A mesoscale vortex occurrence probability prediction method and related device

Abstract

This invention discloses a method and related equipment for predicting the probability of mesoscale eddies. Using continuous ocean reanalysis data as the prediction input, and training the model with SAR-labeled samples, it can be independent of SAR imagery, thus achieving large-scale continuous monitoring of eddies in the Arctic marginal seas. A grid window sample is constructed based on ocean reanalysis data, enabling the model to learn the physical field characteristics associated with mesoscale eddies, effectively characterizing potentially high-incidence areas of eddies that are difficult to represent with traditional altimeters. The probability values ​​are output through a deep learning model, and a sliding window prediction strategy is used to generate a spatially continuous probability field, providing prior spatial information for subsequent fine SAR observations and dynamic process analysis. This method can be widely applied in the field of data processing technology.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and related equipment for predicting the probability of mesoscale vortex occurrence. Background Technology

[0002] Mesoscale eddies are important dynamic processes in the ocean, influencing the transport of heat, salinity, momentum, and biogeochemical substances. In the Arctic marginal seas, mesoscale eddies are closely related to boundary currents, frontal instability, sea ice changes, and water mass exchange. Therefore, large-scale, continuous monitoring of Arctic mesoscale eddies is of great significance.

[0003] Current methods for identifying mesoscale eddies primarily rely on synthetic aperture radar (SAR) imagery and gridded sea surface height data. SAR imagery offers high spatial resolution, clearly displaying sea surface roughness variations caused by eddies, making it suitable for detailed identification of eddy structures in local sea areas. However, SAR imagery is limited by satellite orbits, imaging swath width, and revisit periods, resulting in discontinuous spatiotemporal coverage and making it difficult to support large-scale, long-term eddy monitoring in the Arctic marginal seas.

[0004] Gridded sea surface height data offers extensive spatial coverage, making it suitable for analyzing large-scale ocean dynamic processes. However, traditional altimeter products have low effective spatial resolution, while mesoscale eddies in the high-latitude Arctic seas are typically small, with many eddies measuring only 10–40 km. This makes it difficult for traditional altimeter data to effectively resolve small-scale mesoscale eddies in the Arctic. Therefore, relying solely on altimeter data is insufficient to meet the high-resolution monitoring needs of mesoscale eddies in the Arctic marginal seas.

[0005] In addition, existing methods for manual visual interpretation and deep learning recognition based on SAR images still mainly rely on the high-resolution SAR images themselves. Their results are limited by the SAR coverage area and it is difficult to continuously provide the spatial probability distribution of vortex occurrence across the entire area. Summary of the Invention

[0006] The main objective of this invention is to provide a method, apparatus, electronic device, storage medium, and program product for predicting the probability of mesoscale vortex occurrence, aiming to solve at least one problem in the prior art.

[0007] To achieve the above objectives, one aspect of this invention proposes a method for predicting the probability of mesoscale vortex occurrence, the method comprising: Synthetic Aperture Radar (SAR) images of the study area are acquired, and the SAR images are preprocessed to obtain initial radar images; the SAR images are associated with the imaging date. In response to human visual interpretation of the target object, the spatial positions of vortex samples and non-vortex samples are marked in the initial radar image; Acquire ocean reanalysis data that matches spatial location and imaging date; the parameter types of ocean reanalysis data include sea surface height anomalies, sea surface temperature, sea surface salinity, and current velocity characteristics; Using spatial location as the center, a two-dimensional grid window with a preset spatial range for each parameter type is cropped from the ocean reanalysis data. The two-dimensional grid windows of all parameter types are stacked along the channel dimension to form a multi-channel two-dimensional physical field tensor. Each multi-channel two-dimensional physical field tensor is assigned a category label according to the vortex coverage ratio of the preset central region of the two-dimensional grid window, thus forming a training sample set. The preset central region is a preset region range for mesoscale vortices, and the category labels include positive vortex samples and negative non-vortex samples. The pre-set deep learning model is trained using a training sample set to obtain a vortex prediction model. The deep learning model includes a two-dimensional convolutional layer, a backbone network, an attention module, and a probability output layer. Based on ocean reanalysis data of the target area on the target date, a vortex occurrence probability field of the target area on the target date is generated using a sliding window prediction strategy based on a vortex prediction model.

[0008] In some implementations, preprocessing of synthetic aperture radar imagery includes at least one of the following steps: Radiometric calibration of synthetic aperture radar images; Thermal noise removal is performed on synthetic aperture radar images; Spot filtering is applied to synthetic aperture radar images; The incident angle of the synthetic aperture radar image is normalized.

[0009] In some implementations, acquiring ocean reanalysis data matched with spatial location and imaging date includes the following steps: Use the first marked spatial location as the target location; Acquire information on sea surface height anomalies, sea surface temperature, sea surface salinity, and upper ocean current velocity within a preset range near the target location on the corresponding imaging date; Among them, the upper ocean current velocity information is characterized by velocity profiles of multiple vertical layers; The velocity profile is divided into a surface layer and a subsurface layer, and then principal component analysis is used to extract the principal component features of the surface layer and the subsurface layer as velocity features. The next marked spatial location is used as the target location. The process then returns to the previous step of obtaining information on sea surface height anomalies, sea surface temperature, sea surface salinity, and upper ocean current velocity within a preset range near the target location on the corresponding imaging date, until all spatial locations have been traversed.

[0010] In some implementations, a class label is assigned to each multi-channel two-dimensional physics tensor based on the vortex coverage ratio of a preset central region of the two-dimensional grid window to form a training sample set, including the following steps: Use the first multi-channel two-dimensional physics tensor as a candidate input sample; Extract local images of the region corresponding to the two-dimensional grid window from the initial radar image corresponding to the candidate input sample; When the proportion of the vortex region in the preset central region of the local image is greater than the preset threshold, the category label of the positive vortex sample is associated with the candidate input sample to form a training sample. When the proportion of the vortex region in the preset central region of the local image is less than or equal to the preset threshold, the category label of the non-vortex negative sample is associated with the candidate input sample to form a training sample. The next multi-channel two-dimensional physical field tensor is used as a candidate input sample. The process of extracting the local image of the corresponding region of the two-dimensional grid window from the initial radar image corresponding to the candidate input sample is repeated until all multi-channel two-dimensional physical field tensors are traversed. All training samples are then combined to form a training sample set.

[0011] In some implementations, the method further includes the following steps: The first training sample in the training sample set is used as the candidate training sample. Randomly generate offsets within a preset grid range, and use these offsets to spatially offset the two-dimensional grid window corresponding to the candidate training sample to obtain the offset spatial range. Offset grid windows of the offset spatial range for each parameter type are cropped from ocean reanalysis data. All offset grid windows of all parameter types are stacked along the channel dimension to form a multi-channel augmented physical field tensor. The multi-channel augmented physical field tensor is assigned a category label according to the vortex coverage ratio of the preset central region of the offset grid window to form an augmented sample. Take the next training sample in the training sample set as a candidate training sample, and return to execute the step of randomly generating an offset within the preset grid range until all training samples in the training sample set have been traversed. All augmented samples are added to the training sample set, and then data augmentation operations are performed on all samples in the training sample set to update the training sample set.

[0012] In some implementations, a pre-defined deep learning model is trained using a training sample set to obtain a vortex prediction model, including the following steps: The first feature is obtained by performing channel mapping and initial feature extraction on the multi-channel two-dimensional physical field tensor through a two-dimensional convolutional layer; Based on the initial features, spatial features at different scales are extracted through the backbone network using the moving inverse residual convolution module, and the spatial features output by the last layer of the backbone network are used as the second features. Based on the second feature, the Q matrix, K matrix and V matrix are constructed through the attention module, and then the third feature is obtained after attention weighted aggregation and convolution. The second and third features are residually connected to obtain the fourth feature; Based on the fourth feature, the logic value is output through the probability output layer using global average pooling and fully connected layers, and then the logic value is converted into the probability of vortex occurrence through the sigmoid function; Based on the category label and vortex occurrence probability corresponding to the multi-channel two-dimensional physical field tensor of each training sample in the training sample set, the deep learning model is optimized by tuning the parameters of the binary cross-entropy loss function to obtain the vortex prediction model.

[0013] In some implementations, based on ocean reanalysis data of the target area on a target date, a vortex occurrence probability field for the target area on a target date is generated using a sliding window prediction strategy with a vortex prediction model, including the following steps: Ocean reanalysis data of the target area on the target date will be used as the data to be analyzed; A sliding window prediction strategy is adopted, in which the sliding window slides row by row and column by column on the spatial grid of the target area with a preset sliding step size, and the sub-data to be analyzed at each location of the target area is extracted from the data to be analyzed through the sliding window; The sub-data to be analyzed is input into the vortex prediction model so that the vortex prediction model can predict the vortex prediction probability value at each location in the target area, and then the data is summarized to obtain the vortex occurrence probability field of the target area on the target date.

[0014] To achieve the above objectives, another aspect of the present invention provides a mesoscale vortex occurrence probability prediction device, the device comprising: The first module is used to acquire synthetic aperture radar (SAR) images of the study area, preprocess the SAR images to obtain initial radar images; the SAR images are associated with the imaging date. The second module is used to mark the spatial positions of vortex samples and non-vortex samples in the initial radar image in response to human visual interpretation of the target object. The third module is used to acquire ocean reanalysis data that matches spatial location and imaging date; the parameter types of ocean reanalysis data include sea surface height anomalies, sea surface temperature, sea surface salinity, and current velocity characteristics. The fourth module is used to use the spatial location as the center to cut out a two-dimensional grid window with a preset spatial range for each parameter type from the ocean reanalysis data. The two-dimensional grid windows of all parameter types are stacked along the channel dimension to form a multi-channel two-dimensional physical field tensor. Each multi-channel two-dimensional physical field tensor is assigned a category label according to the vortex coverage ratio of the preset central region of the two-dimensional grid window, which constitutes a training sample set. The preset central region is a preset region range for mesoscale vortices, and the category labels include positive vortex samples and negative non-vortex samples. The fifth module is used to train a pre-defined deep learning model using a training sample set to obtain a vortex prediction model. The deep learning model includes a two-dimensional convolutional layer, a backbone network, an attention module, and a probability output layer. The sixth module is used to generate the vortex occurrence probability field of the target area on the target date based on ocean reanalysis data of the target area on the target date using a sliding window prediction strategy and a vortex prediction model.

[0015] To achieve the above objectives, another aspect of the present invention provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned method.

[0016] To achieve the above objectives, another aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0017] To achieve the above objectives, another aspect of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.

[0018] The embodiments of this invention include at least the following beneficial effects: This invention provides a method, apparatus, electronic device, storage medium, and program product for predicting the probability of mesoscale eddies. This scheme acquires synthetic aperture radar (SAR) images of the study area, preprocesses the SAR images to obtain initial radar images; wherein the SAR images are associated with an imaging date; in response to visual interpretation of the target object, the spatial locations of eddy samples and non-eddy samples are marked in the initial radar images; ocean reanalysis data matching the spatial location and imaging date is acquired; wherein the parameter types of the ocean reanalysis data include sea surface height anomalies, sea surface temperature, sea surface salinity, and current velocity characteristics; using the spatial location as the center, a preset spatial range for each parameter type is cropped from the ocean reanalysis data. A two-dimensional grid window is used to stack all parameter types of two-dimensional grid windows along the channel dimension to form a multi-channel two-dimensional physical field tensor. Each multi-channel two-dimensional physical field tensor is assigned a category label based on the vortex coverage ratio of a preset central region of the two-dimensional grid window, forming a training sample set. The preset central region is a preset area for mesoscale vortices, and the category labels include positive vortex samples and negative non-vortex samples. A preset deep learning model is trained using the training sample set to obtain a vortex prediction model. The deep learning model includes two-dimensional convolutional layers, a backbone network, an attention module, and a probability output layer. Based on ocean reanalysis data of the target region on the target date, the vortex prediction model uses a sliding window prediction strategy to generate a vortex occurrence probability field for the target region on the target date. This invention uses continuous ocean reanalysis data as predictive input. After training the model with SAR-annotated samples, it can be independent of SAR imagery, thereby achieving large-scale continuous monitoring of Arctic marginal sea eddies. Based on ocean reanalysis data, a grid window sample is constructed, enabling the model to learn the physical field characteristics associated with mesoscale eddies, thus effectively characterizing potential high-incidence areas of eddies that are difficult to represent with traditional altimeters. The probability values ​​are output through a deep learning model, and a sliding window prediction strategy is used to generate a spatially continuous probability field, providing prior spatial information for subsequent fine SAR observations and dynamic process analysis. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of an implementation environment for the method for predicting the probability of mesoscale vortex occurrence provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the method for predicting the probability of mesoscale vortex occurrence provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the overall technical flow of the mesoscale vortex occurrence probability prediction method provided in the embodiments of the present invention; Figure 4 This is a schematic diagram illustrating an example of verifying the consistency between the model prediction probability and the relative vorticity of ocean currents provided in this embodiment of the invention. Figure 5 This is a schematic diagram illustrating a comparison of typical regions between the model prediction results provided in this embodiment of the invention and SAR images; Figure 6 This is a schematic diagram illustrating a comparison between SWOT SSHA observation results and model-predicted vortex occurrence probability results provided in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating an example of a predicted probability density distribution based on a SWOT vortex mask provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of the mesoscale vortex occurrence probability prediction device provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of this invention; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this invention as detailed in the appended claims.

[0021] It is understood that the terms "first," "second," etc., used in this invention may be used to describe various concepts, but unless specifically stated otherwise, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of embodiments of this invention, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words "if" or "when" as used herein may be interpreted as "when," "in response to determination," or "in the event of a determination."

[0022] The terms “at least one,” “multiple,” “each,” “any,” etc., used in this invention, “at least one” includes one, two, or more than two; “multiple” includes two or more than two; “each” refers to each of the corresponding multiple; and “any” refers to any one of the multiple.

[0023] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this invention is for descriptive purposes only and is not intended to limit the invention.

[0024] In related technologies, existing methods mostly focus on identifying whether a vortex exists, and there is a lack of a method that can combine SAR labeled samples with continuous marine environmental data to output a large-scale, daily-scale, continuous probability field for predicting the probability of vortex occurrence.

[0025] In view of this, this invention provides a method and related equipment for predicting the probability of mesoscale eddies. This method acquires synthetic aperture radar (SAR) images of the study area, preprocesses these images to obtain initial radar images, and associates the SAR images with imaging dates. In response to visual interpretation of the target object, the spatial locations of eddy samples and non-eddy samples are marked in the initial radar images. Ocean reanalysis data matching the spatial locations and imaging dates is acquired. The parameters in the ocean reanalysis data include sea surface height anomalies, sea surface temperature, sea surface salinity, and current velocity characteristics. Using the spatial location as the center, a two-dimensional grid window with a preset spatial range for each parameter type is cropped from the ocean reanalysis data, and all parameters are... A multi-channel two-dimensional physical field tensor is formed by stacking two-dimensional grid windows of various types along the channel dimension. Each multi-channel two-dimensional physical field tensor is assigned a category label according to the vortex coverage ratio of the preset central region of the two-dimensional grid window, forming a training sample set. The preset central region is a preset area range for mesoscale vortices, and the category labels include positive vortex samples and negative non-vortex samples. The training sample set is used to train a preset deep learning model to obtain a vortex prediction model. The deep learning model includes a two-dimensional convolutional layer, a backbone network, an attention module, and a probability output layer. Based on the ocean reanalysis data of the target area on the target date, the vortex prediction model uses a sliding window prediction strategy to generate the vortex occurrence probability field of the target area on the target date. This invention uses continuous ocean reanalysis data as predictive input. After training the model with SAR-annotated samples, it can be independent of SAR imagery, thereby achieving large-scale continuous monitoring of Arctic marginal sea eddies. Based on ocean reanalysis data, a grid window sample is constructed, enabling the model to learn the physical field characteristics associated with mesoscale eddies, thus effectively characterizing potential high-incidence areas of eddies that are difficult to represent with traditional altimeters. The probability values ​​are output through a deep learning model, and a sliding window prediction strategy is used to generate a spatially continuous probability field, providing prior spatial information for subsequent fine SAR observations and dynamic process analysis.

[0026] It is understood that the mesoscale vortex occurrence probability prediction method provided by this invention can be applied to any computer device with data processing and computing capabilities, and this computer device can be various terminals or servers. When the computer device in the embodiment is a server, the server is an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Optionally, the terminal can be a smartphone, tablet, laptop, or desktop computer, but it is not limited to these.

[0027] like Figure 1 The diagram shown is a schematic representation of an implementation environment provided by an embodiment of the present invention. (Refer to...) Figure 1 The implementation environment includes at least one terminal 102 and a server 101. The terminal 102 and the server 101 can be connected via a network, either wirelessly or via a wired connection, to complete data transmission and exchange.

[0028] Server 101 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0029] Additionally, server 101 can also be a node server in a blockchain network. Blockchain is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.

[0030] Terminal 102 can be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. Terminal 102 and server 101 can be directly or indirectly connected via wired or wireless communication, and this embodiment of the invention does not impose any limitations.

[0031] For example, based on Figure 1 The implementation environment shown in this embodiment of the invention provides a method for predicting the probability of mesoscale vortex occurrence. The following description uses the application of this method in server 101 as an example. It can be understood that this method can also be applied in terminal 102.

[0032] Reference Figure 2 , Figure 2 This is an optional flowchart of the mesoscale vortex occurrence probability prediction method provided in the embodiments of the present invention. The execution subject of the mesoscale vortex occurrence probability prediction method can be any of the aforementioned computer devices (including servers or terminals). Figure 2 The method may include, but is not limited to, steps S100 to S600.

[0033] Step S100: Acquire synthetic aperture radar imagery of the study area, preprocess the synthetic aperture radar imagery to obtain initial radar imagery. Among them, the synthetic aperture radar images are associated with the imaging date; It should be noted that, in some embodiments, preprocessing of synthetic aperture radar (SAR) images may include at least one of the following steps: radiometric calibration of SAR images; thermal noise removal of SAR images; speckle filtering of SAR images; and incident angle normalization of SAR images.

[0034] For example, in some specific implementations, Sentinel-1 SAR images of the study area are first acquired. The study area includes the Greenland Sea, Barents Sea, and Chukchi Sea. The SAR images can be GRDM products in EW mode, containing HH and HV dual polarization information. Before sample annotation, the SAR images undergo uniform preprocessing, including radiometric calibration, thermal noise removal, speckle filtering, and incident angle normalization, to reduce the impact of differences in image imaging conditions on vortex texture recognition.

[0035] Step S200: In response to the human visual interpretation of the target object, mark the spatial positions of vortex samples and non-vortex samples in the initial radar image; For example, in some specific implementations, manual visual interpretation is performed on preprocessed SAR images to label mesoscale vortex samples and non-vortex samples. Vortex samples are mainly identified based on spiral, arc, strip, or closed / semi-closed rotating textures presented in the SAR images; non-vortex samples are selected from adjacent background areas without obvious vortex structures. An initial SAR labeling sample library is constructed in this way, including positive vortex samples and negative non-vortex samples.

[0036] Step S300: Acquire ocean reanalysis data that matches the spatial location and imaging date; Among them, the parameter types of ocean reanalysis data include sea surface height anomalies, sea surface temperature, sea surface salinity, and current velocity characteristics; It should be noted that, in some embodiments, acquiring ocean reanalysis data matching spatial location and imaging date may include the following steps: taking the first marked spatial location as the target location; acquiring sea surface height anomalies, sea surface temperature, sea surface salinity, and upper ocean current velocity information within a preset range near the target location on the corresponding imaging date; wherein, the upper ocean current velocity information represents a velocity profile containing multiple vertical layers; dividing the velocity profile into surface and subsurface layers, and then using principal component analysis to extract the principal component features of the surface and subsurface layers as velocity features; taking the next marked spatial location as the target location, and returning to execute the steps of acquiring sea surface height anomalies, sea surface temperature, sea surface salinity, and upper ocean current velocity information within a preset range near the target location on the corresponding imaging date, until all spatial locations have been traversed.

[0037] For example, in some specific implementations, for each SAR-labeled sample location, diurnal-scale ocean reanalysis data of the Arctic Ocean near the same date and spatial location is extracted. Extracted variables include sea surface height anomalies, sea surface temperature, sea surface salinity, and upper ocean current velocity information. Sea surface height anomalies, sea surface temperature, and sea surface salinity are presented as two-dimensional surface variables; current velocity variables include east-west current velocity u and north-south current velocity v, and may include current velocity profiles from multiple vertical layers within a range of 0–200 m.

[0038] For multi-layer velocity profiles (upper ocean current information), principal component analysis (PCA) can be used for dimensionality reduction to decrease input feature redundancy and computational complexity. Specifically, the 0–200 m velocity profile can be divided into surface and subsurface layers, and PCA can be performed on each layer separately, retaining the main principal component features. The dimensionality-reduced velocity features, along with sea surface height anomalies, sea surface temperature, and sea surface salinity, constitute the multi-channel input features of the deep learning model.

[0039] Step S400: Using the spatial location as the center, cut out a two-dimensional grid window with a preset spatial range for each parameter type from the ocean reanalysis data, stack all the two-dimensional grid windows of all parameter types along the channel dimension to form a multi-channel two-dimensional physical field tensor, and assign a category label to each multi-channel two-dimensional physical field tensor according to the vortex coverage ratio of the preset central region of the two-dimensional grid window to form a training sample set. The preset center region is a preset region range for mesoscale vortices, and the category labels include positive vortex samples and negative non-vortex samples. It should be noted that in some embodiments, each multi-channel two-dimensional physical field tensor is assigned a category label based on the vortex coverage ratio of the preset central region of the two-dimensional grid window to form a training sample set. This may include the following steps: using the first multi-channel two-dimensional physical field tensor as a candidate input sample; extracting the local image of the region corresponding to the two-dimensional grid window from the initial radar image corresponding to the candidate input sample; when the proportion of the vortex region in the preset central region of the local image is greater than a preset threshold, associating the category label of the positive vortex sample with the candidate input sample to form a training sample; when the proportion of the vortex region in the preset central region of the local image is less than or equal to the preset threshold, associating the category label of the negative vortex sample with the candidate input sample to form a training sample; using the next multi-channel two-dimensional physical field tensor as a candidate input sample, returning to the step of extracting the local image of the region corresponding to the two-dimensional grid window from the initial radar image corresponding to the candidate input sample, until all multi-channel two-dimensional physical field tensors have been traversed, and summarizing all training samples to form a training sample set.

[0040] For example, in some specific implementations, a 16×16 grid window is cropped from the spatiotemporally matched reanalysis data, centered on the location of each SAR-labeled sample. Each grid window serves as a model input sample. The spatial resolution of this window is 6.25 km, which can cover the vortex and its surrounding background dynamic environment. For each input variable, a 16×16 window of the same spatial extent is cropped and stacked along the channel dimension to form a multi-channel two-dimensional physics tensor.

[0041] Centered on the location of each SAR-labeled sample, a 16×16 grid window is cropped from the spatiotemporally matched reanalysis data, with each grid window serving as a model input sample. This window has a spatial resolution of 6.25 km and a coverage area of ​​approximately 100 km × 100 km, capable of encompassing the Arctic mesoscale vortex (typically 10–40 km in scale) and its surrounding dynamic environment. For each input variable, a 16×16 window of the same spatial extent is cropped and stacked along the channel dimension to form a multi-channel two-dimensional physics tensor.

[0042] Sample labels are determined based on the proportion of vortex coverage within the central 2×2 grid of the window. When the proportion of the vortex region within the central 2×2 grid is greater than 50%, the sample is labeled as a positive vortex sample; when the proportion is equal to or less than 50%, it is labeled as a negative non-vortex sample. The assignment strategy is fixed and unambiguous. The design of using the central 2×2 region for label determination is based on the fact that information at the window edges may contain fragmented signals from background dynamic structures or nearby vortices, which can interfere with sample category determination. The central region is most likely to represent the vortex presence state at the core of the window. Therefore, label determination based on the central region allows the model to pay more attention to the vortex characteristics of the core area of ​​the window.

[0043] It should be noted that this invention avoids the problem of vortex edges being misclassified as negative samples through the following design: First, the spatial scale of the 16×16 window is approximately 100km, while the typical scale of a mesoscale vortex in the Arctic is only 10~40km. The vortex diameter usually only occupies 1 / 6 to 1 / 3 of the window, and the extreme case of a vortex covering the entire window is statistically very low. Second, the sample label depends only on the central 2×2 grid (approximately 12.5km×12.5km). Even if the window edge contains the outer structure of the vortex, as long as the central region does not meet the vortex proportion threshold, the window is marked as a negative sample—this is consistent with physical intuition: when the center of the window is not located in the core region of the vortex, it should not be used as a positive sample for training. Third, a random spatial offset amplification strategy can be introduced. By offsetting the center position of the window and recalculating the label, different spatial offset samples of the same vortex may generate different positive and negative labels, thereby allowing the model to learn the complete spatial feature gradient of the vortex from the core to the edge, effectively avoiding the edge misclassification problem caused by a single fixed window.

[0044] It should be noted that in some embodiments, the method may further include the following steps: taking the first training sample in the training sample set as a candidate training sample; randomly generating an offset within a preset grid range, using the offset to spatially offset the two-dimensional grid window corresponding to the candidate training sample to obtain the offset spatial range; cropping the offset grid window of the offset spatial range for each parameter type from the ocean reanalysis data, stacking all the offset grid windows of all parameter types along the channel dimension to form a multi-channel augmented physical field tensor, assigning a category label to the multi-channel augmented physical field tensor according to the vortex coverage ratio of the preset central region of the offset grid window to constitute an augmented sample; taking the next training sample in the training sample set as a candidate training sample, returning to the step of randomly generating an offset within the preset grid range, until all training samples in the training sample set have been traversed; supplementing all augmented samples to the training sample set, and then performing data augmentation operations on all samples in the training sample set to update the training sample set.

[0045] For example, in some specific embodiments, to increase the diversity of the sample space and reduce the bias of the annotation position, the present invention employs a data augmentation method of random spatial offset. For each vortex annotation position, 17 additional 16×16 grid samples (a total of 18 samples, including the original center sample) are generated based on the original center position sample. The new samples are obtained by randomly offsetting 1 to 8 grids around the original position, with the offset being randomly sampled in a uniform distribution within the range of 1 to 8 grids.

[0046] Regarding the issue of spatial independence between amplified samples, the specific approach taken in the experiment of this invention is as follows: no additional minimum interval threshold (e.g., offset distance ≥ 3 grids) was set for the Euclidean distance between the centers of the offset samples, nor was a "use in the early stage of training and gradually reduce in the later stage" reduction strategy adopted. Instead, the amplification strategy was continuously used throughout the entire training cycle. Its rationale is as follows: First, the offset range covers 8 grids (approximately 50km), which is sufficiently large relative to the typical scale of the Arctic mesoscale vortex (10~40km). The 17 offset samples are randomly distributed within a 50km×50km range, resulting in a low probability of spatial overlap. Second, each offset is independently and randomly sampled, eliminating artificially introduced constraints on inter-sample correlation. Third, even if the centers of individual offset samples are relatively close, their labels are independently re-determined based on the vortex coverage ratio of their respective new 2×2 grids, ensuring that the correspondence between sample labels and spatial locations remains strictly valid. Fourth, the model training uses batch stochastic gradient descent, and the number of neighboring samples within a single batch is limited, which will not have a dominant influence on the model's convergence direction; therefore, no additional spatial interval constraints are required.

[0047] Because the center position of the window changes after offsetting, the offset samples do not directly inherit the original labels. Instead, their positive or negative class is redefined based on the proportion of vortex coverage within the new 2×2 grid at the center of the window: if the proportion of the vortex region within the new 2×2 grid is greater than 50%, it is labeled as a positive sample; otherwise, it is labeled as a negative sample. This recalibration mechanism ensures that the labels of the offset samples strictly correspond to their new spatial positions, avoiding the introduction of label noise due to the inheritance of the original labels.

[0048] After the training sample set is constructed, data augmentation operations such as rotation, translation, and noise perturbation are performed only on the training samples to improve the model's generalization ability to vortex features under different spatial morphologies and noise conditions. In addition, some preferred implementations can also construct a test set (including multiple test samples) and a validation set (including multiple validation samples) to test, validate, and optimize the trained model. The test samples and validation samples remain unchanged to ensure the objectivity of the model evaluation.

[0049] Step S500: Use the training sample set to perform prediction training on the preset deep learning model to obtain the vortex prediction model. The deep learning model includes a two-dimensional convolutional layer, a backbone network, an attention module, and a probability output layer. It should be noted that in some embodiments, using a training sample set to train a preset deep learning model to obtain a vortex prediction model may include the following steps: First, a first feature is obtained by performing channel mapping and initial feature extraction on a multi-channel two-dimensional physical field tensor through a two-dimensional convolutional layer; based on the initial feature, spatial features at different scales are extracted through a backbone network using a moving inverse residual convolution module, and the spatial features output by the last layer of the backbone network are used as the second feature; based on the second feature, a Q matrix, K matrix, and V matrix are constructed through an attention module, and then recovered through convolution after attention weighted aggregation, resulting in a third feature; a fourth feature is obtained by performing a residual connection between the second and third features; based on the fourth feature, a logical value is output through a probability output layer using global average pooling and a fully connected layer, and then the logical value is converted into a vortex occurrence probability using a sigmoid function; based on the category label and vortex occurrence probability corresponding to the multi-channel two-dimensional physical field tensor of each training sample in the training sample set, the deep learning model is optimized by using a binary cross-entropy loss function to obtain the vortex prediction model.

[0050] For example, in some specific embodiments, the present invention constructs a deep learning model for predicting the probability of mesoscale eddy occurrence, denoted as ENL-Net. This model takes a multi-channel two-dimensional physical field tensor as input, and the input includes sea surface height anomalies, sea surface temperature, sea surface salinity, and dimensionality-reduced upper ocean current velocity characteristics.

[0051] This invention constructs a deep learning model for predicting the probability of mesoscale eddy occurrence, denoted as ENL-Net. The model takes a multi-channel two-dimensional physical field tensor as input, and the input channels include sea surface height anomaly, sea surface temperature, sea surface salinity, and dimensionality-reduced upper ocean current velocity features (a total of 6 channels: SSHA, SST, SSS, and three-dimensional current velocity principal component features obtained after PCA dimensionality reduction).

[0052] ENL-Net first performs channel mapping and initial feature extraction through two-dimensional convolutional layers (3×3 kernel size, stride 1), mapping the number of input channels to the standard input dimension suitable for the backbone network. Subsequently, the feature maps are input into the EfficientNet backbone network (using the EfficientNet-B4 architecture), which utilizes its moving inverse residual convolutional structure (MBConv block) to extract spatial features at different scales.

[0053] To enhance the model's ability to model long-range spatial dependencies, a non-local attention module is introduced in parallel at the output of the last stage of the EfficientNet backbone network (i.e., after the last MBConv block of Stage 8 and before global average pooling). Specifically, the feature map output from Stage 8 is input into two branches. One branch is directly passed to the subsequent global pooling layer, while the other branch, after the non-local module calculates the attention weights between distant pixels, is residually concatenated with the original feature map. The fused feature map is then fed into the global average pooling layer. Inside the non-local module, three 1×1 convolutions are used to compress the number of channels in the input feature map to one-quarter of its original number (e.g., from 320 channels to 80 channels), generating Q, K, and V matrices respectively. Attention weights are then calculated and weighted, and finally, a 1×1 convolution restores the original number of channels, which is then added to the input feature map to achieve residual concatenation. This channel compression strategy effectively controls the number of model parameters and computational complexity while maintaining the ability to model long-range dependencies.

[0054] By introducing a non-local attention module into the EfficientNet feature extraction structure, the model can comprehensively consider the correlation between distant grid points, thereby better distinguishing between vortex and non-vortex regions. After feature extraction, the model outputs a logit value through global average pooling and fully connected layers, and then uses the sigmoid function to convert this logit value into a vortex occurrence probability between 0 and 1. The higher the probability value, the greater the likelihood of a mesoscale vortex occurring in that region.

[0055] Finally, a binary cross-entropy loss function with logits is used as the training objective, and the Adam optimizer is used for parameter updates. During model training, training samples are input into ENL-Net in batches, and the model outputs the probability of vortex occurrence for each sample. After training, metrics such as accuracy, precision, recall, F1-score, and AUC can be calculated on the test or validation set.

[0056] To determine the optimal classification threshold, F1-scores under different probability thresholds are calculated based on the test set results, and the probability that maximizes the F1-score is selected as the binary classification decision threshold. Based on this threshold, the probability field can be further converted into vortex / non-vortex binary results.

[0057] Step S600: Based on the ocean reanalysis data of the target area on the target date, generate the vortex occurrence probability field of the target area on the target date using a sliding window prediction strategy with a vortex prediction model. It should be noted that, in some embodiments, based on ocean reanalysis data of the target area on a target date, the generation of the vortex occurrence probability field of the target area on a target date using a sliding window prediction strategy with a vortex prediction model may include the following steps: using ocean reanalysis data of the target area on a target date as the data to be analyzed; using a sliding window prediction strategy, sliding the sliding window row by row and column by column on the spatial grid of the target area with a preset sliding step size, and extracting the sub-data to be analyzed for each location in the data to be analyzed through the sliding window; inputting the sub-data to be analyzed into the vortex prediction model so that the vortex prediction model predicts the vortex prediction probability value for each location in the target area, and then summarizing and organizing to obtain the vortex occurrence probability field of the target area on the target date.

[0058] For example, in some specific implementations, the trained ENL-Net is applied to diurnal-scale ocean reanalysis data of the Arctic marginal seas. For the target date and target region, a sliding window prediction strategy is adopted: with a sliding step size of 2 grids (i.e., 12.5 km), the network slides row by row and column by column across the spatial grid of the target region, extracting a 16×16 grid window as the model input each time. For each input window, ENL-Net outputs a probability value between 0 and 1, representing the probability of a mesoscale eddy occurring within the central 2×2 grid region of that window. The probability value output by each window is assigned to its corresponding central 2×2 grid region. The output regions of each window are spatially contiguous without overlap, therefore no weighted averaging or overlap processing is required; a complete two-dimensional probability field can be reconstructed simply by splicing the windows according to their spatial locations.

[0059] The design basis of the aforementioned sliding window strategy of "inputting a 16×16 window and outputting 2×2 region probabilities" is as follows: ENL-Net is essentially a binary classifier that takes a 16×16 spatial window as input. The positive or negative label determination during model training depends on the vortex coverage ratio of the 2×2 region at the center of the window (see step four). Therefore, the probability value output by the model naturally corresponds to the vortex existence state of the 2×2 central region rather than the entire 16×16 window. Using a sliding step size of 2 ensures that the output 2×2 regions of adjacent windows are seamlessly spliced ​​in space without overlapping, thus guaranteeing a spatial resolution of 6.25 km for the output probability field, and preventing any grid from being repeatedly predicted or missed. This step size setting matches the typical scale (10~40 km) of Arctic mesoscale vortices, effectively capturing the spatial distribution details of the vortices.

[0060] In addition, in some preferred embodiments, the monthly or annual average vortex occurrence probability map can be generated by averaging the daily probability field over time, which can be used to analyze the seasonal and interannual variations of vortex activity.

[0061] To explain in detail the principle of the technical solution of the present invention, the overall process of the present invention will be described below with reference to some specific embodiments. It is easy to understand that the following is an explanation of the technical principle of the present invention and should not be regarded as a limitation of the present invention.

[0062] First, it should be noted that existing technologies have at least the following shortcomings: (1) SAR images have high resolution but discontinuous coverage, making it difficult to support large-scale, long-term vortex monitoring in the Arctic marginal seas.

[0063] (2) Traditional gridded sea surface height data has a wide coverage, but its effective spatial resolution is insufficient, making it difficult to analyze small-scale and mesoscale eddies in the Arctic.

[0064] (3) Existing SAR manual interpretation and SAR deep learning recognition methods rely on dense SAR observations and are difficult to extend to continuous large-scale probabilistic mapping.

[0065] (4) Existing methods mostly output discrete identification results and lack a continuous probability expression of the possibility of vortex occurrence, making it difficult to provide prior spatial information for subsequent SAR observation and dynamic process analysis.

[0066] In view of this, the present invention aims to solve the above problems and proposes a method for predicting the probability of mesoscale eddies based on multi-source ocean reanalysis data and deep learning models, so as to achieve large-scale and continuous prediction of the probability of mesoscale eddies in Arctic marginal seas such as the Greenland Sea, Barents Sea and Chukchi Sea.

[0067] As shown in Figure 3, this invention provides a method for predicting the probability of mesoscale eddies. In some specific embodiments, this method constructs a deep learning sample database based on manually labeled samples of SAR images and ocean reanalysis data, and generates a mesoscale eddy occurrence probability field in the Arctic marginal seas by introducing an EfficientNet backbone network and a Non-local attention module in the ENL-Net model. The technical solution of this invention may specifically include the following steps: Step 1: Acquire and preprocess SAR image data.

[0068] Sentinel-1 SAR imagery was acquired for the study area, which includes the Greenland Sea, Barents Sea, and Chukchi Sea. GRDM products in EW mode, containing both HH and HV dual polarization information, were used for the SAR imagery. Before sample annotation, the SAR imagery underwent uniform preprocessing, including radiometric calibration, thermal noise removal, speckle filtering, and incident angle normalization, to reduce the impact of differences in image imaging conditions on vortex texture recognition.

[0069] Step 2: Construct initial samples of vortex and non-vortex types.

[0070] Based on the preprocessed SAR imagery, manual visual interpretation was performed to annotate mesoscale vortex and non-vortex samples. Vortex samples were primarily identified based on spiral, arc, stripe, or closed / semi-closed rotating textures present in the SAR imagery; non-vortex samples were selected from nearby background areas without obvious vortex structures. An initial SAR annotation sample library, including positive vortex samples and negative non-vortex samples, was constructed using this method.

[0071] Step 3: Extract marine environmental variables that are spatiotemporally matched with the SAR samples.

[0072] For each SAR-labeled sample location, diurnal ocean reanalysis data of the Arctic Ocean from the same date and spatial location are extracted. Extracted variables include sea surface height anomalies, sea surface temperature, sea surface salinity, and upper ocean current velocity information. Specifically, sea surface height anomalies, sea surface temperature, and sea surface salinity are presented as two-dimensional surface variables; current velocity variables include east-west current velocity u and north-south current velocity v, and may include current velocity profiles from multiple vertical layers within the range of 0–200 m.

[0073] For multi-layer velocity profiles (upper ocean current information), principal component analysis (PCA) can be used for dimensionality reduction to decrease input feature redundancy and computational complexity. Specifically, the 0–200 m velocity profile can be divided into surface and subsurface layers, and PCA can be performed on each layer separately, retaining the main principal component features. The dimensionality-reduced velocity features, along with sea surface height anomalies, sea surface temperature, and sea surface salinity, constitute the multi-channel input features of the deep learning model.

[0074] Step 4: Construct a 16×16 spatial window sample.

[0075] Centered on the location of each SAR-labeled sample, a 16×16 grid window is cropped from the spatiotemporally matched reanalysis data, with each grid window serving as a model input sample. This window has a spatial resolution of 6.25 km, sufficient to cover the vortex and its surrounding dynamic environment. For each input variable, a 16×16 window of the same spatial extent is cropped and stacked along the channel dimension to form a multi-channel two-dimensional physics tensor.

[0076] Centered on the location of each SAR-labeled sample, a 16×16 grid window is cropped from the spatiotemporally matched reanalysis data, with each grid window serving as a model input sample. This window has a spatial resolution of 6.25 km and a coverage area of ​​approximately 100 km × 100 km, capable of encompassing the Arctic mesoscale vortex (typically 10–40 km in scale) and its surrounding dynamic environment. For each input variable, a 16×16 window of the same spatial extent is cropped and stacked along the channel dimension to form a multi-channel two-dimensional physics tensor.

[0077] Sample labels are determined based on the proportion of vortex coverage within the central 2×2 grid of the window. When the proportion of the vortex region within the central 2×2 grid is greater than 50%, the sample is labeled as a positive vortex sample; when the proportion is equal to or less than 50%, it is labeled as a negative non-vortex sample. The assignment strategy is fixed and unambiguous. The design of using the central 2×2 region for label determination is based on the fact that information at the window edges may contain fragmented signals from background dynamic structures or nearby vortices, which can interfere with sample category determination. The central region is most likely to represent the vortex presence state at the core of the window. Therefore, label determination based on the central region allows the model to pay more attention to the vortex characteristics of the core area of ​​the window.

[0078] Step 5: Perform sample amplification.

[0079] To increase spatial diversity and reduce labeling bias, this invention employs a random spatial offset data augmentation method. For each vortex labeling location, 17 additional 16×16 grid samples (a total of 18 samples, including the original center sample) are generated based on the original center sample. The new samples are obtained by randomly offsetting 1 to 8 grids around the original location, with the offset amount randomly sampled in a uniform distribution within the range of 1 to 8 grids.

[0080] Regarding the issue of spatial independence between amplified samples, the specific approach taken in the experiment of this invention is as follows: no additional minimum interval threshold (e.g., offset distance ≥ 3 grids) was set for the Euclidean distance between the centers of the offset samples, nor was a "use in the early stage of training and gradually reduce in the later stage" reduction strategy adopted. Instead, the amplification strategy was continuously used throughout the entire training cycle. Its rationale is as follows: First, the offset range covers 8 grids (approximately 50km), which is sufficiently large relative to the typical scale of the Arctic mesoscale vortex (10~40km). The 17 offset samples are randomly distributed within a 50km×50km range, resulting in a low probability of spatial overlap. Second, each offset is independently and randomly sampled, eliminating artificially introduced constraints on inter-sample correlation. Third, even if the centers of individual offset samples are relatively close, their labels are independently re-determined based on the vortex coverage ratio of their respective new 2×2 grids, ensuring that the correspondence between sample labels and spatial locations remains strictly valid. Fourth, the model training uses batch stochastic gradient descent, and the number of neighboring samples within a single batch is limited, which will not have a dominant influence on the model's convergence direction; therefore, no additional spatial interval constraints are required.

[0081] Because the center position of the window changes after offsetting, the offset samples do not directly inherit the original labels. Instead, their positive or negative class is redefined based on the proportion of vortex coverage within the new 2×2 grid at the center of the window: if the proportion of the vortex region within the new 2×2 grid is greater than 50%, it is labeled as a positive sample; otherwise, it is labeled as a negative sample. This recalibration mechanism ensures that the labels of the offset samples strictly correspond to their new spatial positions, avoiding the introduction of label noise due to the inheritance of the original labels.

[0082] After the training sample set was constructed, data augmentation operations such as rotation, translation, and noise perturbation were performed only on the training samples to improve the model's generalization ability to vortex features under different spatial morphologies and noise conditions. The test and validation samples remained unchanged to ensure the objectivity of the model evaluation. Finally, the training set constructed after the above augmentation contained 102,697 samples (45,780 positive samples and 56,917 negative samples), and the test set contained 6,289 samples (2,788 positive samples and 3,501 negative samples).

[0083] Step 6: Construct the ENL-Net deep learning model.

[0084] This invention constructs a deep learning model, denoted as ENL-Net, for predicting the probability of mesoscale eddies. The model takes a multi-channel two-dimensional physics tensor as input, which includes sea surface height anomalies, sea surface temperature, sea surface salinity, and dimensionality-reduced upper ocean current velocity characteristics.

[0085] This invention constructs a deep learning model for predicting the probability of mesoscale eddy occurrence, denoted as ENL-Net. The model takes a multi-channel two-dimensional physical field tensor as input, and the input channels include sea surface height anomaly, sea surface temperature, sea surface salinity, and dimensionality-reduced upper ocean current velocity features (a total of 6 channels: SSHA, SST, SSS, and three-dimensional current velocity principal component features obtained after PCA dimensionality reduction).

[0086] ENL-Net first performs channel mapping and initial feature extraction through two-dimensional convolutional layers (3×3 kernel size, stride 1), mapping the number of input channels to the standard input dimension suitable for the backbone network. Subsequently, the feature maps are input into the EfficientNet backbone network (using the EfficientNet-B4 architecture), which utilizes its moving inverse residual convolutional structure (MBConv block) to extract spatial features at different scales.

[0087] To enhance the model's ability to model long-range spatial dependencies, a non-local attention module is introduced in parallel at the output of the last stage of the EfficientNet backbone network (i.e., after the last MBConv block of Stage 8 and before global average pooling). Specifically, the feature map output from Stage 8 is input into two branches. One branch is directly passed to the subsequent global pooling layer, while the other branch, after the non-local module calculates the attention weights between distant pixels, is residually concatenated with the original feature map. The fused feature map is then fed into the global average pooling layer. Inside the non-local module, three 1×1 convolutions are used to compress the number of channels in the input feature map to one-quarter of its original number (e.g., from 320 channels to 80 channels), generating Q, K, and V matrices respectively. Attention weights are then calculated and weighted, and finally, a 1×1 convolution restores the original number of channels, which is then added to the input feature map to achieve residual concatenation. This channel compression strategy effectively controls the number of model parameters and computational complexity while maintaining the ability to model long-range dependencies.

[0088] By introducing a non-local attention module into the EfficientNet feature extraction structure, the model can comprehensively consider the correlation between distant grid points, thereby better distinguishing between vortex and non-vortex regions. After feature extraction, the model outputs a logit value through global average pooling and fully connected layers, and then uses the sigmoid function to convert this logit value into a vortex occurrence probability between 0 and 1. The higher the probability value, the greater the likelihood of a mesoscale vortex occurring in that region.

[0089] To verify the effectiveness of the non-local attention module, this invention compared the performance of EfficientNet models with and without the non-local module under the same training strategy. The results show that after introducing the non-local module, the model's F1-score on the test set increased from 72.4% to 74.7%, and the AUC increased from 0.803 to 0.827, verifying the positive contribution of the non-local module to improving the accuracy of vortex recognition. Experimental results are shown in Table 1: Table 1

[0090] After feature extraction, the model outputs a logit value through global average pooling and a fully connected layer, and then uses the sigmoid function to convert this logit value into a probability of vortex occurrence between 0 and 1. The higher the probability value, the greater the likelihood of a mesoscale vortex occurring in that region.

[0091] Step 7: Model training and threshold determination.

[0092] A binary cross-entropy loss function with logits is used as the training objective, and the Adam optimizer is used for parameter updates. During model training, training samples are input into ENL-Net in batches, and the model outputs the vortex occurrence probability for each sample. After training, accuracy, precision, recall, F1-score, and AUC are calculated on the test or validation set.

[0093] To determine the optimal classification threshold, F1-scores under different probability thresholds are calculated based on the test set results, and the probability that maximizes the F1-score is selected as the binary classification decision threshold. Based on this threshold, the probability field can be further converted into vortex / non-vortex binary results.

[0094] Step 8: Generate the vortex occurrence probability field.

[0095] The trained ENL-Net was applied to diurnal ocean reanalysis data from the Arctic marginal seas. For the target date and target region, a sliding window prediction strategy was employed: with a sliding step size of 2 grids (12.5 km), the network slid row-by-row and column-by-column across the spatial grid of the target region, extracting a 16×16 grid window as the model input each time. For each input window, ENL-Net outputs a probability value between 0 and 1, representing the probability of a mesoscale eddy occurring within the central 2×2 grid region of that window. The probability value output by each window was assigned to its corresponding central 2×2 grid region. Since the output regions of each window are spatially contiguous without overlap, no weighted averaging or overlap processing is required; a complete two-dimensional probability field can be reconstructed simply by splicing the spatial locations.

[0096] The design basis of the aforementioned sliding window strategy of "inputting a 16×16 window and outputting 2×2 region probabilities" is as follows: ENL-Net is essentially a binary classifier that takes a 16×16 spatial window as input. The positive or negative label determination during model training depends on the vortex coverage ratio of the 2×2 region at the center of the window (see step four). Therefore, the probability value output by the model naturally corresponds to the vortex existence state of the 2×2 central region rather than the entire 16×16 window. Using a sliding step size of 2 ensures that the output 2×2 regions of adjacent windows are seamlessly spliced ​​in space without overlapping, thus guaranteeing a spatial resolution of 6.25 km for the output probability field, and preventing any grid from being repeatedly predicted or missed. This step size setting matches the typical scale (10~40 km) of Arctic mesoscale vortices, effectively capturing the spatial distribution details of the vortices.

[0097] By averaging the daily probability fields over time, monthly or annual vortex occurrence probability maps can be generated, which can be used to analyze the seasonal and interannual variations of vortex activity.

[0098] Step 9: Model result verification.

[0099] To verify the reliability of the model's predictions, this invention employs multi-source validation using flow field vorticity, SAR imagery, and SWOT observation data. First, relative vorticity is calculated using east-west and north-south flow velocities from the reanalysis data, and the spatial consistency between high-probability regions and strong vorticity regions is compared. Second, high-probability regions output by the model are selected and compared with SAR images from the corresponding dates to determine whether these regions correspond to vortex-like texture structures in the SAR images. For all available SAR images within a certain time period, vortices can be identified through visual interpretation, and the proportion successfully predicted by the model is statistically analyzed. Finally, SSHA data from SWOT satellite observations can be used as independent validation data. Mesoscale vortices are identified from SWOT SSHA observations using the PET algorithm, and spatially matched with ENL-Net prediction results. Accuracy, precision, recall, false alarm rate, and crossover ratio (CROR) are calculated to evaluate the reliability of the probabilistic prediction results in different Arctic marginal sea regions.

[0100] In some specific application scenarios, such as Figure 3 As shown, the technical solution of the present invention mainly includes four stages: vortex sample annotation, sample database construction, deep learning model training, and model prediction and verification.

[0101] First, SAR images of the study area were acquired and preprocessed, including radiometric calibration, thermal noise removal, speckle filtering, and incident angle normalization. Then, vortex locations were recorded visually, forming a vortex sample distribution. Next, the SAR-labeled samples were matched with multivariate ocean reanalysis data from the same date and location to extract input variables including sea surface height anomalies, sea surface temperature, sea surface salinity, and upper ocean current velocity, constructing a training sample database. Then, multi-channel physical fields were input into a deep learning model, and vortex-related features were extracted using the EfficientNet backbone network and a non-local attention module, outputting vortex occurrence probabilities in the range of 0–1. Finally, a sliding window approach was used to predict the target area daily, generating a daily-scale vortex occurrence probability map, which was then validated using SAR imagery, current field vorticity, and SWOT observations.

[0102] In some specific application scenarios, the consistency verification between the model prediction results and the flow field vorticity in the technical solution of this invention is implemented as follows: like Figure 4As shown, this invention utilizes flow field vorticity to verify the physical consistency of model prediction results. First, a vortex occurrence probability map of the study area is generated using the trained ENL-Net model. Simultaneously, the relative vorticity field is calculated based on the east-west and north-south flow velocities in the reanalysis data. Subsequently, the model prediction probability map and the relative vorticity map are spatially compared.

[0103] exist Figure 4 middle, Figure 4 The (a) marked portion represents the spatial distribution of the predicted vortex occurrence probability by the model. Figure 4 The (b) marked portion represents the spatial distribution of relative vorticity of ocean currents on the same date. The magnified local area shows a good spatial correspondence between the high-probability region and the strong vorticity region. Figure 4 The (c) marked section further illustrates the relationship between predicted probability and relative vorticity. High-probability samples generally correspond to higher vorticity values, while low-probability samples are mainly distributed in the low vorticity range. Figure 4 The (d) markers show the daily average vorticity changes of vortex and non-vortex samples in September 2019. The average vorticity of the vortex samples is generally higher than that of the non-vortex samples. This result indicates that the vortex occurrence probability output by this invention is in good agreement with the background ocean current dynamic characteristics.

[0104] In some specific application scenarios, the verification implementation of the technical solution of this invention based on SAR imagery in typical regions is as follows: like Figure 5 As shown, this invention further uses SAR imagery to verify the model's prediction results. First, high-probability regions are selected in the vortex occurrence probability map output by the model, and SAR images for the corresponding dates are extracted. Then, the spatial correspondence between the high-probability regions and the vortex-like texture structures in the SAR images is compared.

[0105] Figure 5 The middle area is a probability map of eddy occurrence in the Greenland Sea region on September 3, 2019. The boxes in the map indicate typical high-probability areas. Figure 5 Sections (a) to (j) show local comparison results for different dates and locations, with the left side displaying the model-predicted probability map and the right side showing the corresponding SAR image. Each local region is labeled with its average predicted probability and average relative vorticity. Figure 5 It is evident that the high-probability regions predicted by the model correspond well to the spiral, arc, or strip-shaped vortex textures in SAR images, indicating that the present invention can provide effective prior spatial information for the identification of mesoscale vortices in SAR images.

[0106] In some specific application scenarios, the independent verification implementation based on SWOT observation for the technical solution of this invention is as follows: like Figure 6 As shown, this invention uses SWOT satellite observation data as independent validation data to verify the model's predicted eddy occurrence probability. First, SWOT Level-2 wide-swath sea surface height anomaly data are acquired within the Greenland Sea, Barents Sea, and Chukchi Sea regions. Then, the PET algorithm is used to extract mesoscale eddy locations from SWOT SSHA observations and these locations are used as reference results. Finally, the model's predicted eddy occurrence probability map is spatially matched and compared with the SWOT observation results.

[0107] Figure 6 The (a), (c) and (e) sections represent the SWOT SSHA observation results for the Greenland Sea, Barents Sea and Chukchi Sea regions, respectively. The red circles in the figure indicate the vortex locations identified by the PET algorithm. Figure 6 The portions marked (b), (d), and (f) are model-predicted vortex occurrence probability maps for the corresponding regions. The comparison results show that the vortex regions identified by SWOT analysis have good spatial consistency with the high-probability regions predicted by the model, indicating that this invention has a certain predictive capability for vortex occurrence probability in different Arctic marginal sea regions.

[0108] In some specific application scenarios, the probability distribution verification of the technical solution of this invention based on SWOT vortex mask is implemented as follows: like Figure 7 As shown, this invention further analyzes the difference in predicted probability distribution between the vortex region and the background region based on SWOT observation results. First, a vortex region mask is generated using SWOT observations and the PET algorithm. Then, the model predicted probability values ​​are extracted from the vortex region and the non-vortex background region, respectively, and the probability density distribution of the two types of regions is statistically analyzed.

[0109] Figure 7 The labeled portions (a), (b), and (c) represent the predicted probability density distributions of vortex regions and background regions within the Greenland Sea, Barents Sea, and Chukchi Sea, respectively. The results show that the predicted probability of vortex regions is generally higher, while the predicted probability of background regions is generally lower, indicating a significant difference in probability distribution between the two types of regions. This demonstrates that the vortex occurrence probability field generated by this invention can effectively distinguish between vortex regions and non-vortex background regions, further proving the effectiveness of the model's prediction results.

[0110] In summary, compared with the prior art, the embodiments of the present invention have at least the following beneficial effects: (1) Achieving large-scale, continuous probabilistic monitoring of eddies in the Arctic marginal seas. Traditional SAR eddy identification relies on single-scene images, which are limited by imaging swath width and revisit period, making it difficult to carry out large-scale, long-term sequence monitoring. This invention uses continuous ocean reanalysis data as input to generate eddy occurrence probability fields at daily, monthly, and annual scales, which can overcome the limitation of discontinuous coverage of single-scene SAR images and achieve large-scale, continuous monitoring of mesoscale eddies in the Arctic marginal seas.

[0111] (2) Improve the ability to characterize potential active areas of small-scale and mesoscale eddies in the Arctic. Arctic mesoscale eddies are small in scale, and the effective resolution of traditional gridded altimeter data is limited, making it difficult to directly analyze many small-scale eddy structures. This invention learns the ocean physical field characteristics related to eddies through SAR labeled samples and generates eddy occurrence probability maps based on 6.25km reanalysis data, which helps to characterize potential high-incidence areas of eddies that are difficult to represent in traditional altimeter data.

[0112] (3) Constructing a probabilistic prediction model that integrates multi-source physical variables and a non-local attention mechanism. This invention comprehensively utilizes sea surface height anomalies, sea surface temperature, sea surface salinity, and upper ocean current velocity characteristics to enhance the model's ability to express the dynamic background of eddies. At the same time, a non-local attention module is introduced to improve the model's ability to model long-distance spatial dependencies, enabling the model to better distinguish between eddy samples and non-eddy samples, and output the probability of continuous eddy occurrence in the range of 0–1, providing prior spatial information for subsequent fine SAR observations and eddy morphology analysis.

[0113] (4) A multi-source verification mechanism is established with good application scalability. This invention combines relative vorticity of the flow field, SAR imagery, and SWOT observation data to verify the prediction results, and can evaluate the reliability of the model from multiple perspectives such as dynamic consistency, remote sensing texture structure, and independent satellite observations. This method is not only applicable to the Greenland Sea, Barents Sea, and Chukchi Sea, but can also be extended to other polar marginal seas or high-latitude seas, providing technical support for mesoscale eddy activity monitoring, SAR observation mission planning, and Arctic Ocean dynamic process research.

[0114] like Figure 8 As shown, this embodiment of the invention also provides a mesoscale vortex occurrence probability prediction device 900, which can implement the above-described method. This device may include: The first module 901 is used to acquire synthetic aperture radar images of the study area, preprocess the synthetic aperture radar images to obtain initial radar images; wherein, the synthetic aperture radar images are associated with the imaging date. The second module 902 is used to mark the spatial positions of vortex samples and non-vortex samples in the initial radar image in response to human visual interpretation of the target object. The third module 903 is used to acquire ocean reanalysis data matched with spatial location and imaging date; the parameter types of the ocean reanalysis data include sea surface height anomaly, sea surface temperature, sea surface salinity, and current velocity characteristics. The fourth module, 904, is used to crop a two-dimensional grid window with a preset spatial range for each parameter type from the ocean reanalysis data, using the spatial location as the center. The two-dimensional grid windows of all parameter types are stacked along the channel dimension to form a multi-channel two-dimensional physical field tensor. Each multi-channel two-dimensional physical field tensor is assigned a category label according to the vortex coverage ratio of the preset central region of the two-dimensional grid window, forming a training sample set. The preset central region is a preset region range for mesoscale vortices, and the category labels include positive vortex samples and negative non-vortex samples. The fifth module 905 is used to train a pre-set deep learning model using a training sample set to obtain a vortex prediction model. The deep learning model includes a two-dimensional convolutional layer, a backbone network, an attention module, and a probability output layer. Module 6, 906, is used to generate the vortex occurrence probability field of the target area on the target date based on ocean reanalysis data of the target area on the target date using a sliding window prediction strategy and a vortex prediction model.

[0115] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0116] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0117] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0118] like Figure 9 As shown, Figure 9 The hardware structure of an electronic device 1000 according to another embodiment is illustrated. The electronic device 1000 includes: The processor 1001 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (aSIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention. The memory 1002 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RaM). The memory 1002 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1002 and is called and executed by the processor 1001. Input / output interface 1003 is used to implement information input and output; The communication interface 1004 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 1005 transmits information between various components of the device (e.g., processor 1001, memory 1002, input / output interface 1003, and communication interface 1004); The processor 1001, memory 1002, input / output interface 1003 and communication interface 1004 are connected to each other within the device via bus 1005.

[0119] The electronic device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0120] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0121] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0122] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0123] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0124] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0125] The mesoscale eddy occurrence probability prediction method, device, electronic device, storage medium, and program product provided in this invention acquire synthetic aperture radar (SAR) images of the study area, preprocess the SAR images to obtain initial radar images; wherein the SAR images are associated with imaging dates; in response to human visual interpretation of the target object, the spatial locations of eddy samples and non-eddy samples are marked in the initial radar images; ocean reanalysis data matching the spatial locations and imaging dates is acquired; wherein the parameter types of the ocean reanalysis data include sea surface height anomalies, sea surface temperature, sea surface salinity, and current velocity characteristics; using the spatial location as the center, a two-dimensional grid window with a preset spatial range for each parameter type is cropped from the ocean reanalysis data, and the... Two-dimensional grid windows with parameter types are stacked along the channel dimension to form a multi-channel two-dimensional physical field tensor. Each multi-channel two-dimensional physical field tensor is assigned a category label according to the vortex coverage ratio of the preset central region of the two-dimensional grid window, forming a training sample set. The preset central region is a preset area range for mesoscale vortices, and the category labels include positive vortex samples and negative non-vortex samples. The training sample set is used to train a preset deep learning model to obtain a vortex prediction model. The deep learning model includes a two-dimensional convolutional layer, a backbone network, an attention module, and a probability output layer. Based on ocean reanalysis data of the target area on the target date, the vortex prediction model uses a sliding window prediction strategy to generate a vortex occurrence probability field of the target area on the target date. This invention uses continuous ocean reanalysis data as predictive input. After training the model with SAR-annotated samples, it can be independent of SAR imagery, thereby achieving large-scale continuous monitoring of Arctic marginal sea eddies. Based on ocean reanalysis data, a grid window sample is constructed, enabling the model to learn the physical field characteristics associated with mesoscale eddies, thus effectively characterizing potential high-incidence areas of eddies that are difficult to represent with traditional altimeters. The probability values ​​are output through a deep learning model, and a sliding window prediction strategy is used to generate a spatially continuous probability field, providing prior spatial information for subsequent fine SAR observations and dynamic process analysis.

[0126] The embodiments described in this invention are for the purpose of more clearly illustrating the technical solutions of the embodiments of this invention, and do not constitute a limitation on the technical solutions provided by the embodiments of this invention. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this invention are also applicable to similar technical problems.

[0127] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present invention, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0128] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0129] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0130] The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the present invention should be within the scope of the claims of the present invention.

Claims

1. A method for predicting the probability of mesoscale vortex occurrence, characterized in that, The method includes the following steps: Synthetic Aperture Radar (SAR) images of the study area are acquired, and the SAR images are preprocessed to obtain initial radar images; wherein, the SAR images are associated with the imaging date. In response to human visual interpretation of the target object, the spatial positions of vortex samples and non-vortex samples are marked in the initial radar image; Acquire ocean reanalysis data that matches the spatial location and the imaging date; wherein the parameter types of the ocean reanalysis data include sea surface height anomalies, sea surface temperature, sea surface salinity, and current velocity characteristics; Using the spatial location as the center, a two-dimensional grid window with a preset spatial range for each parameter type is cropped from the ocean reanalysis data. The two-dimensional grid windows of all parameter types are stacked along the channel dimension to form a multi-channel two-dimensional physical field tensor. Each multi-channel two-dimensional physical field tensor is assigned a category label according to the vortex coverage ratio of the preset central region of the two-dimensional grid window to form a training sample set. The preset central region is a preset region range for mesoscale vortices, and the category labels include positive vortex samples and negative non-vortex samples. The preset deep learning model is trained using the training sample set to obtain a vortex prediction model. The deep learning model includes a two-dimensional convolutional layer, a backbone network, an attention module, and a probability output layer. Based on ocean reanalysis data of the target area on the target date, the eddy prediction model is used to generate the eddy occurrence probability field of the target area on the target date using a sliding window prediction strategy.

2. The method according to claim 1, characterized in that, The preprocessing of the synthetic aperture radar image includes at least one of the following steps: Radiometric calibration is performed on the synthetic aperture radar image; Thermal noise removal is performed on the synthetic aperture radar image; The synthetic aperture radar image is subjected to speckle filtering; The incident angle of the synthetic aperture radar image is normalized.

3. The method according to claim 1, characterized in that, The acquisition of ocean reanalysis data matching the spatial location and the imaging date includes the following steps: The first marked spatial location is taken as the target location; Obtain information on sea surface height anomaly, sea surface temperature, sea surface salinity, and upper ocean current velocity within a preset range near the target location on the corresponding imaging date; The upper ocean current velocity information representation includes velocity profiles of multiple vertical layers; The velocity profile is divided into a surface layer and a subsurface layer, and then principal component analysis is used to extract the principal component features of the surface layer and the subsurface layer as the velocity features. The next marked spatial location is taken as the target location, and the process returns to the step of obtaining information on sea surface height anomalies, sea surface temperature, sea surface salinity, and upper ocean current velocity within a preset range near the target location on the corresponding imaging date, until all spatial locations have been traversed.

4. The method according to claim 1, characterized in that, Assigning category labels to each multi-channel two-dimensional physical field tensor based on the vortex coverage ratio of the preset central region of the two-dimensional grid window to form a training sample set includes the following steps: The first multi-channel two-dimensional physical field tensor is used as a candidate input sample; Extract the local image of the region corresponding to the two-dimensional grid window from the initial radar image corresponding to the candidate input sample; When the proportion of the vortex region in the preset central region of the local image is greater than a preset threshold, the category label of the positive vortex sample is associated with the candidate input sample to form a training sample. When the proportion of the vortex region in the preset central region of the local image is less than or equal to the preset threshold, the category label of the non-vortex negative sample is associated with the candidate input sample to form a training sample; The next multi-channel two-dimensional physical field tensor is used as the candidate input sample. The process of extracting the local image of the corresponding region of the two-dimensional grid window from the initial radar image corresponding to the candidate input sample is repeated until all the multi-channel two-dimensional physical field tensors are traversed. All the training samples are then summarized to form the training sample set.

5. The method according to claim 1 or 4, characterized in that, The method further includes the following steps: The first training sample in the training sample set is used as a candidate training sample. Randomly generate an offset within a preset grid range, and use the offset to spatially offset the two-dimensional grid window corresponding to the candidate training sample to obtain the offset spatial range; The offset grid window of the offset spatial range of each parameter type is cropped from the ocean reanalysis data. The offset grid windows of all parameter types are stacked along the channel dimension to form a multi-channel amplified physical field tensor. The multi-channel amplified physical field tensor is assigned a category label according to the vortex coverage ratio of the preset central region of the offset grid window to form an amplified sample. The next training sample in the training sample set is taken as the candidate training sample, and the step of randomly generating an offset in the preset grid range is returned to be executed until all training samples in the training sample set have been traversed. All the amplified samples are added to the training sample set, and then data augmentation operations are performed on all samples in the training sample set to update the training sample set.

6. The method according to claim 1, characterized in that, The step of using the training sample set to train a preset deep learning model to obtain a vortex prediction model includes the following steps: The first feature is obtained by performing channel mapping and initial feature extraction on the multi-channel two-dimensional physical field tensor through the two-dimensional convolutional layer; Based on the initial features, spatial features at different scales are extracted through the backbone network using the moving inverse residual convolution module, and the spatial features output by the last layer of the backbone network are used as the second features. Based on the second feature, Q matrix, K matrix and V matrix are constructed through attention module, and then recovered by convolution after attention weighted aggregation to obtain the third feature; A fourth feature is obtained by performing a residual connection between the second feature and the third feature; Based on the fourth feature, the probability output layer uses global average pooling and a fully connected layer to output logical values, and then the sigmoid function is used to convert the logical values ​​into the probability of vortex occurrence. Based on the category label and vortex occurrence probability corresponding to the multi-channel two-dimensional physical field tensor of each training sample in the training sample set, the deep learning model is optimized by tuning the parameters of the binary cross-entropy loss function to obtain the vortex prediction model.

7. The method according to claim 1, characterized in that, The method for generating a vortex occurrence probability field for the target area on the target date using ocean reanalysis data of the target area on the target date and employing a sliding window prediction strategy based on the vortex prediction model includes the following steps: The ocean reanalysis data of the target area on the target date is used as the data to be analyzed. A sliding window prediction strategy is adopted, in which the sliding window slides row by row and column by column on the spatial grid of the target area with a preset sliding step size, and the sub-data to be analyzed at each position of the target area is extracted through the sliding window in the data to be analyzed; The sub-data to be analyzed is input into the vortex prediction model so that the vortex prediction model can predict the vortex prediction probability value at each location in the target area, and then the vortex occurrence probability field of the target area on the target date is obtained by summarizing and organizing the data.

8. A device for predicting the probability of mesoscale vortex occurrence, characterized in that, The device includes: The first module is used to acquire synthetic aperture radar (SAR) images of the study area, and preprocess the SAR images to obtain initial radar images; wherein the SAR images are associated with the imaging date. The second module is used to mark the spatial positions of vortex samples and non-vortex samples in the initial radar image in response to human visual interpretation of the target object. The third module is used to acquire ocean reanalysis data that matches the spatial location and the imaging date; wherein, the parameter types of the ocean reanalysis data include sea surface height anomalies, sea surface temperature, sea surface salinity, and current velocity characteristics; The fourth module is used to use the spatial location as the center to cut out a two-dimensional grid window with a preset spatial range for each parameter type from the ocean reanalysis data, stack the two-dimensional grid windows of all parameter types along the channel dimension to form a multi-channel two-dimensional physical field tensor, and assign a category label to each multi-channel two-dimensional physical field tensor according to the vortex coverage ratio of the preset central region of the two-dimensional grid window to form a training sample set; wherein, the preset central region is a preset region range for mesoscale vortices, and the category label includes positive vortex samples and negative non-vortex samples; The fifth module is used to perform prediction training on a preset deep learning model using the training sample set to obtain a vortex prediction model. The deep learning model includes a two-dimensional convolutional layer, a backbone network, an attention module, and a probability output layer. The sixth module is used to generate the vortex occurrence probability field of the target area on the target date based on the ocean reanalysis data of the target area on the target date using the vortex prediction model and a sliding window prediction strategy.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images