Mesoscale eddy identification method based on deep learning

Abstract

The invention discloses a mesoscale eddy identification method based on deep learning, and the method comprises the following steps: S1, obtaining sea surface abnormal data, carrying out the preprocessing of the sea surface abnormal data, and generating a training set; s2, inputting the training set into a PAM-ResNet model, and setting a loss function and an optimizer; s3, hyper-parameters of the PAM-ResNet model are set, the PAM-ResNet model is trained, and a mesoscale vortex recognition model is obtained; and S4, analyzing the abnormal data of the sea surface through the mesoscale vortex recognition model, and completing ocean mesoscale vortex recognition. Compared with an existing neural network learning method, the method has the advantages that the PAM-ResNet model is compared with the existing neural network learning method, the result shows that the PAM-ResNet model has the best recognition effect on the mesoscale vortex, higher feature extraction fusion capability and generalization are achieved, and the method has feasibility and remarkable superiority in specific tasks of mesoscale vortex recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of mesoscale eddy identification, and particularly relates to a mesoscale eddy identification method based on deep learning. Background Art

[0002] As an important ocean phenomenon, mesoscale eddies are widely distributed in the world's oceans, having a profound impact on the marine ecosystem, fishery resources, and hydrographic conditions. They also play a crucial role in the transfer of matter and energy in the ocean, ocean dynamics processes, and biogeochemical cycles. Therefore, accurate identification of mesoscale eddies not only has important practical application significance but also has extremely high scientific research value.

[0003] Traditional mesoscale eddy identification methods rely on physical parameters or geometric features, suffering from problems such as being easily affected by noise and low efficiency. In recent years, machine learning methods have shown unique advantages in the extraction and automatic identification of mesoscale eddy features in the ocean. However, existing machine learning methods are prone to false detection due to insufficient sample training sets and lack of multi-scale feature fusion learning of the target, and are also prone to missing small-boundary mesoscale eddies. Summary of the Invention

[0004] Aiming at the above deficiencies in the prior art, a mesoscale eddy identification method based on deep learning provided by the present invention solves the problem of easy false detection in the prior art for mesoscale eddy identification.

[0005] To achieve the above invention purpose, the technical solution adopted by the present invention is: a mesoscale eddy identification method based on deep learning, comprising the following steps:

[0006] S1. Obtain sea surface anomaly data, preprocess the sea surface anomaly data, and generate a training set;

[0007] S2. Input the training set into the PAM-ResNet model, and set the loss function and optimizer;

[0008] S3. Set the hyperparameters of the PAM-ResNet model, train the PAM-ResNet model, and obtain a mesoscale eddy identification model;

[0009] S4. Analyze the sea surface anomaly data through the mesoscale eddy identification model to complete the identification of ocean mesoscale eddies.

[0010] Furthermore: In the above S1, the method for preprocessing the sea surface anomaly data is specifically:

[0011] S11. Use an automatic identification method based on closed contour lines to generate label data according to the sea level anomaly data;

[0012] S12. Augment the sea level anomaly data and the corresponding label data through data augmentation techniques to construct a training set.

[0013] Further: In the above S11, the label data includes non-cyclones, cyclone vortices, and anticyclone vortices.

[0014] Further: In the above S2, the method of inputting the training set into the PAM-ResNet model to obtain the prediction result is specifically as follows:

[0015] S21. Perform encoder processing and max-pooling layer downsampling processing on the input data three times continuously to obtain a first feature map. And after each encoder processing, input the generated feature map and the upper-layer feature map into the FPN module and the attention module to generate an intermediate feature map;

[0016] S22. Input the first feature map into the intermediate layer to obtain a second feature map;

[0017] S23. Perform upsampling operations and decoder processing on the second feature map three times continuously to obtain a third feature map. And during each decoder processing, splice the intermediate feature map onto the generated feature map through skip connections;

[0018] S24. Obtain the prediction result according to the output feature map.

[0019] Further: In the above S21, residual units are set in the encoder, intermediate layer, and decoder. The residual unit includes a 3*3 convolutional layer, a batch normalization layer, a ReLU activation function layer, and a Dropout layer connected in sequence. Among them, the output of the Dropout layer and the input of the 3*3 convolutional layer are added through skip connections as the output of the residual unit;

[0020] In the above S21, the FPN module structures of each layer are the same and all include:

[0021] The input end of the first 1*1 convolutional layer is connected to the encoder of the current layer, and the output end of the first 1*1 convolutional layer is added to the elements of the upsampled upper-layer feature map as the output of the FPN module;

[0022] Among them, the top FPN module is connected to the intermediate layer through a second 1*1 convolutional layer to create an upper-layer feature map. From the top FPN module downwards, each layer of the FPN module is connected to the attention sub-module and FPN module of the previous layer to create an upper-layer feature map;

[0023] In the above S21, the attention module includes a ReLU activation function layer, a third 1*1 convolutional layer, a sigmoid activation function layer, and a weight layer connected in sequence. The output of the weight layer and the input of the first 1*1 convolutional layer of the current layer are multiplied through skip connections to obtain the output of the attention module.

[0024] Furthermore, in S2, the expression of the loss function Weighted Dice is specifically as follows:

[0025]

[0026] In the formula, w k is the weight of class k, used to reflect the importance or priority of this class, K is the total number of classes, and Dice k is the Dice coefficient, and its expression is specifically as follows:

[0027]

[0028] In the formula, represents the predicted value of the i-th pixel of class k, represents the true value of the i-th pixel of class k, and N represents the total number of all pixels in the image;

[0029] The optimizer is specifically the Adam optimizer, which is used to adjust the parameters of the model during training to minimize the loss function.

[0030] Furthermore, in S3, the method for setting the hyperparameters of the PAM-ResNet model is specifically as follows:

[0031] The training batch epoch is set to 80, the learning rate is set to 5e -5 , the training batch size is set to 8, and the EarlyStopping mechanism, ModelCheckpoint mechanism, and reducecall mechanism are all set during training. The EarlyStopping mechanism is used to prematurely terminate training when the validation loss value does not significantly decrease for 50 consecutive iterations; the ModelCheckpoint mechanism is used to monitor the validation loss and only save the optimal model weights; the ReduceLROnPlateau mechanism dynamically reduces the learning rate during the validation loss to ensure that the model can converge more precisely.

[0032] The beneficial effects of the present invention are as follows:

[0033] (1) The present invention provides a mesoscale vortex identification method based on deep learning, and innovatively proposes to construct a mesoscale vortex identification model using the PAM-ResNet model. In order to verify the effectiveness and superiority of the PAM-ResNet model in the mesoscale vortex identification task, it is compared with existing neural network learning methods. The results show that the PAM-ResNet model has the best identification effect on mesoscale vortices, has stronger feature extraction and fusion capabilities and generalization ability, and thus has feasibility and significant superiority in specific mesoscale vortex identification tasks.

[0034] (2) The present invention comprehensively considers accuracy, recall rate, and Dice coefficient. The PAM-ResNet model shows good comprehensive performance in the mesoscale vortex identification task, not only being able to correctly identify and classify vortices, but also accurately restoring the contour and boundary information of vortices. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a flowchart of a mesoscale vortex identification method based on deep learning according to the present invention.

[0036] Figure 2 It is a structure diagram of PAM-ResNet.

[0037] Figure 3 It is a structure diagram of a residual unit.

[0038] Figure 4 It is a structure diagram of an FPN module.

[0039] Figure 5 It is a structure diagram of an attention module.

[0040] Figure 6 It is a structure diagram of ResNet.

[0041] Figure 7 It is a structure diagram of a ResNet-FPN network model. DETAILED DESCRIPTION OF THE INVENTION

[0042] The following describes the specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the accuracy and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

[0043] As Figure 1 shown, in an embodiment of the present invention, a mesoscale vortex identification method based on deep learning includes the following steps:

[0044] S1. Obtain sea surface anomaly data, preprocess the sea surface anomaly data to generate a training set;

[0045] S2. Input the training set into the PAM-ResNet model and set the loss function and optimizer;

[0046] S3. Set the hyperparameters of the PAM-ResNet model, train the PAM-ResNet model to obtain a mesoscale vortex identification model;

[0047] S4. Analyze the sea surface anomaly data through the mesoscale eddy identification model to complete the identification of ocean mesoscale eddies.

[0048] In this embodiment, the present invention obtains the sea surface anomaly data by accessing the CMEMS website. The spatial range is selected as the South China Sea area (12° - 23°N, 111°E - 121°E), and the time is selected as the data from January 1, 2001 to December 31, 2023 for a total of 23 years as the research data. The time resolution of this data is 1 day, and the spatial resolution is 0.125° × 0.125°.

[0049] In S1, the method for preprocessing the sea surface anomaly data is specifically as follows:

[0050] S11. Use the automatic identification method based on closed contour lines to generate label data according to the sea level anomaly data.

[0051] In this embodiment, the method for generating label data is specifically as follows:

[0052] (1) Initial value setting: Set the range of latitude from 12.0625 to 22.9375 and longitude from 111.0625 to 120.9375 in the South China Sea area, and the size of each pixel is 0.125 degrees. Calculate the area of each pixel by setting the radius of the earth, and set the initial date as January 1, 2001 to prepare for subsequent date processing.

[0053] (2) Initial identification: Identify the extreme points in each SLA snapshot, which are defined as the highest or lowest points among the 9 adjacent grid points in the 3×3 grid.

[0054] (3) Contour search: For each extreme point, starting from this point, gradually adjust the threshold in steps of 1 cm to find the largest closed contour at this threshold. Ensure that only one extreme point is included within the contour to ensure the uniqueness of the vortex.

[0055] (4) Vortex confirmation: The closed contour needs to meet the following criteria: Only one extreme point is included within the contour, the number of pixels (I) included satisfies ≥9, and the vortex amplitude ≥1 cm. When the above criteria are met, it is recognized as a mesoscale eddy.

[0056] (5) Data classification: According to the identified vortex types, the data is divided into three categories, including non - cyclone, cyclone vortex, and anticyclone vortex.

[0057] S12. Augment the sea level anomaly data and the corresponding label data through data augmentation technology to construct a training set.

[0058] In S11, the label data includes non - cyclone, cyclone vortex, and anticyclone vortex.

[0059] AsFigure 2 As shown, in S2, the method of inputting the training set into the PAM-ResNet model to obtain the prediction result is specifically as follows:

[0060] S21. Perform encoder processing and max pooling layer downsampling processing on the input data three times continuously to obtain the first feature map, and input the generated feature map and the upper layer feature map into the FPN module and the attention module after each encoder processing to generate the intermediate feature map;

[0061] S22. Input the first feature map into the intermediate layer to obtain the second feature map;

[0062] S23. Perform upsampling operation and decoder processing on the second feature map three times continuously to obtain the third feature map, and splice the intermediate feature map onto the generated feature map through skip connection during each decoder processing;

[0063] S24. Obtain the prediction result according to the output feature map.

[0064] As Figure 3 shown, in S21, residual units are set in the encoder, intermediate layer and decoder. The residual unit includes a 3*3 convolutional layer, a batch normalization layer, a ReLU activation function layer and a Dropout layer connected in sequence. Among them, the output of the Dropout layer and the input of the 3*3 convolutional layer are added through skip connection as the output of the residual unit;

[0065] As Figure 4 shown, in S21, the FPN module structures of each layer are the same and all include:

[0066] The input end of the first 1*1 convolutional layer is connected to the encoder of the current layer, and the output end of the first 1*1 convolutional layer is added to the elements of the upsampled upper layer feature map as the output of the FPN module;

[0067] Among them, the top FPN module is connected to the intermediate layer through the second 1*1 convolutional layer to create the upper layer feature map. The FPN modules of each layer from the top FPN module down are connected to the attention sub-module and FPN module of the previous layer to create the upper layer feature map;

[0068] As Figure 5 shown, in S21, the attention module includes a ReLU activation function layer, a third 1*1 convolutional layer, a sigmoid activation function layer and a weight layer connected in sequence. The output of the weight layer and the input of the first 1*1 convolutional layer of the current layer are multiplied through skip connection to obtain the output of the attention module.

[0069] As Figure 2As shown on the right side, in S2, the PAM-ResNet model includes a ResNet network with an encoder-decoder architecture and an FPN module, and an attention sub-module is set in each layer of the pyramid in the FPN module.

[0070] In this embodiment, the construction process of the PAM-ResNet model is as follows:

[0071] (1) Construct a ResNet network model

[0072] As Figure 6 shown, in this embodiment, ResNet adopts an encoder-decoder structure as the overall architecture, and residual units are added in the encoder, central block, and decoder. The structure of the residual unit is as Figure 3 shown.

[0073] In the encoder part, three layers of improved residual units are adopted, and a max-pooling layer is added after each residual unit for downsampling. This residual unit is designed to alleviate the problem of gradient disappearance in the training of deep networks and accelerate convergence at the same time. The structure of the residual unit is as follows: it contains two layers of convolutional operations, and the input is directly added to the output through a skip connection to ensure the lossless transmission of information. The application of batch normalization (BatchNormalization) and the ReLU activation function ensures stable gradient flow and non-linear expression ability. To enhance the generalization performance of the model, a Dropout mechanism is introduced to prevent overfitting and improve robustness. Finally, after reconstructing the input features into the same shape as the output features, residual addition is performed.

[0074] The central block is similar to the encoder, and one layer of residual unit is adopted to further extract global features and serve as a bridge connecting the encoder and the decoder.

[0075] In the decoder part, UpSampling2D is used for upsampling. Then the upsampling result is concatenated with the feature map of the corresponding layer in the encoder to fuse detailed information. Finally, the fused features are processed using an improved residual unit.

[0076] (2) Construction of the ResNet-FPN network model

[0077] The ResNet-FPN network model is an image segmentation architecture that adds a Feature Pyramid Network (FPN) on the basis of ResNet, and its network structure is as Figure 7 shown. This network constructs a feature pyramid and fuses features at different levels, significantly enhancing the detection ability for multi-scale objects. This fusion strategy not only enables the network to simultaneously focus on global and local features, but also particularly improves the detection accuracy for small objects.

[0078] After the central block of ResNet and before the decoder, the present invention adds an FPN module, and its structure is as Figure 4 shown. FPN starts from the output of the central layer of ResNet and creates the top-level feature map of FPN through a 1×1 convolutional layer. Then, by creating lateral connections, features at different levels are fused together. Specifically, these connections first use a 1x1 convolutional layer to unify the dimensions of the higher-level features, then upsample them to the same spatial size as the lower-level features, and finally add them together to obtain a new feature map. Each connection point in this process uses 1x1 convolution and upsampling techniques to match the spatial size and number of channels, ensuring the effective fusion of features at different levels, retaining more spatial details while introducing high-level semantic information, effectively solving the problem of poor small object detection often encountered in single-scale feature extraction, and at the same time retaining the accuracy of large objects.

[0079] (3) Construction of the PAM-ResNet network model

[0080] Although ResNet-FPN is effective in multi-scale feature representation, it cannot aggregate more discriminative features for segmentation. To further improve the segmentation ability of the model, the present invention proposes a PAM module and fuses it with ResNet to obtain the PAM-ResNet model. Its network structure is as Figure 2 shown, and the structure of the attention sub-module is as Figure 5 shown.

[0081] In the attention mechanism module, first, the number of channels of the upper-layer features and the current-layer features are respectively adjusted through the FPN modules of each layer. Then, the adjusted top-level features and the current-layer features are added together and passed through the ReLU activation function to obtain the fused features. After that, the fused features are passed through a 1×1 convolution to generate a single-channel attention weight, and it is mapped to between [0,1] through the Sigmoid activation function. Finally, the generated attention weight is applied to the features of the current layer.

[0082] In this embodiment, the core goal of the attention sub-module is to enhance the model's attention to key information while reducing attention to irrelevant information. By adding an attention module to each layer of FPN, before fusing the top-layer features (top_layer) and the current-layer features (lateral_layer), the two features are adaptively weighted to improve the model's ability to capture important features, so that the feature maps at each scale can be better fused with each other.

[0083] In step S2, the specific expression of the loss function Weighted Dice is:

[0084]

[0085] where w k is the weight of class k, used to reflect the importance or priority of this class, K is the total number of classes, and Dice k is the Dice coefficient, and its specific expression is:

[0086]

[0087] where represents the predicted value of the i-th pixel of class k, represents the true value of the i-th pixel of class k, and N represents the total number of all pixels in the image;

[0088] In this embodiment, the present invention uses a weighted Dice coefficient as a loss function to handle the data imbalance problem. The weighted Dice coefficient is an improved form of the standard Dice coefficient, mainly used for multi-class segmentation tasks, especially when the class distribution is imbalanced. For example, the targets of some classes appear with a very low frequency or occupy a small proportion of pixels in the image. In this case, higher weights need to be assigned to these small classes to prevent the model from ignoring them.

[0089] In this embodiment, the number of pixels of each class is first counted, and then the reciprocal of the frequency of each class is calculated. The reciprocal of the frequency of each class divided by the sum of the reciprocals of the frequencies of all classes is the weight w of each class. The number of pixels and weights of each class are shown in Table 1.

[0090] Table 1 Number of pixels and weights of each class

[0091]

[0092] The optimizer is specifically the Adam optimizer, which is used to adjust the parameters of the model during training to minimize the loss function.

[0093] The optimizer is used to adjust the parameters of the model during training to minimize the loss function. In this embodiment, the Adam optimizer is selected, which has good performance and adaptability. The learning rate can be set relatively large at the beginning to achieve rapid convergence; in the later stage, the learning rate is reduced to finely adjust the model parameters and avoid oscillation.

[0094] To comprehensively evaluate the performance of the model, this embodiment uses the classification accuracy to measure the proportion of each pixel point predicted by the model that exactly matches the true class, and further evaluates the overall accuracy of the model classification. Calculate the average Dice coefficient of all classes to evaluate the overall segmentation performance of the model. Calculate the average Dice coefficient weighted according to the importance of the classes to obtain the weighted average Dice coefficient, and further evaluate the performance of the model in dealing with the class imbalance problem.

[0095] In S3, the method for setting the hyperparameters of the PAM-ResNet model is specifically as follows:

[0096] The number of training epochs is set to 80, the learning rate is set to 5e -5 , the batch size is set to 8, and the EarlyStopping mechanism, ModelCheckpoint mechanism, and ReduceLROnPlateau mechanism are all set during training. The EarlyStopping mechanism is used to prematurely terminate training when the validation loss value does not significantly decrease for 50 consecutive iterations; the ModelCheckpoint mechanism is used to monitor the validation loss and only save the optimal model weights; the ReduceLROnPlateau mechanism dynamically reduces the learning rate during the validation loss to ensure that the model can converge more precisely.

[0097] In this embodiment, in order to verify the recognition effect of the mesoscale vortex recognition model constructed by the present invention, the present invention provides the following comparative experimental data.

[0098] In this embodiment, the classification accuracies, recall rates, average dice coefficients, weighted dice coefficients, and average accuracies of different types of cyclones of the ResNet model, ResNet-FPN model, and PAM-ResNet model are comparatively analyzed. The data selects the SLA and label data in the South China Sea region in 2023, and the classification accuracy, recall rate, Dice coefficient, average Dice coefficient, and weighted Dice coefficient are used as the evaluation indicators of the model, as shown in Tables 2 to 4.

[0099] Table 2 Statistical results of ResNet segmentation

[0100]

[0101] Table 3 Statistical results of ResNet-FPN segmentation

[0102]

[0103] Table 4 Statistical results of PAM-ResNet segmentation

[0104]

[0105] In terms of vortex classification, the accuracy of PAM-ResNet in cyclone and anticyclone categories reached 81% and 76%, respectively, which was 6% and 2% higher than that of ResNet before improvement. At the same time, compared with the classification accuracy of ResNet-FPN, PAM-ResNet has a slight improvement of 6% and 1% respectively. Through the above comparison, it is analyzed that in vortex classification, the PAM-ResNet proposed in the present invention has successfully improved the ResNet model in two steps, and has significantly improved the vortex classification function of the ResNet model.

[0106] In terms of vortex recall, the similar recall rates of the ResNet model, ResNet-FPN model, and PAM-ResNet model indicate that during the improvement process, both ResNet-FPN and PAM-ResNet inherited the low false negative advantage of the ResNet model for vortex recognition.

[0107] In terms of the similarity between the vortex area and the real vortex area, the dice coefficients of PAM-ResNet for cyclones, anticyclones and non-cyclones reached 82%, 82% and 97% respectively, which are significantly improved compared with the ResNet and ResNet-FPN before the improvement. The above comparative analysis shows that in terms of the similarity between the vortex area predicted by the model and the real vortex area, the PAM-ResNet model can more accurately identify the boundaries and areas of the vortex than the ResNet model and the ResNet-FPN model.

[0108] Based on the above analysis, we can draw the following conclusions:

[0109] (1) Significant improvement: The PAM-ResNet model has significantly improved the classification accuracy of cyclones and anticyclones compared with the original ResNet model, which shows the effectiveness and superiority of the PAM-ResNet model in the vortex classification task.

[0110] (2) Overall performance optimization: The PAM-ResNet model has shown effective improvements over the ResNet model in multiple evaluation indicators such as classification accuracy, recall rate, and Dice coefficient. This not only improves the prediction accuracy of the model, but also enhances the reliability and practicality of the model in practical applications.

[0111] (3) Verification of model improvement direction: By comparing the performance of ResNet, ResNet-FPN and PAM-ResNet models, it is verified that the direction of improvement based on ResNet is correct, and the PAM-ResNet model is more effective in vortex classification and recognition tasks.

[0112] In addition, the present invention selects the recognition results of the commonly used U-Net model for mesoscale eddy recognition, the advanced DeeplabV3plus model for mesoscale eddy recognition, and the PAM-ResNet model innovatively used in the present invention for comparative analysis. The data selects the SLA and label data in the South China Sea region in 2023, and uses classification accuracy, recall rate, Dice coefficient, average Dice coefficient, and weighted Dice coefficient as the evaluation indicators of the model, as shown in Tables 5 to 7.

[0113] Table 5 Statistical results of U-Net segmentation

[0114]

[0115] Table 6 Statistical results of DeeplabV3plus segmentation

[0116]

[0117] Table 7 Statistical results of PAM-ResNet segmentation

[0118]

[0119] In terms of vortex classification, the 94.5% of the PAM-ResNe model is higher than the 93% of the U-Net model and the 93.5% of the DeeplabV3plus model in terms of average accuracy. In the categories of cyclones and anticyclones, the accuracy differences among the U-Net model, DeeplabV3plus model, and PAM-ResNet model do not exceed 4%. In the non-cyclone category, the accuracy of the PAM-ResNet model reaches 98%, which is significantly higher than the 95% of the U-Net model and the 95% of the DeeplabV3plus model. Compared with the U-Net model, the accuracy of the PAM-ResNet model in the cyclone category increases by 4%, and the accuracy in the anticyclone category decreases by 5%. Compared with DeepLabV3plus, the accuracy of the PAM-ResNet model in the cyclone category increases by 1%, and the accuracy in the anticyclone category decreases by 3%. Although the average accuracy of DeepLabV3plus is 1% lower than that of PAM-ResNet, the recognition accuracy differences of PAM-ResNet for cyclones and anticyclones are relatively high. While the recognition accuracies of DeepLabV3plus for cyclones and anticyclones are almost the same. This indicates that compared with U-Net and PAM-ResNet, DeepLabV3plus has more stable classification ability for cyclones and anticyclones.

[0120] In terms of vortex recall, the recall rates of the PAM-ResNet model for cyclone and anticyclone categories are significantly higher than those of the U-Net model and the DeeplabV3plus model. As shown in Table 7, the recall rates of the PAM-ResNet model for cyclones and anticyclones are both 86%, which are 9% and 16% higher than those of the U-Net model, respectively. At the same time, compared with the DeeplabV3plus model, the recall rates are 10% and 8% higher, respectively. The above comparison shows that the PAM-ResNet model has a significant advantage of low false negative rate among existing models for vortex identification.

[0121] In terms of the similarity between the vortex region and the true vortex region, the Dice coefficients of PAM-ResNet for cyclones, anticyclones, and non-cyclones reach 82%, 82%, and 97%, respectively, showing a slight improvement compared with the 75%, 75%, 96% of the commonly used U-Net model for mesoscale vortex identification and the 76%, 78%, 96% of the advanced DeeplabV3plus model for mesoscale vortex identification. Through the above comparative analysis, it shows that in terms of the similarity between the vortex region predicted by the model and the true vortex region, the PAM-ResNet model can identify the boundary and region of the vortex more accurately than the U-Net model and the DeeplabV3plus model.

[0122] Based on the above analysis, the following conclusions can be drawn:

[0123] (1) Model stability: The DeepLabV3plus model has almost the same recognition accuracy for cyclones and anticyclones, showing good stability. This indicates that the DeepLabV3plus model can maintain relatively consistent performance when dealing with cyclone and anticyclone categories.

[0124] (2) Model recall performance: The recall rates of the PAM-ResNet model for cyclone and anticyclone categories are significantly higher than those of the U-Net and DeeplabV3plus models, showing a lower false negative rate. This indicates that the PAM-ResNet model can cover the target area more comprehensively and reduce omissions when identifying mesoscale vortices.

[0125] (3) Similarity of vortex boundary and region: The PAM-ResNet model has an improvement in the Dice coefficient compared with the U-Net and DeeplabV3plus models, which indicates that the PAM-ResNet model is more accurate in predicting the similarity between the vortex region and the true vortex region and can better identify the boundary and region of the vortex.

[0126] (4) Overall performance: Considering accuracy, recall rate, and Dice coefficient comprehensively, the PAM-ResNet model shows good overall performance in the mesoscale vortex identification task.

[0127] The beneficial effects of the present invention are as follows: The present invention provides a method for identifying mesoscale vortices based on deep learning, and innovatively proposes a PAM-ResNet model to construct a mesoscale vortex identification model. In order to verify the effectiveness and superiority of the PAM-ResNet model in the task of mesoscale vortex identification, it is compared with existing neural network learning methods. The results show that the PAM-ResNet model has the best identification effect on mesoscale vortices, has stronger feature extraction and fusion capabilities and generalization, and thus has feasibility and significant superiority in specific tasks of mesoscale vortex identification.

[0128] The present invention comprehensively considers accuracy, recall rate, and Dice coefficient. The PAM-ResNet model shows good comprehensive performance in the task of mesoscale vortex identification, and can not only correctly identify and classify vortices, but also accurately restore the contour and boundary information of vortices.

[0129] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by terms such as "center", "thickness", "upper", "lower", "horizontal", "top", "bottom", "inner", "outer", "radial", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be understood as a limitation of the present invention. In addition, the terms "first", "second", and "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of technical features. Therefore, the features defined by "first", "second", and "third" may explicitly or implicitly include one or more of such features.

Claims

1. A mesoscale eddy identification method based on deep learning, characterized in that: The following steps are involved: S1. Acquire sea surface anomaly data, pre-process the sea surface anomaly data, and generate a training set; S2. Input the training set into the PAM-ResNet model and set the loss function and optimizer. S3, setting the hyperparameters of the PAM-ResNet model, training the PAM-ResNet model, and obtaining a mesoscale eddy recognition model; S4. Analyze the abnormal data on the sea surface through the mesoscale eddy identification model to complete the identification of ocean mesoscale eddies.

2. The mesoscale vortex identification method based on deep learning according to claim 1, characterized in that: In S1, the method for preprocessing the sea surface abnormal data is specifically as follows: S11, using an automatic recognition method based on closed contour lines to generate label data according to sea level anomaly data; S12. Expand the sea level anomaly data and the corresponding label data through data enhancement technology to build a training set.

3. The mesoscale vortex identification method based on deep learning according to claim 2, characterized in that: In S11, the label data includes non-cyclonic, cyclonic vortex and anticyclonic vortex.

4. The mesoscale vortex identification method based on deep learning according to claim 1, characterized in that: In S2, the training set is input into the PAM-ResNet model to obtain the prediction result as follows: S21, performing encoder processing and maximum pooling layer downsampling processing on the input data three times in a row to obtain a first feature map, and after each encoder processing, inputting the generated feature map and the upper layer feature map into the FPN module and the attention module to generate an intermediate feature map; S22, input the first feature map into the middle layer to obtain a second feature map; S23, performing upsampling operations and decoder processing on the second feature map three times in a row to obtain a third feature map, and splicing the intermediate feature map to the generated feature map through a jump connection during each decoder processing; S24. Obtain prediction results based on the output feature map.

5. The mesoscale vortex identification method based on deep learning according to claim 4 is characterized in that: In the S21, a residual unit is provided in the encoder, the intermediate layer and the decoder, and the residual unit includes a 3*3 convolution layer, a batch normalization layer, a ReLU activation function layer and a Dropout layer connected in sequence, wherein the output of the Dropout layer and the input of the 3*3 convolution layer are added through a skip connection as the output of the residual unit; In S21, the FPN modules of each layer have the same structure, and all include: The input of the first 1*1 convolutional layer is connected to the encoder of the current layer, and the output of the first 1*1 convolutional layer is added to the upsampled upper feature map elements as the output of the FPN module; Among them, the top-level FPN module is connected to the middle layer through the second 1*1 convolutional layer to create the upper-level feature map. The FPN modules of each layer below the top-level FPN module are connected to the attention submodule and FPN module of the previous layer to create the upper-level feature map. In S21, the attention module includes a ReLU activation function layer, a third 1*1 convolution layer, a sigmoid activation function layer and a weight layer connected in sequence. The output of the weight layer and the input of the first 1*1 convolution layer of the current layer are multiplied by a jump connection to obtain the output of the attention module.

6. The mesoscale vortex identification method based on deep learning according to claim 5 is characterized in that: In S2, the expression of the loss function Weighted Dice is specifically: In the formula, w k is the weight of category k, which is used to reflect the importance or priority of the category. K is the total number of categories. k is the Dice coefficient, and its specific expression is: In the formula, p i k represents the predicted value of the i-th pixel of category k, g i k represents the true value of the i-th pixel of category k, and N represents the total number of all pixels in the image; The optimizer is specifically the Adam optimizer, which is used to adjust the parameters of the model during the training process to minimize the loss function.

7. The mesoscale vortex identification method based on deep learning according to claim 1 is characterized in that: In S3, the method for setting the hyperparameters of the PAM-ResNet model is specifically as follows: The training batch epoch is set to 80 and the learning rate is set to 5e -5 The batch size of training is set to 8. The EarlyStopping mechanism, ModelCheckpoint mechanism and reducecall mechanism are set during training. The EarlyStopping mechanism is used to terminate the training early when the validation loss value has not decreased significantly for 50 consecutive iterations; the ModelCheckpoint mechanism is used to monitor the validation loss and save only the optimal model weights; the ReduceLROnPlateau mechanism dynamically reduces the learning rate during the validation loss to ensure that the model can converge more finely.

Images