A spatiotemporal downscaling method and device based on marine environmental data

Abstract

The application discloses a kind of space-time downscaling method and device based on marine environment data, the method includes obtaining marine environment data information to be processed, marine environment training dataset and marine environment real dataset;The marine environment training dataset and the marine environment real dataset are processed, and generator model is obtained;Using the generator model, the marine environment data information to be processed is processed, and target marine environment data information is obtained. Visible, this embodiment can generate high-resolution data while considering the complexity and nonlinear characteristics of marine environment, overcome the limitations of traditional methods in processing marine data, so as to provide more accurate data support in marine environment simulation and prediction.

Description

Technical Field

[0001] This invention relates to the field of marine environmental data processing, and in particular to a spatiotemporal downscaling method and apparatus based on marine environmental data. Background Technology

[0002] Spatiotemporal downscaling of marine environmental data is a key technology in marine science and engineering, widely applied in areas such as marine climate monitoring, marine ecological research, maritime safety, and disaster early warning. With the intensification of global climate change, the demand for marine environmental data is constantly increasing, especially in early warning of marine disasters (such as tsunamis and storm surges) and the protection of marine ecosystems, where high-precision marine environmental data is crucial. Currently, with advancements in data acquisition methods such as remote sensing, satellite observation, and ocean buoys, the acquisition of marine environmental data has become more extensive and frequent. However, these datasets often suffer from low spatiotemporal resolution, limiting their application in high-precision simulation and early warning systems.

[0003] Traditional interpolation methods are widely used to convert low-resolution marine environmental data into high-resolution data. These methods include nearest neighbor interpolation, bilinear interpolation, and kriging interpolation. While these methods perform well in some simple cases, they are inadequate when dealing with complex nonlinear data and marine environments with significant spatiotemporal variations. Interpolation methods typically assume the smoothness of data variations and cannot effectively handle the nonlinear characteristics present in marine environmental data, such as the complex effects of ocean currents, wind-induced drift, and other factors. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a spatiotemporal downscaling method and apparatus based on marine environmental data, which can take into account the complexity and nonlinear characteristics of the marine environment while generating high-resolution data, overcome the limitations of traditional methods in processing marine data, and thus provide more accurate data support in marine environment simulation and prediction.

[0005] To address the aforementioned technical problems, a first aspect of this invention discloses a spatiotemporal downscaling method based on marine environmental data, the method comprising:

[0006] S1, acquire marine environmental data information to be processed, including marine environmental training dataset and real marine environmental dataset;

[0007] S2, Process the marine environment training dataset and the marine environment real dataset to obtain the generator model;

[0008] S3. Using the generator model, the marine environmental data information to be processed is processed to obtain the target marine environmental data information.

[0009] As an optional implementation, in the first aspect of the present invention, the process of processing the marine environment training dataset and the marine environment real dataset to obtain a generator model includes:

[0010] S21, Using a marine environment downsampling model, the marine environment training dataset is downsampled to obtain a preprocessed marine environment training dataset;

[0011] The marine environment downsampling model is as follows:

[0012]

[0013] In the formula, JCY is the preprocessed marine environment training dataset, JCY(t) is the preprocessed marine environment training data information at time point t, and ω i1 (i1=0,1,2) are weight coefficients, HJ is the marine environment training dataset, t is the time point in the marine environment training dataset, HJ(t) is the marine environment training data information of the marine environment training dataset at time point t, Δt is the time interval of the marine environment training dataset at the time resolution, and N1 is the number of time points in the marine environment training dataset.

[0014] S22, the preprocessed marine environment training dataset and the marine environment real dataset are processed to obtain the first marine environment training dataset and the first marine environment real dataset;

[0015] S23, process the first marine environment training dataset and the first marine environment real dataset to obtain the generator model.

[0016] As an optional implementation, in the first aspect of the present invention, processing the preprocessed marine environment training dataset and the real marine environment dataset to obtain a first marine environment training dataset and a first real marine environment dataset includes:

[0017] S221, perform raster data vectorization processing on the preprocessed marine environment training dataset and the real marine environment dataset to obtain the second marine environment training dataset and the second real marine environment dataset.

[0018] S222, perform outlier processing on the second marine environment training dataset and the second marine environment real dataset to obtain the third marine environment training dataset and the third marine environment real dataset;

[0019] S223, perform missing value processing on the third marine environment training dataset and the third marine environment real dataset to obtain the fourth marine environment training dataset and the fourth marine environment real dataset.

[0020] S224, perform format unification processing on the fourth marine environment training dataset and the fourth marine environment real dataset to obtain the fifth marine environment training dataset and the fifth marine environment real dataset.

[0021] S225, perform spatiotemporal alignment processing on the fifth marine environment training dataset and the fifth marine environment real dataset to obtain the first marine environment training dataset and the first marine environment real dataset.

[0022] As an optional implementation, in the first aspect of the present invention, the process of processing the first marine environment training dataset and the first marine environment real dataset to obtain a generator model includes:

[0023] S231, default s = 1;

[0024] S232, using the generator initial model and the discriminator initial model, process the first marine environment training dataset and the first marine environment real dataset to obtain the loss function value;

[0025] S233, determine whether the loss function value is less than a preset loss function threshold, and obtain a first determination result;

[0026] When the first judgment result is yes, the generator initial model is determined to be a generator model, and S3 is executed;

[0027] When the first judgment result is negative, the parameters of the generator initial model and the discriminator initial model are updated according to the loss function value to obtain the updated generator initial model and the updated discriminator initial model.

[0028] S234, determine the updated generator initial model as the generator initial model, and determine the updated discriminator initial model as the discriminator initial model;

[0029] S235, increment s by 1, and execute S232.

[0030] As an optional implementation, in the first aspect of the present invention, the step of processing the first marine environment training dataset and the first marine environment real dataset using the generator initial model and the discriminator initial model to obtain the loss function value includes:

[0031] S2321, Using the generator's initial model, the first marine environment training dataset is trained to obtain generator data information;

[0032] S2322, Using the discriminator initial model, the generator data information and the first real marine environment dataset are processed to obtain the discrimination result information;

[0033] S2323, The generator data information and the first real marine environment dataset are processed to obtain marine environment data difference information; the marine environment data difference information includes first difference information, second difference information and third difference information.

[0034] S2324, Process the discrimination result information and the marine environmental data difference information to obtain the loss function value.

[0035] As an optional implementation, in the first aspect of the present invention, the processing of the discrimination result information and the marine environmental data difference information to obtain a loss function value includes:

[0036] The difference between the discrimination result information and the marine environment data is processed using a marine environment loss function model to obtain the loss function value;

[0037] The marine environment loss function model is as follows:

[0038]

[0039] θ1+θ2+θ3+θ4=1;

[0040] 0≤θ1,θ2,θ3,θ4≤1;

[0041] In the formula, SS is the loss function value, and PB is... i3 Let θ1, θ2, and θ3 be the i3rd probability value in the discrimination result information, DY, DE, and DS be the first difference information, the second difference information, and the third difference information, respectively, and θ1, θ2, and θ3 be the first weighting factor, the second weighting factor, and the third weighting factor, respectively.

[0042] As an optional implementation, in the first aspect of the present invention, the step of processing the marine environmental data information to be processed using the generator model to obtain target marine environmental data information includes:

[0043] S31, using the generator model, the marine environmental data information to be processed is processed to obtain marine environmental data information;

[0044] S32, perform multi-scale decomposition processing on the marine environmental data information to obtain the first scale component and the second scale component;

[0045] S33, using the first-scale calculation model of the marine environment, the first-scale component is enhanced to obtain the third-scale component;

[0046] The first-scale calculation model for the marine environment is as follows:

[0047]

[0048] In the formula, CDS is the third scale component, CDY is the first scale component, CD is the mean filter of the first scale component, and δ4, δ5 and δ6 are the fourth weight parameter, the fifth weight parameter and the sixth weight parameter, respectively.

[0049] S34, optimize the second scale component to obtain the fourth scale component;

[0050] S35, the third-scale component and the fourth-scale component are reconstructed to obtain target marine environmental data information.

[0051] A second aspect of this invention discloses a spatiotemporal downscaling device based on marine environmental data, the device comprising:

[0052] The acquisition module is used to acquire marine environmental data information to be processed, including marine environmental training datasets and real marine environmental datasets.

[0053] The first computing module is used to process the marine environment training dataset and the marine environment real dataset to obtain a generator model;

[0054] The second calculation module is used to process the marine environmental data information to be processed using the generator model to obtain the target marine environmental data information.

[0055] A third aspect of this invention discloses another spatiotemporal downscaling device based on marine environmental data, the device comprising:

[0056] processor;

[0057] A memory coupled to the processor stores executable program code;

[0058] The processor calls the executable program code stored in the memory to execute some or all of the steps of the spatiotemporal downscaling method based on marine environmental data disclosed in the first aspect of the present invention.

[0059] The fourth aspect of this invention discloses a computer-readable storage medium storing computer instructions, which, when invoked, are used to execute some or all of the steps of the spatiotemporal downscaling method based on marine environmental data disclosed in the first aspect of this invention.

[0060] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0061] In this embodiment of the invention, marine environmental data to be processed, a marine environmental training dataset, and a real marine environmental dataset are acquired; the marine environmental training dataset and the real marine environmental dataset are processed to obtain a generator model; the generator model is then used to process the marine environmental data to be processed to obtain target marine environmental data. It is evident that this embodiment can generate high-resolution data while considering the complexity and nonlinear characteristics of the marine environment, overcoming the limitations of traditional methods in processing marine data, thereby providing more accurate data support in marine environment simulation and prediction. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a flowchart illustrating a spatiotemporal downscaling method based on marine environmental data disclosed in an embodiment of the present invention.

[0064] Figure 2 This is a statistical graph showing the model training results and measured data accuracy of the spatiotemporal downscaling method of the present invention;

[0065] Figure 3 This is a comparison chart of the results before and after downscaling at different times in the spatiotemporal downscaling method of the present invention;

[0066] Figure 4 This is a spatiotemporal downscaling result diagram for September 2024 in an embodiment of the present invention;

[0067] Figure 5 This is a schematic diagram of a spatiotemporal downscaling device based on marine environmental data disclosed in an embodiment of the present invention;

[0068] Figure 6 This is a schematic diagram of another spatiotemporal downscaling device based on marine environmental data disclosed in an embodiment of the present invention. Detailed Implementation

[0069] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0071] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0072] This invention discloses a spatiotemporal downscaling method and apparatus based on marine environmental data. It can generate high-resolution data while taking into account the complexity and nonlinear characteristics of the marine environment, overcoming the limitations of traditional methods in processing marine data, and thus providing more accurate data support for marine environmental simulation and prediction. These will be described in detail below.

[0073] Example 1

[0074] Please see Figure 1-4 , Figure 1 This is a flowchart illustrating a spatiotemporal downscaling method based on marine environmental data disclosed in an embodiment of the present invention. Figure 1 The described spatiotemporal downscaling method based on marine environmental data is applied to a spatiotemporal downscaling device based on marine environmental data, such as a local server or cloud server for spatiotemporal downscaling optimization management based on marine environmental data. This invention does not limit the application of this method. Figure 1 As shown, this spatiotemporal downscaling method based on marine environmental data can include the following operations:

[0075] S1, acquire marine environmental data information to be processed, including marine environmental training dataset and real marine environmental dataset;

[0076] It should be noted that spatiotemporal downscaling refers to the process of converting low-resolution spatiotemporal data into high-resolution spatiotemporal data.

[0077] It should be noted that the marine environment training dataset is the ERA5 dataset, while the real marine environment dataset is the CMEMS dataset. The ERA5 dataset has a temporal resolution of 1 hour and a spatial resolution of 0.25°; the CMEMS dataset has a temporal resolution of 3 hours and a spatial resolution of 0.083°.

[0078] The marine environment training dataset and the marine environment real dataset both include other variables related to the marine environment, such as ocean current data, sea breeze data, and simulated ocean wave data. Specifically, the embodiments of the present invention do not limit this.

[0079] It should be noted that the marine environmental data to be processed is any marine environmental data information in the ERA5 dataset. For example, using the spatiotemporal downscaling method of this invention, the marine environmental data to be processed is set to ocean wave marine environmental data from January 2024 to December 2024 in the ERA5 dataset. The resulting spatiotemporal downscaling result of the target marine environmental data is shown in the figure below. Figure 4 As shown.

[0080] It should be noted that although the ERA5 dataset has high timeliness, its low spatial resolution makes it difficult to meet the needs of accurate modeling and prediction of local areas. The CMEMS dataset provides higher spatial resolution (0.083°) and lower temporal resolution (3 hours), but due to its low temporal resolution, it still cannot fully capture short-term fluctuations and changes in the marine environment, especially in high-frequency dynamic monitoring.

[0081] It should be noted that this application uses the CMEMS dataset as the real dataset. Its high spatial resolution of 0.083° provides a more refined representation of the spatial characteristics of the marine environment. This data has been systematically observed and verified and is widely used in marine scientific research and operational forecasting. Although its 3-hour temporal resolution has certain limitations in capturing short-term changes, its advantage in spatial accuracy makes it an ideal target dataset for training generator models. By combining it with ERA5 data, the deficiencies of each in spatiotemporal resolution are made up for, providing a reliable basis for generating high-quality marine environmental data.

[0082] This application acquires the ERA5 marine environment training dataset and the CMEMS marine environment real dataset, performs time alignment processing on the two, trains a generator model to learn the mapping relationship from low spatial resolution to high spatial resolution, and finally uses the model to convert the ERA5 data into target marine environment data with both high temporal resolution (1 hour) and high spatial resolution (0.083°).

[0083] This solution successfully combines the advantages of ERA5's high temporal resolution and CMEMS's high spatial resolution, overcoming the limitations of a single dataset, providing flexibility in processing various marine environmental data, enhancing the ability to capture short-term fluctuations in the marine environment, and achieving data augmentation through model training rather than additional observation equipment, thus providing a cost-effective and high-precision data solution for marine environmental monitoring and forecasting.

[0084] S2, Process the marine environment training dataset and the marine environment real dataset to obtain the generator model;

[0085] S3. Using the generator model, the marine environmental data information to be processed is processed to obtain the target marine environmental data information.

[0086] As can be seen, the spatiotemporal downscaling method based on marine environmental data described in the embodiments of the present invention can generate high-resolution data while taking into account the complexity and nonlinear characteristics of the marine environment, thereby providing more accurate data support in marine environment simulation and prediction.

[0087] In an optional embodiment, the process of processing the marine environment training dataset and the marine environment real dataset to obtain the generator model includes:

[0088] S21, Using a marine environment downsampling model, the marine environment training dataset is downsampled to obtain a preprocessed marine environment training dataset;

[0089] The marine environment downsampling model is as follows:

[0090]

[0091] In the formula, JCY is the preprocessed marine environment training dataset, JCY(t) is the preprocessed marine environment training data information at time point t, and ω i1(i1=0,1,2) are weight coefficients, HJ is the marine environment training dataset, t is the time point in the marine environment training dataset, HJ(t) is the marine environment training data information of the marine environment training dataset at time point t, Δt is the time interval of the marine environment training dataset at the time resolution, and N1 is the number of time points in the marine environment training dataset.

[0092] It should be noted that the marine environment downsampling model, by introducing a weighted averaging mechanism to downsample the marine environment training dataset, not only adjusts the 1-hour time resolution of the marine environment training dataset to a 3-hour resolution matching CMEMS, but more importantly, it achieves this through the weighting coefficient ω. i1 The design of this method allows the downsampling process to preserve key temporal features and trends in the original data. Compared to simple sampling or average downsampling, this weighted averaging method better preserves data continuity and physical meaning, reduces information loss, and effectively minimizes distortion and bias that may occur in time-series data during downsampling. Simultaneously, this preprocessing method provides higher-quality training samples with more consistent time scales for the subsequent generator model training, thereby improving the final generator model's learning ability and prediction accuracy of spatiotemporal features, laying the foundation for achieving high-quality spatiotemporal downscaling results.

[0093] It should be noted that the weight coefficients can be optimized based on the characteristics of the marine environmental data. Specifically, this embodiment of the invention does not impose limitations. For example, a non-uniform weight distribution (such as a Gaussian distribution) can be used, assigning a higher weight to the current time point t (e.g., ω0 = 0.5), a medium weight to adjacent time points (e.g., ω1 = 0.3), and a lower weight to further back time points (e.g., ω2 = 0.2). The advantages of this non-uniform weight allocation are: on the one hand, it can highlight the data characteristics of important time points and retain key temporal dynamic information; on the other hand, by considering the influence of adjacent time points, it achieves a smooth transition of the time series, reducing information gaps and abrupt changes that may occur during downsampling; at the same time, this flexible weight design can also be adjusted according to the temporal characteristics (e.g., the rate of change) of different marine variables, improving the adaptability and accuracy of the downsampling model and providing a more reliable training foundation for the subsequent generation of high-quality spatiotemporal downscaling data.

[0094] S22, the preprocessed marine environment training dataset and the marine environment real dataset are processed to obtain the first marine environment training dataset and the first marine environment real dataset;

[0095] S23, process the first marine environment training dataset and the first marine environment real dataset to obtain the generator model.

[0096] As can be seen, the spatiotemporal downscaling method based on marine environmental data described in the embodiments of the present invention can generate high-resolution data while taking into account the complexity and nonlinear characteristics of the marine environment, thereby providing more accurate data support in marine environment simulation and prediction.

[0097] In another optional embodiment, the step of processing the preprocessed marine environment training dataset and the real marine environment dataset to obtain a first marine environment training dataset and a first real marine environment dataset includes:

[0098] S221, perform raster data vectorization processing on the preprocessed marine environment training dataset and the real marine environment dataset to obtain the second marine environment training dataset and the second real marine environment dataset.

[0099] It should be noted that the above-mentioned raster data vectorization processing can be performed using Geographic Information System (GIS) tools, professional marine data processing software, or Python-based geographic data processing libraries (such as GeoPandas, GDAL, etc.). Specifically, the embodiments of the present invention do not limit the specific processing.

[0100] It should be noted that by vectorizing raster data, the original raster-format marine environmental data can be converted into a vector representation, giving the data better geometric characteristics and topological relationships. This makes it easier to capture the spatial distribution characteristics and variation patterns of marine environmental elements, while reducing data redundancy, improving data processing efficiency, providing a more suitable input format for subsequent deep learning models, and enhancing the model's ability to learn spatial features.

[0101] It should be noted that the above raster data vectorization processing is performed on the preprocessed marine environment training dataset to obtain the second marine environment training dataset, and on the real marine environment dataset to obtain the second real marine environment dataset.

[0102] S222, perform outlier processing on the second marine environment training dataset and the second marine environment real dataset to obtain the third marine environment training dataset and the third marine environment real dataset;

[0103] It should be noted that the above outlier handling can be implemented through statistical methods (such as the 3σ rule and box plot method), machine learning methods (such as isolated forest and cluster analysis) or domain knowledge constraints (such as physical law constraints), etc. The specific implementation of this invention is not limited.

[0104] It should be noted that outlier handling can identify and process abnormal observations or measurement errors in marine environmental data, improve data quality and reliability, avoid the negative impact of outliers on model training, and ensure that the generator model can learn the true distribution patterns and variation characteristics of marine environmental data, thereby improving the accuracy and stability of spatiotemporal downscaling results.

[0105] It should be noted that the above outlier processing involves processing outliers on the second marine environment training dataset to obtain the third marine environment training dataset, and processing outliers on the second real marine environment dataset to obtain the third real marine environment dataset.

[0106] S223, perform missing value processing on the third marine environment training dataset and the third marine environment real dataset to obtain the fourth marine environment training dataset and the fourth marine environment real dataset.

[0107] It should be noted that the above missing value processing can be implemented by interpolation methods (such as linear interpolation, spline interpolation), statistical models (such as expectation-maximization algorithm), spatiotemporal correlation analysis (such as considering the information of surrounding spatiotemporal points), or deep learning methods (such as autoencoders, GAN networks), etc. The specific implementation of this invention is not limited.

[0108] It should be noted that missing value processing can fill in data gaps in marine environmental data caused by observation limitations, sensor failures, or data transmission problems, ensuring the integrity and continuity of the data, maintaining the temporal and spatial consistency of marine environmental elements, providing uninterrupted training samples for the generator model, and improving the model's understanding and learning ability of the global features of the marine environment.

[0109] It should be noted that the missing value processing described above involves processing the missing values ​​of the third marine environment training dataset to obtain the fourth marine environment training dataset, and processing the missing values ​​of the third marine environment real dataset to obtain the fourth marine environment real dataset.

[0110] S224, perform format unification processing on the fourth marine environment training dataset and the fourth marine environment real dataset to obtain the fifth marine environment training dataset and the fifth marine environment real dataset.

[0111] It should be noted that the above-mentioned format unification process can be achieved through data structure transformation, coordinate system unification, data format standardization, and adjustment of variable names and units, so that datasets from different sources have a consistent data organization form and expression method. Specifically, the embodiments of the present invention do not limit this.

[0112] It should be noted that by unifying the format, the differences between the ERA5 and CMEMS datasets in terms of data storage format, variable definition, coordinate representation, etc., can be eliminated, and an accurate correspondence between the two datasets can be established. This makes it easier for the generator model to directly learn the mapping rules from low resolution to high resolution, reduce training bias caused by inconsistent formats, and simplify the subsequent model training and data processing process, thereby improving the efficiency and stability of the entire spatiotemporal downscaling system.

[0113] It should be noted that the above format unification processing is performed on the fourth marine environment training dataset to obtain the fifth marine environment training dataset, and on the fourth marine environment real dataset to obtain the fifth marine environment real dataset.

[0114] S225, perform spatiotemporal alignment processing on the fifth marine environment training dataset and the fifth marine environment real dataset to obtain the first marine environment training dataset and the first marine environment real dataset.

[0115] It should be noted that the above spatiotemporal alignment processing can be achieved through methods such as timestamp matching, spatial interpolation resampling, coordinate transformation, and projection transformation to ensure that the two datasets have an accurate correspondence in the spatial dimension. Specifically, the embodiments of the present invention do not limit this.

[0116] It's important to note that spatiotemporal alignment addresses the spatial resolution differences (0.25° and 0.083°) between the ERA5 and CMEMS datasets, establishing a point-to-point mapping between the two datasets. This ensures the generator model learns accurate spatiotemporal feature transformation patterns, avoiding spurious correlations caused by data misalignment. This step is crucial in the entire spatiotemporal downscaling process, providing scientifically sound paired samples for subsequent generator model training and guaranteeing the model's ability to correctly learn the mapping from low to high resolution.

[0117] It should be noted that the above spatiotemporal alignment process involves performing spatiotemporal alignment on the fifth marine environment training dataset to obtain the first marine environment training dataset, and performing spatiotemporal alignment on the fifth marine environment real dataset to obtain the first marine environment real dataset. In this process, the time points of the CMEMS dataset (3-hour intervals), which has a lower temporal resolution, are used as a benchmark. Corresponding time point data from the ERA5 dataset are selected for pairing. Simultaneously, spatially, the two datasets are unified to the same geographic grid system, establishing a strict spatial correspondence and providing high-quality training sample pairs for the generator model.

[0118] As can be seen, the spatiotemporal downscaling method based on marine environmental data described in the embodiments of the present invention can generate high-resolution data while taking into account the complexity and nonlinear characteristics of the marine environment, thereby providing more accurate data support in marine environment simulation and prediction.

[0119] In another optional embodiment, the processing of the first marine environment training dataset and the first marine environment real dataset to obtain the generator model includes:

[0120] S231, default s = 1;

[0121] S232, using the generator initial model and the discriminator initial model, process the first marine environment training dataset and the first marine environment real dataset to obtain the loss function value;

[0122] It should be noted that the large model used in this application is a GAN model, and the initial generator model and the initial discriminator model are the untrained generator and discriminator in the GAN model.

[0123] For example, the generator aims to transform low-resolution marine environmental data (such as ERA5 data) into high-resolution data. The generator employs a deep neural network design based on the U-Net architecture, with the following structure:

[0124] Input layer: Receives low-resolution marine environmental data (such as ERA5 data). Each input data point is a 64×64×5 matrix, where 64×64 represents the spatial dimension (corresponding to a 0.25° resolution marine region grid), and 5 represents the number of channels (including environmental variables such as ocean current U-component, ocean current V-component, sea breeze U-component, sea breeze V-component, and wave height). This multivariate input design fully utilizes the interrelationships among marine environmental elements to improve the model's predictive ability.

[0125] Encoding stage (downsampling path): Extract multi-scale spatial features of low-resolution data through a 4-layer convolutional neural network.

[0126] The first convolutional layer consists of 64 3×3 convolutional kernels with a stride of 1 and padding of 1. It uses the LeakyReLU (α = 0.2) activation function and has an output size of (64×64×64).

[0127] The second convolutional layer consists of 128 3×3 convolutional kernels with a stride of 2 and padding of 1. It uses the LeakyReLU (α = 0.2) activation function and has an output size of (32×32×128).

[0128] The third convolutional layer consists of 256 3×3 convolutional kernels with a stride of 2 and padding of 1. It uses the LeakyReLU (α = 0.2) activation function and has an output size of (16×16×256).

[0129] The fourth convolutional layer consists of 512 3×3 convolutional kernels with a stride of 2 and padding of 1. It uses the LeakyReLU activation function (α = 0.2) and has an output size of (8×8×512).

[0130] This progressive downsampling design can effectively capture spatial features from local to global perspectives and learn the performance patterns of marine environmental data at different scales.

[0131] Decoding stage (upsampling path): The spatial resolution of the data is gradually improved through 4 deconvolution layers.

[0132] The first deconvolutional layer consists of 256 3×3 deconvolutional kernels with a stride of 2, padding of 1, and ReLU activation function. The output size is (16×16×256).

[0133] The second deconvolution layer consists of 128 3×3 deconvolution kernels with a stride of 2 and padding of 1. It uses the ReLU activation function and has an output size of (32×32×128).

[0134] The third deconvolutional layer consists of 64 3×3 deconvolutional kernels with a stride of 2, padding of 1, and ReLU activation function. The output size is (64×64×64).

[0135] The fourth deconvolutional layer consists of 64 3×3 deconvolution kernels with a stride of 3 and padding of 1. It uses the ReLU activation function and has an output size of (192×192×64), achieving a precise spatial resolution improvement from 0.25° to 0.083°.

[0136] This multi-stage upsampling strategy ensures a smooth improvement in spatial resolution, avoiding information loss and artifacts that may result from direct upsampling.

[0137] Batch normalization layer: Batch normalization is used after each convolutional and deconvolutional layer (before the activation function), with parameters set to momentum 0.9 and epsilon 1e-5. This design significantly improves training stability, accelerates model convergence, and effectively prevents the vanishing / exploding gradient problem.

[0138] Residual connections (skip connections): Three sets of skip connections are implemented between the corresponding coding and decoding layers.

[0139] The first layer encoding output (64×64×64) is connected to the third layer decoding input; the second layer encoding output (32×32×128) is connected to the second layer decoding input; and the third layer encoding output (16×16×256) is connected to the first layer decoding input.

[0140] These skip connections are achieved through feature fusion, which ensures the direct transmission of low-level features (such as texture and edges), greatly alleviating the information bottleneck problem and enabling the generated high-resolution data to retain both global semantic information and rich local details.

[0141] Attention mechanism: A channel attention module is introduced at each skip connection to adaptively adjust the weights of different feature channels. This approach enhances the model's ability to identify key ocean features and improves the accuracy and physical consistency of the downscaling results.

[0142] Output layer: The last layer uses a 1×1 convolutional kernel with 5 output channels. Then, the Tanh activation function is applied to generate high-resolution data, with a size of (192×192×5). Here, 192×192 corresponds to the spatial dimension of the target resolution (0.083°×0.083°), and 5 represents the number of channels in the original input. The use of the Tanh activation function ensures that the output value is within the range [-1,1], matching the normalization range of the training data and producing more stable results.

[0143] The generator structure in this application can effectively learn the mapping relationship between the ERA5 and CMEMS datasets, realizing the conversion from low spatiotemporal resolution to high spatiotemporal resolution. The generated high-resolution marine environmental data not only maintains the temporal continuity of the ERA5 data (1-hour resolution), but also obtains spatial fineness comparable to CMEMS (0.083° resolution), which greatly improves the application value of the data while preserving physical consistency.

[0144] For example, the structure of the discriminator is as follows:

[0145] Input layer: Simultaneously receives real high-resolution data (CMEMS dataset) and high-resolution data generated by the generator. Each input is a matrix of size (192×192×5), where 192×192 represents the spatial dimension (corresponding to a 0.083° resolution ocean region grid), and 5 represents the number of channels (including environmental variables such as ocean current U-component, ocean current V-component, sea breeze U-component, sea breeze V-component, and wave height). This approach allows the discriminator to directly compare the feature differences between real and generated data at the same resolution.

[0146] Convolutional layer sequence: Extracting feature levels from the input data through 5 progressively downsampled convolutional layers:

[0147] The first convolutional layer consists of 64 4×4 convolutional kernels with a stride of 2 and padding of 1. It uses the LeakyReLU (α = 0.2) activation function and has an output size of (96×96×64).

[0148] The second convolutional layer consists of 128 4×4 convolutional kernels with a stride of 2 and padding of 1. It uses the LeakyReLU (α = 0.2) activation function and has an output size of (48×48×128).

[0149] The third convolutional layer consists of 256 4×4 convolutional kernels with a stride of 2 and padding of 1. It uses the LeakyReLU (α = 0.2) activation function and has an output size of (24×24×256).

[0150] The fourth convolutional layer consists of 512 4×4 convolutional kernels with a stride of 2 and padding of 1. It uses the LeakyReLU (α = 0.2) activation function and has an output size of (12×12×512).

[0151] The fifth convolutional layer consists of 1024 4×4 convolutional kernels with a stride of 2 and padding of 1. It uses the LeakyReLU (α = 0.2) activation function and has an output size of (6×6×1024).

[0152] This progressive dimensionality reduction convolutional structure enables the discriminator to effectively capture multi-scale features of marine environmental data, from local texture to global structure, and accurately distinguish between genuine and false details in high-resolution marine data.

[0153] Spectral normalization: Applying spectral normalization after each convolutional layer limits the spectral norm of the weight matrix in the discriminator network, effectively improving training stability, preventing mode collapse, and making the GAN training process more stable and reliable.

[0154] Feature extraction module: A self-attention mechanism is introduced after the third convolutional layer, which enables the discriminator to pay attention to long-range spatial dependencies in marine environmental data and improve the ability to distinguish global consistency.

[0155] Fully connected layer: Flattens the output (6×6×1024) of the last convolutional layer into a vector, and processes it through two fully connected layers:

[0156] The first fully connected layer has an input dimension of 6×6×1024 and an output dimension of 1024, using the LeakyReLU (α=0.2) activation function.

[0157] Second fully connected layer: Input dimension 1024, output dimension 1;

[0158] This approach compresses the complex high-dimensional feature space into a single discrimination score, providing a basis for the final true / false judgment.

[0159] Output layer: The sigmoid activation function is used to map the final network output to the interval [0,1], representing the probability that the input data is real data. The closer the output value is to 1, the more real the discriminator considers the data; the closer it is to 0, the more generated the data is considered.

[0160] This multi-level discriminator design progressively guides the generator during training to produce high-resolution marine environmental data that more closely approximates the real CMEMS data distribution. The discriminator's discrimination capability and the generator's generation capability mutually reinforce each other during adversarial training, ultimately achieving precise spatiotemporal downscaling from low-resolution ERA5 data to high-resolution CMEMS data. The generated data not only resembles real data visually but also maintains the continuity of the physical field and the intrinsic correlation between marine environmental parameters.

[0161] S233, determine whether the loss function value is less than a preset loss function threshold, and obtain a first determination result;

[0162] It should be noted that the preset loss function threshold ranges from [0.01, 0.2], but the specific values ​​are not limited in this embodiment of the invention.

[0163] When the first judgment result is yes, the generator initial model is determined to be a generator model, and S3 is executed;

[0164] When the first judgment result is negative, the parameters of the generator initial model and the discriminator initial model are updated according to the loss function value to obtain the updated generator initial model and the updated discriminator initial model.

[0165] It should be noted that the above parameter updates are processed by the Adam optimizer, and the specific implementation of this invention is not limited thereto.

[0166] For example, the above parameters are updated as follows: In each training batch, the generator parameters are first fixed, and the discriminator is updated using only the discriminant loss to improve its ability to distinguish between real and fake data; then, the discriminator parameters are fixed, and the generator is updated using the full loss function, taking into account both the authenticity of the generated data (through adversarial loss) and its spatiotemporal characteristics and physical consistency (through three types of difference losses). The entire optimization process is implemented using the Adam optimizer (learning rate 0.0002, β1 = 0.5, β2 = 0.999), supplemented by a learning rate scheduling strategy and gradient pruning techniques to ensure training stability. This multi-objective optimization framework enables the generator to produce high-quality marine environmental data that not only deceives the discriminator but also satisfies the constraints of temporal continuity, spatial granularity, and physical laws. It achieves accurate spatiotemporal downscaling from ERA5 to CMEMS resolution, ensuring the practical value of the generated data in scientific applications.

[0167] S234, determine the updated generator initial model as the generator initial model, and determine the updated discriminator initial model as the discriminator initial model;

[0168] S235, increment s by 1, and execute S232.

[0169] As can be seen, the spatiotemporal downscaling method based on marine environmental data described in the embodiments of the present invention can generate high-resolution data while taking into account the complexity and nonlinear characteristics of the marine environment, thereby providing more accurate data support in marine environment simulation and prediction.

[0170] In another optional embodiment, the step of processing the first marine environment training dataset and the first marine environment real dataset using the generator initial model and the discriminator initial model to obtain the loss function value includes:

[0171] S2321, Using the generator's initial model, the first marine environment training dataset is trained to obtain generator data information;

[0172] It should be noted that the above processing was performed using the generator initial model of GAN. Since this is a common technique, it will not be elaborated further.

[0173] S2322, Using the discriminator initial model, the generator data information and the first real marine environment dataset are processed to obtain the discrimination result information;

[0174] It should be noted that the above processing was achieved using the initial model of the GAN discriminator. Since this is a conventional technique, it will not be elaborated further.

[0175] S2323, The generator data information and the first real marine environment dataset are processed to obtain marine environment data difference information; the marine environment data difference information includes first difference information, second difference information and third difference information.

[0176] S2324, Process the discrimination result information and the marine environmental data difference information to obtain the loss function value.

[0177] As can be seen, the spatiotemporal downscaling method based on marine environmental data described in the embodiments of the present invention can generate high-resolution data while taking into account the complexity and nonlinear characteristics of the marine environment, thereby providing more accurate data support in marine environment simulation and prediction.

[0178] In another optional embodiment, the step of processing the generator data information and the first real marine environment dataset to obtain marine environment data difference information includes:

[0179] S23231, The generator data information and the first real marine environment dataset are processed to obtain the first difference information;

[0180] It should be noted that the above calculations can be performed using conventional techniques such as mean squared error (MSE) and mean absolute error (MAE), or they can be obtained using the first difference calculation model of the marine environment. Specifically, the embodiments of the present invention do not limit the specifics.

[0181] The calculation model for the first difference in the marine environment is as follows:

[0182]

[0183] 1≤i2≤N2, 1≤j2≤M2;

[0184] In the formula, DY represents the first difference information, and SC... i2,j2 ZS is the j2nd feature value of the i2th sample in the generator data information. i2,j2 α is the j2nd feature value of the i2th sample in the first real marine environment dataset. j2 N is the feature weight coefficient, N2 is the number of samples in the generator data information, and M2 is the number of feature values ​​in any sample of the generator data information.

[0185] It should be noted that the first difference calculation model for the marine environment accurately quantifies the feature differences between generated data and real data, and introduces a feature weighting coefficient α. j2It allows different marine environmental variables to be assigned differentiated importance, enabling the model to prioritize the accuracy of key physical quantities; its normalization design ensures that the calculation results are independent of the sample size, while its sensitivity to outliers enables the generator to more accurately reproduce extreme marine events, thus comprehensively improving the overall numerical accuracy and application value of spatiotemporal downscaling results.

[0186] It should be noted that the feature weight coefficient enables the model to distinguish the importance of different marine environmental features (such as ocean currents, sea breezes, and waves), and assign higher weights to key physical quantities. For example, wave height can be assigned a higher weight (e.g., feature weight coefficient of 0.4), while relatively minor features are assigned lower weights.

[0187] S23232, The generator data information and the first real marine environment dataset are processed to obtain the second difference information;

[0188] It should be noted that the above processing can be performed using the Structural Similarity Index (SSIM), but the specific implementation of this invention is not limited thereto.

[0189] It should be noted that the above processing can assess the similarity between generated data and real data in terms of structure, texture, and local features. By considering the similarity of three dimensions—brightness, contrast, and structure—SSIM is more in line with how the human perception system evaluates image quality. When applied to marine environmental data, it can effectively capture important characteristics such as spatial distribution patterns, gradient features, and local correlations.

[0190] S23233, The generator data information and the first real marine environment dataset are processed to obtain the third difference information;

[0191] It should be noted that the above calculations can be performed using conventional techniques such as frequency domain correlation coefficients and wavelet coherence analysis, or they can be obtained using the third difference calculation model of the marine environment. In particular, the embodiments of the present invention do not limit the specifics.

[0192] The calculation model for the third difference in the marine environment is as follows:

[0193]

[0194] 1≤i2≤N2, 1≤j2≤M2;

[0195] In the formula, DS represents the third difference information, MA represents the maximum value of all feature values ​​in all samples of the first real marine environment dataset, DP represents the energy value of the low-frequency component in the generator data information, GP represents the energy value of the high-frequency component in the generator data information, and δ1, δ2 and δ3 represent the first weight parameter, the second weight parameter and the third weight parameter, respectively.

[0196] It should be noted that the first weight parameter, the second weight parameter, and the third weight parameter can be set by the user or obtained from historical data. Specifically, this embodiment of the invention does not limit the specifics.

[0197] It should be noted that the third difference calculation model for the marine environment constructs a composite loss mechanism based on frequency domain analysis and statistical characteristics. It evaluates whether the spectral characteristics of the generated data conform to the laws of marine physics by using the low-frequency to high-frequency energy ratio (DP / GP). The logarithmic design enables it to simultaneously focus on large-scale marine phenomena and small-scale turbulence characteristics, while the ingenious combination of the maximum term and the mean square error constructs an adaptive loss structure. This design enables the generator to not only focus on numerical matching during the learning process, but also to pay more attention to physical consistency and frequency domain characteristics, which significantly improves the reliability and practicality of the generated data in marine science applications.

[0198] It should be noted that the energy values ​​of low-frequency components and high-frequency components can be obtained by combining Fourier transform (FFT) with frequency domain energy analysis algorithms. Specifically, the embodiments of the present invention are not limited.

[0199] As can be seen, the spatiotemporal downscaling method based on marine environmental data described in the embodiments of the present invention can generate high-resolution data while taking into account the complexity and nonlinear characteristics of the marine environment, thereby providing more accurate data support in marine environment simulation and prediction.

[0200] It should be noted that the first weight parameter ranges from [0.1, 0.5] and controls the overall contribution of the third difference information to the loss function. When the first weight parameter is small (e.g., 0.2), the sensitivity to extreme values ​​is moderate, maintaining training stability; when it is large (e.g., 0.4), it significantly enhances the focus on extreme marine phenomena, prompting the generator to reproduce extreme events more accurately. The second weight parameter ranges from [0.5, 2.0] and adjusts the influence of the low-frequency to high-frequency energy ratio (DP / GP) in the loss calculation. When the second weight parameter is small (e.g., 0.8), the model's requirements for spectral distribution are relatively relaxed, which is conducive to fast convergence; when it is large (e.g., 1.5), it forces the generator to strictly follow the spectral characteristics of real data, and the generated results can more accurately maintain the energy distribution pattern of the marine environment at multiple scales. The third weight parameter ranges from [0.01, 0.1] and serves as the coefficient of the mean squared error term, affecting the sensitivity of the loss function to the overall numerical accuracy. When the third weight parameter is set to a small value (e.g., 0.03), the influence of the mean squared error is amplified, making the model focus more on numerical accuracy. When it is set to a large value (e.g., 0.08), the penalty for numerical error is reduced, allowing the model greater freedom to explore downscaling mapping relationships while maintaining physical consistency. The combination and adjustment of these three weight parameters construct a highly flexible physical consistency evaluation mechanism. By finely adjusting their values, an optimal balance can be achieved across the three dimensions of numerical accuracy, spectral characteristics, and extreme value reproducibility. This ensures that the generated high-resolution marine environmental data conforms to physical laws and meets the accuracy requirements of specific application scenarios, greatly improving the scientific value and practicality of the spatiotemporal downscaling results.

[0201] In another optional embodiment, the processing of the discrimination result information and the marine environmental data difference information to obtain the loss function value includes:

[0202] The difference between the discrimination result information and the marine environment data is processed using a marine environment loss function model to obtain the loss function value;

[0203] The marine environment loss function model is as follows:

[0204]

[0205] θ1+θ2+θ3+θ4=1;

[0206] 0≤θ1,θ2,θ3,θ4≤1;

[0207] In the formula, SS is the loss function value, and PB is... i3Let θ1, θ2, and θ3 be the i3rd probability value in the discrimination result information, DY, DE, and DS be the first difference information, the second difference information, and the third difference information, respectively, and θ1, θ2, and θ3 be the first weighting factor, the second weighting factor, and the third weighting factor, respectively.

[0208] It should be noted that the first weighting factor, the second weighting factor, and the third weighting factor can be set by the user or obtained from historical data. Specifically, this embodiment of the invention does not limit the specifics.

[0209] It should be noted that the marine environment loss function model constructs a multi-dimensional, multi-scale marine environment data evaluation system. Through adversarial loss (the first term), the generated data is driven to closely approximate the real data in terms of overall distribution. The first difference information ensures numerical accuracy, the second difference information captures spatial features and texture patterns, and the third difference information ensures the consistency of energy distribution and extreme value characteristics. These four complementary loss components achieve optimal synergy under the fine control of weighting factors (θ1 to θ4), enabling the generator to simultaneously learn the statistical characteristics, spatial structure, temporal evolution, and physical laws of marine environment data. The resulting high-resolution marine environment data is not only visually highly similar to real data, but more importantly, it maintains physical consistency and scientific validity. This provides a high-quality data foundation for marine environment monitoring, forecasting, and research, and significantly improves the value and reliability of spatiotemporal downscaling results in practical applications.

[0210] As can be seen, the spatiotemporal downscaling method based on marine environmental data described in the embodiments of the present invention can generate high-resolution data while taking into account the complexity and nonlinear characteristics of the marine environment, thereby providing more accurate data support in marine environment simulation and prediction.

[0211] In another optional embodiment, the step of processing the marine environmental data information to be processed using the generator model to obtain the target marine environmental data information includes:

[0212] S31, using the generator model, the marine environmental data information to be processed is processed to obtain marine environmental data information;

[0213] It should be noted that the above processing involves using a trained generator model to directly input the ERA5 low-resolution marine environment data (1 hour / 0.25°) to be processed into the generator model, and finally outputting spatiotemporally downscaled marine environment data that simultaneously possesses the characteristics of high ERA5 temporal resolution (1 hour) and high CMEMS spatial resolution (0.083°), i.e., marine environment data information.

[0214] S32, perform multi-scale decomposition processing on the marine environmental data information to obtain the first scale component and the second scale component;

[0215] It should be noted that multi-scale decomposition processing can be implemented using wavelet transform or Laplacian pyramid algorithms; specifically, the embodiments of this invention are not limited to any particular method. Through multi-scale decomposition, marine environmental data information is decomposed into a first-scale component representing low-frequency, large-scale features and a second-scale component representing high-frequency, detailed features. This decomposition allows the system to process large-scale structures (such as ocean current circulation and large-scale temperature fields) and small-scale details (such as turbulence and local fluctuations) in marine environmental data separately, effectively improving the multi-scale characteristic reconstruction capability of the spatiotemporal downscaling results. This ensures that the final generated marine environmental data maintains both the accuracy of the macroscopic structure and rich microscopic details.

[0216] S33, using the first-scale calculation model of the marine environment, the first-scale component is enhanced to obtain the third-scale component;

[0217] The first-scale calculation model for the marine environment is as follows:

[0218]

[0219] In the formula, CDS is the third scale component, CDY is the first scale component, CD is the mean filter of the first scale component, and δ4, δ5 and δ6 are the fourth weight parameter, the fifth weight parameter and the sixth weight parameter, respectively.

[0220] It should be noted that the fourth, fifth, and sixth weight parameters can be set by the user or obtained from historical data. Specifically, this embodiment of the invention does not limit the specific parameters.

[0221] It should be noted that the first-scale marine environment calculation model can perform targeted enhancements on the first-scale components. Mean filtering can be achieved using a Gaussian filter (with a window size approximately 1 / 5 of the first-scale component resolution). The fourth weight parameter (range [0.5, 1.5]) controls the enhancement intensity; larger values, such as 1.2, significantly enhance large-scale features. The fifth weight parameter (range [0.01, 0.1]) is a stability coefficient to prevent the denominator from being zero; it is typically set to 0.05. The sixth weight parameter (range [0.2, 1.0]) controls the decay rate of the exponential term; smaller values, such as 0.3, concentrate nonlinear enhancements more in regions with significant feature differences. This model effectively enhances important structural features in the large-scale components while suppressing noise, improving the reconstruction accuracy and physical consistency of large-scale marine environment models.

[0222] S34, optimize the second scale component to obtain the fourth scale component;

[0223] It should be noted that the above optimization process can be performed using conventional techniques such as nonlinear enhancement filtering, adaptive detail preservation algorithm, and high-frequency information reconstruction technology, or it can be obtained using a second-scale calculation model of the marine environment. Specifically, the embodiments of the present invention do not limit the specifics.

[0224] The second-scale calculation model for the marine environment is as follows:

[0225]

[0226] In the formula, C is the fourth scale component, D is the second scale component, DQ is the regional average value of the second scale component, DL is the preset reference scale component, and δ7, δ8 and δ9 are the seventh weight parameter, the eighth weight parameter and the ninth weight parameter, respectively.

[0227] It should be noted that the seventh, eighth, and ninth weight parameters can be set by the user or obtained from historical data. Specifically, this embodiment of the invention does not limit the specific parameters.

[0228] It should be noted that the second-scale marine environment calculation model can adaptively optimize high-frequency detail components. The regional average is obtained by dividing the second-scale component into several sub-regions (e.g., a 10×10 grid) and calculating the average value of each sub-region. The preset reference scale component is typically constructed using the statistical characteristics of high-frequency components from high-resolution data of the same region in historical data. The seventh weight parameter (range [0.1, 0.5]) controls the degree of global detail enhancement; a larger value, such as 0.4, enhances the overall texture. The eighth weight parameter (range [0.3, 1.0]) controls the dependence on the reference scale; a larger value, such as 0.8, emphasizes historical statistical characteristics. The ninth weight parameter (range [0.05, 0.2]) is a sensitivity parameter for the nonlinear function; a smaller value, such as 0.08, concentrates the enhancement on specific regions. This model can effectively reconstruct the subtle structures and local changes in marine environmental data, enhancing the expressiveness of high-resolution features while maintaining physical plausibility.

[0229] It should be noted that the regional average value can be obtained through the adaptive zoning statistical method, and the preset reference scale component can be obtained through a hybrid modeling method that combines historical sample statistics with physical constraints. First, high-frequency component features under similar conditions, seasons and weather conditions in the same region are extracted from high-quality historical CMEMS datasets. A statistical model (such as EOF analysis or principal component analysis) is established to extract typical patterns. Then, physical constraints are applied in conjunction with ocean dynamic equations (such as shallow water equations, geostrophic balance, etc.) to make adjustments, ensuring that the reference component conforms to ocean physical laws. Specifically, the embodiments of the present invention are not limited.

[0230] S35, the third-scale component and the fourth-scale component are reconstructed to obtain target marine environmental data information.

[0231] It should be noted that the reconstruction process employs either inverse wavelet transform or Laplace pyramid reconstruction algorithms; the specific implementation of this invention is not limited to these methods. The enhanced and optimized third-scale component (the enhanced low-frequency component) and fourth-scale component (the optimized high-frequency component) are fused and combined to reconstruct complete marine environmental data. This multi-scale decomposition and reconstruction method enables the system to independently optimize marine environmental data features at different scales, while maintaining the physical correlation and overall consistency between features at each scale during the reconstruction phase. The resulting target marine environmental data possesses both the high temporal resolution (1 hour) of ERA5 data and the high spatial resolution (0.083°) of CMEMS data, while exhibiting richer and more accurate multi-scale feature representation, significantly enhancing the practical value and application potential of the downscaling results.

[0232] As can be seen, the spatiotemporal downscaling method based on marine environmental data described in the embodiments of the present invention can generate high-resolution data while taking into account the complexity and nonlinear characteristics of the marine environment, thereby providing more accurate data support in marine environment simulation and prediction.

[0233] See Figure 2-4 ,in, Figure 2 This is a statistical chart showing the model training results and measured data accuracy of the spatiotemporal downscaling method of the present invention. Figure 3 This is a comparison chart of the results before and after downscaling at different times in the spatiotemporal downscaling method of the present invention. Figure 4 This is a spatiotemporal downscaling result image for September 2024 in an embodiment of the present invention. (Through...) Figures 2-4 It can be seen that the spatiotemporal downscaling method of this invention exhibits significant advantages in processing marine environmental data: Figure 2 The accuracy statistics chart shows that the model output and the measured data are in high agreement; Figure 3 The comparison chart visually demonstrates the quality improvement before and after downscaling, with the processed data exhibiting richer spatial details and more accurate local features; Figure 4 The complete downscaling results in September 2024 fully verified the effectiveness of this method in practical applications, and fully demonstrated the excellent performance and practical value of this invention in improving the spatiotemporal resolution of marine environmental data while maintaining physical consistency.

[0234] Example 2

[0235] Please see Figure 5 , Figure 5 This is a schematic diagram of a spatiotemporal downscaling device based on marine environmental data disclosed in an embodiment of the present invention. Figure 5 The described spatiotemporal downscaling device based on marine environmental data is applied to a spatiotemporal downscaling optimization system based on marine environmental data, such as a local server or cloud server for spatiotemporal downscaling based on marine environmental data. This invention does not limit the application to such applications. Figure 5 As shown, the spatiotemporal downscaling device based on marine environmental data includes:

[0236] The acquisition module 201 is used to acquire marine environmental data information to be processed, marine environmental training dataset, and marine environmental real dataset;

[0237] The first computing module 202 is used to process the marine environment training dataset and the marine environment real dataset to obtain a generator model;

[0238] The second calculation module 203 is used to process the marine environmental data information to be processed using the generator model to obtain the target marine environmental data information.

[0239] As can be seen, the spatiotemporal downscaling device based on marine environmental data described in the embodiments of the present invention can generate high-resolution data while taking into account the complexity and nonlinear characteristics of the marine environment, thereby providing more accurate data support in marine environment simulation and prediction.

[0240] Example 3

[0241] Please see Figure 6 , Figure 6 This is a schematic diagram of another spatiotemporal downscaling device based on marine environmental data disclosed in an embodiment of the present invention. Figure 6 The described spatiotemporal downscaling device based on marine environmental data is applied to a spatiotemporal downscaling optimization system based on marine environmental data, such as a local server or cloud server for spatiotemporal downscaling based on marine environmental data. This invention does not limit the application to such applications. Figure 6 As shown, the spatiotemporal downscaling device based on marine environmental data includes:

[0242] Processor 301;

[0243] A memory 302 containing executable program code is coupled to the processor 301;

[0244] The processor 301 calls the executable program code stored in the memory 302 to execute some or all of the steps of the spatiotemporal downscaling method based on marine environmental data in Embodiment 1.

[0245] As can be seen, the spatiotemporal downscaling device based on marine environmental data described in the embodiments of the present invention can generate high-resolution data while taking into account the complexity and nonlinear characteristics of the marine environment, thereby providing more accurate data support in marine environment simulation and prediction.

[0246] Example 4

[0247] This invention discloses a computer-readable storage medium storing computer instructions. When the computer instructions are invoked, they are used to execute some or all of the steps of the spatiotemporal downscaling method based on marine environmental data in Embodiment 1.

[0248] Example 5

[0249] This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform some or all of the steps in the spatiotemporal downscaling method based on marine environmental data described in Embodiment 1.

[0250] The system embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0251] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0252] Finally, it should be noted that the spatiotemporal downscaling method and apparatus based on marine environmental data disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A spatiotemporal downscaling method based on marine environmental data, characterized in that, The method comprises: S1, obtaining to-be-processed marine environment data information, a marine environment training data set, and a marine environment real data set; S2, processing the marine environment training data set and the marine environment real data set to obtain a generator model; S3, processing the to-be-processed marine environment data information by using the generator model to obtain target marine environment data information; S2 comprises: S21, performing down-sampling processing on the marine environment training data set by using a marine environment down-sampling model to obtain a preprocessed marine environment training data set; The marine environment down-sampling model is: ； ； wherein is the pre-processed marine environment training data set, is pre-processed marine environment training data information of the pre-processed marine environment training data set at a time point t, is a weight coefficient, , is the marine environment training data set, t is a time point in the marine environment training data set, is marine environment training data information of the marine environment training data set at a time point t, is a time interval on a time resolution of the marine environment training data set, is a number of time points in the marine environment training data set; S22, processing the preprocessed marine environment training data set and the marine environment real data set to obtain a first marine environment training data set and a first marine environment real data set; S23, processing the first marine environment training data set and the first marine environment real data set to obtain a generator model, comprising: S231, presetting s=1; S232, processing the first marine environment training data set and the first marine environment real data set by using a generator initial model and a discriminator initial model to obtain a loss function value; S233, determining whether the loss function value is less than a preset loss function threshold to obtain a first determination result; When the first determination result is yes, determining that the generator initial model is the generator model, and performing S3; When the first determination result is no, updating parameters of the generator initial model and the discriminator initial model according to the loss function value to obtain an updated generator initial model and an updated discriminator initial model; S234, determining that the updated generator initial model is the generator initial model and determining that the updated discriminator initial model is the discriminator initial model; S235, increasing s by 1 and performing S232; S232 comprises: S2321, training the first marine environment training data set by using the generator initial model to obtain generator data information; S2322, processing the generator data information and the first marine environment real data set by using the discriminator initial model to obtain discrimination result information; S2323, performing calculation processing on the generator data information and the first marine environment real data set to obtain marine environment data difference information; the marine environment data difference information comprises first difference information, second difference information, and third difference information; S2324, processing the discrimination result information and the marine environment data difference information to obtain a loss function value; The marine environment first difference calculation model is: The processing of the preprocessed marine environment training data set and the marine environment real data set to obtain the first marine environment training data set and the first marine environment real data set comprises: In the formula, is the first difference information, is the j2th feature value of the i2th sample in the generator data information, is the j2th feature value of the i2th sample in the first marine environment real data set, is the feature weight coefficient, N2 is the number of samples in the generator data information, and M2 is the number of feature values in any sample in the generator data information.

2. The spatiotemporal downscaling method based on marine environmental data according to claim 1, characterized in that, ​ S221, performing raster data vectorization processing on the pretreated marine environment training data set and the marine environment real data set to obtain a second marine environment training data set and a second marine environment real data set; S222, performing outlier processing on the second marine environment training data set and the second marine environment real data set to obtain a third marine environment training data set and a third marine environment real data set; S223, performing missing value processing on the third marine environment training data set and the third marine environment real data set to obtain a fourth marine environment training data set and a fourth marine environment real data set; S224, performing format unification processing on the fourth marine environment training data set and the fourth marine environment real data set to obtain a fifth marine environment training data set and a fifth marine environment real data set; S225, performing spatio-temporal alignment processing on the fifth marine environment training data set and the fifth marine environment real data set to obtain a first marine environment training data set and a first marine environment real data set.

3. The spatiotemporal downscaling method based on marine environmental data according to claim 1, characterized in that, The processing of the discrimination result information and the marine environment data difference information to obtain a loss function value comprises: processing the discrimination result information and the marine environment data difference information by using a marine environment loss function model to obtain a loss function value; The marine environment loss function model is: In the formula, is the loss function value, is the i3th probability value in the discrimination result information, and are the first difference information, the second difference information and the third difference information respectively, and are a first weight factor, a second weight factor and a third weight factor respectively, and N3 is the number of the probability values in the discrimination result information, is a fourth weight factor.

4. The spatiotemporal downscaling method based on marine environmental data according to claim 1, characterized in that, The processing of the to-be-processed marine environment data information by using the generator model to obtain target marine environment data information comprises: S31, processing the to-be-processed marine environment data information by using the generator model to obtain marine environment data information; S32, performing multi-scale decomposition processing on the marine environment data information to obtain a first scale component and a second scale component; S33, performing enhancement processing on the first scale component by using a marine environment first scale calculation model to obtain a third scale component; The marine environment first scale calculation model is: wherein is the third scale component, is the first scale component, is the mean filtering of the first scale component, and are fourth, fifth and sixth weight parameters, respectively. S34, performing optimization processing on the second scale component to obtain a fourth scale component; S35, performing reconstruction processing on the third scale component and the fourth scale component to obtain target marine environment data information.

5. A spatiotemporal downscaling device based on marine environmental data, characterized in that, The device comprises: a processor; a memory coupled with the processor and storing executable program codes; The processor invokes the executable program codes stored in the memory to execute the spatio-temporal downscaling method based on marine environment data according to any one of claims 1-4.

6. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions, which when invoked, are used to execute the spatio-temporal downscaling method based on marine environment data according to any one of claims 1-4.

Images