A fast reconstruction method of regional marine environment field based on ViT neural network

Abstract

The present invention discloses a method for rapid reconstruction of a regional marine environment field based on a ViT neural network, belonging to the field of environmental field reconstruction technology. The method comprises acquiring and preprocessing multi-source marine data, calculating the density field from the preprocessed multi-source marine data, and constructing a multi-factor coupling field based on the temperature field, salinity field, and density field; preparing a deep learning sample set, constructing and training a ViT environmental field generation model, and using the trained ViT environmental field generation model to reconstruct the regional marine environment field. The method comprehensively considers the coupling of multiple marine factor fields and achieves the reconstruction of a three-dimensional environmental field of regional ocean temperature, salinity, and density. The method is accurate, high-resolution, and rapid, and can play a role in data support and application assurance in fields such as marine disaster prevention and mitigation, marine ecosystems, and marine climate change.

Description

Technical Field

[0001] The invention discloses a method for quickly reconstructing a regional ocean environment field based on a ViT neural network, and belongs to the technical field of environment field reconstruction. Background Art

[0002] Traditional ocean environmental field reconstruction methods primarily rely on data interpolation and assimilation techniques based on statistical principles or variational methods, such as stepwise correction, kriging interpolation, optimal interpolation, and variational assimilation. These methods typically achieve good results when ocean observation data are abundant and evenly distributed. However, a notable characteristic of ocean observation data is its small total volume and uneven temporal and spatial distribution. This is particularly true below the ocean surface, where remote sensing data is scarce. Acquisition of ocean observation data relies on a limited number of in-situ observation devices, such as buoys, Argo floats, underwater gliders, and underwater drones. Under these circumstances, the effectiveness of traditional ocean environmental field reconstruction methods is uncertain, resulting in large reconstruction errors and low reconstruction resolution. Consequently, they struggle to reflect submesoscale and small-scale ocean dynamic processes, hindering the observation of important ocean phenomena such as submesoscale eddies, fronts, internal waves, internal tides, and typhoons. Summary of the Invention

[0003] The purpose of the present invention is to provide a method for rapid reconstruction of regional ocean environment fields based on ViT neural network, so as to solve the problem of inaccurate results of ocean environment field reconstruction in the prior art.

[0004] A rapid reconstruction method for regional marine environmental fields based on the ViT neural network includes obtaining multi-source ocean data and preprocessing them, calculating the density field from the preprocessed multi-source ocean data, and constructing a multi-factor coupling field based on the temperature field, salinity field, and density field; producing a deep learning sample set, constructing a training ViT environmental field generation model, and using the trained ViT environmental field generation model to reconstruct the regional marine environmental field.

[0005] The elements of multi-source ocean data include temperature data, salinity data and density data.

[0006] Types of multi-source ocean data include observational data, reanalysis data, and numerical simulation data.

[0007] Preprocessing includes unified data format, depth correction, spatiotemporal interpolation and quality control.

[0008] The deep learning sample set includes training set, validation set and test set.

[0009] The ViT environment field generation model includes a generator and a discriminator.

[0010] The input of the generator is the reconstructed field position information, random noise and multiple observation values. The random noise is input into the fully connected layer to generate a noise vector. The multiple observation values ​​are input into multiple multi-layer perceptrons to generate multiple observation feature vectors. These are then input into the Transformer encoder together with the noise vector. The reconstructed field position information is input into the multi-layer perceptron to generate a reconstructed field position vector.

[0011] The Transformer encoder outputs an environmental field feature vector equal to the number of observation values, adds a reconstructed field position vector to each environmental field feature vector, and then inputs them into multiple multi-layer perceptrons to generate a reconstructed environmental field block. The reconstructed environmental field block is a multi-layer structure.

[0012] The input of the discriminator is the reconstructed environment field block. Each element in the reconstructed environment field block is input into multiple multi-layer perceptrons respectively. The reconstructed field position vector is added to the output of the multi-layer perceptron, and then input into the Transformer encoder. Finally, it passes through a multi-layer perceptron for true or false discrimination.

[0013] The Transformer encoder has a 6-layer structure, each layer includes the first normalization layer, the multi-head self-attention layer, the first feature fusion layer, the second normalization layer, the multi-layer perceptron and the second feature fusion layer. The first normalization layer is quickly connected to the first feature fusion layer, and the second normalization layer is quickly connected to the second feature fusion layer.

[0014] The regional ocean environment reconstruction using the trained ViT environment field generation model includes:

[0015] S1 designs the loss function and constructs the loss function Loss according to the difference between the original ocean environment field and the reconstructed ocean environment field:

[0016]

[0017] Where N is the number of training samples, i is the i-th training sample, y is a binary label, and the binary label is 0 or 1. i is the i-th binary label, p(y i ) is the output belonging to y i The probability of the label;

[0018] S2 alternately iteratively trains the neural network. In the alternating training, the generator is first fixed to train the discriminator, and then the discriminator is fixed to train the generator, and this cycle continues until the training converges.

[0019] After the S3 neural network training is completed, the discriminator is removed and the generator is retained. When performing true and false discrimination, the observation value of the ocean environment field is input into the generator, and the output is the reconstructed ocean environment field;

[0020] The observed values ​​of the ocean environment field are any one or a combination of temperature data, salinity data and density data.

[0021] Compared with the existing technology, the present invention has the following beneficial effects: it comprehensively considers the coupling of multiple ocean factors, realizes the reconstruction of the three-dimensional environmental field of regional ocean temperature, salinity and density, and has the characteristics of accuracy, high resolution and speed. It can play a role in data support and application guarantee in the fields of marine disaster prevention and mitigation, marine ecosystems, marine climate change, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a technical flow chart of the present invention;

[0023] Figure 2 This is a structural diagram of the ViT environment field generation model of the present invention;

[0024] Figure 3 It is a structural diagram of the Transformer encoder of the present invention. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely below. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] A rapid reconstruction method for regional marine environmental fields based on the ViT neural network includes obtaining and preprocessing multi-source marine data, calculating the density field from the preprocessed multi-source marine data, and constructing a multi-factor coupling field based on the density field; preparing a deep learning sample set, constructing and training a ViT environmental field generation model, and using the trained ViT environmental field generation model to reconstruct the regional marine environmental field.

[0027] The elements of multi-source ocean data include temperature data, salinity data and density data.

[0028] Types of multi-source ocean data include observational data, reanalysis data, and numerical simulation data.

[0029] Preprocessing includes unified data format, depth correction, spatiotemporal interpolation and quality control.

[0030] The deep learning sample set includes training set, validation set and test set.

[0031] The ViT environment field generation model includes a generator and a discriminator.

[0032] The input of the generator is the reconstructed field position information, random noise and multiple observation values. The random noise is input into the fully connected layer to generate a noise vector. The multiple observation values ​​are input into multiple multi-layer perceptrons to generate multiple observation feature vectors. These are then input into the Transformer encoder together with the noise vector. The reconstructed field position information is input into the multi-layer perceptron to generate a reconstructed field position vector.

[0033] The Transformer encoder outputs an environmental field feature vector equal to the number of observation values, adds a reconstructed field position vector to each environmental field feature vector, and then inputs them into multiple multi-layer perceptrons to generate a reconstructed environmental field block. The reconstructed environmental field block is a multi-layer structure.

[0034] The input of the discriminator is the reconstructed environment field block. Each element in the reconstructed environment field block is input into multiple multi-layer perceptrons respectively. The reconstructed field position vector is added to the output of the multi-layer perceptron, and then input into the Transformer encoder. Finally, it passes through a multi-layer perceptron for true or false discrimination.

[0035] The Transformer encoder has a 6-layer structure, each layer includes the first normalization layer, the multi-head self-attention layer, the first feature fusion layer, the second normalization layer, the multi-layer perceptron and the second feature fusion layer. The first normalization layer is quickly connected to the first feature fusion layer, and the second normalization layer is quickly connected to the second feature fusion layer.

[0036] The regional ocean environment reconstruction using the trained ViT environment field generation model includes:

[0037] S1 designs the loss function and constructs the loss function Loss according to the difference between the original ocean environment field and the reconstructed ocean environment field:

[0038]

[0039] Where N is the number of training samples, i is the i-th training sample, y is a binary label, and the binary label is 0 or 1. i is the i-th binary label, p(y i ) is the output belonging to y i The probability of the label;

[0040] S2 alternately iteratively trains the neural network. In the alternating training, the generator is first fixed to train the discriminator, and then the discriminator is fixed to train the generator, and this cycle continues until the training converges.

[0041] After the S3 neural network training is completed, the discriminator is removed and the generator is retained. When performing true and false discrimination, the observation value of the ocean environment field is input into the generator, and the output is the reconstructed ocean environment field;

[0042] The observed values ​​of the ocean environment field are any one or a combination of temperature data, salinity data and density data.

[0043] The technical process of the present invention is as follows Figure 1 As shown, it includes obtaining multi-source ocean temperature and salinity data, preprocessing ocean temperature and salinity data, constructing a multi-factor coupling field, making a deep learning sample set, constructing and training a ViT environment field generation model, and reconstructing the regional ocean environment field.

[0044] The ViT environment field generation model structure of the present invention is as follows Figure 2 As shown, it includes a generator and a discriminator. The input of the generator is the reconstructed field position information, random noise and multiple observations. The random noise is input into the fully connected layer to generate a noise vector. The multiple observations are input into multiple multi-layer perceptrons to generate multiple observation feature vectors. The observations are then input into the Transformer encoder together with the noise vector. The reconstructed field position information is input into the multi-layer perceptron to generate a reconstructed field position vector. The Transformer encoder outputs an environment field feature vector equal to the number of observations. The reconstructed field position vector is added to each environment field feature vector, and then input into multiple multi-layer perceptrons to generate a reconstructed environment field block. The reconstructed environment field block has a multi-layer structure. The input of the discriminator is the reconstructed environment field block. Each element in the reconstructed environment field block is input into multiple multi-layer perceptrons. The reconstructed field position vector is added to the output of the multi-layer perceptron, and then input into the Transformer encoder. Finally, it passes through a multi-layer perceptron to perform true or false discrimination.

[0045] The Transformer encoder structure of the present invention is as follows Figure 3 As shown in the figure, the Transformer encoder has a 6-layer structure, each layer includes the first normalization layer, the multi-head self-attention layer, the first feature fusion layer, the second normalization layer, the multi-layer perceptron and the second feature fusion layer in sequence. The first normalization layer is quickly connected to the first feature fusion layer, and the second normalization layer is quickly connected to the second feature fusion layer.

[0046] In this invention, observation data includes in-situ observation data from buoys, submersibles, underwater gliders, wave gliders, Argo floats, and GHRSST and SMAP-SSS satellite remote sensing data products; reanalysis data includes GOFS 3.1, SODA, and GODAS reanalysis data; and numerical simulation data includes HYCOM and OFES data. Data preprocessing involves standardizing data formats, units, precision, and data types, including latitude and longitude formats, the selection and format of standard time, the units and precision of physical quantities such as temperature, salinity, and depth, and the data types of various data types stored in computers. Specifically, first, data standards are established based on international standards, industry practices, and specific project requirements. Second, programs are developed for raw observation data from various sources to enable automated data processing and batch processing in large-scale data scenarios. Data correction and calculation of indirect physical quantities include depth correction and density calculation. Specifically, first, a unified process and algorithm are developed to correct ocean pressure sensor data using atmospheric pressure and calculate depth; second, density is calculated using temperature, salinity, and depth. Data cleaning and quality control involve spatiotemporal interpolation, redundancy identification, and outlier detection for heterogeneous, multi-source ocean data. This approach aims to supplement missing values, remove duplicates, flag outliers, and eliminate erroneous values, resulting in high-quality and consistent data. Specifically, the team programed automatic longitudinal interpolation to compensate for missing data at certain depths; developed algorithms to identify data redundancy caused by repeated counting and other factors, and removed duplicate data; and used statistical principles to identify outliers and manually set thresholds to remove clearly erroneous data and flag outliers.

[0047] For a given moment in time, let T, S, and D represent the temperature, salinity, and density fields, respectively. T, S, and D are three-dimensional matrices of equal size, d × x × y, where the first dimension, d, represents depth, the second dimension, x, represents longitude, and the third dimension, y, represents latitude. The temperature, salinity, and density fields are concatenated to construct a multi-element coupled field, C, where C = [T, S, D]. C is a four-dimensional matrix of n × d × x × y, where the first dimension, n, represents the coupled element, the second dimension, d, represents depth, the third dimension, x, represents longitude, and the fourth dimension, y, represents latitude.

[0048] In terms of time, data from 2012 to 2022 were selected to improve the temporal diversity of the samples; in terms of spatial scale, sea areas of different sizes, such as 100km×100km, 200km×200km, and 300km×300km, were selected at a ratio of 1:1:1 to improve the spatial scale diversity of the samples; in terms of quantity, a total of 100,000 samples were produced to improve the richness of the samples; in terms of sample set division, the training set, validation set, and test set were divided at a ratio of 7:2:1, that is, the number of samples in the training set, validation set, and test set were 70,000, 20,000, and 10,000, respectively.

[0049] The ViT environmental field generation model is designed based on the generator-discriminator architecture of a generative adversarial network. The generator input is the environmental field observation o, a 5×n matrix where the first dimension represents the temperature, salinity, density, abscissa, and ordinate of a point, respectively, and n is the number of observation points. The input variable is mapped into an observation feature vector x of length 512 via an embedding layer, where x = MLP_1(o). The Transformer encoder module receives the observation feature vector x and a noise vector w as input, where w = FC(z). It outputs 512 reconstructed environmental field feature vectors y, where y = TransformerEncoder_1(x,w). The noise vector w is obtained by mapping randomly sampled normally distributed Gaussian noise z through a fully connected network. At the same time, the reconstructed field position information p0 is encoded in a one-hot codding manner and mapped to 512 coordinate feature vectors p through the MLP multilayer perceptron, p = MLP_2(p0), and the reconstructed environment field feature vector y output by the Transformer encoder module is connected with the coordinate feature vector p to obtain the reconstructed environment field feature vector y′, that is, y′ = [y, p], thereby explicitly emphasizing the reconstructed field position feature. The reconstructed environment field feature vector y′ is then passed through the multilayer perceptron to obtain 512 reconstructed environment field blocks b, b = MLP_3(y′), and the environment field blocks are mapped to Figure 2The discriminator takes the dimensional concatenation c of the original environment field a and the reconstructed environment field b as input, i.e., c = [a, b]. This input field is spatially divided into 8×8×8 blocks, totaling 512 blocks. A multilayer perceptron (MLP) then flattens the blocks into one-dimensional vectors d, d = MLP_4(c). This vector, along with the coordinate feature vector p, is fed into the Transformer encoder module. The output of the Transformer encoder module is mapped to a true / false binary judgment through the MLP, resulting in a true / false binary classification result R, R = MLP_5(TransformerEncoder_2(d′)). The Transformer encoder module consists of a sample normalization layer, a multi-head self-attention layer, a sample normalization layer, and a multilayer perceptron, forming a block with two skip connections. This block is repeated six times to form the Transformer encoder module.

[0050] The discriminator is trained using gradient ascent, while the generator is trained using gradient descent. The discriminator is trained first, and then the generator and discriminator are trained alternately. This avoids large performance differences between the generator and discriminator, which can lead to poor training results. Neural network training is achieved by calculating the partial derivatives of the loss function with respect to specific neural network parameters and backpropagating the gradient.

[0051] The present invention also designs a value function, which is constructed based on real environment field samples and reconstructed environment fields as an optimization criterion in neural network supervised learning. The value function is denoted as V(D,G), and the optimization goal is as follows:

[0052]

[0053] Among them, a is the real environment field sample, represents the expectation of random sampling in the real environment field sample space, o is the observed sample, represents the expectation of random sampling in the observation sample space, D(·) represents the discriminator neural network function, and G(·) represents the generator neural network function.

[0054] Training the neural network specifically includes: T1 sampling m observation value samples o, sampling m real environment field samples a.

[0055] T2 calculates the gradient of the discriminator parameters and updates the discriminator using the gradient ascent method, namely:

[0056]

[0057] Where θ d Represents the parameters of the discriminator neural network.

[0058] T3 repeats steps T1-T2 until the discriminator effect improves and the training converges.

[0059] T4 samples m observation samples o.

[0060] T5 calculates the gradient of the generator parameters and updates the generator using the gradient descent method, namely:

[0061]

[0062] Where θ g Represents the parameters of the generator neural network.

[0063] T6 Repeat steps T4-T5 until the generator effect improves and the training converges.

[0064] T7 repeats steps T1-T6, so that the discriminator network and the generator network are trained alternately and improved synchronously until the training of the neural network converges.

[0065] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for rapid reconstruction of regional marine environment fields based on ViT neural network, characterized by: This includes acquiring and preprocessing multi-source ocean data, calculating the density field from the preprocessed multi-source ocean data, and constructing a multi-factor coupling field based on the temperature field, salinity field, and density field; preparing a deep learning sample set, building and training a ViT environmental field generation model, and using the trained ViT environmental field generation model to reconstruct the regional ocean environmental field; The ViT environment field generation model includes a generator and a discriminator; The input of the generator is the reconstructed field position information, random noise and multiple observation values. The random noise is input into the fully connected layer to generate a noise vector. The multiple observation values ​​are input into multiple multi-layer perceptrons to generate multiple observation feature vectors. These are then input into the Transformer encoder together with the noise vector. The reconstructed field position information is input into the multi-layer perceptron to generate a reconstructed field position vector. The Transformer encoder outputs an environmental field feature vector equal to the number of observations, adds a reconstruction field position vector to each environmental field feature vector, and then inputs them into multiple multi-layer perceptrons to generate a reconstructed environmental field block. The reconstructed environmental field block is a multi-layer structure. The input of the discriminator is the reconstructed environment field block. Each element in the reconstructed environment field block is input into multiple multi-layer perceptrons respectively. The reconstructed field position vector is added to the output of the multi-layer perceptron, and then input into the Transformer encoder. Finally, it passes through a multi-layer perceptron to perform true or false discrimination; The Transformer encoder has a 6-layer structure. Each layer includes the first normalization layer, the multi-head self-attention layer, the first feature fusion layer, the second normalization layer, the multi-layer perceptron, and the second feature fusion layer. The first normalization layer is connected to the first feature fusion layer, and the second normalization layer is connected to the second feature fusion layer. The regional ocean environment reconstruction using the trained ViT environment field generation model includes: S1 designs the loss function and constructs the loss function Loss according to the difference between the original ocean environment field and the reconstructed ocean environment field: Where N is the number of training samples, i is the i-th training sample, y is a binary label, and the binary label is 0 or 1. i is the i-th binary label, p(y i ) is the output belonging to y i The probability of the label; S2 alternately iteratively trains the neural network. In the alternating training, the generator is first fixed to train the discriminator, and then the discriminator is fixed to train the generator, and this cycle continues until the training converges. After the S3 neural network training is completed, the discriminator is removed and the generator is retained. When performing true and false discrimination, the observation value of the ocean environment field is input into the generator, and the output is the reconstructed ocean environment field; The observed values ​​of the ocean environment field are any one or a combination of temperature data, salinity data and density data.

2. The method for rapid reconstruction of regional marine environment field based on ViT neural network according to claim 1 is characterized in that: The elements of multi-source ocean data include temperature data, salinity data and density data.

3. The method for rapid reconstruction of regional marine environment field based on ViT neural network according to claim 2 is characterized in that: Types of multi-source ocean data include observational data, reanalysis data, and numerical simulation data.

4. The method for rapid reconstruction of regional marine environment field based on ViT neural network according to claim 3 is characterized in that: Preprocessing includes unified data format, depth correction, spatiotemporal interpolation and quality control.

5. The method for rapid reconstruction of regional marine environment field based on ViT neural network according to claim 4 is characterized in that: The deep learning sample set includes training set, validation set and test set.

Images