A Deep Learning-Based Mesoscale Eddy Recognition Method

Abstract

This invention discloses a deep learning-based method for identifying mesoscale eddies, comprising the following steps: S1, acquiring sea surface anomaly data, preprocessing the data, and generating a training set; S2, inputting the training set into a PAM-ResNet model, and setting the loss function and optimizer; S3, setting the hyperparameters of the PAM-ResNet model, training the model to obtain a mesoscale eddy identification model; S4, analyzing the sea surface anomaly data using the mesoscale eddy identification model to complete the identification of oceanic mesoscale eddies. This invention compares the PAM-ResNet model with existing neural network learning methods, showing that the PAM-ResNet model has the best identification effect on mesoscale eddies, with stronger feature extraction and fusion capabilities and generalization ability, thus demonstrating feasibility and significant superiority in specific tasks related to mesoscale eddy identification.

Description

Technical Field

[0001] This invention belongs to the field of mesoscale eddy recognition technology, specifically relating to a mesoscale eddy recognition method based on deep learning. Background Technology

[0002] Mesoscale eddies, as an important ocean phenomenon, are widely distributed in the world's oceans, exerting a profound impact on marine ecosystems, fishery resources, and hydrological conditions. They also play a crucial role in ocean mass and energy transport, ocean dynamics, and biogeochemical cycles. Therefore, accurate identification of mesoscale eddies is not only of significant practical application value but also of extremely high scientific research value.

[0003] Traditional mesoscale eddy identification methods rely on physical parameters or geometric features, which are susceptible to noise and inefficient. In recent years, machine learning methods have shown unique advantages in the feature extraction and automatic identification of ocean mesoscale eddies. However, existing machine learning methods are prone to false detections and missed detections of small-boundary mesoscale eddies due to insufficient training sets and lack of multi-scale feature fusion learning of targets. Summary of the Invention

[0004] To address the aforementioned shortcomings in existing technologies, this invention provides a deep learning-based mesoscale eddy identification method that solves the problem of false detections in existing mesoscale eddy identification technologies.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a mesoscale eddy recognition method based on deep learning, comprising the following steps:

[0006] S1. Acquire 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 the mesoscale eddy recognition model.

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

[0010] Further: In step S1, the method for preprocessing the sea surface anomaly data specifically includes:

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

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

[0013] Furthermore, in S11, the tag data includes non-cyclones, cyclonic vortices, and anticyclonic vortices.

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

[0015] S21. Perform encoder processing and max pooling downsampling processing on the input data three times in a row to obtain the first feature map. 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 the intermediate feature map.

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

[0017] S23. Perform three consecutive upsampling operations and decoder processing on the second feature map to obtain the third feature map, and in each decoder processing process, the intermediate feature map is spliced ​​onto the generated feature map through skip connections.

[0018] S24. Obtain the prediction result based on the output feature map.

[0019] Further: In S21, residual units are provided 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. The output of the Dropout layer and the input of the 3*3 convolutional layer are added through a skip connection as the output of the residual unit.

[0020] In S21, the FPN module structure is the same for each layer, and all include:

[0021] 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 feature map elements of the upper layer as the output of the FPN module.

[0022] The top-level FPN module is connected to the intermediate 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 sub-module and FPN module of the previous layer to create the upper-level feature map.

[0023] 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 by a skip connection to obtain the output of the attention module.

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

[0025]

[0026] In the formula, w k Here, K represents the weight of category k, reflecting the importance or priority of that category, where K is the total number of categories. k The Dice coefficient is expressed as follows:

[0027]

[0028] In the formula, Let represent the predicted value of the i-th pixel in category k. Let represent the true value of the i-th pixel in category k, and N represent the total number of pixels in the image;

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

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

[0031] The training batch epoch is set to 80, and the learning rate is set to 5e. -5 The training batch size was set to 8. During training, EarlyStopping, ModelCheckpoint, and ReduceCall mechanisms were set. The EarlyStopping mechanism was used to terminate training early when the validation loss value did not decrease significantly after 50 consecutive iterations. The ModelCheckpoint mechanism was used to monitor the validation loss and save only the optimal model weights. The ReduceLROnPlateau mechanism dynamically reduced the learning rate during the validation loss period to ensure that the model could converge more finely.

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

[0033] (1) This invention provides a mesoscale eddy identification method based on deep learning. It innovatively proposes the PAM-ResNet model to construct a mesoscale eddy identification model. In order to verify the effectiveness and superiority of the PAM-ResNet model in the mesoscale eddy 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 eddies, with stronger feature extraction and fusion capabilities and generalization ability. Therefore, it has feasibility and significant superiority in the specific task of mesoscale eddy identification.

[0034] (2) The present invention takes into account accuracy, recall and Dice coefficient. The PAM-ResNet model has shown good comprehensive performance in the mesoscale vortex identification task. It can not only correctly identify and classify vortices, but also accurately restore the contour and boundary information of vortices. Attached Figure Description

[0035] Figure 1 This is a flowchart of a mesoscale eddy recognition method based on deep learning according to the present invention.

[0036] Figure 2 This is a diagram of the PAM-ResNet architecture.

[0037] Figure 3 This is a structural diagram of the residual unit.

[0038] Figure 4 This is a diagram of the FPN module structure.

[0039] Figure 5 This is a structural diagram of the attention module.

[0040] Figure 6 This is a diagram of the ResNet architecture.

[0041] Figure 7 This is a diagram of the ResNet-FPN network model structure. Detailed Implementation

[0042] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the precision and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0043] like Figure 1 As shown, in one embodiment of the present invention, a mesoscale eddy recognition method based on deep learning includes the following steps:

[0044] S1. Acquire sea surface anomaly data, preprocess the sea surface anomaly data, and 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, and obtain the mesoscale eddy recognition model.

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

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

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

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

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

[0052] (1) Initial value settings: Set the latitude range of the South China Sea region to 12.0625 to 22.9375 and the longitude range to 111.0625 to 120.9375, with each pixel size being 0.125 degrees. Calculate the area of ​​each pixel by setting the Earth's radius, and set the initial date to January 1, 2001, to prepare for subsequent date processing.

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

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

[0055] (4) Vortex confirmation: The closed profile needs to meet the following criteria: the profile contains only one extreme point, the number of pixels (I) contained is ≥9, and the vortex amplitude is ≥1cm. When the above criteria are met, it is identified as a mesoscale vortex.

[0056] (5) Data classification: Based on the identified vortex type, the data is divided into three categories, including non-cyclonic, cyclonic vortex and anticyclonic vortex.

[0057] S12. Expand the sea level anomaly data and corresponding label data using data augmentation techniques to construct a training set.

[0058] In S11, the tag data includes non-cyclones, cyclonic vortices, and anticyclonic vortices.

[0059] like Figure 2 As shown, in step S2, the method for inputting the training set into the PAM-ResNet model to obtain the prediction results is as follows:

[0060] S21. Perform encoder processing and max pooling downsampling processing on the input data three times in a row to obtain the first feature map. 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 the intermediate feature map.

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

[0062] S23. Perform three consecutive upsampling operations and decoder processing on the second feature map to obtain the third feature map, and in each decoder processing process, the intermediate feature map is spliced ​​onto the generated feature map through skip connections.

[0063] S24. Obtain the prediction result based on the output feature map.

[0064] like Figure 3 As shown, in S21, residual units are provided 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. The output of the Dropout layer and the input of the 3*3 convolutional layer are added through skip connections and used as the output of the residual unit.

[0065] like Figure 4 As shown, in S21, the FPN module structures of each layer are identical, all including:

[0066] 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 feature map elements of the upper layer as the output of the FPN module.

[0067] The top-level FPN module is connected to the intermediate 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 sub-module and FPN module of the previous layer to create the upper-level feature map.

[0068] like Figure 5 As 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 by a skip connection to obtain the output of the attention module.

[0069] like Figure 2As shown on the right, in S2, the PAM-ResNet model includes a ResNet network with an encoder-decoder architecture and an FPN module. Each layer of the pyramid in the FPN module has an attention submodule.

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

[0071] (1) Constructing the ResNet network model

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

[0073] In the encoder section, a three-layer improved residual unit is employed, with a max-pooling layer added after each residual unit for downsampling. This residual unit is designed to alleviate the vanishing gradient problem during deep network training and accelerate convergence. The residual unit structure is as follows: it contains two convolutional layers, and skip connections directly add the input and output to ensure lossless information transfer. The application of batch normalization and ReLU activation functions guarantees stable gradient flow and non-linear expressive power. To enhance the model's generalization performance, a dropout mechanism is introduced to prevent overfitting and improve robustness. Finally, the input features are reconstructed to the same shape as the output features before residual summation.

[0074] Similar to the encoder, the central block uses a layer of residual units to further extract global features, serving as a bridge connecting the encoder and decoder.

[0075] The decoder uses UpSampling2D for upsampling. The upsampled result is then concatenated with the feature map of the corresponding layer in the encoder to fuse detailed information. Finally, an improved residual unit is used to process the fused features.

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

[0077] The ResNet-FPN network model is an image segmentation architecture that adds a Feature Pyramid Network (FPN) to ResNet. Its network structure is as follows: Figure 7 As shown, this network significantly enhances its ability to detect objects at multiple scales by constructing a feature pyramid and fusing features from different levels. This fusion strategy not only allows the network to focus on both global and local features simultaneously, but also particularly improves the detection accuracy for small objects.

[0078] This invention adds an FPN module after the ResNet center block and before the decoder, with the following structure: Figure 4 As shown, FPN starts from the output of the ResNet central layer and creates the top-level feature map of FPN through a 1×1 convolutional layer. Then, it fuses features from different levels by creating lateral connections. Specifically, these connections first use 1x1 convolutional layers to unify the dimensions of higher-level features, then upsample them to the same spatial size as the lower-level features, and finally sum them 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 effective fusion of features from different levels. This preserves more spatial details and introduces higher-level semantic information, effectively solving the problem of poor detection of small objects often encountered in single-scale feature extraction, while maintaining the accuracy of large objects.

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

[0080] While ResNet-FPN is effective in multi-scale feature representation, it cannot aggregate more discriminative features for segmentation. To further improve the model's segmentation capability, this invention proposes a PAM module and fuses it with ResNet to obtain the PAM-ResNet model. Its network structure is as follows: Figure 2 As shown, the attention submodule structure is as follows: Figure 5 As shown.

[0081] In the attention mechanism module, the number of channels of the upper-layer features and the current-layer features are first adjusted by the FPN modules of each layer. Then, the adjusted top-layer features and the current-layer features are added together and then processed by the ReLU activation function to obtain the fused features. Next, the fused features are processed by 1×1 convolution to generate single-channel attention weights, and then processed by the Sigmoid activation function to map them to the range [0,1]. Finally, the generated attention weights are applied to the features of the current layer.

[0082] In this embodiment, the core objective of the attention submodule is to enhance the model's focus on key information while reducing its focus on irrelevant information. By adding an attention module to each layer of the FPN, the two features are adaptively weighted before fusing the top-level features (top_layer) and the current-level features (lateral_layer) to improve the model's ability to capture important features, thereby allowing feature maps at each scale to be better fused together.

[0083] In S2, the expression for the loss function Weighted Dice is specifically as follows:

[0084]

[0085] In the formula, w k Here, K represents the weight of category k, reflecting the importance or priority of that category, where K is the total number of categories. k The Dice coefficient is expressed as follows:

[0086]

[0087] In the formula, Let represent the predicted value of the i-th pixel in category k. Let represent the true value of the i-th pixel in category k, and N represent the total number of pixels in the image;

[0088] In this embodiment, the present invention uses a weighted Dice coefficient as the 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, some classes of targets appear very infrequently in the image or occupy a small proportion of pixels. In this case, it is necessary to assign higher weights to these smaller classes to prevent the model from ignoring them.

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

[0090] Table 1. Number and Weight of Pixels in Each Category

[0091]

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

[0093] The optimizer is used to adjust the model's parameters during training to minimize the loss function. This example uses the Adam optimizer, which offers good performance and adaptability. Starting with a larger learning rate allows for faster convergence; decreasing the learning rate in later stages allows for fine-tuning of model parameters and avoids oscillations.

[0094] To comprehensively evaluate the model's performance, this embodiment uses classification accuracy to measure the proportion of times the model's predicted class for each pixel perfectly matches the true class, thereby evaluating the model's overall classification accuracy. The average Dice coefficient for all classes is calculated to evaluate the model's overall segmentation performance. The weighted average Dice coefficient is then calculated based on the importance of each class to further evaluate the model's performance in handling class imbalance problems.

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

[0096] The training batch epoch is set to 80, and the learning rate is set to 5e. -5 The training batch size was set to 8. During training, EarlyStopping, ModelCheckpoint, and ReduceCall mechanisms were set. The EarlyStopping mechanism was used to terminate training early when the validation loss value did not decrease significantly after 50 consecutive iterations. The ModelCheckpoint mechanism was used to monitor the validation loss and save only the optimal model weights. The ReduceLROnPlateau mechanism dynamically reduced the learning rate during the validation loss period to ensure that the model could converge more finely.

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

[0098] This embodiment compares and analyzes the classification accuracy, recall, average Dice coefficient, weighted Dice coefficient, and average accuracy of different types of cyclones using the ResNet model, ResNet-FPN model, and PAM-ResNet model. SLA and label data from the South China Sea region in 2023 were selected as the data. Classification accuracy, recall, Dice coefficient, average Dice coefficient, and weighted Dice coefficient were used as the model evaluation metrics, as shown in Tables 2 to 4.

[0099] Table 2 Statistics of ResNet segmentation results

[0100]

[0101] Table 3. Statistics of ResNet-FPN segmentation results

[0102]

[0103] Table 4. Statistics of PAM-ResNet segmentation results

[0104]

[0105] In vortex classification, PAM-ResNet achieved 81% and 76% accuracy for cyclones and anticyclones, respectively, representing improvements of 6% and 2% compared to the original ResNet. Furthermore, PAM-ResNet also showed slight improvements over ResNet-FPN, at 6% and 1% accuracy, respectively. These comparisons demonstrate that PAM-ResNet, proposed in this invention, successfully improves the ResNet model in two steps, significantly enhancing its vortex classification capabilities.

[0106] Regarding vortex recall, the nearly identical recall rates of the ResNet, ResNet-FPN, and PAM-ResNet models indicate that, during the improvement process, ResNet-FPN and PAM-ResNet both inherited the low false negative rate advantage of the ResNet model for vortex identification.

[0107] Regarding the similarity between the predicted and actual vortex regions, PAM-ResNet achieved Dice coefficients of 82%, 82%, and 97% for cyclones, anticyclones, and non-cyclones, respectively. These figures represent significant improvements over both the unmodified ResNet and ResNet-FPN models. This comparative analysis demonstrates that, compared to ResNet and ResNet-FPN, the PAM-ResNet model more accurately identifies the boundaries and regions of vortices in terms of similarity between the predicted and actual vortex regions.

[0108] Based on the above analysis, the following conclusions can be drawn:

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

[0112] In addition, this invention compares and analyzes the recognition results of the commonly used U-Net model for mesoscale eddy identification, the advanced DeeplabV3plus model for mesoscale eddy identification, and the PAM-ResNet model innovatively used in this invention. Data selected is the SLA and label data of the South China Sea region in 2023. Classification accuracy, recall, Dice coefficient, average Dice coefficient, and weighted Dice coefficient are used as evaluation indicators for the model, as shown in Tables 5 to 7.

[0113] Table 5. Statistics of U-Net segmentation results

[0114]

[0115] Table 6. Statistics of DeeplabV3plus segmentation results

[0116]

[0117] Table 7. Statistics of PAM-ResNet segmentation results

[0118]

[0119] In vortex classification, the average accuracy of the PAM-ResNet model (94.5%) is higher than that of the U-Net model (93%) and the DeepLabV3plus model (93.5%). For cyclones and anticyclones, the accuracy differences between the U-Net, DeepLabV3plus, and PAM-ResNet models are less than 4%. For non-cyclone categories, the PAM-ResNet model achieves an accuracy of 98%, significantly higher than the U-Net and DeepLabV3plus models (95%). Compared to the U-Net model, the PAM-ResNet model improves cyclone accuracy by 4% and decreases anticyclone accuracy by 5%. Compared to DeepLabV3plus, the PAM-ResNet model improves cyclone accuracy by 1% and decreases anticyclone accuracy by 3%. Although the average accuracy of DeepLabV3plus is 1% lower than that of PAM-ResNet, PAM-ResNet's accuracy for cyclones and anticyclones is significantly higher. DeepLabV3plus achieved almost identical accuracy in identifying cyclones and anticyclones. This indicates that DeepLabV3plus demonstrates more stable classification capabilities for cyclones and anticyclones compared to U-Net and PAM-ResNet.

[0120] In terms of vortex recall, the PAM-ResNet model significantly outperformed the U-Net and DeeplabV3plus models in both cyclone and anticyclone categories. As shown in Table 7, the PAM-ResNet model achieved a recall of 86% for both cyclones and anticyclones, representing improvements of 9% and 16% respectively compared to the U-Net model. Furthermore, it improved recall by 10% and 8% respectively compared to the DeeplabV3plus model. These comparisons demonstrate that the PAM-ResNet model exhibits a significant advantage in reducing false negatives among existing vortex identification models.

[0121] Regarding the similarity between the predicted and actual vortex regions, PAM-ResNet achieved DICE coefficients of 82%, 82%, and 97% for cyclones, anticyclones, and non-cyclones, respectively. These figures represent a slight improvement over the commonly used mesoscale vortex identification models U-Net (75%, 75%, 96%) and DeeplabV3plus (76%, 78%, 96%). The above comparative analysis indicates that, compared to U-Net and DeeplabV3plus, PAM-ResNet more accurately identifies the boundaries and regions of vortices in terms of similarity between the predicted and actual vortex regions.

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

[0123] (1) Model stability: The DeepLabV3plus model showed almost the same accuracy in identifying cyclones and anticyclones, demonstrating good stability. This indicates that the DeepLabV3plus model can maintain relatively consistent performance when handling cyclone and anticyclone categories.

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

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

[0126] (4) Overall performance: Considering accuracy, recall and Dice coefficient, the PAM-ResNet model showed good overall performance in the mesoscale eddy recognition task.

[0127] The beneficial effects of this invention are as follows: This invention provides a mesoscale eddy identification method based on deep learning. It innovatively proposes the PAM-ResNet model to construct a mesoscale eddy identification model. In order to verify the effectiveness and superiority of the PAM-ResNet model in the mesoscale eddy 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 eddies, with stronger feature extraction and fusion capabilities and generalization ability. Therefore, it has feasibility and significant superiority in the specific task of mesoscale eddy identification.

[0128] This invention comprehensively considers accuracy, recall, and Dice coefficient. The PAM-ResNet model demonstrates good overall performance in mesoscale vortex identification tasks, not only correctly identifying and classifying vortices but also accurately reconstructing the contour and boundary information of vortices.

[0129] In the description of this invention, it should be understood that the terms "center," "thickness," "upper," "lower," "horizontal," "top," "bottom," "inner," "outer," and "radial," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying the relative importance or the number of technical features implicitly specified. Therefore, a feature defined by "first," "second," and "third" may explicitly or implicitly include one or more of that feature.

Claims

1. A mesoscale eddy identification method based on deep learning, characterized in that, Includes the following steps: S1. Acquire sea surface anomaly data, preprocess 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; In S2, the method for inputting the training set into the PAM-ResNet model to obtain the prediction results is as follows: S21. Perform encoder processing and max pooling downsampling processing on the input data three times in a row to obtain the first feature map. 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 the intermediate feature map. In S21, residual units are provided 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. The output of the Dropout layer and the input of the 3*3 convolutional layer are added through a skip connection and used as the output of the residual unit. In S21, the FPN module structure is the same for each layer, 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 feature map elements of the upper layer as the output of the FPN module. The top-level FPN module is connected to the intermediate 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 sub-module 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 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 by a skip connection to obtain the output of the attention module. S22. Input the first feature map into the intermediate layer to obtain the second feature map; S23. Perform three consecutive upsampling operations and decoder processing on the second feature map to obtain the third feature map, and in each decoder processing process, the intermediate feature map is spliced ​​onto the generated feature map through skip connections. S24. Obtain the prediction result based on the output feature map. S3. Set the hyperparameters of the PAM-ResNet model, train the PAM-ResNet model, and obtain the mesoscale eddy recognition model. S4. Analyze sea surface anomaly data using a mesoscale eddy identification model to complete the identification of ocean mesoscale eddies.

2. The deep learning-based mesoscale eddy recognition method according to claim 1, characterized in that, In step S1, the method for preprocessing sea surface anomaly data is as follows: S11. Use an automatic identification method based on closed contour lines to generate label data from sea level anomaly data; S12. Expand the sea level anomaly data and corresponding label data using data augmentation techniques to construct a training set.

3. The deep learning-based mesoscale eddy recognition method according to claim 2, characterized in that, In S11, the tag data includes non-cyclones, cyclonic vortices, and anticyclonic vortices.

4. The deep learning-based mesoscale eddy recognition method according to claim 1, characterized in that, In S2, the expression for the loss function Weighted Dice is specifically as follows: In the formula, w k For category k The weights are used to reflect the importance or priority of the category. K The total number of categories, Dice k The Dice coefficient is expressed as follows: In the formula, p i k Indicate category k The i The predicted value of each pixel. g i k Indicate category k The i The actual value of each pixel. N This represents the total number of pixels in the image; The optimizer is specifically the Adam optimizer, which is used to adjust the model's parameters during training to minimize the loss function.

5. The deep learning-based mesoscale eddy recognition method according to claim 1, characterized in that, In S3, the method for setting the hyperparameters of the PAM-ResNet model is as follows: The training batch epoch is set to 80, and the learning rate is set to 5e. -5 The training batch size was set to 8. During training, EarlyStopping, ModelCheckpoint, and ReduceCall mechanisms were set. The EarlyStopping mechanism was used to terminate training early when the validation loss value did not decrease significantly after 50 consecutive iterations. The ModelCheckpoint mechanism was used to monitor the validation loss and save only the optimal model weights. The ReduceLROnPlateau mechanism dynamically reduced the learning rate during the validation loss period to ensure that the model could converge more finely.

Images