A method for constructing a composite vortex identification model based on Helmholtz decomposition and deep learning

Abstract

The present invention discloses a method for constructing a composite vortex identification model based on Helmholtz decomposition and deep learning, which comprises the following steps: obtaining the east-west flow velocity U and the north-south flow velocity V data of seawater to construct a velocity vector field image; performing Helmholtz decomposition on the velocity vector field image to obtain a potential field F d Image and rotation field F c Image; According to the rotation field F c The potential function and UV data in the image are used to mark single vortex cyclonic vortex, single vortex anticyclonic vortex, compound vortex cyclonic vortex and compound vortex anticyclonic vortex, and the vortex field images before and after the marking are used as deep learning data sets; the deep learning data sets are trained using a neural network to obtain a compound vortex recognition model. The present invention uses the Helmholtz decomposition method to process the derived data of the satellite altimeter to improve the recognition accuracy of the compound vortex, and uses a deep learning method to improve the recognition efficiency of the compound vortex, and finally realizes the application of the model in the entire sea area.

Description

Technical Field

[0001] The present invention belongs to the field of compound vortex identification, and in particular relates to a method for constructing a compound vortex identification model based on Helmholtz decomposition and deep learning. Background Art

[0002] Mesoscale vortices are a type of eddy in the ocean, widespread across the globe. Ocean vortices carry immense kinetic energy, which plays a key role in global ocean material transport and energy transfer. Previously, we often identified single vortices, but paid less attention to compound vortices. Compound vortices are a combination of single vortices. When two single vortices collide to form a compound vortex, the motion of the original single vortices is affected. Furthermore, the energy exchange between two independent single vortices differs from that between the two single vortices in a compound vortex. Therefore, identifying compound vortices has significant scientific significance and application value. Current methods for identifying compound vortices primarily rely on physical methods, such as the improved HD method. This method is based on raw satellite altimeter data. However, using this unprocessed data to identify compound vortices results in low efficiency and accuracy. Summary of the Invention

[0003] An object of the present invention is to provide a method for constructing a composite vortex identification model based on Helmholtz decomposition and deep learning and to provide at least the advantages to be described later.

[0004] Another object of the present invention is to provide a composite vortex identification method based on Helmholtz decomposition and deep learning. The Helmholtz decomposition method is used to process the derivative data of the satellite altimeter to improve the recognition accuracy, and the deep learning method is used to improve the recognition efficiency of composite vortices, ultimately realizing the application of the model in the entire sea area.

[0005] The technical solutions of the present invention are as follows:

[0006] A composite vortex identification model construction method based on Helmholtz decomposition and deep learning includes the following steps:

[0007] 1) Obtain seawater velocity data (U) in the east-west direction and V in the north-south direction, and construct a velocity vector field image;

[0008] 2) Perform Helmholtz decomposition on the velocity vector field image to obtain the potential field F d Image and rotation field F c image;

[0009] 3) According to the rotation field F c Potential function and Uv data annotation deep learning dataset in;

[0010] 4) Using a neural network to train the deep learning data set to obtain a composite vortex identification model.

[0011] Preferably, in step 2), the composite vortex is identified by the swirling field to improve the recognition accuracy of the composite vortex, F d , F c The definition of is shown in formula 1 and 2;

[0012]

[0013]

[0014] Among them, V is a scalar function, and the gradient of V can be used to represent a potential field. A is a vector field, and the curl of A can be used to represent a rotation field.

[0015] Preferably, the velocity vector field is calculated from sea level anomaly satellite altimeter data, wherein the spatial resolution is 1 / 12°×1 / 12°, the temporal resolution is 3 hours, and the spatial coverage range is from 80 degrees south latitude to 90 degrees north latitude.

[0016] Preferably, in step 3), the deep learning data set is four types of vortices, namely, single vortex cyclonic vortex, single vortex anticyclonic vortex, compound vortex cyclonic vortex and compound vortex anticyclonic vortex.

[0017] Preferably, the neural network is an EddyNet neural network.

[0018] Preferably, in step 4), the deep learning dataset is trained using a neural network to obtain a compound vortex identification model, comprising the following steps:

[0019] 1) Dividing the deep learning dataset into a training set and a validation set;

[0020] 2) Establish the initial model of EddyNet neural network;

[0021] 3) setting initial training parameters, and using the training set to train the EddyNet neural network initial model to obtain a trained EddyNet neural network model and optimized training parameters;

[0022] 4) Using the verification set and the optimized training parameters to verify and identify the trained EddyNet neural network model to obtain a final compound vortex identification model.

[0023] Preferably, in the method for constructing a composite vortex identification model based on Helmholtz decomposition and deep learning, the initial training parameters include a learning rate of 1e-3, 50 training rounds, and each training is performed on the training set in batches, with 16 samples in each batch.

[0024] Preferably, the semantic segmentation refers to classifying each pixel in the image.

[0025] The present invention has the following beneficial effects:

[0026] 1) The Helmholtz decomposition method is used to extract the vortex field from the original velocity field, which reduces the occurrence of non-vortex parts in the original velocity field being mistaken for vortices, and improves the accuracy of identifying compound vortices and single vortices;

[0027] 2) The addition of EddyNet addresses the inefficiency of existing physical methods for vortex identification, particularly improving the accuracy of identifying compound vortices, which are less studied. This neural network approach can be applied across the entire ocean, overcoming shortcomings of physical methods, such as the artificial threshold settings required to identify compound vortices in different ocean regions.

[0028] 3) The velocity field data constructed by UV is used to identify single vortices and compound vortices, so that we can see not only the location of the vortex but also the outline of the vortex, which plays an important role in studying the energy exchange between vortices and the motion trajectory of the vortex;

[0029] 4) Other advantages, objectives and features of the present invention will be reflected in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A flow chart of the method for constructing a composite vortex identification model based on Helmholtz decomposition and deep learning provided by the present invention;

[0031] Figure 2 A cloud map of seawater flow velocity in the east-west direction obtained in one embodiment of the present invention;

[0032] Figure 3 A cloud map of seawater flow velocity in the north-south direction obtained in one embodiment of the present invention;

[0033] Figure 4 is a velocity vector field image in one embodiment of the present invention;

[0034] Figure 5 for Figure 4 The potential field F obtained after Helmholtz decomposition is d image;

[0035] Figure 6 for Figure 4 The rotation field F obtained after Helmholtz decomposition c image;

[0036] Figure 7 is the rotation field image before annotation;

[0037] Figure 8 for Figure 7 Grayscale image of the vortex after annotation;

[0038] Figure 9 Schematic diagram of the EddyNet neural network model structure in the present invention;

[0039] Figure 10 The figure shows how the accuracy of the composite vortex identification model constructed by the present invention on the training set and the validation set changes with the increase of epochs. DETAILED DESCRIPTION

[0040] The present invention will be described in further detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0041] It should be understood that terms such as “having,” “including,” and “comprising” used herein do not prescribe the existence or addition of one or more other elements or combinations thereof.

[0042] A mesoscale vortex is an eddy in the ocean with a radius of 10 to 100 km. Its life cycle, that is, the time from its formation to its decay, can last for tens of days or even years. Mesoscale vortices are divided into single vortices and compound vortices based on their morphology. A single vortex is a vortex in the ocean formed by closed streamlines with one vortex center, while a compound vortex is a vortex formed by closed streamlines with two to three single vortices. Based on the direction of rotation of the mesoscale vortex, it can be further divided into cyclonic vortices (rotating clockwise in the Southern Hemisphere) and anticyclonic vortices (rotating counterclockwise in the Southern Hemisphere). Due to the characteristics of mesoscale vortices, such as their high-speed rotation in the ocean, they play a very important role in ocean circulation, ocean temperature and salinity structure, and global climate change.

[0043] Satellite altimeters play a crucial role in ocean dynamics due to their high observation accuracy, large coverage area, and vast information collection capabilities. They can measure ocean dynamic parameters (such as sea surface height and sea level anomalies) and obtain information on ocean dynamic parameters (such as waves and tides). Data from satellite altimeters can be used to identify mesoscale eddies and other related research areas, and are also helpful in studying other aspects of ocean dynamics.

[0044] Currently, methods for identifying compound vortices are divided into physical methods and artificial intelligence methods. The physical method mainly includes the improved HD method. In addition to the original HD method, which identifies the vortex center and vortex boundary and then determines whether it is a compound vortex based on the number of vortex centers within the vortex boundary, this method adds a new judgment condition for compound vortices. That is, if the ratio of the sum of the radius lengths of two single vortices to the distance between the vortex centers of the two single vortices is greater than or equal to 1, it is a compound vortex. The addition of the judgment condition makes the identification of compound vortices more accurate. However, this method only considers the physical perspective and cannot be applied in a wide range of sea areas. In addition, the recognition efficiency is low and the overall accuracy is not high. The artificial intelligence method mainly performs target detection on compound vortices and can only identify the location of the compound vortex, with a low recognition accuracy.

[0045] Currently, most of the identification of compound vortices is based on data directly measured by satellite altimeters. Few researchers have considered processing satellite altimeter data. For example, in "Characteristic Ocean Flow Visualization using Helmholtz Decomposition", in order to improve the accuracy of single vortex identification, Cuicui Zhang processed satellite altimeter data using the Helmholtz decomposition method, and then used the HD method to identify single vortices based on the processed data.

[0046] Currently, compound vortex identification is mostly based on satellite altimeter data, such as sea level anomalies and sea surface height data. However, if compound vortices are identified directly based on this data through physical methods such as the HD method, non-vortices may be identified as vortices, affecting the accuracy of vortex identification. Furthermore, vortex identification based on traditional physical methods is generally inefficient, and because different thresholds are required to identify vortices in different sea areas, such identification methods cannot be widely applied to the entire ocean. In the ocean, compound vortices, which are constantly in motion, take into account the interconnectedness between vortices more than single vortices. This means that each vortex that makes up the compound vortex may undergo deformation, separation, and merging of new vortices during its motion. Furthermore, when two single vortices combine to form a compound vortex, the original motion trajectory of the single vortex is altered. This phenomenon requires careful consideration when studying vortex trajectory prediction.

[0047] like Figure 1 As shown, the present invention provides a method for constructing a composite vortex identification model based on Helmholtz decomposition and deep learning, which includes the following steps:

[0048] 1) Obtain seawater velocity data (U) in the east-west direction and V in the north-south direction, and construct a velocity vector field image;

[0049] 2) Perform Helmholtz decomposition on the velocity vector field image to obtain the potential field F d Image and rotation field F c image;

[0050] 3) According to the rotation field F c The potential function and UV data in the image are used to mark single vortex cyclonic vortex, single vortex anticyclonic vortex, compound vortex cyclonic vortex and compound vortex anticyclonic vortex, and the cyclonic field images before and after marking are used as deep learning datasets;

[0051] 4) Using a neural network to train the deep learning data set to obtain a composite vortex identification model.

[0052] In one embodiment of the method for constructing a composite vortex identification model based on Helmholtz decomposition and deep learning provided by the present invention, the velocity vector field is calculated from sea level anomaly satellite altimeter data, wherein the spatial resolution is 1 / 12°×1 / 12°, the temporal resolution is 3h, and the spatial coverage range is from 80 degrees south latitude to 90 degrees north latitude.

[0053] In one embodiment of the method for constructing a composite vortex identification model based on Helmholtz decomposition and deep learning provided by the present invention, the neural network is an EddyNet neural network.

[0054] In one embodiment of the method for constructing a compound vortex identification model based on Helmholtz decomposition and deep learning provided by the present invention, the deep learning dataset is trained using a neural network to obtain a compound vortex identification model, comprising the following steps:

[0055] 1) Dividing the deep learning dataset into a training set and a validation set;

[0056] 2) Establish the initial model of EddyNet neural network;

[0057] 3) setting initial training parameters, and using the training set to train the EddyNet neural network initial model to obtain a trained EddyNet neural network model and optimized training parameters;

[0058] 4) Using the verification set and the optimized training parameters to verify and identify the trained EddyNet neural network model to obtain a final compound vortex identification model.

[0059] In one embodiment of the method for constructing a composite vortex identification model based on Helmholtz decomposition and deep learning provided by the present invention, the initial training parameters include a learning rate of 1e-3, 50 training rounds, and each training is divided into batches for training, with the number of samples in each batch being 16.

[0060] Specifically, the method for constructing a composite vortex identification model based on Helmholtz decomposition and deep learning provided by the present invention includes the following steps:

[0061] (1) Download satellite altimeter data:

[0062] The identification of composite eddies was conducted based on the sea level anomaly (SLA) and corresponding reanalysis data from 2013 to 2015. The data comes from HYCOM, with a spatial resolution of 1 / 12°×1 / 12° and a temporal resolution of 3 hours. The spatial coverage range is from 80°S to 90°N. We performed the corresponding composite eddy identification based on the two sets of reanalysis data provided: the east-west velocity U and the north-south velocity V of the seawater. Figure 2 ,The U,V cloud map shown in 3, the pure gray area at the boundary is ,the land, and the middle part is the ocean. The darker the color, the greater the seawater flow rate.

[0063] (2) Perform Helmholtz decomposition on the velocity vector field composed of U and V:

[0064] In order to improve the accuracy of compound vortex identification and reduce the occurrence of ocean phenomena such as tides being mistaken for vortex events, the Helmholtz decomposition method is added to the compound vortex identification task. Since any vector field has non-zero divergence and curl, and our identification of vortices mainly relies on curl, we can use the Helmholtz decomposition method to decompose the vector field F into a pure irrotational field F d (with potential field), pure no scattered field F c (with vortex field), and then identify the compound vortex only through the vortex field. This approach can improve the accuracy of compound vortex identification. d , F c The definition of is shown in formula 1 and 2.

[0065]

[0066]

[0067] Where V is a scalar function, and the gradient of V can be used to represent a potential field. A is a vector field, and the curl of A can be used to represent a rotation field. The decomposition example obtained based on a set of UV data from 29 to 36 degrees south latitude and 89 to 96 degrees west longitude is as follows: Figure 4 、 5 As shown in 6, Figure 4 、 5 and 6 represent the original velocity field, vortex field, and potential field respectively. It can be seen that there are a large number of vortices in the vortex field.

[0068] (3) Annotated dataset:

[0069] After the original velocity field is decomposed by Helmholtz, only the results of the vortex field are taken, including the potential function corresponding to the vortex field and the UV data of the vortex field. Based on these results, we use them to annotate the data set required for deep learning. A total of 4 types of vortices are annotated, namely: single vortex (cyclonic vortex, anticyclonic vortex), compound vortex (cyclonic vortex, anticyclonic vortex). Figure 7 and Figure 8 A set of labeled data instances shown, Figure 7 To mark the original image, Figure 8 is the truth map required by the annotated neural network.

[0070] (4) Establishing EddyNet neural network:

[0071] Based on the labeled data obtained above, the improved U-Net is used to perform semantic segmentation on the composite vortex. The so-called semantic segmentation refers to the classification of each pixel in the image. In this experiment, the work content of the model that realizes semantic segmentation is to identify the position and outline of various vortices in the image. The input of the model is the vortex field image after Helmholtz decomposition, and the output is the grayscale image of the various vortices identified. The specific structure of the neural network model is as follows Figure 9 As shown in the figure, in the encoder stage, the input image is downsampled three times. Each downsampling operation includes two 3x3 convolutions, two Relu + regularization operations, one dropout operation, and one 2x2 max pooling operation. In the decoder stage, the upsampled feature map is combined with the downsampled feature map of the corresponding layer on the left. Finally, a 1x1 convolution is performed to obtain the final semantic segmentation result.

[0072] (5) Set training parameters:

[0073] The learning rate of the EddyNet network model is set to 1e-3, the number of training epochs is set to 50, and all samples are trained in batches each time, with the batch_size of 16 samples in each batch.

[0074] (6) Training EddyNet neural network:

[0075] The EddyNet neural network was trained using labeled input and output data, and the parameters were adjusted. This adjustment resulted in a better semantic segmentation performance on the training set and improved eddy recognition accuracy on the validation set. This network model uses cross entropy (Dice Loss) as the loss function, defined as Equation (3).

[0076]

[0077] After parameter adjustment and other work, the accuracy of the final version of the model on the training set and validation set changes with the increase of epochs as follows Figure 10 As shown in the figure, the accuracy on the training set ultimately reached approximately 85%, indicating that the model was sufficiently trained on the training set, and the accuracy results on the validation set are credible. The final model achieved an accuracy of approximately 80% for identifying single and compound vortices on the validation set.

[0078] The present invention has the following characteristics:

[0079] (1) Apply the Helmholtz decomposition method to the identification of compound vortices to improve the accuracy of compound vortex identification;

[0080] (2) The deep learning method EddyNet is used in the identification of single vortices and compound vortices, which improves the efficiency of identifying compound vortices and realizes the identification of compound vortices in the entire sea area.

[0081] The present invention provides a method for constructing a compound vortex identification model based on Helmholtz decomposition and deep learning. First, the velocity field data obtained through sea level anomalies is Helmholtz decomposition to extract the vortex part and improve the accuracy of vortex identification; second, the neural network data set is labeled based on the vortex field part obtained after Helmholtz decomposition to prepare for the supervised learning process of compound vortices; finally, the labeled data set is trained through the EddyNet neural network to realize the identification of compound vortices. This approach of identifying compound vortices based on Helmholtz decomposition and deep learning can improve the recognition accuracy and efficiency of compound vortices, and the method can be applied in the entire sea area.

[0082] In the present invention,

[0083] 1. The UV velocity field can be calculated using satellite altimeter data such as sea level anomaly (SLA). The present invention directly uses the UV velocity field provided by HYCOM.

[0084] 2. The result of the rotation field part obtained after Helmholtz decomposition is not unique.

[0085] 3. The EddyNet neural network can be any other neural network.

[0086] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to specific details.

Claims

1. A method for constructing a composite vortex identification model based on Helmholtz decomposition and deep learning, characterized in that: The following steps are involved: 1) Obtain the data of seawater velocity U in the east-west direction and velocity V in the north-south direction, and construct a velocity vector field image; 2) Perform Helmholtz decomposition on the velocity vector field image to obtain the potential field F d Image and rotation field F c image; 3) According to the rotation field F c Potential function and UV data annotation deep learning dataset in; 4) Using a neural network to train the deep learning data set to obtain a compound vortex identification model; In step 2), the compound vortex is identified by using a swirling field to improve the accuracy of identifying the compound vortex. , The definition of is shown in formula (1), (2); Among them, V is a scalar function, the gradient of V is used to represent a potential field, A is a vector field, and the curl of A is used to represent a rotation field; In step 4), the deep learning data set is trained using a neural network to obtain a compound vortex identification model, comprising the following steps: 4.1) Dividing the deep learning dataset into a training set and a validation set; 4.2) Establish the initial model of EddyNet neural network; 4.3) Setting initial training parameters, and using the training set to train the EddyNet neural network initial model to obtain a trained EddyNet neural network model and optimized training parameters; 4.4) Using the verification set and the optimized training parameters to verify and identify the trained EddyNet neural network model, and obtain a final compound vortex identification model.

2. The method for constructing a composite vortex identification model based on Helmholtz decomposition and deep learning according to claim 1, characterized in that: In the step 3), the deep learning data set is 4 types of vortices, single vortex cyclonic vortex, single vortex anticyclonic vortex, compound vortex cyclonic vortex and compound vortex anticyclonic vortex.

3. The method for constructing a composite vortex identification model based on Helmholtz decomposition and deep learning according to claim 1, characterized in that: The velocity vector field is calculated from sea level anomaly satellite altimeter data, wherein the spatial resolution is 1 / 12°× 1 / 12°, the temporal resolution is 3h, and the spatial coverage range is from 80° south latitude to 90° north latitude.

4. The method for constructing a composite vortex identification model based on Helmholtz decomposition and deep learning according to claim 1, characterized in that: The initial training parameters include a learning rate of 1e-3, a number of training rounds of 50, and each training is divided into batches for training the training set, with 16 samples in each batch.

Images