A marine-oriented disaster forecast revision method, device, equipment and medium

Abstract

This invention relates to the interdisciplinary field of physical oceanography and artificial intelligence, and particularly to a method, apparatus, equipment, and medium for correcting ocean disaster forecasts. The method includes: acquiring GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing field; obtaining a global baseline forecast using a global ocean forecasting model; cropping a surface field image of a preset ocean region based on the global baseline forecast; acquiring a corrector, which is optimized using a weighted loss function. This weighted loss function guides the corrector to address different concerns in ocean disaster prevention and mitigation; and performing correction based on the surface field image of the preset ocean region using the corrector to obtain a corrected baseline forecast for the preset ocean region. By correcting according to various concerns in ocean disaster prevention and mitigation, the accuracy of disaster prediction for the preset ocean region is improved.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of physical oceanography and artificial intelligence, and in particular to a method, apparatus, equipment and medium for correcting marine disaster forecasts. Background Technology

[0002] Frequent typhoons, storm surges, and marine heat waves pose serious threats to the lives and property of coastal residents and the safety of marine engineering projects. Accurate prediction of multivariate fields such as sea surface temperature (SST), sea surface salinity (SSS), meridional velocity (VO), zonal velocity (UO), and sea surface height (SSH) can enhance early warning capabilities for marine disasters.

[0003] There are two existing types of disaster prediction: One approach is to simulate ocean processes by solving fluid dynamics and thermodynamic equations. However, this method is extremely inefficient and computationally expensive, making it difficult to meet the rapid forecasting needs of sudden disasters such as typhoons and thus hindering the marine disaster prevention and mitigation emergency response capabilities.

[0004] Another approach is based on deep learning (large-scale ocean forecasting models), such as Xihe, Wenhai, and Fuxi. These models, trained on massive amounts of data, can quickly generate global-scale ocean forecasts. However, their forecast accuracy is insufficient in certain near-shore areas, especially under extreme weather conditions such as typhoons. They struggle to accurately capture dramatic changes in sea surface temperature and height, resulting in weak forecasting capabilities for secondary disasters such as storm surges and marine heat waves.

[0005] Therefore, how to accurately predict disasters in various marine areas is a technical problem that urgently needs to be solved. Summary of the Invention

[0006] In view of the above problems, the present invention provides a marine-oriented disaster forecasting correction method, apparatus, equipment and medium that overcomes or at least partially solves the above problems.

[0007] In a first aspect, the present invention provides a disaster forecast correction method for the ocean, comprising: Acquire GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing field; Based on GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing field, a global baseline forecast is obtained using a large global ocean forecasting model. Based on the global baseline forecast, a surface field image of a preset ocean region is obtained by cropping. A corrector is obtained, which is optimized by a weighted loss function. The weighted loss function is used to guide the corrector to focus on different concerns in marine disaster prevention and mitigation. Based on the surface field image of the preset ocean region, a corrector is used to correct the baseline forecast of the preset ocean region, and the corrected result is obtained.

[0008] Preferably, the preset ocean area surface field image includes: The superimposed images of sea surface salinity, sea surface temperature, zonal current, meridional current, and sea surface height.

[0009] Preferably, the corrector comprises: The feature extraction and encoding processing layer is used to perform image segmentation, image block merging, and deep semantic feature extraction operations on the surface field image of the preset ocean region to obtain a feature map. The prediction correction layer is used to learn the dependency relationship between any two deep semantic features on the feature map through a multi-head attention mechanism, and further refine the features through a multilayer perceptron to obtain correction values. The correction values ​​are then superimposed on the surface field image of the preset ocean region through residual connections to obtain the baseline prediction correction result for the preset ocean region.

[0010] Preferably, the weighted loss function is determined as follows: Obtain physical prior weights, semantic attention weights, and variable adaptive weights; The comprehensive weight is obtained based on physical prior weights, semantic attention weights, and variable adaptive weights. Based on the combined weights and the original loss function, a weighted loss function is obtained.

[0011] Preferably, the comprehensive weight is obtained based on physical prior weights, semantic attention weights, and variable adaptive weights in the following manner: in, For comprehensive weighting, For physical prior weights, For semantic attention weights, For adaptive weights of variables, The first weighting coefficient of the physical prior weights. This is the second weight coefficient of the semantic attention weight. The third weighting coefficient is used for the adaptive weighting of variables.

[0012] Preferably, a weighted loss function is obtained based on the comprehensive weights and the original loss function, specifically as follows: in, For the baseline forecast correction results of the pre-defined marine area, The true values ​​for the GLORYS reanalysis of the predefined ocean region. For the purpose of navigating the sea route, The preset number of pixels for the ocean area.

[0013] Preferably, when obtaining the physical prior weights, the typhoon-affected area is determined based on the NCEI typhoon trajectory data, and the physical prior weights are increased to the corresponding typhoon-affected area. ,in, This refers to the distance from the typhoon. The strength coefficient, This is a scale parameter for controlling the spatial extent of the area affected by typhoons.

[0014] Secondly, the present invention also provides a disaster forecast correction device for the ocean, comprising: The first acquisition module is used to acquire GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing field. The first module is used to obtain a global baseline forecast based on GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing field through a global ocean forecasting model. The cropping module is used to crop a surface field image of a preset ocean region based on the global baseline forecast. The second acquisition module is used to acquire the Swin-Transformer corrector, which is obtained by optimizing the corrector with a weighted loss function. The weighted loss function is used to guide the corrector to different concerns in marine disaster prevention and mitigation. The second module is used to correct the baseline forecast of the preset ocean region based on the surface field image of the preset ocean region using a corrector, so as to obtain the correction result of the baseline forecast of the preset ocean region.

[0015] Thirdly, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in the first aspect.

[0016] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0017] One or more technical solutions in the embodiments of the present invention have at least the following technical effects or advantages: This invention provides a disaster forecast correction method for the ocean, comprising: acquiring GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing field; obtaining a global baseline forecast using a global ocean forecasting model; cropping a surface field image of a preset ocean region based on the global baseline forecast; acquiring a corrector, which is optimized using a weighted loss function. This weighted loss function guides the corrector to address different concerns in ocean disaster prevention and mitigation; and performing correction based on the surface field image of the preset ocean region using the corrector to obtain a correction result for the baseline forecast of the preset ocean region. By correcting according to various concerns in ocean disaster prevention and mitigation, the accuracy of disaster prediction for the preset ocean region is improved. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating the disaster forecast correction method for the ocean in an embodiment of the present invention is shown; Figure 2 A schematic diagram of the corrector structure in an embodiment of the present invention is shown; Figure 3 A comparative schematic diagram of the root mean square error of sea surface temperature in the typhoon Haikui-affected area is shown in an embodiment of the present invention. Figure 4 A schematic diagram of the structure of a disaster forecast correction device for the ocean in an embodiment of the present invention is shown; Figure 5 A schematic diagram of a computer device for implementing a disaster forecast correction method for the ocean is shown in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0020] Rather, these embodiments are provided so that the invention may be more thoroughly understood and its scope may be fully conveyed to those skilled in the art.

[0021] Example 1: Embodiments of the present invention provide a disaster forecast correction method for the ocean, such as... Figure 1 As shown, it includes: S101, acquire GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing field; S102, based on GLORYS reanalysis data, OSTIA sea surface temperature data and ERA5 atmospheric forcing field, obtains a global baseline forecast through a large global ocean forecasting model; S103, based on global baseline forecasts, cropped to obtain surface field images of a preset ocean region; S104, Obtain the corrector, which is obtained through optimization of the weighted loss function. The weighted loss function is used to guide the corrector to different concerns in marine disaster prevention and mitigation. S105, based on the surface field image of the preset ocean area, the correction is performed by a corrector to obtain the correction result of the baseline forecast of the preset ocean area.

[0022] First, in S101, these three types of data are all core reanalysis data in the ocean and atmosphere fields, which can be obtained through some official data service platforms and are known and available data.

[0023] Next, S102 was executed, and based on GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing field, a global baseline forecast was obtained using a large global ocean forecasting model.

[0024] Specifically, the large-scale global ocean forecasting model can be a global deep learning model such as Xihe, Wenhai, or Fuxi. These models are trained using historical ocean data and can predict global baseline forecasts based on GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing fields.

[0025] After obtaining a global baseline forecast using a large global ocean forecast model, step S103 is executed to crop a surface field image of a preset ocean region based on the global baseline forecast.

[0026] The preset ocean region here is specifically defined by cropping, such as the South China Sea region or the East China Sea region, etc. The surface field image specifically refers to the image corresponding to the spatial distribution of temperature field, salinity field, height field, and meridional and zonal currents in the uppermost water body of the preset ocean region.

[0027] Each variable field's spatial distribution corresponds to a primary color, thus dividing the surface field image into 5 images.

[0028] Therefore, the pre-defined ocean surface field image includes: superimposed sea surface salinity field image, sea surface temperature field image, zonal current field image, meridional current field image, and sea surface height field image.

[0029] After cropping to obtain the surface field image of the preset marine area, S104 is executed to obtain the corrector. The corrector is obtained by optimizing the weighted loss function, which is used to guide the corrector to focus on different concerns in marine disaster prevention and mitigation.

[0030] Below, as Figure 2 The diagram shows the specific structure of the corrector. This corrector is based on the Swin-Transformer, meaning it integrates a Swin-Transformer backbone network. This optimizes the prediction of key ocean variables and improves prediction accuracy, especially under extreme weather conditions such as typhoons.

[0031] The Swin-Transformer-based corrector (STC) includes a feature extraction and encoding layer, and a prediction correction layer. The feature extraction and encoding layer performs image segmentation, image patch merging, and deep semantic feature extraction on the surface field image of a preset ocean region to obtain a feature map. This prediction correction layer is used to learn the dependency relationship between any two deep semantic features on the feature map through a multi-head attention mechanism, and refine the features through a multilayer perceptron to obtain correction values. The correction values ​​are then superimposed on the surface field image of the preset ocean region through residual connections to obtain the baseline prediction correction result for the preset ocean region.

[0032] The processing procedures for these two structural layers are described in detail below: The feature extraction and encoding processing layer, specifically the Swin-Transformer backbone network, extracts features from the input data. The input data consists of a predefined surface field image of a marine region. The input data is represented as (B, C, H, W), which includes five key ocean variables: sea surface salinity, sea surface temperature, zonal current, meridional current, and sea surface height. B represents the batch size, C represents the number of channels (in this case, 5, corresponding to the 5 ocean variables), H is the height of the resampled spatial network, and W is the width of the resampled spatial network. Then, through image block processing, the input data (B, C, H, W) is transformed into (B, D, H / 4, W / 4), where D is the hidden layer dimension of the corrector (i.e., the number of channels output after block processing).

[0033] After image segmentation, the system sequentially enters the first, second, third, and fourth layers. The first layer includes a linear layer and a first window attention layer; the second layer includes first image patch merging and second window attention; the third layer includes second image patch merging and third window attention; and the fourth layer includes third image patch merging and third window attention.

[0034] The linear layer projects the data into a suitable representation. Then, multiple window attention mechanisms in the first, second, third, and fourth layers capture local dependencies within each non-overlapping window. A relative positional bias is introduced into the attention calculation to enhance the corrector's perception of spatial relationships. The attention mechanism within each window is defined as follows: in, , For query, As key, For value, For feature dimension, The relative position bias matrix is ​​a learnable matrix, and using this relative position bias can significantly improve the performance of the corrector.

[0035] In the image patch merging from the second to the fourth layer, spatial resolution is progressively reduced, and higher-level features are captured at each layer. More complex features are gradually captured through window attention at each layer.

[0036] After multiple forward and backward propagations, the features of the highest-resolution first layer out of the four layers are used as the feature representation of the corrector. This first layer captures key low-level spatial features, such as local spatial models, edges, and gradient information, which are crucial for the accuracy of ocean forecasts. This feature not only contains rich spatial information but also maintains sufficient resolution to capture fine-grained changes in the ocean variable field.

[0037] Next, for the prediction correction layer, the input is the feature map output from the feature extraction and encoding processing layers, specifically... ,in, , After entering the forecast correction layer, spatial dimensions are flattened to form a labeled sequence, preparing for subsequent global context modeling. For the labeled sequence.

[0038] Next, normalization is performed, followed by the Multi-Head Attention (MHA) module. The MHA mechanism enables the corrector to learn the dependencies between any two labels in the entire space, thereby capturing long-range interactions, which are crucial for identifying consistent error patterns. This is the result of normalization.

[0039] This is the result after multiple heads have paid attention.

[0040] Next, normalization is performed, and the features are further refined using a multilayer perceptron with residual connections. This is the result of the second normalization.

[0041] To refine the features, It was then restructured into a two-dimensional structure and upsampled to restore the original input resolution. The result is to restore the original input resolution.

[0042] Finally, the correction value is obtained from this. : The corrected values ​​are then superimposed onto the original baseline forecast via residual concatenation to obtain the final baseline forecast correction result: in, As the baseline forecast correction result, The baseline forecast is the original input.

[0043] This corrector focuses on learning complex, nonlinear error residuals, which helps stabilize the training process and enhances the interpretability of the corrector. The end-to-end correction framework effectively combines the local window-based attention mechanism and the global self-attention mechanism in the Swin-Transformer backbone network.

[0044] Next, we will introduce the construction process of the weighted loss function: Obtain physical prior weights, semantic attention weights, and variable adaptive weights; The comprehensive weight is obtained based on physical prior weights, semantic attention weights, and variable adaptive weights.

[0045] The physical prior weights are determined based on prior knowledge of ocean physics. On the one hand, by focusing on high-impact nearshore areas and assigning them higher weights (such as 1.5 to 2.0 times), the corrector can focus on these densely populated and economically active nearshore areas during training, thereby improving its ability to forecast nearshore disasters such as storm surges and seawater intrusion.

[0046] On the other hand, focusing on historically high-risk disaster areas is based on the statistical error distribution of historical typhoon paths and red tide occurrence areas, which strengthens the corrector's learning of disaster-prone areas.

[0047] On the other hand, we focus on the dynamic typhoon path zone, which is dynamically generated based on the real-time typhoon forecast path, providing dynamic weight support for responding to typhoon disasters.

[0048] When obtaining the physical prior weights, for the typhoon-affected area, the physical prior weights need to be redefined. Specifically, based on the NCEI typhoon trajectory data, the typhoon-affected area is determined, and the physical prior weight for that area is increased to [a higher value]. ,in, This refers to the distance from the typhoon. An adjustable intensity coefficient is used to increase the physical prior weights within the controlled typhoon's influence area. The larger the value, the higher the weight of the typhoon center and its surrounding areas in the loss function. The model will pay more attention to the forecast error of these extreme weather areas during training, thereby improving the ability to correct changes in marine elements under the influence of typhoons. This is a scale parameter used to control the spatial extent of the typhoon's influence area, adjusting the decay rate of physical prior weights with distance from the typhoon center. The larger the typhoon, the gentler its attenuation, and the wider its effective impact area; The smaller the physical prior weight, the faster the decay, and the physical prior weight is concentrated only in a small area near the typhoon center. By increasing the physical prior weight, the corrector can pay attention to the error in the typhoon-affected area during training, thus solving the problem of marine disaster prevention and mitigation under extreme weather conditions such as typhoons.

[0049] Semantic attention weights are generated by passing the first-layer high-resolution feature map of the corrector through a lightweight attention head, and automatically learning regions where the current prediction error may be large (such as fronts and vortex edges).

[0050] Adaptive weighting of variables is used because different ocean variable fields converge at different speeds in the early stages of training. This is achieved by adjusting the adaptive weights of the variable dimensions to avoid the problem of sea surface temperature field dominating the loss and sea surface height converging too slowly.

[0051] The weights determined from multiple points of interest are used to obtain the comprehensive weight as follows: in, For comprehensive weighting, For physical prior weights, For semantic attention weights, For adaptive weights of variables, The first weighting coefficient of the physical prior weights. This is the second weight coefficient of the semantic attention weight. The third weighting coefficient is used for the adaptive weighting of variables.

[0052] in, , as well as The allocation is based on the degree of attention, with a focus on physical prior weights; accordingly, The value is set relatively high. + + =1.

[0053] After obtaining the comprehensive weights, the weighted loss function is derived based on the comprehensive weights and the original loss function, specifically as follows: in, For the baseline forecast correction results of the pre-defined marine area, The true values ​​for the GLORYS reanalysis of the predefined ocean region. For the purpose of navigating the sea route, The preset number of pixels for the ocean area.

[0054] Optimizing the corrector according to the weighted loss function can improve the correction accuracy of the corrector, thereby enabling the corrector to focus the baseline prediction on various concerns of marine disaster prevention and mitigation.

[0055] Finally, two standard metrics can be used to evaluate the correction performance of the corrector: root mean square error and bias. These metrics are calculated on the effective grid points within the mask area.

[0056] The root mean square error (RMSE) is used to quantify the average magnitude of the forecast error. A lower RMSE indicates higher forecast accuracy. The formula for calculating the RMSE for the corrected result field and the actual field is as follows: in, For the corrected result field, For the real scene, The number of effective grid points is used to evaluate the correction performance of a single ocean variable field by calculating the root mean square error of each channel.

[0057] Bias measures the systematic deviation of the corrected result from the true value. The ideal value is zero. A positive bias indicates that the corrected result is too high, and a negative bias indicates that the corrected result is too low. The formula for calculating bias is: The summation process also applies to valid grid points, and this metric helps identify whether the corrector has a consistent tendency to overestimate or underestimate.

[0058] Finally, step S105 is executed, and based on the surface field image of the preset ocean region, the correction is performed by the corrector to obtain the correction result of the baseline forecast for the preset ocean region.

[0059] The trained corrector is then used to input the surface field image of the preset ocean region (such as the South China Sea region) acquired at this time into the corrector, thereby obtaining the correction result of the baseline forecast for the South China Sea region.

[0060] By focusing on the South China Sea (SCS) region, we will conduct an experiment using this region as an example.

[0061] The South China Sea, located in the semi-enclosed marginal sea of ​​the Northwest Pacific, is characterized by strong monsoon forcing, complex ocean depth measurements, and active mesoscale variability. Within this regional window, sea surface variables were corrected for five core ocean state variables: sea surface salinity (SSS), sea surface temperature (SST), zonal currents (UO), meridional currents (VO), and sea surface height (SSH). This region frequently experiences intense air-sea interactions and the influence of tropical cyclones, providing a challenging and practically significant testing ground for improving sea surface ocean forecasting techniques.

[0062] First, generate a global XiHe forecast for the period 2021–2023, configuring the required initial and forcing fields, including: GLORYS12 Reanalysis Field (a next-generation version of GLORYS reanalysis data): GLORYS12, a global ocean physics reanalysis product provided by the Copernicus Ocean Service, is used as the primary ocean initial conditions. This dataset has a horizontal resolution of 1 / 12° (approximately 8 km) and 50 vertical strata (ranging from 0.49 m to 573.57 m). It provides a globally gridded field of fundamental variables, including sea surface height, sea surface salinity, potential temperature, salinity, and the zonal and meridional components of ocean currents, thereby extracting the following sea surface variables: sea surface height, sea surface salinity, sea surface temperature, zonal currents, and meridional currents. These variables serve as the primary physics fields. The GLORYS12 reanalysis field is also used as a reference "true value" for model evaluation because it integrates various observational data through data assimilation, representing the best estimate of ocean state.

[0063] OSTIA Sea Surface Temperature: The Operational Sea Surface Temperature and Ice Analysis (OSTIA) product, produced by the UK Met Office, is used as an independent sea surface temperature forcing field. OSTIA provides daily global gridded sea surface temperature analysis at a resolution of 1 / 20° (approximately 5 km) by integrating satellite and field observation data. It incorporates XiHe inputs to provide an accurate characterization of sea surface thermal boundary conditions.

[0064] ERA5 Atmospheric Forcing Field: Sea surface wind forcing was provided by data from the ERA5 global atmospheric reanalysis from the European Centre for Medium-Range Weather Forecasts (ECMWF). We used 10-meter zonal and meridional wind components with a raw horizontal resolution of 31 km. The gridded wind field is crucial for driving the mixing processes in the ocean surface circulation.

[0065] By formatting this data and generating global forecasts using the XiHe model, and then cropping the global forecasts, we obtain surface field images of the South China Sea, including superimposed sea surface salinity field images, sea surface temperature field images, zonal current field images, meridional current field images, and sea surface height field images.

[0066] The surface field image of the South China Sea region obtained above is input into the corrector to obtain the correction result of the baseline forecast for the South China Sea region.

[0067] The correction was compared with seven baseline models to assess its accuracy. These baseline models included the original XiHe forecast and models that applied different correction techniques, specifically: XiHe: It uses a marine-specific Transformer architecture with a marine-land masking mechanism to generate predictions, and the prediction results are not corrected for regional bias.

[0068] WenHai: Another modern deep learning-driven global ocean forecasting model, developed to forecast eddy oceans using deep neural networks and generate multivariate global ocean forecasts.

[0069] Empirical quantile mapping (EQM): A statistical bias correction baseline method that adjusts the distribution of XiHe forecasts to better match reference reanalysis data. It performs correction by mapping the empirical quantiles of the forecast field to the empirical quantiles of the target field and is widely used as a simple and effective post-processing method.

[0070] PredRNN: A deep learning-based recurrent neural network (RNN) model designed specifically for spatiotemporal forecasting. It excels at capturing complex spatiotemporal features and using these learned patterns to refine XiHe forecasts.

[0071] BiLSTM: A bidirectional long short-term memory network model that processes sequences from two directions, enabling it to learn temporal and spatial dependencies more effectively. This allows it to correct XiHe forecasts by identifying and adjusting spatiotemporal patterns in ocean variables.

[0072] ConvGRU combines convolutional layers for spatial feature extraction with GRU units for temporal modeling. The convolutional layers enhance spatial feature learning, while the GRU units capture temporal evolution, enabling the model to handle both dimensions simultaneously.

[0073] CNN: Employs a convolution-based baseline correction method to perform spatial post-processing on XiHe forecasts. It uses convolution operations to learn local spatial patterns and refine the regional forecast field.

[0074] These baseline models provide complementary references from different methodological perspectives, including original deep learning prediction systems, recurrent spatiotemporal correction, simple convolutional correction, and Transformer-based region correction. The experimental results obtained are as follows:

[0075] As shown in the table, the correction method employed in this invention significantly improves upon the baseline forecast (XiHe) in all five key indicators. The largest absolute improvement is observed in sea surface temperature (SST), with the corresponding root mean square error (RMSE) decreasing from 0.7712 to 0.5810. The largest relative improvement is seen in sea surface height (SSH), with the corresponding RMSE decreasing from 0.0687 to 0.0428, representing a relative improvement of 37.70%. Furthermore, substantial progress is demonstrated in bias, significantly outperforming the performance of other correction models. The original XiHe forecast performed worse than WenHai in the South China Sea region. However, after applying the corrector (STC) of this invention, the corrected XiHe forecast outperforms WenHai overall. This result further demonstrates the effectiveness of the STC (in this invention) scheme, indicating that it can significantly improve the regional forecast skill of the baseline global forecast system.

[0076] This invention not only improves correction accuracy but also maintains a significant computational efficiency advantage. Specifically, under the same testing environment, XiHe completed global inference over 10 days in just 3.17 seconds, while traditional numerical oceanographic forecasting systems like PSY4 typically require several hours to extrapolate ocean conditions. The STC scheme provided by this invention requires only 0.46 seconds for daily forecasts. Therefore, this invention can improve forecasting skills for a predetermined region with minimal additional computation time, making it highly efficient and practical for real-world applications.

[0077] Furthermore, the frequent occurrence of typhoons in the South China Sea region has a significant impact on local livelihoods and economic activities. Due to the complex interaction between typhoons and ocean dynamic processes, accurate sea surface temperature (SST) forecasting under such extreme weather conditions has always been a challenging task. Therefore, we examine the performance of our corrector (STC) under typhoon conditions using Typhoon Haikui in 2023 as an example. Typhoon Haikui was a significant event affecting the South China Sea region in September 2023. We obtained typhoon track data for 2023 from the National Center for Environmental Information (NCEI). Specifically, we identified records of Typhoon Haikui in the relevant regions. For each day, multiple time points existed; specific latitude and longitude of the typhoon were selected, and the affected area was defined by extending the region outward by 1 / 6 degree. The union of these regions within a day was used to calculate the root mean square error (RMSE) of the SST forecast.

[0078] As shown in Figure 3, the corrector (STC) employed in this invention significantly reduces the root mean square error (RMSE) of sea surface temperature (SST) predictions within typhoon-affected areas. Notably, on September 3, 2023, the corrector (STC) of this invention achieved a substantial reduction in RMSE, far exceeding the performance of other correction models. This indicates that the corrector (STC) of this invention can effectively identify and adapt to the changing patterns of sea surface temperature (SST) caused by typhoons. Conversely, on September 2, 2023, the corrected forecasts generated by the BiLSTM and ConvGRU models were worse than the original XiHe forecast, indicating that they failed to accurately identify the typhoon-driven patterns, leading to erroneous corrections. This comparison further highlights the superiority of the corrector (STC) of this invention in handling extreme weather scenarios.

[0079] One or more technical solutions in the embodiments of the present invention have at least the following technical effects or advantages: This invention provides a disaster forecast correction method for the ocean, comprising: acquiring GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing field; obtaining a global baseline forecast using a global ocean forecasting model; cropping a surface field image of a preset ocean region based on the global baseline forecast; acquiring a corrector, which is optimized using a weighted loss function. This weighted loss function guides the corrector to address different concerns in ocean disaster prevention and mitigation; and performing correction based on the surface field image of the preset ocean region using the corrector to obtain a correction result for the baseline forecast of the preset ocean region. By correcting according to various concerns in ocean disaster prevention and mitigation, the accuracy of disaster prediction for the preset ocean region is improved.

[0080] Example 2 Based on the same inventive concept, embodiments of the present invention provide a disaster forecast correction device for the ocean, such as... Figure 4 As shown, it includes: The first acquisition module 401 is used to acquire GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing field. The first module 402 is used to obtain a global baseline forecast based on GLORYS reanalysis data, OSTIA sea surface temperature data and ERA5 atmospheric forcing field through a global ocean forecasting model. The cropping module 403 is used to crop a surface field image of a preset ocean region based on the global baseline forecast. The second acquisition module 404 is used to acquire a corrector, which is obtained by optimization of a weighted loss function. The weighted loss function is used to guide the corrector to different concerns in marine disaster prevention and mitigation. The second module 405 is used to correct the preset ocean area surface field image using a corrector to obtain the baseline forecast correction result for the preset ocean area.

[0081] In one optional implementation, the preset ocean area surface field image includes: The superimposed images of sea surface salinity, sea surface temperature, zonal current, meridional current, and sea surface height.

[0082] In one alternative implementation, the corrector includes: The feature extraction and encoding processing layer is used to perform image segmentation, image block merging, and deep semantic feature extraction operations on the surface field image of the preset ocean region to obtain a feature map. The prediction correction layer is used to learn the dependency relationship between any two deep semantic features on the feature map through a multi-head attention mechanism, and further refine the features through a multilayer perceptron to obtain correction values. The correction values ​​are then superimposed on the surface field image of the preset ocean region through residual connections to obtain the baseline prediction correction result for the preset ocean region.

[0083] In one optional implementation, the weighted loss function is determined as follows: Obtain physical prior weights, semantic attention weights, and variable adaptive weights; The comprehensive weight is obtained based on physical prior weights, semantic attention weights, and variable adaptive weights. Based on the combined weights and the original loss function, a weighted loss function is obtained.

[0084] In one alternative implementation, the weighted loss function obtains the comprehensive weights as follows: in, For comprehensive weighting, For physical prior weights, For semantic attention weights, For adaptive weights of variables, The first weighting coefficient of the physical prior weights. This is the second weight coefficient of the semantic attention weight. The third weighting coefficient is used for the adaptive weighting of variables.

[0085] In one alternative implementation, the weighted loss function is obtained as follows: in, For the baseline forecast correction results of the pre-defined marine area, The true values ​​for the GLORYS reanalysis of the predefined ocean region. For the purpose of navigating the sea route, The preset number of pixels for the ocean area.

[0086] In one alternative implementation, the physical prior weights are adjusted as follows: Based on NCEI typhoon trajectory data, the typhoon-affected areas are determined. For each typhoon-affected area, the physical prior weights are increased to [a specific value]. ,in, This represents the distance from the typhoon.

[0087] Example 3

[0088] Based on the same inventive concept, embodiments of the present invention provide a computer device, such as... Figure 5 As shown, it includes a memory 504, a processor 502, and a computer program stored in the memory 504 and executable on the processor 502. When the processor 502 executes the program, it implements the steps of the above-described marine disaster forecast correction method.

[0089] Among them, Figure 5 In this document, a bus architecture (represented by bus 500) is used. Bus 500 may include any number of interconnected buses and bridges, linking various circuits including one or more processors represented by processor 502 and memory represented by memory 504. Bus 500 may also link various other circuits such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and therefore will not be described further herein. Bus interface 506 provides an interface between bus 500 and receiver 501 and transmitter 503. Receiver 501 and transmitter 503 may be the same element, i.e., a transceiver, providing a unit for communicating with various other devices over a transmission medium. Processor 502 is responsible for managing bus 500 and general processing, while memory 504 can be used to store data used by processor 502 during operation.

[0090] Example 4

[0091] Based on the same inventive concept, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described marine disaster forecast correction method.

[0092] The algorithms and displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, this invention is not directed to any particular programming language. It should be understood that the contents of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0093] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0094] Similarly, it should be understood that, in order to simplify the invention and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of the invention, various features of the invention are sometimes grouped together in a single embodiment, figure, or description thereof. However, this disclosure should not be construed as reflecting an intention that the claimed invention requires more features than are explicitly recited in each embodiment. Rather, as reflected in each embodiment, inventive aspects lie in fewer than all features of the single foregoing disclosed embodiment. Therefore, the claims, following the detailed description, are hereby expressly incorporated into this detailed description, wherein each claim itself is a separate embodiment of the invention.

[0095] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0096] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of the invention and form different embodiments. For example, in the specific implementation, any of the claimed embodiments can be used in any combination.

[0097] The various component embodiments of the present invention can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some or all of the components in the marine disaster forecasting correction device or computer equipment according to embodiments of the present invention. The present invention can also be implemented as a device or apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such programs implementing the present invention can be stored on a computer-readable medium or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0098] It should be noted that the above embodiments are illustrative of the invention and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names.

Claims

1. A disaster forecast correction method for the ocean, characterized in that, include: Acquire GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing field; Based on GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing field, a global baseline forecast is obtained using a large global ocean forecasting model. Based on the global baseline forecast, a surface field image of a preset ocean region is obtained by cropping. A corrector is obtained, which is optimized by a weighted loss function. The weighted loss function is used to guide the corrector to focus on different concerns in marine disaster prevention and mitigation. Based on the surface field image of the preset ocean region, a corrector is used to correct the baseline forecast of the preset ocean region, and the corrected result is obtained.

2. The method as described in claim 1, characterized in that, The preset ocean region surface field image includes: The superimposed images of sea surface salinity, sea surface temperature, zonal current, meridional current, and sea surface height.

3. The method as described in claim 1, characterized in that, The corrector includes: The feature extraction and encoding processing layer is used to perform image segmentation, image block merging, and deep semantic feature extraction operations on the surface field image of the preset ocean region to obtain a feature map. The prediction correction layer is used to learn the dependency relationship between any two deep semantic features on the feature map through a multi-head attention mechanism, and further refine the features through a multilayer perceptron to obtain correction values. The correction values ​​are then superimposed on the surface field image of the preset ocean region through residual connections to obtain the baseline prediction correction result for the preset ocean region.

4. The method as described in claim 1, characterized in that, The weighted loss function is determined as follows: Obtain physical prior weights, semantic attention weights, and variable adaptive weights; The comprehensive weight is obtained based on physical prior weights, semantic attention weights, and variable adaptive weights. Based on the combined weights and the original loss function, a weighted loss function is obtained.

5. The method as described in claim 4, characterized in that, Based on physical prior weights, semantic attention weights, and variable adaptive weights, the comprehensive weights are obtained as follows: ； in, For comprehensive weighting, For physical prior weights, For semantic attention weights, For adaptive weights of variables, The first weighting coefficient of the physical prior weights. This is the second weight coefficient of the semantic attention weight. The third weighting coefficient is used for the adaptive weighting of variables.

6. The method as described in claim 5, characterized in that, Based on the combined weights and the original loss function, the weighted loss function is obtained as follows: ； in, For the baseline forecast correction results of the pre-defined marine area, The true values ​​for the GLORYS reanalysis of the predefined ocean region. For the purpose of navigating the sea route, The preset number of pixels for the ocean area.

7. The method as described in claim 4, characterized in that, When obtaining the physical prior weights, based on the NCEI typhoon trajectory data, the typhoon-affected area is determined, and the physical prior weights are increased to the corresponding typhoon-affected areas. ,in, This refers to the distance from the typhoon. The strength coefficient, This is a scale parameter for controlling the spatial extent of the area affected by typhoons.

8. A disaster forecast correction device for the ocean, characterized in that, include: The first acquisition module is used to acquire GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing field. The first module is used to obtain a global baseline forecast based on GLORYS reanalysis data, OSTIA sea surface temperature data, and ERA5 atmospheric forcing field through a global ocean forecasting model. The cropping module is used to crop a surface field image of a preset ocean region based on the global baseline forecast. The second acquisition module is used to acquire a corrector, which is obtained by optimization of a weighted loss function. The weighted loss function is used to guide the corrector to different concerns in marine disaster prevention and mitigation. The second module is used to correct the baseline forecast of the preset ocean region based on the surface field image of the preset ocean region using a corrector, so as to obtain the correction result of the baseline forecast of the preset ocean region.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images