Internal solitary wave identification and amplitude inversion method in pattern data based on machine learning

Abstract

The present invention relates to a method for identifying and inverting internal solitary waves in pattern data based on machine learning, and belongs to the field of ocean observation technology. The present invention combines Mask-RCNN and variational autoencoders, first preprocesses and images the sea surface height data, extracts texture features through a local binary pattern algorithm, constructs a labeled data set and performs model training, and identifies internal solitary wave stripes and their positions therefrom; then, through the extraction and processing of temperature field data, uses a variational autoencoder to invert the temperature field at depth; and extracts the amplitude of the internal solitary wave in the temperature field based on the internal wave identification result. The present invention can efficiently and accurately process large-scale, refined ocean pattern data, significantly improving the efficiency and accuracy of data processing. The present invention has broad application prospects, especially in marine disaster warning, marine scientific research, and internal wave dynamics research, and can greatly improve the automatic identification and amplitude extraction capabilities of internal waves.

Description

Technical Field

[0001] The present invention belongs to the technical field of ocean observation, and in particular relates to a method for identifying internal solitary waves and inverting amplitudes in pattern data based on machine learning. Background Art

[0002] Internal waves occur in stably stratified seawater, with maximum amplitude near the jump layer. They are a widespread dynamic process within the ocean. They play a key role in marine ecosystems, resource development, and maritime military and navigation activities. Numerical internal wave model data, with its high temporal and spatial resolution, wide regional coverage, long-term data stability, strong scalability, and ease of integration with advanced technologies such as machine learning, provides data support for analyzing internal solitary waves in key sea areas.

[0003] However, current research on internal solitary waves still faces the following challenges: The sheer volume and richness of model data make it difficult for traditional data analysis methods to quickly and easily extract information about the location and shape of internal solitary waves. The sheer volume of model data, combined with the complex underwater ocean structure and the frequent influence of other ocean waves, makes it difficult to quickly and easily extract the amplitude of internal solitary waves from the model data. Furthermore, due to the difficulty of observing internal solitary waves, the size of the available internal solitary wave datasets is relatively limited. While obtaining information about internal solitary waves from numerical model data can enrich internal wave databases to a certain extent, in order to efficiently construct high-quality, large-scale, and refined internal solitary wave datasets, it is still necessary to develop technical means for identifying and extracting internal solitary waves from model data to ensure the accuracy and completeness of the datasets. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for identifying internal solitary waves and inverting amplitudes in pattern data based on machine learning, so as to make up for the deficiencies of the prior art.

[0005] This invention combines deep learning methods with pattern data. Starting solely from sea surface height data, this method can identify and extract internal solitary waves (ISWs). Compared to traditional methods that rely on computational analysis of massive amounts of pattern data, this strategy omits this complex step, significantly saving computing resources and improving processing efficiency. It can be used to automatically extract the shape, position, and amplitude of ISWs from refined pattern data, saving computing resources and improving processing efficiency.

[0006] In order to achieve the above objectives, the specific technical solutions provided by the present invention are:

[0007] A method for identifying and inverting internal solitary waves in pattern data based on machine learning includes the following steps:

[0008] S1: Extract the sea surface height data at the required location from the model data, and perform preprocessing and visualization;

[0009] S2: Annotate the internal wave fringes on the image obtained in S1 to construct a data set;

[0010] S3: Process and correct the internal wave fringe annotation data in S2;

[0011] S4: Construct a Mask-RCNN image segmentation model. Building on Faster R-CNN, Mask R-CNN divides object detection into two parts: region generation and classification, and introduces a parallel Mask branch for pixel-level segmentation. By using ResNet and FPN as feature extraction networks and proposing a RoIAlign operation that precisely aligns the spatial position of the feature map with the original image through bilinear interpolation, the accuracy of both detection and segmentation tasks is significantly improved.

[0012] S5: training the Mask-RCNN image segmentation model to obtain a trained internal wave fringe recognition model;

[0013] S6: Use the internal wave recognition model obtained in S4 to recognize internal waves in the pattern data, and record the output internal wave position and shape;

[0014] S7: extract temperature field data from the pattern data and construct a data set;

[0015] S8: Train the variational autoencoder algorithm VAE, perform performance analysis on the model with the optimal parameters, and obtain a trained temperature field inversion model;

[0016] S9: Using the internal wave identification result obtained in S6 as the input of the temperature field inversion model, and then extracting the amplitude of the internal solitary wave from the temperature field output by the trained temperature field inversion model.

[0017] Furthermore, the S1 includes:

[0018] S1-1: The preprocessing includes interpolating the acquired sea surface height and filling in blank areas;

[0019] S1-2: The data is processed using a local binary pattern (LBP) to image the data, and then a uniform resolution adjustment is performed to uniformly convert the data into sub-images with a resolution of 1080*1080.

[0020] Furthermore, the S3 includes:

[0021] S3-1: Extract the annotations made in S2 and create a json file;

[0022] S3-2: Manual corrections are made to problems such as the adhesion or incompleteness of multiple internal solitary waves in the annotations;

[0023] S3-3: Split the dataset into training dataset and validation dataset.

[0024] Furthermore, in S4, Mask R-CNN is an efficient instance segmentation algorithm that improves on Faster R-CNN. It extends the object detection task to instance segmentation by introducing a parallel mask branch for pixel-level segmentation. Mask R-CNN follows the principles of Faster R-CNN, dividing object detection into two parts: region generation and classification, while further improving the feature alignment problem of Faster R-CNN. By adopting ResNet and FPN (Feature Pyramid Network) as feature extraction networks, Mask R-CNN not only effectively captures multi-scale features but also significantly improves the accuracy of object detection and segmentation.

[0025] Mask R-CNN adds a parallel branch to object detection, specifically for identifying pixel-level segmentation masks for objects. This mask branch, a fully convolutional network (FCN), outputs pixel-level identification for each region of interest (ROI). Compared to traditional bounding box detection, the mask branch achieves a more granular description of the object region. Mask R-CNN uses ResNet or ResNet-FPN networks as the backbone feature extraction network. ResNet is a deep residual network with enhanced feature representation capabilities; FPN effectively enhances multi-scale object detection by constructing a feature pyramid structure. The Region of Interest Pooling (RoIPool) operation in Faster R-CNN can lead to accuracy loss due to misalignment between feature maps and the original image. Mask R-CNN introduces the RoIAlign operation, which addresses this issue by precisely aligning the feature map with the original image through bilinear interpolation, significantly improving the accuracy of both detection and segmentation tasks.

[0026] Furthermore, the process of constructing FPN includes three main steps: first, the input image is preprocessed through the bottom-top feature extraction network and then sent to the pre-trained feature extraction network, where features are gradually extracted from the bottom to the top, feature maps of different levels are generated, and the multi-level information of the image is captured; second, a top-down feature fusion strategy is adopted to gradually integrate deep features into shallow features; specifically, the shallow features are first reduced in dimension using 1×1 convolution, and then element-wise added to the up-sampled deep features, and finally 3×3 convolution is applied to further integrate multi-scale information; finally, region generation and target detection are performed on the generated multi-scale feature map, and the 4th, 5th, and 6th layers of the feature map respectively perform region proposal network (RPN) operations; RPN uses a sliding window mechanism to generate multi-scale anchor boxes, calculates target scores and coordinate offsets through parallel classification and regression branches, and combines non-maximum suppression to generate high-quality candidate regions.

[0027] Furthermore, the S7 includes:

[0028] S7-1: Extract the sea surface height data and the corresponding 400m water depth temperature field data from the model data.

[0029] S7-2: All sea surface height and temperature data are interpolated and preprocessed by filling in blank areas, and then randomly shuffled after being processed by cropping, rotating, etc.

[0030] S7-3: Divide the dataset into training set and test set.

[0031] Furthermore, the S8 includes:

[0032] S8-1: The VAE architecture consists of an encoder and a decoder. The encoder receives input data, extracts features through multiple hidden layers (such as fully connected layers or multi-convolutional layers), and outputs two vectors: the mean vector μ and the variance vector σ. These two vectors describe the probability distribution of the input data in the latent space. The decoder receives the latent variable z, gradually restores it to a structure similar to the input data through a series of hidden layers (such as deconvolution layers or fully connected layers), and generates reconstructed data through the output layer.

[0033] S8-2: During the training process, the variational lower bound (ELBO, Evidence Lower Bound) is used as the loss function. The goal of VAE is to maximize the lower bound so that the generated temperature field is as close to the true value as possible. The specific calculation formula of ELBO is:

[0034] ;

[0035] ELBO consists of two parts: one is the reconstruction error (reconstruction loss), which is the difference between the generated temperature field and the true temperature field; the other is the KL divergence.

[0036] S8-3: Adjust the model weights and parameters based on the training dataset to minimize the loss function. Specifically, use stochastic gradient descent to optimize the model and gradually update the network parameters. During training, use batch processing to accelerate training and perform multiple iterations until the loss converges.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] This paper proposes an efficient and automated method for identifying and extracting internal solitary waves (ISWs) by combining deep learning algorithms with ocean numerical model data. This method automatically identifies the location and morphology of ISWs from sea surface height data and extracts their amplitudes from the inverted temperature field.

[0039] Compared with traditional methods, this invention offers significant advantages: First, it can efficiently and accurately process large-scale, high-precision data. Second, it implements a fully automated process from sea surface height data to amplitude extraction, laying the foundation for subsequent internal wave research based on observational data. Finally, by combining machine learning with physical models, it significantly improves computational efficiency and processing speed, saving significant computing resources and possessing strong practical application value. This invention is suitable for internal wave disaster warning and marine science research, significantly improving the processing efficiency and accuracy of refined internal solitary wave data. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is the overall process of the present invention Figure 1 .

[0041] Figure 2 This is the overall process of the present invention Figure 2 .

[0042] Figure 3 This is the flow chart of internal wave fringe annotation and dataset construction.

[0043] Figure 4 It is a schematic diagram of the Mask R-CNN network structure and principle.

[0044] Figure 5 This is the principle diagram of FPN structure.

[0045] Figure 6 This is the schematic diagram of the RPN structure.

[0046] Figure 7 This is a diagram of the structural principle of the variational autoencoder.

[0047] Figure 8 The following is a demonstration of the recognition results, where (a), (c), and (e) are the original data of sea surface height, and (b), (d), and (f) are the recognition results.

[0048] Figure 9 These are two demonstration images of internal wave amplitude extraction at randomly selected different locations. DETAILED DESCRIPTION

[0049] The present invention is further explained and illustrated below through specific embodiments in conjunction with the accompanying drawings.

[0050] Example 1:

[0051] like Figure 1 、 Figure 2 As shown, a method for identifying internal solitary waves and inverting amplitudes in pattern data based on machine learning includes the following steps:

[0052] Step 1: First, extract sea surface height data for the target area from the selected ocean numerical model data. This data is typically in a three-dimensional format, containing time, space (latitude and longitude), and corresponding sea surface height values. Interpolation is then used to fill in missing portions or blank values ​​in the data. Texture features are then extracted from the processed sea surface height data using the LBP algorithm.

[0053] The LBP algorithm extracts texture features by comparing the relative sizes of pixel grayscale values ​​in a local area of ​​an image. Specifically, given a central pixel p c and its surrounding n neighboring pixels p1, p2, ..., p n , the LBP algorithm compares the gray value p of each neighborhood pixel with the central pixel i and p c The size relationship between them generates a binary pattern. If the grayscale value of the neighboring pixel is greater than or equal to the grayscale value of the central pixel, the binary value of the pixel is 1, otherwise it is 0. The specific formula is as follows:

[0054] ;

[0055] Here, B i It is The binary value of all neighboring pixels B1, B2, ..., Bn is then concatenated in clockwise (or counterclockwise) order to form a binary string B1, B2, ..., Bn of length n, thus obtaining a binary number. This binary pattern is then converted into a decimal number as the texture descriptor of the area. The specific form of is obtained as follows:

[0056]

[0057] Here, n is the number of pixels in the neighborhood, typically set to 8 (for example, in a 3x3 neighborhood). This value is the texture feature of the local area, reflecting the grayscale relationship between the central pixel and its neighboring pixels. This process effectively captures the local texture structure of the image, exhibits good noise immunity, and is sensitive to local details.

[0058] After LBP processing, the resulting image is a two-dimensional image representing the local texture features of the sea surface height. All of these images are resized to a standard 1080×1080 pixel image to ensure consistent resolution for the deep learning model's input requirements. This resolution adjustment is performed using an image scaling algorithm to preserve image detail and maintain physical consistency.

[0059] Step 2: Annotate the internal wave stripes in the sea surface height image processed in step 1. First, through manual observation and image analysis, locate the internal solitary wave area in each image and accurately annotate it. The annotation process includes: determining the starting position, crest position, trough position and the outline of the entire waveform of the internal wave stripes. Since internal solitary waves usually appear as stripes propagating in the horizontal direction, they are specifically manifested in the data as striped areas where the sea surface height changes dramatically. During the annotation process, according to the internal wave morphology in different images, we use image annotation tools to select the internal solitary wave area and mark the boundaries of each band. After the annotation is completed, all annotated images are organized into an annotated dataset containing approximately 1,600 images.

[0060] Step 3: If Figure 3 As shown in the figure, the internal solitary wave adhesion and edge incompleteness that occurred during the annotation process were manually corrected, and the dataset was then converted into a JSON file. The annotation results of this dataset include the specific location and morphology of the internal solitary wave streaks in each image, providing the labeled samples necessary for training the machine learning model, thereby helping the model learn how to accurately identify and extract the characteristics of internal solitary waves from sea surface height images. With this precisely labeled data, the model can better identify and segment internal solitary wave regions. During the dataset partitioning stage, the corrected dataset was divided into training and validation datasets according to a certain ratio. The training set accounted for 80%, and the validation set accounted for 20%.

[0061] Step 4: Input the training data set obtained in step 3 into the Mask-RCNN image segmentation algorithm for training to generate a recognition model for internal solitary wave fringes. The Mask-RCNN structure is as follows: Figure 4As shown in the figure. First, the Mask-RCNN network structure is configured. In the specific implementation, this embodiment selects the Mask R-CNN network architecture based on ResNet-50-FPN. Compared with ResNet-101, the ResNet-50 network has a shallower depth and fewer parameters, thus reducing computational complexity and training time. In the internal solitary wave detection task, since a large amount of high-resolution sea surface height data needs to be processed, the use of ResNet-50 can significantly shorten the training time and improve the training efficiency of the model, while still meeting the task requirements in terms of accuracy.

[0062] In addition, combined with the design of FPN, ResNet-50 can capture multi-scale features more efficiently. The process of building FPN includes three main steps. First, the input image is preprocessed through the bottom-top feature extraction network and then sent to the pre-trained feature extraction network (such as ResNet). The network gradually extracts features from the bottom to the top, generates feature maps of different levels, and captures the multi-level information of the image. Secondly, a top-down feature fusion strategy is adopted to gradually fuse deep features into shallow features. Specifically, the shallow features are first reduced in dimension using 1×1 convolution, and then added element-by-element to the deep features after upsampling, and finally 3×3 convolution is applied to further integrate multi-scale information. In this way, each layer of feature map combines shallow and deep features, greatly improving the expression ability of multi-scale features. Finally, region generation and target detection are performed on the generated multi-scale feature map, and the 4th, 5th, and 6th layers of the feature map perform region proposal network (RPN) operations respectively. After the above operations, the structural principle diagram of FPN is as follows. Figure 6 As shown. This multi-scale feature capture capability is crucial for internal solitary wave detection in complex backgrounds, because the spatial distribution of internal solitary waves often has diversity and different scale characteristics. By combining ResNet-50 with FPN, Mask R-CNN can enhance the detection and segmentation capabilities of internal solitary wave stripes of different scales while ensuring the lightweight model. Combined with FPN to enhance the multi-scale feature extraction capability, and set up a region generation network (RPN) to identify candidate regions for internal solitary waves, RPN uses a sliding window mechanism to generate multi-scale anchor boxes, calculates target scores and coordinate offsets through parallel classification and regression branches, and combines non-maximum suppression to generate high-quality candidate regions. Its design is efficient and flexible, providing a key basic module for target detection and significantly improving the accuracy and efficiency of detection. The RPN structural principle diagram is shown below. Figure 7 shown.

[0063] Next, we begin model training using the training dataset by setting an appropriate learning rate, batch size, and number of training rounds. During training, we employ data augmentation techniques to improve the model's generalization capabilities and train the model to minimize both the object detection loss and the mask segmentation loss. After each round of training, we evaluate model performance using the validation set, optimize hyperparameters, and ultimately select the optimal model for storage.

[0064] The evaluation indicators of the model are as follows: Accuracy refers to the proportion of samples correctly identified by the model to the total number of samples, and the calculation formula is:

[0065] ;

[0066] Compared with other indicators, accuracy is intuitive and concise, and is suitable for cases where the categories are balanced. However, in datasets with unbalanced categories, accuracy may be dominated by high-frequency categories, resulting in inaccurate evaluation results.

[0067] Precision refers to the proportion of samples that are actually positive among the samples identified as positive by the model. The calculation formula is:

[0068] ;

[0069] Precision is suitable for tasks that are sensitive to false positives (FPs), focusing on measuring the accuracy of the model in identifying the positive class. However, it only focuses on the recognition of the positive class and ignores the performance of the negative class, which may lead to biased recognition of one class.

[0070] Recall refers to the proportion of all actual positive classes that the model correctly identifies as positive. The formula is:

[0071] ;

[0072] Recall measures the model's ability to identify positive samples and is suitable for tasks that are sensitive to false negatives (FN). An increase in recall may lead to a decrease in precision because it will identify more positive samples, which may contain incorrect recognition results.

[0073] The F1-Score is the harmonic mean of precision and recall, defined as:

[0074] ;

[0075] The F1 score takes into account both precision and recall and is suitable for cases with imbalanced categories. Although it balances precision and recall, it gives equal weight to both and may not be flexible enough to adapt to different tasks.

[0076] After training, the optimal internal wave fringe recognition model obtained can efficiently and automatically identify internal solitary wave areas from new sea surface height data, and output their position and morphological characteristics, providing important data support for subsequent internal wave amplitude extraction.

[0077] Step 5: Apply the optimal internal wave fringe recognition model trained in step 4 to the new ocean pattern data to achieve automatic recognition and position extraction of internal solitary waves.

[0078] In step 6, sea surface height (SSH) data and the corresponding 400-meter water depth temperature field data are extracted, processed, and organized for subsequent temperature field inversion model training. This data needs to cover a sufficiently long timeframe and a wide spatial area to ensure data representativeness and training diversity. The extracted SSH and temperature data undergo preprocessing operations such as cropping and rotation. This helps remove irrelevant areas and ensures that only useful regional data is included in the dataset. Furthermore, to enhance the robustness and generalization of the model, all data are randomly processed using data augmentation methods, such as random rotation, scaling, and cropping. Next, all SSH and temperature data are randomly shuffled. The dataset is divided into training and validation datasets according to a certain ratio, with the training set accounting for 80% and the validation set accounting for 20%.

[0079] Subsequently, the data is input into the VAE model for training. The VAE model structure is as follows Figure 5 As shown in the figure, sea surface height data and temperature field data at a depth of 400 meters are used as input features. The VAE model learns the latent spatial distribution of these data to invert the temperature field. This provides the temperature field inversion model with a sufficient, preprocessed, and enhanced dataset to improve the model's training and predictive capabilities.

[0080] In step 7, the training set obtained in step 6 is fed into a variational autoencoder algorithm for training. A variational autoencoder is a deep learning generative model that learns the probability distribution of a latent space from data and generates outputs similar to the input data. In this process, the VAE will be used to invert the temperature field at a depth of 400 meters from the sea surface height data.

[0081] The VAE architecture consists of an encoder and a decoder. The variational autoencoder (VAE) network structure is a deep generative model that expands and improves upon traditional autoencoders. Its core structure consists of three parts: an encoder, a reparameterization technique, and a decoder. After receiving input data, the encoder extracts features through multiple hidden layers (such as fully connected layers or convolutional layers) and outputs two vectors: a mean vector μ and a variance vector σ. These two vectors describe the probability distribution of the input data in the latent space. Unlike autoencoders, which directly output low-dimensional feature vectors, VAEs use a reparameterization technique to sample μ and σ from a standard normal distribution to obtain the latent variable z, thereby introducing randomness. The decoder receives the latent variable z and gradually restores it to a structure similar to the input data through a series of hidden layers (such as deconvolutional layers or fully connected layers). The decoder then generates reconstructed data through the output layer.

[0082] During the training process, the variational lower bound (ELBO, Evidence Lower Bound) is used as the loss function. The goal of VAE is to maximize the lower bound so that the generated temperature field is as close to the true value as possible. The specific calculation formula of ELBO is:

[0083] .

[0084] ELBO consists of two components: the reconstruction loss, which is the difference between the generated temperature field and the true temperature field; and the KL divergence, which measures the difference between the distribution of the latent variable and the prior distribution. This serves as a regularizer to prevent model overfitting. During training, we adjust the model weights and parameters based on the training dataset to minimize the loss function. Specifically, stochastic gradient descent is used to optimize the model, gradually updating the network parameters. During training, batch processing is used to accelerate training, and multiple iterations are performed until the loss converges.

[0085] After each round of training, the model is evaluated using the test set to check the accuracy of the generated temperature field. Evaluation indicators usually include mean square error (MSE), mean absolute error (MAE), etc. The specific calculation formula for MSE is:

[0086] ;

[0087] The specific calculation formula of MAE is:

[0088] .

[0089] By analyzing the performance of the test set, hyperparameters (such as the learning rate and latent space dimension) were adjusted to achieve optimal training results. Ultimately, after training, an optimal temperature field inversion model was obtained, capable of generating the temperature field at a depth of 400 meters based on the input sea surface height data. This model can infer the underwater temperature distribution from the sea surface height data, providing accurate temperature field data for subsequent internal solitary wave amplitude extraction.

[0090] In step 8, the internal wave identification result obtained in step 5 is used as input, and the temperature field inversion model trained in step 7 is applied to extract the amplitude of the internal solitary wave from the inverted temperature field.

[0091] First, using the internal wave identification results obtained in step 5, the location information of each internal solitary wave is extracted. Next, the identified internal solitary wave locations are fed as input to the variational autoencoder model trained in step 7. This model can map the corresponding sea surface height data to temperature field data.

[0092] To extract the amplitude, the region containing the internal solitary wave is first demarcated in the inverted temperature field. By combining the positional and morphological information from the internal wave identification results, the internal solitary wave's location range in the temperature field can be determined. The amplitude of the internal solitary wave is then calculated by calculating the maximum vertical variation of the temperature field within this region. Specifically, the amplitude can be estimated by the difference between the local maximum and minimum values ​​of the temperature field.

[0093] Example 2

[0094] A practical verification was conducted based on the method of Example 1. This example selected sea surface height data from March 22, 2014, to March 24, 2014, at 12:00 CST (UTC+8). During the internal solitary wave identification process, the identification results were mapped onto a graph of unprocessed sea surface height data to facilitate visualization. The original data graph and the identification results are also displayed.

[0095] Among them, the automatic identification and position extraction of internal solitary waves, the identification results are demonstrated as follows Figure 8As shown in the figure, the purple box accurately identifies the area where the internal solitary wave is located, ensuring the precise positioning of the target area. In this way, the model can effectively determine the specific location of the internal solitary wave, providing reliable regional information for further analysis and application. At the same time, the red line clearly outlines the shape of the internal solitary wave, providing a clear outline of the internal solitary wave characteristics and helping to further understand the spatial distribution characteristics of the internal solitary wave. These recognition results show that the adopted method can not only efficiently locate the internal solitary wave area (box result) but also accurately extract its morphological characteristics (mask result), fully demonstrating the efficiency, accuracy, and robustness of the model in the task of internal solitary wave detection, providing strong support for marine environmental monitoring and related research.

[0096] Amplitude extraction demonstration Figure 9 As shown in the figure, two internal waves were randomly selected at different locations within the target time for amplitude extraction. The isotherm analysis reveals a clear internal solitary wave phenomenon, manifested by the isotherms exhibiting significant curvature or fluctuations in certain regions. To extract the amplitude of the internal solitary wave, the method is to select the region with the most dramatic isotherm fluctuations and calculate the difference between the maximum amplitude in that region and the maximum isotherm within a certain range to the west. This range is then extracted from the internal wave identification results.

[0097] Specifically, the maximum amplitude corresponds to the deepest point of the isotherm fluctuation, while the western maximum of the isotherm represents the temperature value of the stable area before the fluctuation. By calculating the difference between the two, the amplitude of the fluctuation can be obtained. This amplitude value reflects the strength of the internal solitary wave and its impact on the underwater temperature field.

[0098] Finally, the extracted internal solitary wave amplitudes are stored and recorded along with other analysis results to support further research and data visualization. This amplitude extraction process provides key data support for a deeper understanding of the dynamic characteristics of internal solitary waves and their relationship with environmental factors, thereby promoting the scientific research and application of internal solitary waves.

[0099] The present invention is based on a deep learning method for rapidly identifying internal solitary waves and extracting their amplitudes in refined pattern data. By combining the Mask-RCNN target detection model and a variational autoencoder, the present invention can rapidly and automatically identify the position and shape of internal solitary waves in refined pattern data, and accurately invert their amplitudes.

[0100] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects disclosed in the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for identifying and inverting internal solitary waves in pattern data based on machine learning, characterized in that: The following steps are involved: S1: Extract the sea surface height data at the required location from the model data, and perform preprocessing and visualization; S2: Annotate the internal wave fringes on the image obtained in S1 to construct a data set; S3: Process and correct the internal wave fringe annotation data in S2; S4: Build a Mask-RCNN image segmentation model. Mask R-CNN is based on Faster R-CNN, which divides object detection into two parts: region generation and classification, and introduces a parallel Mask branch for pixel-level segmentation tasks; By using ResNet and FPN as feature extraction networks and proposing the RoIAlign operation, the spatial position of the feature map and the original image is accurately aligned through bilinear interpolation; S5: training the Mask-RCNN image segmentation model to obtain a trained internal wave fringe recognition model; S6: Use the internal wave recognition model obtained in S4 to recognize internal waves in the pattern data, and record the output internal wave position and shape; S7: extract temperature field data from the pattern data and construct a data set; S8: Train the variational autoencoder algorithm VAE, perform performance analysis on the model with the optimal parameters, and obtain a trained temperature field inversion model; S9: Using the internal wave identification result obtained in S6 as the input of the temperature field inversion model, and then extracting the amplitude of the internal solitary wave from the temperature field output by the trained temperature field inversion model.

2. The method for identifying internal solitary waves and inverting amplitudes in pattern data according to claim 1, wherein: Said S1 comprises: S1-1: The preprocessing includes interpolating the acquired sea surface height and filling in blank areas; S1-2: The data is processed using a local binary pattern (LBP) to image the data, and then a uniform resolution adjustment is performed to uniformly convert the data into sub-images with a resolution of 1080*1080.

3. The method for identifying internal solitary waves and inverting amplitudes in pattern data according to claim 1, wherein: The S3 includes: S3-1: Extract the annotations made in S2; S3-2: Manual corrections are made to the problems of multiple internal solitary waves being stuck or incomplete in the annotations; S3-3: Split the dataset into training dataset and validation dataset.

4. The method for identifying internal solitary waves and inverting amplitudes in pattern data according to claim 1, wherein: In S4, the Mask branch is a fully convolutional network that can output corresponding pixel-level recognition based on each target region ROI.

5. The method for identifying internal solitary waves and inverting amplitudes in pattern data according to claim 1, wherein: In the S4, the process of constructing FPN includes three main steps: first, the input image is preprocessed through the bottom-top feature extraction network and then sent to the pre-trained feature extraction network, where features are gradually extracted from the bottom to the top, feature maps of different levels are generated, and the multi-level information of the image is captured; second, a top-down feature fusion strategy is adopted to gradually fuse deep features into shallow features; specifically, the shallow features are first subjected to dimensionality reduction processing using 1×1 convolution, and then element-wise addition is performed with the deep features after upsampling, and finally 3×3 convolution is applied to further integrate multi-scale information; finally, region generation and target detection are performed on the generated multi-scale feature map, and the 4th, 5th, and 6th layers of the feature map respectively perform region proposal network (RPN) operations, which use a sliding window mechanism to generate multi-scale anchor boxes, calculate target scores and coordinate offsets through parallel classification and regression branches, and generate high-quality candidate regions in combination with non-maximum suppression.

6. The method for identifying internal solitary waves and inverting amplitudes in pattern data according to claim 1, wherein: The S7 includes: S7-1: Extract sea surface height data and corresponding water depth temperature field data from the model data; S7-2: All sea surface height and temperature data are interpolated and preprocessed by filling in blank areas, and then cropped, rotated, and randomly shuffled; S7-3: Divide the dataset into training set and test set.

7. The method for identifying internal solitary waves and inverting amplitudes in pattern data according to claim 1, wherein: The S8 includes: S8-1: The VAE architecture consists of an encoder and a decoder. The encoder receives input data, extracts features through multiple hidden layers, and outputs two vectors: the mean vector μ and the variance vector σ. These two vectors describe the probability distribution of the input data in the latent space. The decoder receives the latent variable z, gradually restores it to a structure similar to the input data through a series of hidden layers, and generates reconstructed data through the output layer. S8-2: During the training process, the variational lower bound ELBO is used as the loss function. The goal of VAE is to maximize the lower bound. The specific calculation formula of ELBO is: ； ELBO consists of two parts: the first is the reconstruction error, which is the difference between the generated temperature field and the true temperature field; the second is the KL divergence; S8-3: Adjust the model weights and parameters based on the training dataset to minimize the loss function. Specifically, use stochastic gradient descent to optimize the model and gradually update the network parameters. During training, use batch processing to accelerate training and perform multiple iterations until the loss converges.

Images