the real low-resolution multivariate wave field sequence at different scales. Based on the learned propagation law and the extracted physical structure of the real low-resolution multivariate wave field sequence at different scales, the wave effective wave height field is reconstructed to obtain the target high-resolution wave effective wave height field at time τ. Based on the data difference between the target high-resolution wave significant height field and the actual high-resolution wave significant height field at time τ, the model parameters of the wave significant height field reconstruction model are adjusted until the data difference satisfies the preset loss function constraint. The wave effective wave height field reconstruction model corresponding to the data difference satisfying the preset loss function constraint is determined as the target wave effective wave height field reconstruction model.

[0006] Secondly, embodiments of this application provide a super-resolution reconstruction device for the significant wave height of ocean waves, wherein the device includes: The acquisition module is used to acquire a low-resolution multivariate wave field sequence for T consecutive time points from the current time t, wherein the low-resolution multivariate field sequence at each time point includes the following physical variables: significant wave height, mean wave period, wave direction sine component and wave direction cosine component; The reconstruction module is used to call the target wave significant wave height field reconstruction model to reconstruct the wave significant wave height field based on the low-resolution multivariate wave field sequence of the continuous T time periods, so as to obtain the target high-resolution wave significant wave height field at the current time t. The model training module is used to construct a training dataset using real low-resolution multivariate wave field sequences of T consecutive times at several times τ and real high-resolution effective wave height fields of waves at various times τ. The training dataset is input into the initial wave effective wave height field reconstruction model. The wave effective wave height field reconstruction model learns the wave propagation law along the wave direction and extracts the physical structure of the real low-resolution multivariate wave field sequence at different scales. Based on the learned propagation law and the extracted physical structure of the real low-resolution multivariate wave field sequence at different scales, the wave effective wave height field is reconstructed to obtain the target high-resolution wave effective wave height field at time τ. Based on the data difference between the target high-resolution wave significant height field and the actual high-resolution wave significant height field at time τ, the model parameters of the wave significant height field reconstruction model are adjusted until the data difference satisfies the preset loss function constraint. The wave effective wave height field reconstruction model corresponding to the data difference satisfying the preset loss function constraint is determined as the target wave effective wave height field reconstruction model.

[0007] Thirdly, embodiments of this application provide an electronic device, wherein the electronic device includes: a processor; and a memory storing a program; wherein the program includes instructions, which, when executed by the processor, cause the processor to perform the ocean wave effective wave height super-resolution reconstruction method described in the first aspect.

[0008] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the super-resolution reconstruction method for effective wave height of ocean waves described in the first aspect.

[0009] The beneficial effects of this application are: This application provides a method, apparatus, and electronic device for super-resolution reconstruction of significant wave height. The method involves pre-constructing a training dataset using real low-resolution multivariate wave field sequences at several consecutive times τ and real high-resolution significant wave height fields at each time τ. This dataset is then used to train an initial significant wave height reconstruction model. The model learns the wave propagation laws along the wave direction from the physical quantities of significant wave height, mean wave period, wave direction sine component, and wave direction cosine component in the low-resolution multivariate wave field sequence, and extracts the physical structures of the real low-resolution multivariate wave field sequence at different scales. Based on the learned propagation laws and extracted physical structures, the significant wave height field is reconstructed to obtain the target high-resolution significant wave height field. Then, based on the data difference between the target high-resolution significant wave height field and the real high-resolution significant wave height field, the model parameters of the significant wave height reconstruction model are adjusted according to the constraints of a preset loss function until the data difference satisfies the constraints of the loss function.

[0010] Thus, when it is necessary to reconstruct the effective wave height field of ocean waves in practice, the target effective wave height field reconstruction model can be used to quickly reconstruct the effective wave height field of ocean waves by using the low-resolution multivariate ocean wave field sequence of T consecutive time steps at the current time t according to the learned propagation law and feature extraction method, so as to obtain a high-resolution target high-resolution effective wave height field at the current time t.

[0011] This application's embodiments utilize a target wave effective wave height field reconstruction model, based on T time points from the current moment, including significant wave height, average period, wave direction sine component, and wave direction cosine component. This model, pre-trained and possessing knowledge of wave propagation along the wave direction and the ability to extract physical structures at different scales, accurately reconstructs the wave effective wave height field according to the known wave propagation laws and the extracted physical structures of the wave field at different scales. The method provided in this application overcomes the limitations of traditional methods that rely solely on single-frame data, failing to accurately depict the propagation and evolution of the wave field over time. It also overcomes the shortcomings of traditional methods that only apply single-scale features to construct the effective wave field, failing to capture detailed information within the wave field. Therefore, the resolution of the wave effective wave height field constructed using the method provided in this application is significantly improved. Attached Figure Description

[0012] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This paper presents a schematic flowchart of a super-resolution reconstruction method for ocean wave effective wave height provided in this application. Figure 2 This paper illustrates a model training process for the effective wave height field reconstruction model of ocean waves provided in this application. Figure 3 This paper shows a schematic diagram of a model structure for the ocean wave effective wave height field reconstruction model provided in this application; Figure 4 This paper illustrates a data processing flowchart for the ocean wave effective wave height field reconstruction model provided in this application. Figure 5 This paper shows a schematic diagram of the structure of an ultra-high resolution reconstruction device for ocean wave significant wave height provided in this application. Figure 6 A structural block diagram of an exemplary electronic device that can be used to implement embodiments of this application is shown. Detailed Implementation

[0013] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While some embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this application. It should be understood that the drawings and embodiments of this application are for illustrative purposes only and are not intended to limit the scope of protection of this application.

[0014] It should be understood that the steps described in the method embodiments of this application may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this application is not limited in this respect.

[0015] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc., mentioned in this application are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0016] It should be noted that the terms "a" and "a plurality of" used in this application are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0017] Before describing the method, apparatus, and electronic equipment for super-resolution reconstruction of ocean wave significant wave height provided in this application, a brief description of commonly used ocean wave significant wave height field reconstruction methods is provided below: As described in the background art, due to the limitations of satellite observation data and the spatial resolution of numerical simulations, the acquisition cost of high-resolution SWH data is high. Traditional methods use interpolation methods or deep learning-based image super-resolution methods to improve the resolution of SWH, but they are difficult to recover high-frequency details, and the SWH reconstruction results are often overly smoothed.

[0018] Traditional interpolation methods, such as bicubic interpolation and radial basis function interpolation (RBF), upsample by using known local weighted relationships between pixels or grid points. They are simple to implement and computationally efficient, but they are essentially mathematical fitting methods. They are difficult to recover true high-frequency details, and the reconstruction results are usually relatively smooth, with large errors in peaks, troughs and high gradient regions.

[0019] Deep learning-based image super-resolution methods include EDSR (Enhanced Deep Residual Networks for Single Image Super-Resolution), SwinIR (Swin Transformer for Image Restoration), and CAT (Cross Aggregation Transformer). EDSR enhances reconstruction capabilities through deep residual convolutional networks; SwinIR introduces the Swin Transformer into super-resolution tasks, modeling long-range dependencies through hierarchical structures and shifting window mechanisms; CAT further introduces a cross-scale attention mechanism, demonstrating strong performance in image super-resolution. While these three techniques excel in natural images, most are still based on single-time, single-image input, primarily targeting visual texture reconstruction rather than ocean data with significant physical and spatiotemporal evolution characteristics, such as wave fields.

[0020] Existing deep learning super-resolution methods are primarily designed for single-moment wave images, and their application to wave significant wave height reconstruction tasks remains significantly insufficient. Specifically, the closest implementation to this application directly applies existing single-moment super-resolution networks to wave significant wave height field reconstruction, such as Transformer or Convolutional Neural Network (CNN) type super-resolution methods that take a single frame of wave height field as input and the current high-resolution wave height field as output. While these methods can outperform traditional interpolation methods to some extent, they still suffer from the following significant limitations: First, the single-moment method cannot effectively utilize the spatiotemporal evolution information of the wave field. Specifically, waves exhibit patterns such as propagation along specific directions, continuous periodic evolution, and gradual energy transfer. Reconstructing based solely on the spatial distribution of a single frame makes it difficult to fully recover true high-resolution details.

[0021] Secondly, existing methods typically use only single-scale features or default to using deep features from the last stage for reconstruction, making it difficult to simultaneously take into account multi-scale structures such as local high-frequency fluctuations, mesoscale wave groups, and large-scale background surges.

[0022] Furthermore, existing methods typically only include wave height information as input, without explicitly utilizing physical variables such as wave period, wave direction sine component, and wave direction cosine component, making it difficult to establish the coupling relationship between wave height, period, and wave direction. In addition, existing methods generally lack physical constraints for ocean field reconstruction tasks, which can easily lead to problems such as negative wave height, unreasonable extrema, spatial gradient distortion, and overall energy inconsistency.

[0023] In view of the above limitations, this application provides a method, apparatus, and electronic device for super-resolution reconstruction of significant wave height. Specifically, in a first aspect, this application provides a method for super-resolution reconstruction of significant wave height, applicable to any electronic device capable of reconstructing significant wave height. The significant wave height field reconstruction model can be trained in any electronic device with model training capabilities. Electronic devices include, but are not limited to, personal mobile terminals, computers, supercomputers, industrial computers, or servers, etc.

[0024] In some possible embodiments, it may be as follows Figure 1 As shown, through steps S11 and S12, the significant wave height field of the target ocean waves is reconstructed using the significant wave height field reconstruction model: S11. Obtain the low-resolution multivariate wave field sequence for T consecutive time points from the current time t, wherein the low-resolution multivariate field sequence at each time point includes the following physical variables: significant wave height, average wave period, wave direction sine component and wave direction cosine component. S12. The target wave significant wave height field reconstruction model is invoked to reconstruct the wave significant wave height field based on the low-resolution multivariate wave field sequence of the continuous T time periods, so as to obtain the target high-resolution wave significant wave height field at the current time t.

[0025] In some possible embodiments, it may be as follows Figure 2 As shown, the target wave significant wave height field reconstruction model can be obtained through training steps S21-S24: S21. Construct a training dataset using the real low-resolution multivariate wave field sequence of several consecutive times τ and the real high-resolution effective wave height field of each time τ. S22. Input the training dataset into the initial wave effective wave height field reconstruction model. The wave effective wave height field reconstruction model learns the wave propagation law along the wave direction and extracts the physical structure of the real low-resolution multivariate wave field sequence at different scales. Based on the learned propagation law and the extracted physical structure of the real low-resolution multivariate wave field sequence at different scales, the wave effective wave height field is reconstructed to obtain the target high-resolution wave effective wave height field at time τ. S23. Based on the data difference between the target high-resolution wave significant wave height field and the real high-resolution wave significant wave height field at time τ, adjust the model parameters of the wave significant wave height field reconstruction model until the data difference satisfies the preset loss function constraint. S24. The wave effective wave height field reconstruction model corresponding to the data difference satisfying the preset loss function constraint is determined as the target wave effective wave height field reconstruction model.

[0026] This application embodiment learns the wave propagation law along the wave direction from the significant wave height, mean wave period, and wave direction (i.e., wave direction sine component and wave direction cosine component) in a low-resolution multivariate wave field sequence of T consecutive time moments by training a wave significant wave height field reconstruction model, and extracts the physical structure of the real low-resolution multivariate wave field sequence at different scales. This is to improve the reconstruction quality of the spatial resolution of the wave significant wave height field at the current time by utilizing historical information, and to display the physical coupling relationship between significant wave height, wave direction, and period during the reconstruction of the wave significant wave height field, which can help obtain a high-resolution wave significant wave height field that conforms to actual wave propagation.

[0027] It is understood that the wave significant wave height field reconstruction model constructed in this application is the core key to achieving high-resolution reconstruction of the wave significant wave height field. This reconstruction model is divided into two stages: a training stage and an application stage. In the training stage, the model is trained using sample data to learn the physical coupling relationship between significant wave height, wave direction, and period. This allows the model to learn the propagation law of waves along the wave direction and extract the physical structure of waves at different sizes. During training, the model parameters are continuously adjusted based on the difference between the target high-resolution wave significant wave height field output by the model and the actual high-resolution wave significant wave height field, so that the accuracy of the trained wave significant wave height field reconstruction model approaches that of the actual high-resolution wave significant wave height field.

[0028] In the application phase, a high-precision wave significant height field reconstruction model, trained from prior training, is used to directly reconstruct a high-resolution wave significant height field based on a low-resolution multivariate wave field sequence of T consecutive time intervals (including the current time t). Leveraging the learned propagation patterns and feature extraction capabilities, the model rapidly reconstructs a high-resolution wave significant height field. The method provided in this application overcomes the limitations of traditional methods that rely solely on single-frame data, failing to accurately depict the propagation and evolution of the wave field over time. Furthermore, it overcomes the shortcomings of traditional methods that only apply single-scale features to construct the wave significant height field, thus failing to capture the detailed information within the wave field. Therefore, the resolution of the wave significant height field constructed using the method provided in this application is significantly improved.

[0029] The following will provide a detailed explanation of steps 11, S12, and S21 to S24 above, with specific examples: In this embodiment, the effective wave height field refers to the joint distribution of effective wave height, mean wave period, and wave direction components at various locations within a spatial grid of the target sea area at a single observation time, characterizing the instantaneous physical state of the wave field at that moment. Specifically, at a specific time t, after spatially discretizing the target sea area, the two-dimensional distribution field formed by the values ​​of the effective wave height (SWH) and its related physical variables (mean wave period, wave direction sine / cosine components) at each grid point is the effective wave height field. It can be understood that the effective wave height field is the spatial distribution of wave energy in a certain sea area at a certain moment.

[0030] A multivariate wave field sequence is a time-ordered sequence containing four physical variables: effective wave height (SWH value), mean wave period (Tp), wave direction sinusoidal component (sinθ), and wave direction cosine component (cosθ). The mean wave period is the average time required for one complete wave cycle, measured in seconds (s). More precisely, it usually refers to the spectral peak period (Tp), representing the period corresponding to the frequency where wave energy is most concentrated. Wave direction refers to the direction of wave propagation, typically defined as the direction of wave arrival or departure. The wave direction cosine component (cosθ) and wave direction sinusoidal component (sinθ) represent the projected components of wave energy flux in the east-west and north-south directions, respectively.

[0031] In this application, since the effective wave height field reconstruction model is divided into a training phase and an application phase, for ease of distinction, the time corresponding to the application phase is named time t, and the time corresponding to the model training phase is named time τ. Both refer to a specific point in time, but t represents the current time, and τ represents a historical time. Step S11 acquires a low-resolution multivariate field sequence of T consecutive times from the current time t. Step S21 acquires a true low-resolution multivariate field sequence of T consecutive times from a historical time τ. Here, T times represents T acquisition times (the acquisition time depends on the acquisition period of the physical quantity, i.e., the period arrival time corresponds to one acquisition time). T consecutive times from the current time t refers to a total of T acquisition times including the current time t, and T consecutive times from time τ refers to a total of T acquisition times including time τ. That is, the acquired field sequence includes the low-resolution multivariate wave field corresponding to a total of T acquisition times, and in addition to the wave field corresponding to the current time t (or time τ), it also includes T-1 wave fields.

[0032] Here, the true low-resolution multivariate wave field sequence for T consecutive moments from the current time t refers to the true low-resolution multivariate wave field sequence for a total of T moments, tracing back T-1 acquisition moments from the current time t. Similarly, the true low-resolution multivariate wave field sequence for T consecutive moments from time τ refers to the true low-resolution multivariate wave field sequence for a total of T moments, tracing back T-1 acquisition moments from the time τ node.

[0033] Here, the low-resolution multivariable field refers to a field with a resolution below a certain threshold (this threshold can be flexibly set according to actual needs) that includes four types of physical variables: significant wave height, average wave period, wave direction (through the sine and cosine components of the wave direction). The low-resolution multivariable field sequence refers to a sequence of low-resolution multivariable fields at T time points ordered by time. As one implementation method, this low-resolution multivariable field sequence can be denoted as a sequence. X t = ,in This indicates the target time at which reconstruction is currently needed. Among them, C represents the number of input physical variables. It refers to the grid size in the significant wave height field of ocean waves.

[0034] Assuming the target wave significant wave height field reconstruction model trained in this application is considered as a black box, then the target wave significant wave height field reconstruction model according to the functional expression... In model parameters Under the influence of low-resolution multivariable field sequences Based on the mapping of function f, the target high-resolution effective wave height field at the current time t is reconstructed. : Formula (1); in, Indicates time Low-resolution multivariable wave field, For the number of input variables, This refers to the spatial grid size; This represents the high-resolution significant wave height field at the current time t, output by the ocean wave significant wave height field reconstruction model. During the model training phase, real high-resolution ocean wave significant wave height fields are used... As a reference result for the reconstruction model of the significant wave height field of ocean waves, the model continuously learns from the low-resolution multivariate ocean wave field sequence the significant wave height, mean wave period, and wave direction (through the sine and cosine components of the wave direction) and the true high-resolution significant wave height field of ocean waves. The causal relationship between them is established, thus enabling the application of the model of the significant wave height field in the low-resolution multivariate wave field sequence based on the learned significant wave height, mean wave period, and wave direction (through the wave direction sine and cosine components) and the true high-resolution significant wave height field. The causal relationship between them is used to reconstruct a high-resolution effective wave height field at time t based on the significant wave height, mean wave period, and wave direction (via the sine and cosine components of the wave direction) from the input low-resolution multivariate wave field sequence. And output it.

[0035] As an example, suppose T This involves inputting the low-resolution multivariate wave fields at the current time t and the previous four historical times, for a total of five consecutive time points, representing the low-resolution multivariate wave fields. Assume... The four variables are the significant wave height VHM0, the average wave period VTM10, the sine component of the wave direction, and the cosine component of the wave direction; assuming the input space size is... The output space size is The method provided in this application can achieve Super-resolution reconstruction.

[0036] At this point, the input low-resolution multivariate ocean wave field can be represented as:

[0037] The output target high-resolution ocean wave significant wave height field can be expressed as:

[0038] in, The batch size for participating in the reconstruction of the significant wave height field of ocean waves is determined by the number of waves, such as reconstructing 10 significant wave height fields at once, where B=10.

[0039] In some possible embodiments, the model architecture of the ocean wave significant wave height field reconstruction model provided in this application can be as follows: Figure 3 As shown, it includes: a shallow feature extraction and temporal coding module, a multi-scale spatial feature extraction and fusion module, a hierarchical residual super-resolution reconstruction module, and a physical constraint training module. Specifically, the shallow feature extraction and temporal coding module includes a shallow feature extraction unit and a temporal coding unit; the multi-scale spatial feature extraction and fusion module includes a multi-scale spatial feature extraction unit and a multi-scale spatial feature fusion unit; and the hierarchical residual super-resolution reconstruction module includes a local residual connection unit and a global residual connection unit. The shallow feature extraction and temporal coding module is used to extract shallow features from the real low-resolution multivariate field sequence at time t (or time τ) for T consecutive times and to perform temporal coding on the extracted shallow features to obtain the first temporal feature at time t (or time τ).

[0040] The multi-scale spatial feature extraction and integration module is used to extract physical structure features at multiple scales based on the first temporal features at time t (or time τ), obtain shallow embedding features at time t (or time τ), and fuse the extracted features at multiple scales to obtain the target fused features.

[0041] The hierarchical residual super-resolution reconstruction module is used to perform local residual connections and global residual connections based on the shallow embedding features and target fusion features at time t (or time τ), and output the target high-resolution effective wave height field of ocean waves at time t (or time τ).

[0042] The physical constraint training module is used to calculate the loss based on the real high-resolution effective wave height field of ocean waves at each time t (or time τ) and the target high-resolution effective wave height field of ocean waves at time t (or time τ) output by the global residual connection unit according to the built-in multi-objective loss function. The module adjusts the model parameters according to the difference between the target high-resolution effective wave height field of ocean waves at time t (or time τ) and the real high-resolution effective wave height field of ocean waves until the target high-resolution effective wave height field of ocean waves at time t (or time τ) is close to or even equal to the real high-resolution effective wave height field of ocean waves.

[0043] It is understood that the wave significant wave height field reconstruction model trained in this application extracts a series of features from a low-resolution multivariate field sequence, then learns the wave propagation law along the wave direction and the physical structure at different scales based on the features, and then reconstructs the wave significant wave height field based on the learned propagation law and physical structure and other important detailed information. In some possible embodiments, when performing step S12 or S22, the following steps S22A and S22B may also be included: S22A. The shallow feature extraction unit performs local spatial encoding on the real low-resolution multivariate wave field sequence at time τ for T consecutive times, so as to map each physical variable to a unified shallow feature space. In this application, local spatial coding refers to independently performing a convolutional mapping on the low-resolution multivariate wave field at each historical moment in the low-resolution multivariate wave field sequence, transforming the heterogeneous effective wave height, mean wave period, and wave direction sine and cosine components from the original observation space to a feature space of uniform dimension. While eliminating the differences in physical dimensions, it extracts the physical coupling features between each grid and its neighborhood, providing a physically consistent input basis for subsequent learning of wave propagation laws.

[0044] As one implementation method, the entire data processing flow of the ocean wave significant wave height field reconstruction model constructed in this application can be as follows: Figure 4 As shown, a shallow feature extraction unit performs shallow feature extraction (local spatial coding) based on a low-resolution multivariate wave field sequence of T consecutive time moments X. t (or X) τ Mapped to a unified shallow feature space (or F) τ )middle.

[0045] Since the low-resolution multivariate wave field sequence contains four physical variables, its spatial dimension is: ,in, represents the spatial grid size of the wave field. This is achieved by analyzing the low-resolution multivariate wave field sequence X. t (or X) τ Local spatial encoding can be performed by convolutional mapping of the low-resolution multivariate wave field at each time step using a 3×3 convolution kernel, i.e., for each historical time step... First, perform low-resolution input on the four channels. Convolutional mapping yields shallow features at each time step. :

[0046] in, It includes significant wave height, average wave period, sine wave direction component, and cosine wave direction component; among which, during the model training phase, , representing any historical moment τ preceding a given historical moment in the sample data, within the T timeframes. During the model application phase... , represents any historical moment among the T moments before the current moment t.

[0047] The shallow features obtained after this convolution mapping The corresponding spatial resolution is: As an example, suppose Therefore, this shallow feature The corresponding spatial resolution is By using this shallow feature extraction unit with the aid of shallow convolution, it is possible to first perform local spatial encoding of multiple physical variables at each time step to obtain the shallow features corresponding to each time step. Then, variables with different dimensions and different physical meanings are uniformly mapped to a high-dimensional feature space (such as the 64-dimensional feature space mentioned above) to obtain the shallow feature space. (or F) τ ),in, Then, the shallow feature sequences in the obtained shallow feature space are... Input timing coding unit.

[0048] S22B. Based on the shallow feature sequence in the shallow feature space, the temporal coding unit generates a first temporal feature at time τ using a convolutional gate control memory mechanism. The first temporal feature is used to characterize the propagation law between wave height amplitude, wave propagation direction and wave period in the real low-resolution multivariate wave field sequence at time T consecutive before time τ.

[0049] In the model application phase, the temporal coding unit generates the first temporal feature at time t (which can be denoted as t) based on the shallow feature sequence in the shallow feature space and using a convolutional gate control memory mechanism. During the model training phase, the temporal coding unit is based on the shallow feature sequence in shallow feature control, and uses a convolutional gate control memory mechanism to generate the first temporal feature at time τ (which can be denoted as...). ).

[0050] As one implementation, the timing coding unit is a convolutional long short-term memory (ConvLSTM) unit. The ConvLSTM unit used in this application includes two layers, which are described here as follows: Indicates the layer number of the ConvLSTM. ;use Represents the time index in the input sequence, satisfying At that moment The input, hidden state, and memory unit are denoted as follows: , and .

[0051] The first layer input These are shallow features, i.e. Second layer input The hidden states output by the first layer at the same time ,Right now .

[0052] Based on the ConvLSTM gated update mechanism, an expression is used based on the above input. , will the current input Compared to the previous hidden state By splicing along the channel dimension, shallow splicing features are obtained. Furthermore, according to the input gate expression in the ConvLSTM gating update mechanism: Forget gate expression: Output gate expression: And candidate memory expressions: Solve separately , , and Then use the solution obtained , , and Update the expression according to the memory unit. Hidden state update expression Memory cells in ConvLSTM and the hidden states of each layer Update.

[0053] in, This is the input gate, used to control the writing of new information. This is the Forget Gate, used to control the retention of historical memories. This is an output gate used to control the output of memory information to the hidden state. Candidate memories.

[0054] For the sigmoid function, For the bias parameters of the input gate, For the bias parameters of the forget gate, For the bias parameters of the output gate, The bias parameters for candidate memories. It is the hyperbolic tangent function; This is element-wise multiplication; This indicates channel dimension splicing.

[0055] The four gated branches described above can be obtained by generating four sets of channels through a single convolution and then splitting them, or by using an equivalent multi-branch convolution. Since each of the two ConvLSTM layers has an input gate, a forget gate, an output gate, a candidate memory, a memory unit update, and a hidden state output, for the first historical moment... The initial hidden states and initial memory cells of the two ConvLSTM layers can be set to zero. Subsequently, the wave significant height field reconstruction model can be performed according to... The order is recursively processed time-by-time. After processing all... After a certain time, take the second layer of ConvLSTM at the current time. The final hidden state is used as the encoding of the first temporal feature, that is If H=W=8, then the spatial resolution of the first temporal feature at the current time t is: .

[0056] It is understood that the temporal coding unit (ConvLSTM) in this application does not directly solve the explicit wave dynamics governing equations, but instead learns the statistical mapping relationships related to wave propagation from continuous historical wave fields through a convolutional gating memory mechanism. Since the input variables include not only the significant wave height, but also the sine and cosine components of the mean wave period and wave direction, the network simultaneously obtains joint information of "wave height amplitude - propagation direction - period scale" in each grid point and its neighborhood.

[0057] For wave direction information, this application decomposes the angular wave direction into sine and cosine components, avoiding the discontinuity problem of the angular variable near (0° / 360°). The convolutional gating of ConvLSTM shares the convolution kernel between adjacent time steps and adjacent grids, thus enabling it to learn the pattern of wave height structure migrating to neighboring regions along the wave direction indication direction in historical time steps.

[0058] For periodic information, the mean wave period reflects the time scale of the wave system. Under the same wave height, different periods often correspond to different wavelengths, propagation speeds, and energy distribution characteristics. The memory cells of ConvLSTM retain historical stable components through forget gates and write current changing components through input gates, thus enabling a certain degree of distinction between the persistent large-scale swell background and the short-term changing local wind and wave structure.

[0059] Therefore, in this embodiment, the historical joint evolution relationship of wave height, period, and wave direction is encoded as the first temporal feature at the current moment. or This provides prior information that conforms to the continuity of wave propagation for the subsequent super-resolution reconstruction of the effective wave height field of ocean waves.

[0060] Furthermore, through the multi-scale spatial feature extraction and fusion module, the physical structure of the real low-resolution multivariate wave field sequence at different scales is extracted based on the first temporal feature at time τ. As one implementation method, it can be as follows: Figure 4 As shown, the multi-scale spatial feature extraction unit includes: a first-stage feature extractor, a second-stage feature extractor, a third-stage feature extractor, and a fourth-stage feature extractor. Based on this, when performing step S12 or step S22 to extract the physical structure of the real low-resolution multivariate wave field sequence at different scales, the following steps S22C-S22F may also be included: S22C. Extract the first-stage feature at time τ based on the first temporal feature at time τ using the first-stage feature extractor; the resolution of the first-stage feature is higher than the resolution of the first temporal feature. S22D: The first stage features are processed by the second stage feature extractor to reduce resolution and expand channels to obtain the second stage features at time τ. The resolution of the second stage features is lower than that of the first stage features, and the number of channels of the second stage features is higher than that of the first stage features. The number of channels is positively correlated with the range of sea area covered by the waves. S22E: The second-stage features are processed by the third-stage feature extractor to reduce resolution and expand channels, resulting in the third-stage features at time τ. The resolution of the third-stage features is lower than that of the second-stage features, and the number of channels of the third-stage features is higher than that of the second-stage features. S22F: The fourth-stage feature extractor reduces the resolution and expands the channels of the third-stage features to obtain the fourth-stage features at time τ. The resolution of the fourth-stage features is lower than that of the third-stage features, and the number of channels of the fourth-stage features is higher than that of the third-stage features.

[0061] In simple terms, this application processes the first temporal features at different scales (resolution + channels) using different stage feature extractors to obtain features with different spatial resolutions. Before executing step S22C, the multi-scale spatial feature extraction unit can perform convolutional mapping based on the first temporal features at time t (or time τ) to obtain shallow embedding features at time τ. Then, the first-stage feature extractor performs feature extraction based on these shallow embedding features to obtain the first-stage features. As one implementation, the multi-scale feature extraction unit can be a Swin Transformer hierarchical spatial encoder. By performing convolutional mapping on the first temporal features to obtain shallow embedding features, the input of the Swin Transformer hierarchical spatial encoder can be matched.

[0062] Specifically, the first temporal coding features output by ConvLSTM can be... pass Convolution mapping to the embedding space required by the Swin Transformer hierarchical spatial encoder In, that is, according to the expression Mapping yields shallow embedding features .

[0063] For example, the spatial resolution of the first temporal feature at the current time t is... For example, using this convolutional mapping method, the 64-channel first temporal feature can be mapped to a 96-channel shallow embedded feature, i.e. Thus, it is possible... Figure 4 As shown, this shallow embedding feature On one hand, it can be used as input to the Swin Transformer hierarchical spatial encoder to perform stage feature extraction at each stage. On the other hand, it can be retained to participate in subsequent local residual connections.

[0064] The multi-scale spatial feature extraction unit, corresponding to the Swin Transformer hierarchical spatial encoder, can extract multi-scale spatial features. This encoder includes feature extractors for four stages, each stage consisting of several Swin Transformer blocks. The encoder composed of Swin Transformer blocks corresponding to each stage is the stage feature extractor for that stage. Patch merging operations are used between adjacent stages to reduce spatial resolution and increase channel dimensions. As an example, the depth of the four stages can be set to... The number of heads for multi-head attention can be set to... .

[0065] The feature extraction process of the four stages corresponding to the first-stage feature extractor, the second-stage feature extractor, the third-stage feature extractor, and the fourth-stage feature extractor can be represented as follows:

[0066] Among them, the model application stage Represents the first time step t at the current time. Spatial features output at each stage, model training stage The Spatial characteristics output at each stage.

[0067] Input dimensions For example, the stages can form the following hierarchical characteristics: Phase 1: Maintaining the same spatial resolution as the shallow embedding features, the output features can be represented as... The corresponding spatial resolution is The first-stage features extracted in the first stage It mainly preserves the local peaks, troughs, spatial gradients, and detailed variations between adjacent grids.

[0068] The second stage: Building upon the features from the first stage, the resolution is reduced and the channel size is increased through Patch Merging, thus reducing the spatial dimensions to [specific value]. With the number of channels expanded to 192, the output characteristics can be represented as follows: The corresponding spatial resolution is The second-stage features extracted in the second stage Used to express the morphology of wave groups within a larger neighborhood.

[0069] The third stage: Building upon the features from the second stage, further reduction in resolution and expansion of channels are achieved through Patch Merging, thus reducing the spatial size to [specific value missing]. With the number of channels expanded to 384, the output characteristics can be expressed as follows: The corresponding spatial resolution is The third-stage features extracted in this third stage Used to express regional scale wave organization structure.

[0070] The fourth stage: Building upon the features from the third stage, further patch merging is used to reduce resolution and expand the number of channels, resulting in a coarser-scale global representation. The number of channels is expanded to 768, and the output features can be represented as follows: The corresponding spatial resolution is The fourth-stage features extracted in the fourth stage Used to encode large-scale background surge patterns and regional statistics within the input window.

[0071] In this embodiment, the scale division of each stage is not determined solely by the number of parameters, but by the changes in spatial resolution and receptive field within the hierarchical coding structure. The first stage maintains high spatial resolution, with each feature unit corresponding to a relatively small sea area, making it more suitable for representing local details. As Patch Merging progresses layer by layer, the spatial resolution of subsequent stages gradually decreases, but the original sea area corresponding to a single feature unit increases, enabling the aggregation of wave height, period, and direction information from a larger neighborhood. Therefore, the second and third stages are more suitable for representing wave group structures with a certain spatial extension, while the fourth stage is more suitable for representing the background swell pattern and regional statistical characteristics within the entire input window.

[0072] Thus, by dividing the space into stages and extracting features at different resolutions, the physical structure at different scales in the entire wave field sequence can be obtained based on this multi-scale spatial feature extraction unit. This provides information that can accurately reflect local details and thus provides a solid data foundation for the reconstruction of the effective wave height field of ocean waves.

[0073] Furthermore, in some possible embodiments, it may be as follows: Figure 4 As shown, the stage features extracted from each stage are obtained through the multi-scale spatial feature fusion unit. The target fusion feature X is obtained by performing fusion processing according to the steps of upsampling, normalization, concatenation and channel compression. t-fused (or X during the model training phase) τ-fused ).

[0074] In one implementation, the first-stage features, second-stage features, third-stage features, and fourth-stage features can be upsampled according to the spatial resolution corresponding to the first-stage features to obtain upsampled features corresponding to each stage. The upsampled features corresponding to each stage are then spliced ​​and fused to obtain target fused features. The target fused features are used to characterize the physical structure of the real low-resolution multivariate wave field sequence at the levels of local details, mesoscale wave groups, and large-scale background information.

[0075] Here, it is assumed that the set of stages that need to participate in the integration is as follows: That is, the first, second, third, and fourth stages mentioned above. Using the aforementioned size 8... Taking an 8-resolution example as an example, all four stages can be selected to participate in the fusion process, i.e. The multi-scale fusion process includes the following steps: First, the characteristics of each stage According to the expression Unified upsampling to shallow embedding features The same Resolution, to obtain upsampled features .

[0076] Second, through The convolution according to the expression Upsampling features corresponding to different stages Channel normalization is performed to uniformly map the features to 96 channels of normalized multi-scale features. .

[0077] Third, normalized multi-scale features According to the expression By concatenating the features along the channel dimension, we obtain concatenated multi-scale features. .

[0078] Finally, through Convolution according to the expression Learning multi-scale features of different splicing The combination relationships between them are analyzed, and the number of channels is compressed back to 96 dimensions to obtain the final fused features. : Through the fusion process provided in this application, features at different scales can be aligned at the same low-resolution grid location through upsampling, and then through... Convolution eliminates dimensionality inconsistencies caused by varying channel numbers at different stages, and further improves dimensionality through channel concatenation and... Convolution enables the network to adaptively learn the combined weights between local details, mesoscale wave groups, and large-scale background information within each grid neighborhood. Since the actual effective wave height field of ocean waves is simultaneously affected by local wave crests and troughs, neighboring wave groups, and regional background swells, the embodiment of this application is beneficial for recovering a high-resolution wave height field that preserves local details while conforming to the overall regional situation.

[0079] Furthermore, such as Figure 4 As shown, the local residual connection unit in the hierarchical residual super-resolution reconstruction module is based on the shallow embedded feature X0 and the target fusion feature. By performing local residual connections, a single-channel high-frequency correction term is obtained at time t (or time τ), where the high-frequency correction term can be denoted as... .

[0080] As one implementation, the target fusion feature and the shallow embedding feature can be connected through the local residual connection unit, and deep feature fusion can be performed to obtain the deep feature at time τ; upsampling pre-convolution and sub-pixel operation are performed on the deep feature at time τ to obtain the single-channel high-frequency correction term at time τ.

[0081] Specifically, it can be done according to the expression. The target fusion features are concatenated with the shallow embedding features. Among them, The target fusion features are obtained after multi-scale fusion. For ConvLSTM output through The 96-channel shallow embedding features obtained by convolution. Through local residual connections, local spatial details can be prevented from being lost during deep Swin Transformer encoding and multi-scale fusion, so that the fused features can retain local wave height variation information on the original low-resolution grid while obtaining large-scale context.

[0082] After adding the fused features and the shallow embedding features, then according to the expression After several Convolution and LeakyReLU nonlinear activation are used to integrate deep features, resulting in deep features at time t (or time τ). .

[0083] Furthermore, in this deep feature Based on the expression Perform pre-upsampling convolution and subpixel PixelShuffle operations to scale up to the target high-resolution size. Using the aforementioned T=5... In the super-resolution embodiment, the PixelShuffle subpixel operation can... Feature magnification Features. Then, expressions are used. The features obtained from the pre-upsampling convolution and sub-pixel PixelShuffle operations Perform convolution mapping again to obtain the single-channel high-frequency correction term. .in, This represents the high-frequency residual correction term learned by the backbone network, which is mainly used to supplement local peaks, troughs and spatial gradient details outside the low-frequency baseline.

[0084] Furthermore, the first time-series feature is amplified by bicubic interpolation using a global residual connection unit to obtain the low-frequency baseline of a single channel at time τ. The high-frequency correction term of the single channel at time τ is then added to the low-frequency baseline of the single channel at time τ to obtain the target high-resolution effective wave height field of the ocean wave at time τ.

[0085] One input to the global residual connection unit bypasses the deep spatial encoding of the Swing Transformer and directly inputs the low-resolution first temporal features output by the ConvLSTM. By examining this first time-series feature Perform bicubic interpolation to magnify to the target resolution, then follow the expression. through The convolutional mapping is a bypass branch of the single-channel waveform height baseline, which yields the first temporal features. Low-frequency baseline obtained by upsampling The low-frequency baseline The spatial variation is relatively smooth, which can well preserve the large-scale background distribution of the current wave field. Furthermore, this global residual connection unit follows the expression... By adding the single-channel high-frequency correction term to the global residual low-frequency baseline, the reconstructed target high-resolution effective wave height field can be obtained. .

[0086] In this embodiment, based on the concept of residual super-resolution, a high-resolution wave height field can be decomposed into a low-frequency background field and high-frequency detail correction. The low-frequency baseline provides a stable large-scale wave height distribution, while the backbone network focuses on learning the fine-scale structures that are difficult to recover from the low-frequency baseline. This reduces the difficulty of generating a complete high-resolution wave height field from scratch, improving training stability and the quality of reconstructed details.

[0087] In this embodiment of the application, to improve the physical rationality of the reconstruction results, a multi-objective constraint loss function is built into the physical constraint training module. ; The multi-objective loss function adjusts the model parameters of the wave significant wave height reconstruction model in reverse based on the data difference between the target high-resolution wave significant wave height field at time τ and the actual high-resolution wave significant wave height field at time τ, thereby adjusting the model accuracy of the wave significant wave height reconstruction model.

[0088] Each loss term is defined as follows: in, The reconstruction loss function is expressed as follows: ,in, The reconstruction loss function is used to directly constrain the target high-resolution effective wave height field. Wave height in the image and the actual high-resolution wave height field The error between them is used to improve the overall numerical accuracy. These are the weighting coefficients for the reconstruction loss function.

[0089] The SSIM structural similarity loss function is expressed as follows: ,in, The SSIM structural similarity loss function is used to constrain the structural similarity between the target high-resolution effective wave height field and the true high-resolution wave height field in terms of local mean, variance, and covariance, in order to maintain the spatial structure and regional distribution morphology of the wave height field. These are the weight coefficients of the SSIM structural similarity loss function.

[0090] The gradient consistency loss function is expressed as follows: The gradient consistency loss function This is used to constrain the consistency of the spatial gradient between the effective wave height field and the true high-resolution wave height field of the target high-resolution ocean waves in the x and y directions, and to maintain the spatial variation characteristics of wave crest boundaries, wave trough transition zones, and high-gradient regions. These are the weight coefficients of the gradient consistency loss function.

[0091] The spatial smoothness constraint loss function is expressed as follows: This is used to constrain the wave height variation of the target high-resolution ocean wave effective wave height field between adjacent grids, which can suppress unreasonable spikes, noise, and local oscillations generated by network reconstruction. These are the weighting coefficients of the spatial smoothness constraint function.

[0092] The physical extent constraint loss function is expressed as follows: This is used to constrain the marine statistical boundary between the target high-resolution wave height field and the true high-resolution wave height field. Specifically, based on the physical fact that significant wave heights are non-negative and that there is a reasonable upper limit to the significant wave height in the target sea area, negative wave heights and abnormally large values ​​can be penalized. In other sea areas or other datasets, the upper limit threshold can be adjusted according to the historical statistical range. These are the weighting coefficients of the physical range constraint loss function.

[0093] Let the energy consistency constraint loss function be expressed as follows: This is used to constrain the global energy conservation between the target high-resolution ocean wave effective wave height field and the true high-resolution wave height field. Wherein, This represents the spatial average. Since wave energy is approximately correlated with the square of the effective wave height, this energy consistency constraint loss function can be used to ensure that the overall energy level of the reconstructed result remains consistent with the background energy level reflected by the low-resolution input. These are the weighting coefficients of the energy consistency constraint loss function.

[0094] in, , , , , and The values ​​can be flexibly set according to practical experience. As an example, the weight coefficients of each loss function can be set as follows: The weights mentioned above can also be adjusted based on the dataset, sea area, and training phase.

[0095] It's important to note that the various loss functions don't individually update the parameters of a single fixed module. Instead, they collectively update the learnable parameters in the shallow convolution, ConvLSTM, Swin Transformer, multi-scale fusion, upsampling reconstruction, and residual branches through backpropagation of the total loss. The difference between the different loss functions lies in the angle of constraint they impose on the output. Specifically: The main improvement is in pixel-level numerical accuracy. Mainly maintain spatial structural similarity, Mainly maintain the local gradient and the boundaries of peaks and troughs. The main purpose is to suppress unreasonable high-frequency oscillations. The main focus is to avoid high negative wave heights and abnormal extreme values. The main constraint is the overall wave energy level. Therefore, this multi-objective loss function achieves a balance between numerical accuracy, spatial structure, local gradient, smoothness, reasonable range, and energy consistency through weighted summation, so that the reconstruction results not only approach the true high-resolution field in terms of numerical error, but also better conform to the physical constraints of the effective wave height field of ocean waves.

[0096] In this embodiment, the target wave significant wave height field model trained can be applied to any scenario requiring operations based on wave significant wave height, such as marine monitoring, maritime navigation, and coastal engineering. In such scenarios, the wave significant wave height field data obtained by traditional wave significant wave height analysis methods for the T consecutive moments preceding the current time t can be obtained as a low-resolution multivariate wave field sequence by executing step S11. Alternatively, data collected by image acquisition devices, sensors, satellites, or other acquisition devices with wave significant wave height data set up along the coastline can also be used as a low-resolution multivariate wave field sequence. The specific acquisition method is not strictly limited in this application.

[0097] The target wave significant wave height field reconstruction model can be deployed in the cloud or in an electronic device that performs wave significant wave height field reconstruction. If deployed in the cloud, when executing the method provided in this application, the pre-trained target wave significant wave height field reconstruction model can be invoked from the cloud to execute the above step S12 via cloud invocation. If the target wave significant wave height field reconstruction model is deployed in an electronic device that performs the above steps S11 and S12, when executing step S12, the pre-trained target wave significant wave height field reconstruction model can be invoked locally to execute the above step S12 via local invocation.

[0098] Compared to traditional methods for reconstructing the effective wave height field of ocean waves, this application has the following core differences: Key Difference 1: This application possesses a causal temporal super-resolution modeling mechanism. Unlike traditional single-moment super-resolution methods, this application uses low-resolution multivariate wave field sequences from T consecutive historical moments as input, leveraging historical information to improve the spatial resolution reconstruction quality of the current moment.

[0099] Core Difference 2: This application is based on ConvLSTM wave evolution modeling. This application uses a convolutional gated memory structure to model the propagation direction, periodic evolution, and energy variation of the wave field, enabling temporal information fusion while preserving the spatial structure.

[0100] Core Difference 3: This application provides multi-scale feature fusion for ocean wave fields. In this application, multi-scale features are extracted hierarchically using SwinTransformer, and multi-scale wave structure information is fused through upsampling, (1×1) convolution normalization, concatenation, and (3×3) convolution, which helps to obtain the physical structure at different scales in the ocean wave field sequence.

[0101] Core Difference 4: This application provides the idea of ​​hierarchical residual super-resolution reconstruction, and introduces local residual and global residual connections, which can enable the network to restore high-frequency details while preserving a stable low-frequency baseline.

[0102] Core Difference 5: The input used in this application is a multivariate physically coupled input. Specifically, the input data used in this application includes not only wave height, but also the average wave period, the sine component of the wave direction, and the cosine component of the wave direction, thereby explicitly utilizing the physical coupling relationship between wave height, wave direction, and period to construct a higher resolution effective wave height field.

[0103] Core Difference 6: This application sets up a multi-objective constraint loss function for ocean physics. Based on the conventional reconstruction loss function, it introduces gradient consistency, spatial smoothness, range constraint and energy consistency constraint to improve the physical consistency of the reconstruction results.

[0104] Data verification revealed that the target high-resolution wave significant wave height field reconstructed using the wave significant wave height field reconstruction model provided in this application can overcome the limitation of single-moment super-resolution methods in utilizing insufficient historical information. Specifically, existing methods are mostly based on single-frame input, making it difficult to characterize the propagation and evolution of the wave field over time. The embodiments of this application, by modeling continuous historical moment information using ConvLSTM, reduce the RMSE (Root Mean Square Error) of the reconstructed target high-resolution wave significant wave height field from 0.0245m to 0.0185m, a reduction of 24.5%. This demonstrates that the propagation direction, periodic evolution, and energy change information in historical moments can significantly enhance the reconstruction capability at the current moment.

[0105] Furthermore, the target high-resolution wave significant wave height field reconstructed by the wave significant wave height field reconstruction model provided in this application can improve the characterization ability of multi-scale spatial structures. Specifically, existing methods typically only use single-scale features or default to using deep features from the final stage, making it difficult to simultaneously take into account local details, mesoscale wave groups, and large-scale background structures. In the embodiments of this application, after further introducing multi-scale fusion, the RMSE is further reduced from 0.0185m to 0.0134m, an additional reduction of 27.6%, indicating that the method provided in this application can more fully recover multi-scale wave structure information.

[0106] Furthermore, the target high-resolution wave significant wave height field reconstructed by the wave significant wave height field reconstruction model provided in this application can enhance the multivariate physical coupling modeling capability. Specifically, compared with using only a single variable SWH input, the multivariate input scheme provided in this application can reduce RMSE from 0.0151m to 0.0134m, MAE (Mean Absolute Error) from 0.0089m to 0.0079m, and PSNR (Peak Signal-to-Noise Ratio) from 36.97dB to 37.89dB. This indicates that in the method of this application, the joint modeling of information such as wave height, period, and wave direction can effectively improve the reconstruction accuracy.

[0107] Furthermore, the target high-resolution wave significant wave height field reconstructed by the wave significant wave height field reconstruction model provided in this application can improve the physical rationality of the reconstruction results. Specifically, the complete physical constraint loss reduces the RMSE from 0.0162m to 0.0134m, a reduction of 17.3%, compared to the baseline configuration using only the L1 reconstruction loss function. It can also better constrain the spatial smoothness, value range rationality, and overall energy consistency of the results.

[0108] Furthermore, the target high-resolution wave significant wave height field reconstructed by the wave significant wave height field reconstruction model provided in this application exhibits overall superior reconstruction performance compared to existing comparative methods. For instance, in a 5x super-resolution experiment in the Black Sea, the embodiment of this application achieves an RMSE of 0.0134m and a PSNR of 37.89dB. Compared to the single-time method CAT, the RMSE is reduced by 23.9%, the MAE is reduced by 28.8%, and the PSNR is increased by 2.69dB, all demonstrating that the method provided in this application has significant advantages in both accuracy and stability.

[0109] Secondly, this application provides a super-resolution reconstruction device for the effective wave height of ocean waves, wherein, as Figure 5 As shown, the device 50 includes: The acquisition module 501 is used to acquire a low-resolution multivariate wave field sequence for T consecutive time points from the current time t, wherein the low-resolution multivariate field sequence at each time point includes the following physical variables: significant wave height, mean wave period, wave direction sine component and wave direction cosine component. Reconstruction module 502 is used to call the target wave significant wave height field reconstruction model to reconstruct the wave significant wave height field based on the low-resolution multivariate wave field sequence of the continuous T time periods, so as to obtain the target high-resolution wave significant wave height field at the current time t. Model training module 503 is used to train models at several time points. The training dataset is constructed by using a sequence of real low-resolution multivariate ocean wave fields at consecutive T time points and real high-resolution effective wave height fields at each time point τ. The training dataset is input into the initial wave effective wave height field reconstruction model. The wave effective wave height field reconstruction model learns the wave propagation law along the wave direction and extracts the physical structure of the real low-resolution multivariate wave field sequence at different scales. Based on the learned propagation law and the extracted physical structure of the real low-resolution multivariate wave field sequence at different scales, the wave effective wave height field is reconstructed to obtain the target high-resolution wave effective wave height field at time τ. Based on the data difference between the target high-resolution wave significant height field and the actual high-resolution wave significant height field at time τ, the model parameters of the wave significant height field reconstruction model are adjusted until the data difference satisfies the preset loss function constraint. The wave effective wave height field reconstruction model corresponding to the data difference satisfying the preset loss function constraint is determined as the target wave effective wave height field reconstruction model.

[0110] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in this application comply with relevant laws and regulations and do not violate public order and good morals.

[0111] The names of the messages or information exchanged between multiple devices in the embodiments of this application are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0112] Thirdly, exemplary embodiments of this application also provide an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the electronic device to perform a method according to an embodiment of this application.

[0113] An exemplary embodiment of this application also provides a non-transitory computer-readable storage medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform a method according to an embodiment of this application.

[0114] An exemplary embodiment of this application also provides a computer program product, including a computer program, wherein, when executed by a computer's processor, the computer program is used to cause the computer to perform a method according to an embodiment of this application.

[0115] refer to Figure 6The present invention describes a structural block diagram of an electronic device 600 that can serve as a server or client of this application, which is an example of a hardware device that can be applied to various aspects of this application. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the application described and / or claimed herein.

[0116] like Figure 6 As shown, the electronic device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM 602) or a computer program loaded from a storage unit 608 into a random access memory (RAM 603). The RAM 603 may also store various programs and data required for the operation of the electronic device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output interface (I / O interface 605) is also connected to the bus 604.

[0117] Multiple components in electronic device 600 are connected to I / O interface 605, including: input unit 606, output unit 607, storage unit 608, and communication unit 609. Input unit 606 can be any type of device capable of inputting information to electronic device 600. Input unit 606 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of electronic device. Output unit 607 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 608 may include, but is not limited to, disks and optical discs. Communication unit 609 allows electronic device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, wireless communication transceivers, and / or chipsets, such as Bluetooth™ devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0118] The computing unit 601 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above. For example, in some embodiments, the aforementioned wave significant wave height super-resolution reconstruction method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 600 via ROM 602 and / or communication unit 609. In some embodiments, the computing unit 601 can be configured to perform the aforementioned wave significant wave height super-resolution reconstruction method by any other suitable means (e.g., by means of firmware).

[0119] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0120] In the context of this application, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0121] As used in this application, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0122] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0123] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0124] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other.

Claims

1. A super-resolution reconstruction method for the effective wave height of ocean waves, characterized in that, The method includes: Obtain a low-resolution multivariate wave field sequence for T consecutive time points from the current time t, wherein the low-resolution multivariate field sequence at each time point includes the following physical variables: significant wave height, mean wave period, wave direction sine component, and wave direction cosine component; The target wave significant wave height field reconstruction model is invoked to reconstruct the wave significant wave height field based on the low-resolution multivariate wave field sequence of the continuous T time periods, so as to obtain the target high-resolution wave significant wave height field at the current time t. The target wave significant wave height field reconstruction model is pre-trained using the following method: A training dataset is constructed using a sequence of real low-resolution multivariate ocean wave fields at several times τ and real high-resolution effective wave height fields at each time τ. The training dataset is input into the initial wave effective wave height field reconstruction model. The wave effective wave height field reconstruction model learns the wave propagation law along the wave direction and extracts the physical structure of the real low-resolution multivariate wave field sequence at different scales. Based on the learned propagation law and the extracted physical structure of the real low-resolution multivariate wave field sequence at different scales, the wave effective wave height field is reconstructed to obtain the target high-resolution wave effective wave height field at time τ. Based on the data difference between the target high-resolution wave significant height field and the actual high-resolution wave significant height field at time τ, the model parameters of the wave significant height field reconstruction model are adjusted until the data difference satisfies the preset loss function constraint. The wave effective wave height field reconstruction model corresponding to the data difference satisfying the preset loss function constraint is determined as the target wave effective wave height field reconstruction model.

2. The method according to claim 1, characterized in that, The defined wave height field reconstruction model includes: a shallow feature extraction and temporal coding module, and a multi-scale spatial feature extraction and fusion module; the shallow feature extraction and temporal coding module includes: a shallow feature extraction unit and a temporal coding unit. The process of learning the wave propagation law along the wave direction and extracting the physical structure of the real low-resolution multivariate wave field sequence at different scales using the defined wave height field reconstruction model includes: The shallow feature extraction unit performs local spatial encoding on the real low-resolution multivariate wave field sequence at time τ for T consecutive time points, so as to map each physical variable to a unified shallow feature space. Based on the shallow feature sequence in the shallow feature space, the temporal coding unit generates a first temporal feature at time τ using a convolutional gate control memory mechanism. The first temporal feature is used to characterize the propagation law between wave height amplitude, wave propagation direction and wave period in the real low-resolution multivariate wave field sequence at time T consecutive times before time τ. The multi-scale spatial feature extraction and fusion module extracts the physical structure of the real low-resolution multivariate wave field sequence at different scales based on the first temporal feature at time τ.

3. The method according to claim 2, characterized in that, The multi-scale spatial feature extraction and fusion module includes: a multi-scale spatial feature extraction unit and a multi-scale spatial feature fusion unit. The multi-scale spatial feature extraction unit includes: a first-stage feature extractor, a second-stage feature extractor, a third-stage feature extractor, and a fourth-stage feature extractor. The first temporal feature extraction based on time τ to extract the physical structure of the real low-resolution multivariate ocean wave field sequence at different scales includes: The first-stage feature extractor extracts the first-stage feature at time τ based on the first temporal feature at time τ; the resolution of the first-stage feature is higher than that of the first temporal feature. The second-stage feature extractor reduces the resolution and expands the channels of the first-stage features to obtain the second-stage features at time τ. The resolution of the second-stage features is lower than that of the first-stage features, and the number of channels of the second-stage features is higher than that of the first-stage features. The number of channels is positively correlated with the range of sea area covered by the waves. The third-stage feature extractor reduces the resolution and expands the channels of the second-stage features to obtain the third-stage features at time τ. The resolution of the third-stage features is lower than that of the second-stage features, and the number of channels of the third-stage features is higher than that of the second-stage features. The fourth-stage feature extractor reduces the resolution and expands the channels of the third-stage features to obtain the fourth-stage features at time τ. The resolution of the fourth-stage features is lower than that of the third-stage features, and the number of channels of the fourth-stage features is higher than that of the third-stage features.

4. The method according to claim 3, characterized in that, The method further includes: The multi-scale spatial feature fusion unit upsamples the first-stage, second-stage, third-stage, and fourth-stage features according to the spatial resolution corresponding to the first-stage features, respectively, to obtain the upsampled features corresponding to each stage. The upsampled features corresponding to each stage are then spliced ​​and fused to obtain the target fused features. The target fused features are used to characterize the physical structure of the real low-resolution multivariate wave field sequence at the levels of local details, mesoscale wave groups, and large-scale background information.

5. The method according to claim 4, characterized in that, The effective wave height field reconstruction model further includes: a layered residual super-resolution reconstruction module, which includes: local residual connection units; the method further includes: The shallow embedding features at time τ are obtained by performing convolutional mapping based on the first temporal features at time τ through the multi-scale spatial feature extraction unit. The target fusion feature is connected to the shallow embedding feature through the local residual connection unit, and deep feature fusion is performed to obtain the deep feature at time τ; upsampling pre-convolution and sub-pixel operation are performed on the deep feature at time τ to obtain the single-channel high-frequency correction term at time τ.

6. The method according to claim 5, characterized in that, The hierarchical residual super-resolution reconstruction module further includes a global residual connection unit, and the method further includes: The first time-series feature is amplified by bicubic interpolation using the global residual connection unit to obtain the low-frequency baseline of a single channel at time τ. The high-frequency correction term of the single channel at time τ is added to the low-frequency baseline of the single channel at time τ to obtain the target high-resolution effective wave height field of ocean waves at time τ.

7. The method according to claim 1, characterized in that, The wave significant wave height field reconstruction model further includes: a physical constraint training module, which has a pre-set multi-objective constraint loss function. The adjustment of the model parameters of the wave significant wave height field reconstruction model based on the data difference between the target high-resolution wave significant wave height field at time τ and the true high-resolution wave significant wave height field, until the data difference satisfies the pre-set loss function constraints, includes: Based on the data difference between the target high-resolution wave significant height field and the actual high-resolution wave significant height field at time τ, the model parameters of the wave significant height field reconstruction model are adjusted until the following multi-objective constrained loss function converges: ； in, The reconstruction loss function is used to constrain the error between the wave height in the target high-resolution ocean wave effective wave height field and the true high-resolution wave height. The weighting coefficients for the reconstruction loss function; SSIM structural similarity loss function is used to constrain the structural similarity between the target high-resolution wave height field and the true high-resolution wave height field in terms of local mean, variance, and covariance. These are the weight coefficients of the SSIM structural similarity loss function; The gradient consistency loss function is used to constrain the spatial gradient consistency between the target high-resolution wave effective wave height field and the true high-resolution wave height field in the x and y directions. These are the weight coefficients of the gradient consistency loss function; The spatial smoothness constraint loss function is used to constrain the wave height variation of the target high-resolution ocean wave effective wave height field between adjacent grids. These are the weighting coefficients of the spatial smoothness constraint function; The physical range constraint loss function is used to constrain the marine statistical boundary between the target high-resolution wave effective wave height field and the true high-resolution wave height field; These are the weighting coefficients of the physical extent constraint loss function; The energy consistency constraint loss function is used to constrain the global energy conservation between the target high-resolution ocean wave effective wave height field and the true high-resolution wave height field. These are the weighting coefficients of the energy consistency constraint loss function.

8. A super-resolution reconstruction device for the effective wave height of ocean waves, characterized in that, The device includes: The acquisition module is used to acquire a low-resolution multivariate wave field sequence for T consecutive time points from the current time t, wherein the low-resolution multivariate field sequence at each time point includes the following physical variables: significant wave height, mean wave period, wave direction sine component and wave direction cosine component; The reconstruction module is used to call the target wave significant wave height field reconstruction model to reconstruct the wave significant wave height field based on the low-resolution multivariate wave field sequence of the continuous T time periods, so as to obtain the target high-resolution wave significant wave height field at the current time t. The model training module is used to construct a training dataset using real low-resolution multivariate wave field sequences of T consecutive times at several times τ and real high-resolution effective wave height fields of waves at various times τ. The training dataset is input into the initial wave effective wave height field reconstruction model. The wave effective wave height field reconstruction model learns the wave propagation law along the wave direction and extracts the physical structure of the real low-resolution multivariate wave field sequence at different scales. Based on the learned propagation law and the extracted physical structure of the real low-resolution multivariate wave field sequence at different scales, the wave effective wave height field is reconstructed to obtain the target high-resolution wave effective wave height field at time τ. Based on the data difference between the target high-resolution wave significant height field and the actual high-resolution wave significant height field at time τ, the model parameters of the wave significant height field reconstruction model are adjusted until the data difference satisfies the preset loss function constraint. The wave effective wave height field reconstruction model corresponding to the data difference satisfying the preset loss function constraint is determined as the target wave effective wave height field reconstruction model.

9. An electronic device, characterized in that, The electronic device includes: a processor and a memory storing a program; wherein the program includes instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-7.

Images