Mode data internal solitary wave identification and amplitude inversion method based on machine learning

Abstract

The invention relates to an internal solitary wave identification and amplitude inversion method in mode data based on machine learning, and belongs to the technical field of ocean observation. In combination with a Mask-RCNN and a variational auto-encoder, firstly, sea surface height data is preprocessed and imaged, texture features are extracted through a local binary pattern algorithm, a labeled data set is constructed, model training is performed, and internal solitary wave stripes and positions thereof are identified from the labeled data set; through extraction and processing of temperature field data, the temperature field at the water depth is inversed by using a variational auto-encoder; and extracting the amplitude of the internal solitary wave in the temperature field based on the internal wave identification result. According to the method, large-scale and refined ocean mode data can be efficiently and accurately processed, and the efficiency and accuracy of data processing are remarkably improved. The method has a wide application prospect, and especially in marine disaster early warning, marine scientific research and internal wave mechanics research, the automatic identification and amplitude extraction capability of internal waves can be greatly improved.

Description

Technical Field

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

[0002] Internal waves in the ocean are wave phenomena that occur in seawater with stable density stratification and have the maximum amplitude near the thermocline, and are a dynamic process widely existing in the ocean interior. Ocean internal solitary waves play a key role in activities such as the ocean ecosystem, ocean resource development, and maritime military and navigation. Due to the advantages of high spatio-temporal resolution, wide regional coverage, long-term data stability, strong scalability, and easy integration with advanced technologies such as machine learning, internal wave numerical model data provides data support for analyzing internal solitary waves in key sea areas.

[0003] However, the current research on internal solitary waves still has the following problems: due to the large amount of model data and rich data elements, it is difficult for traditional data analysis methods to quickly and conveniently extract the position and shape information of internal solitary waves from it; due to the large amount of model data and the complex underwater ocean structure, often accompanied by the influence of other ocean waves, it is difficult to quickly and conveniently extract the amplitude of internal solitary waves from the model data; due to the difficulty of observing internal solitary waves, the scale of the current internal solitary wave dataset available for research is relatively limited. Although obtaining information on internal solitary waves from numerical model data can enrich the internal wave database to a certain extent, in order to efficiently construct a high-quality, large-scale, and refined internal solitary wave dataset, it is still necessary to develop technical means for identifying and extracting internal solitary waves in model data to ensure the accuracy and integrity of the dataset. 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 model data based on machine learning to make up for the deficiencies of the existing technology.

[0005] The present invention combines the deep learning method with model data. This method can complete the identification and amplitude extraction of internal solitary waves only starting from the sea surface height data. Compared with the traditional method of calculating and analyzing the large amount of model data, the strategy of combining machine learning with model data omits this complex step, greatly saving computing resources and improving processing efficiency. It is applied to automatically extract the shape, position, and amplitude information of internal solitary waves in refined model data, saving computing resources and improving processing efficiency.

[0006] To achieve the above purpose, the specific technical solution provided by the present invention is as follows: A method for identifying internal solitary waves and inverting amplitudes in model data based on machine learning, comprising the following steps: S1: Extract the sea surface height data at the required positions from the pattern data, and perform preprocessing and imaging; S2: Perform internal wave fringe annotation on the image obtained in S1 to construct a dataset; S3: Process and correct the internal wave fringe annotation data in S2; S4: Construct a Mask-RCNN image segmentation model. The Mask R-CNN divides object detection into two parts: region generation and classification on the basis of Faster R-CNN, 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 positions of the feature map and the original image are accurately aligned through bilinear interpolation, significantly improving the accuracy of detection and segmentation tasks; S5: Train 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 identify the internal waves in the pattern data, and record the output positions and shapes of the internal waves; S7: Extract the temperature field data from the pattern data to construct a dataset; S8: Train the variational autoencoder algorithm VAE, perform performance analysis on the model with the optimal parameters obtained, and obtain a trained temperature field inversion model; S9: Use the internal wave recognition results obtained in S6 as the input of the temperature field inversion model, and then extract the amplitude of the internal solitary wave from the temperature field output by the trained temperature field inversion model.

[0007] Further, the S1 includes: S1-1: The preprocessing includes interpolating the obtained sea surface height and filling in blank areas; S1-2: Image the data through the processing method of local binary pattern (LBP), and then perform unified resolution adjustment to uniformly convert the data into sub-images with a resolution of 1080*1080.

[0008] Further, the S3 includes: S3-1: Extract the annotations made in S2 and make them into a json file; S3-2: Manually correct problems such as adhesion or incompleteness of multiple internal solitary waves in the annotations; S3-3: Divide the dataset into a training dataset and a validation dataset.

[0009] Furthermore, in S4, Mask R-CNN is an efficient instance segmentation algorithm improved based on Faster R-CNN. It expands the object detection task to an instance segmentation task by introducing a parallel Mask branch for pixel-level segmentation tasks. Mask R-CNN follows the idea of Faster R-CNN, divides object detection into region generation and classification parts, and further improves the feature alignment problem of Faster R-CNN. By using ResNet and FPN (Feature Pyramid Network) as the feature extraction network, Mask R-CNN can not only effectively capture multi-scale features but also significantly improve the accuracy of object detection and segmentation.

[0010] Mask R-CNN adds a parallel branch on the basis of object detection to identify the pixel-level segmentation mask of the object. This Mask branch is a fully convolutional network (FCN), which can output corresponding pixel-level recognition based on each region of interest (ROI). Compared with traditional bounding box detection, the Mask branch realizes a finer-grained description of the target region. Mask R-CNN uses the ResNet or ResNet-FPN network as the feature extraction backbone network. ResNet is a deep residual network with stronger feature expression ability; FPN effectively enhances the detection ability for multi-scale objects by constructing a feature pyramid structure. The region of interest pooling (RoIPool) operation in Faster R-CNN may cause accuracy loss due to the mis-alignment between the feature map and the original image. In response to this, Mask R-CNN proposes the RoIAlign operation, which accurately aligns the spatial positions of the feature map and the original image through bilinear interpolation, solves this problem, and thus significantly improves the accuracy of the detection and segmentation tasks.

[0011] Furthermore, the process of constructing the FPN includes three main steps: First, through a bottom-up feature extraction network, the input image is preprocessed and then fed into a pre-trained feature extraction network to gradually extract features from the bottom layer to the top layer, generating feature maps of different levels to capture multi-level information of the image; Second, a top-down feature fusion strategy is adopted to gradually fuse deep features into shallow features; Specifically, first, the shallow features are dimensionally reduced using 1×1 convolutions, then added element-wise to the deep features after upsampling, and finally, 3×3 convolutions are applied to further integrate multi-scale information; Finally, region generation and object detection are performed on the generated multi-scale feature maps, and the region proposal network (RPN) operations are respectively executed on the 4th, 5th, and 6th layers of the feature maps; The RPN uses a sliding window mechanism to generate multi-scale anchor boxes, calculates the object scores and coordinate offsets through parallel classification and regression branches, and combines non-maximum suppression to generate high-quality candidate regions.

[0012] Furthermore, the S7 includes: S7-1: Extract sea surface height data and the corresponding 400m water depth temperature field data from the pattern data respectively.

[0013] S7-2: Interpolate all the sea surface height data and temperature data, perform preprocessing such as filling blank areas, and then randomly shuffle after cropping, rotating, etc.; S7-3: Divide the dataset into a training set and a test set.

[0014] Furthermore, the S8 includes: S8-1: The architecture of the VAE includes an encoder and a decoder. After receiving the input data, the encoder extracts features through multiple hidden layers (such as fully connected layers or multiple convolutional layers) and outputs two vectors: the mean vector μ and the variance vector σ, which describe the probability distribution of the input data in the latent space; The decoder part receives the latent variable z, gradually restores it to a structure similar to the input data through a series of hidden layers (such as transposed convolutional layers or fully connected layers), and generates the reconstructed data through the output layer; S8-2: During the training process, the variational lower bound (ELBO, Evidence Lower Bound) is used as the loss function. The goal of the VAE is to maximize the lower bound, so that the generated temperature field is as close as possible to the real value. The specific calculation formula of the ELBO is: ; The ELBO consists of two parts: one is the reconstruction error, that is, the difference between the generated temperature field and the real temperature field; the other is the KL divergence; S8-3: Adjust the weights and parameters of the model according to the training dataset to minimize the loss function. Specifically, use stochastic gradient descent to optimize the model and gradually update the parameters of the network. During training, use batch processing to accelerate training and perform multiple iterations until the loss converges.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: By combining deep learning algorithms with ocean numerical model data, the present invention proposes an efficient and automated method for identifying internal solitary waves and extracting amplitudes. This method can automatically identify the positions and shapes of internal solitary waves from sea surface height data, and use the inverted temperature field to extract the amplitudes of internal solitary waves.

[0016] Compared with traditional methods, the present invention has significant advantages: First, the present invention can efficiently and accurately process large-scale and high-precision data; Second, it realizes a complete automated process from sea surface height data to amplitude extraction, laying a foundation for subsequent research on internal waves starting from observational data; Finally, by combining machine learning with physical models, the calculation efficiency and processing speed are greatly improved, saving a large amount of computing resources, and having strong practical application value. The present invention is applicable to internal wave disaster warning and ocean science research, and can significantly improve the processing efficiency and accuracy of refined internal solitary wave data. Description of the Drawings

[0017] Figure 1 is the overall process of the present invention Figure 1 .

[0018] Figure 2 is the overall process of the present invention Figure 2 .

[0019] Figure 3 is the flow chart of internal wave fringe annotation and dataset construction.

[0020] Figure 4 is the schematic diagram of the Mask R-CNN network structure and principle.

[0021] Figure 5 is the schematic diagram of the FPN structure principle.

[0022] Figure 6 is the schematic diagram of the RPN structure principle.

[0023] Figure 7 is the schematic diagram of the variational autoencoder structure principle.

[0024] Figure 8 is the demonstration diagram of the recognition result, where (a), (c), and (e) are the original sea surface height data, and (b), (d), and (f) are the recognition results.

[0025] Figure 9 These are two demonstration diagrams for extracting the internal wave amplitudes at different randomly selected positions. Specific implementation manners

[0026] The present invention will be further explained and illustrated below through specific embodiments in conjunction with the accompanying drawings.

[0027] Embodiment 1: As Figure 1 、 Figure 2 shown, a method for identifying internal solitary waves and amplitude inversion in pattern data based on machine learning includes the following steps: Step 1: First, extract the sea surface height data in the target area from the selected ocean numerical model data. The data is usually in a three-dimensional format, including time, space (longitude and latitude), and the corresponding sea surface height values. Subsequently, use the interpolation method to fill the missing parts or blank values in the data. Then, use the LBP algorithm to extract the texture features from the processed sea surface height data.

[0028] The LBP algorithm extracts texture features by comparing the relative magnitudes of pixel gray values within 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 generates a binary pattern by comparing the gray values p i and p c between each neighboring pixel and the central pixel. If the neighboring pixel is greater than or equal to the gray value of the central pixel, the binary value of this pixel is 1, otherwise it is 0. The specific formula is as follows: ; Here, B i is the binarization result of the th neighboring pixel. Then, the binary values B1, B2,..., Bn of all neighboring pixels are connected in a 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 this area is specifically obtained as follows:

[0029] where n is the number of pixels in the neighborhood, usually taking the value of 8 (for example, in a neighborhood of size 3x3). This value is the texture feature of this local area, which reflects the gray relationship pattern between the central pixel and its neighboring pixels. This process can effectively capture the local texture structure of the image and has good anti-noise performance and sensitivity to local details.

[0030] After the LBP processing is completed, the resulting image is a two-dimensional image that represents the local texture features of the sea surface height. All these images will be uniformly adjusted to standard images of 1080×1080 pixels to ensure that each image has a consistent resolution to meet the input requirements of the deep learning model. The resolution adjustment is performed using an image scaling algorithm to ensure that the details of the image are retained and the physical consistency of the image is maintained.

[0031] Step 2: Annotate the internal wave fringes on the processed sea surface height image in Step 1. First, through manual observation and image analysis, locate the internal solitary wave regions in each image and accurately annotate them. The annotation process includes: determining the starting position, wave crest position, wave trough position, and the contour of the entire waveform of the internal wave fringes. Since internal solitary waves usually appear as fringes propagating in the horizontal direction, specifically manifested as fringe regions with drastic changes in sea surface height in the data. During the annotation process, for the internal wave forms in different images, we use an image annotation tool to select the internal solitary wave regions and mark the boundaries of each wave band. After the annotation is completed, all the annotated images are organized into an annotated data set containing approximately 1600 images.

[0032] Step 3: As Figure 3 shown, manually correct the adhesion phenomenon of internal solitary waves and the incomplete phenomenon of the edges that occur during the annotation process, and then convert the data set into a json file. The annotation results of this data set contain the specific positions and forms of the internal wave fringes of each image, providing annotated samples required for training the machine learning model, and thus helping the model learn how to accurately identify and extract the features of internal solitary waves from the sea surface height image. Through these accurately annotated data, the model can better identify and segment the internal solitary wave regions. In the data set division stage, the corrected data set is divided into a training data set and a validation data set according to a certain ratio. Among them, the training set accounts for 80%, and the validation set accounts for 20%.

[0033] Step 4: Input the training data set obtained in Step 3 into the Mask-RCNN image segmentation algorithm for training to generate an identification model for internal wave fringes. The Mask-RCNN structure is as Figure 4 shown. First, configure the network structure of Mask-RCNN. 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, so the computational complexity and training time are lower. In the internal solitary wave detection task, since a large amount of high-resolution sea surface height data needs to be processed, using ResNet-50 can significantly shorten the training time, improve the training efficiency of the model, and still meet the task requirements in terms of accuracy.

[0034] In addition, combined with the design of FPN, ResNet-50 can capture multi-scale features more efficiently. The process of constructing FPN includes three main steps. First, through the bottom-up feature extraction network, the input image is preprocessed and then fed into a pre-trained feature extraction network (such as ResNet). This network gradually extracts features from the bottom layer to the top layer, generates feature maps at different levels, and captures the multi-level information of the image. Second, 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, then added element-wise 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 enhancing the expression ability for multi-scale features. Finally, region generation and object detection are performed on the generated multi-scale feature maps. The 4th, 5th, and 6th layers of the feature maps respectively perform Region Proposal Network (RPN) operations. After the above operations, the structural schematic diagram of FPN is as Figure 6 shown. This multi-scale feature capture ability is crucial for the detection of internal solitary waves 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 ability of internal solitary wave stripes at different scales while ensuring the lightweight of the model. Combining FPN to enhance the multi-scale feature extraction ability and setting up a Region Proposal Network (RPN) to identify candidate regions of internal solitary waves, RPN uses a sliding window mechanism to generate multi-scale anchor boxes, calculates the object 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 object detection and significantly improving the accuracy and efficiency of detection. The structural schematic diagram of RPN is as Figure 7 shown.

[0035] Next, by setting appropriate learning rate, batch size, and number of training epochs, the model training starts using the training dataset. During the training process, data augmentation techniques are adopted to improve the generalization ability of the model, and the model is trained to minimize the object detection loss and mask segmentation loss. After each round of training, the validation set is used to evaluate the model performance, optimize the hyperparameters, and finally select the optimal model for saving.

[0036] Among them, the evaluation metrics of the model are as follows: Accuracy refers to the proportion of the number of samples correctly identified by the model to the total number of samples, and the calculation formula is: ; Accuracy is intuitive and concise compared to other metrics and is suitable for cases with balanced classes. However, in datasets with imbalanced classes, accuracy may be dominated by high-frequency classes, leading to inaccurate evaluation results.

[0037] Precision is the proportion of samples actually classified as positive among those identified as positive by the model. The calculation formula is: ; Precision is suitable for tasks that are sensitive to false positives (FP) and mainly measures the accuracy of the model in identifying positive classes. However, it only focuses on the identification of positive classes and ignores the performance of negative classes, which may lead to a bias towards a certain class in identification.

[0038] Recall is the proportion of all actual positive samples correctly identified as positive by the model. The formula is: ; Recall measures the model's ability to identify positive classes 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, some of which may be incorrect identification results.

[0039] The F1-score is the harmonic mean of precision and recall and is defined as: ; The F1-score comprehensively considers precision and recall and is suitable for cases of class imbalance. Although it balances precision and recall, it assigns the same weight to both, and may not be able to be flexibly adjusted to meet the requirements of different tasks.

[0040] After training is completed, the obtained optimal internal wave fringe recognition model can efficiently automatically identify the internal solitary wave region from new sea surface height data and output its position and morphological characteristics, providing important data support for subsequent internal wave amplitude extraction.

[0041] Step 5: Apply the optimal internal wave fringe recognition model obtained in Step 4 to new ocean model data to achieve automatic identification and position extraction of internal solitary waves.

[0042] Step 6, extract sea surface height (SSH) data and the corresponding 400-meter water depth temperature field data respectively, and process and organize them for use in subsequent training of the temperature field inversion model. These data need to cover a sufficiently long time range and a wide spatial area to ensure the representativeness of the data and the diversity of training. The extracted sea surface height data and temperature data will undergo preprocessing operations such as cropping and rotation. These processes help remove irrelevant regions and ensure that the dataset only contains useful regional data. In addition, to enhance the robustness and generalization ability of the model, all data will be randomly processed through data augmentation methods, such as random rotation, scaling, cropping, etc. Then, all the sea surface height data and temperature data will be randomly shuffled. We will divide the dataset into a training dataset and a validation dataset according to a certain ratio. Among them, the training set accounts for 80% and the validation set accounts for 20%.

[0043] Subsequently, the data will be input into the VAE model for training. The VAE model structure is as Figure 5 shown. During this process, the sea surface height data and the temperature field data at a depth of 400 meters will be used as input features, and the VAE model will learn the latent space distribution of these data and then achieve the inversion of the temperature field. In this way, we provide a sufficient, preprocessed and enhanced dataset for the temperature field inversion model to improve the training effect and prediction ability of the model.

[0044] Step 7, input the training set obtained in Step 6 into the variational autoencoder algorithm for training. The variational autoencoder is a deep learning generative model mainly used to learn the probability distribution of the latent space from data and generate 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.

[0045] The architecture of the VAE includes an encoder and a decoder. The network structure of the variational autoencoder (VAE) is a deep generative model, which is an extension and improvement based on the traditional autoencoder. Its core structure includes three parts: an encoder, a reparameterization trick, and a decoder. After receiving the input data, the encoder extracts features through multiple hidden layers (such as fully connected layers or multiple convolutional layers) and outputs two vectors: the mean vector μ and the variance vector σ, which describe the probability distribution of the input data in the latent space. Different from the autoencoder that directly outputs a low-dimensional feature vector, the VAE introduces randomness through the reparameterization trick, using μ and σ to sample the latent variable z from the standard normal distribution. The decoder part 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), and generates the reconstructed data through the output layer.

[0046] During the training process, the variational lower bound (ELBO, Evidence Lower Bound) is used as the loss function. The goal of the VAE is to maximize the lower bound so that the generated temperature field is as close as possible to the true value. The specific calculation formula of the ELBO is as follows: .

[0047] The ELBO consists of two parts: one is the reconstruction error, which is the difference between the generated temperature field and the true temperature field; the other is the KL divergence, which measures the difference between the distribution of the latent variables and the prior distribution and plays a regularization role to prevent the model from overfitting. During the training process, we adjust the weights and parameters of the model according to the training dataset to minimize the loss function. Specifically, stochastic gradient descent is used to optimize the model and gradually update the parameters of the network. During training, batch processing is used to accelerate training, and multiple iterations are performed until the loss converges.

[0048] After each round of training, the test set is used to evaluate the model and check the accuracy of the generated temperature field. The evaluation metrics usually include the mean squared error (MSE), mean absolute error (MAE), etc. Among them, the specific calculation formula of the MSE is as follows: ; The specific calculation formula of the MAE is as follows: .

[0049] Through the performance analysis of the test set, hyperparameters (such as the learning rate, latent space dimension, etc.) are adjusted to obtain the best training effect. Finally, an optimal temperature field inversion model is obtained after training. This model can generate the temperature field at a depth of 400 meters according to the input sea surface height data. Through this model, the underwater temperature distribution can be inferred from the sea surface height data, providing accurate temperature field data support for the subsequent extraction of the amplitude of internal solitary waves.

[0050] Step 8: Use the internal wave identification result obtained in Step 5 as the input, apply the temperature field inversion model trained in Step 7, and extract the amplitude of the internal solitary wave from the inverted temperature field.

[0051] First, use the internal wave identification result obtained in Step 5 to extract the position information of each internal solitary wave. Next, use the identified position of the internal solitary wave as the input and feed it into the variational autoencoder model trained in Step 7. Through the model, the corresponding sea surface height data can be mapped into temperature field data.

[0052] To extract the amplitude, the region where the internal solitary wave is located is first calibrated in the inverted temperature field. By combining the position and morphology information in the internal wave identification results, the position range of the internal solitary wave in the temperature field can be determined. Then, the amplitude of the internal solitary wave is obtained by calculating the maximum vertical change of the temperature field in this region. Specifically, the amplitude can be estimated by the difference between the local maximum and minimum values of the temperature field.

[0053] Example 2 The actual verification is carried out based on the method of Example 1. This example selects the sea surface height data of 12:00 CST (UTC+8) for three days from March 22, 2014 to March 24, 2014. In the process of internal solitary wave identification, in order to facilitate the visualization of the results, the identification results are mapped onto the unprocessed sea surface height data map, and the original data map and the identification results are displayed at the same time.

[0054] Among them, the automatic identification and position extraction of internal solitary waves, the identification results are demonstrated as follows Figure 8 As 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, 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, and providing strong support for marine environmental monitoring and related research.

[0055] Amplitude extraction demonstration Figure 9 As shown in the figure, two internal waves at different locations were randomly selected for amplitude extraction within the target time. In the isotherm analysis, obvious internal solitary wave phenomenon can be observed, which is mainly manifested in that the isotherms show obvious bending or fluctuation in some areas. In order to extract the amplitude of the internal solitary wave, the method is to select the area where the isotherm fluctuates most violently and calculate the difference between the maximum amplitude of the area and the maximum value of the isotherm within a certain range to the west. This range will be extracted from the internal wave identification results.

[0056] Specifically, the maximum amplitude corresponds to the deepest point of the isotherm fluctuation, while the maximum value of the isotherm on the west side 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 intensity of the internal solitary wave and its impact on the underwater temperature field.

[0057] Finally, the extracted internal solitary wave amplitude is saved and recorded together with other analysis results to support further research and data visualization. This amplitude extraction process provides crucial data support for in-depth understanding of the dynamic characteristics of internal solitary waves and their relationships with environmental factors, thus promoting the scientific research and application of internal solitary waves.

[0058] The present invention is a method for quickly identifying internal solitary waves and extracting amplitudes in refined model data based on deep learning. Combining the Mask-RCNN object detection model and variational autoencoder, it can quickly and automatically identify the positions and shapes of internal solitary waves in refined model data and accurately invert their amplitudes.

[0059] The specific embodiments described above further elaborate on the objectives, technical solutions, and beneficial effects of the present disclosure. It should be understood that the above are only specific embodiments of the present disclosure and are not intended to limit the present disclosure. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present disclosure shall be included within the protection scope of the present disclosure.

Claims

1. A method for identifying internal solitary waves and amplitude inversion in pattern data based on machine learning, characterized in that, It includes the following steps: S1: Extract the sea surface height data at the required positions from the pattern data, and perform preprocessing and imaging; S2: Perform internal wave fringe annotation on the image obtained in S1 to construct a dataset; S3: Process and correct the internal wave fringe annotation data in S2; S4: Construct a Mask-RCNN image segmentation model. The Mask R-CNN divides object detection into two parts: region generation and classification based on Faster R-CNN, 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 positions of the feature map and the original image are accurately aligned through bilinear interpolation; S5: Train 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 identify the internal waves in the pattern data, and record the output internal wave positions and shapes; S7: Extract the temperature field data from the pattern data to construct a dataset; S8: Train the variational autoencoder algorithm VAE, perform performance analysis on the model with the obtained optimal parameters, and obtain a trained temperature field inversion model; S9: Use the internal wave recognition results obtained in S6 as the input of the temperature field inversion model, and extract 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 amplitude inversion in pattern data according to claim 1, wherein S1 includes: S1-1: The preprocessing includes interpolating the obtained sea surface height and filling in blank areas; S1-2: Image the data through the processing method of local binary pattern (LBP), and then perform unified resolution adjustment to uniformly convert the data into sub-images with a resolution of 1080*1080.

3. The method for identifying internal solitary waves and amplitude inversion in the mode data according to claim 1, characterized in that, S3 includes: S3-1: Extract the annotations made in S2; S3-2: Manually correct the problems of adhesion or incompleteness of multiple internal solitary waves in the annotations; S3-3: Divide the dataset into a training dataset and a validation dataset.

4. The method for identifying internal solitary waves and amplitude inversion in the 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 amplitude inversion in pattern data according to claim 1, wherein In S4, the process of constructing the FPN includes three main steps: First, through the bottom-up feature extraction network, the input image is preprocessed and then sent into the pre-trained feature extraction network to gradually extract features from the bottom layer to the high layer, generating feature maps of different levels and capturing the multi-level information of the image. Second, a top-down feature fusion strategy is adopted to gradually fuse the deep features into the shallow features. Specifically, first, the shallow features are dimensionally reduced using 1×1 convolution, then element-wise added to the deep features after upsampling, and finally, 3×3 convolution is applied to further integrate the multi-scale information. Finally, region generation and object detection are performed on the generated multi-scale feature maps. The 4th, 5th, and 6th layers of the feature maps respectively perform the Region Proposal Network (RPN) operation. The RPN uses the sliding window mechanism to generate multi-scale anchor boxes, calculates the object scores and coordinate offsets through parallel classification and regression branches, and combines non-maximum suppression to generate high-quality candidate regions.

6. The method for identifying internal solitary waves and amplitude inversion in pattern data according to claim 1, wherein S7 includes: S7-1: Extract the sea surface height data and the corresponding water depth temperature field data from the pattern data respectively; S7-2: Interpolate all the sea surface height data and temperature data, preprocess to fill in the blank areas, and then randomly shuffle after cropping and rotating; S7-3: Divide the data set into a training set and a test set.

7. The method for identifying internal solitary waves and amplitude inversion in pattern data according to claim 1, wherein S8 includes: S8-1: The architecture of the VAE includes an encoder and a decoder. After receiving the input data, the encoder extracts features through multiple hidden layers and outputs two vectors: the mean vector μ and the variance vector σ, which describe the probability distribution of the input data in the latent space. The decoder part receives the latent variable z, gradually restores it to a structure similar to the input data through a series of hidden layers, and generates the 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 the VAE is to maximize the lower bound. The specific calculation formula of the ELBO is: ； The ELBO consists of two parts: one is the reconstruction error, that is, the difference between the generated temperature field and the real temperature field; the other is the KL divergence; S8-3: Adjust the weights and parameters of the model according to the training data set to minimize the loss function. Specifically, stochastic gradient descent is used to optimize the model and gradually update the parameters of the network; during training, batch processing is used to accelerate the training, and multiple iterations are performed until the loss converges.

Images