Inversion method and system of submarine topography

Abstract

The invention discloses a submarine topography inversion method and system, and relates to the technical field of submarine topography inversion. The method comprises the following steps: acquiring ocean gravity data, and preprocessing the ocean gravity data; constructing a submarine topography inversion model, dividing the preprocessed ocean gravity data into a training set and a verification set, performing submarine topography inversion training on the submarine topography inversion model by using the training set, verifying the trained submarine topography inversion model by using the verification set, and obtaining a trained submarine topography inversion model after verification is passed; and inputting the preprocessed ocean gravity prediction data into the trained submarine topography inversion model for submarine topography inversion, and outputting an inverted submarine topography result. Through the method provided by the invention, the influence of the nonlinear term error of the region with relatively large submarine topography fluctuation can be effectively reduced, so that the precision of the submarine topography obtained by inversion is improved, and the requirement on the calculation performance is reduced to improve the operation efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of seabed topography inversion, and in particular to a seabed topography inversion method and system. Background Art

[0002] The ocean covers 71% of the Earth's surface area. The study of seabed topography provides a scientific basis for the development of marine resources, ocean circulation, and national defense and military, and lays the foundation for the sustainable development of the ocean.

[0003] At present, among the methods of inverting seafloor topography using ocean gravity data, the common traditional methods are Gravity Geologic Method (GGM) and Smith and Sandwell Method (SAS). GGM separates long-wave and short-wave gravity anomalies from observed gravity anomalies and uses short-wave gravity anomalies to invert seafloor topography. SAS method uses the high correlation between water depth and gravity anomaly in a certain band and uses linear regression to determine the proportional relationship between water depth and gravity anomaly.

[0004] Both the traditional GGM and SAS methods rely on linear approximate calculations to invert the seabed topography from ocean gravity data. However, linear approximate calculations introduce errors in the nonlinear terms of the seabed topography. The impact of nonlinear term errors is greater in areas with large seabed topography fluctuations, resulting in lower accuracy of the inverted seabed topography. Summary of the invention

[0005] In order to overcome the problem that the existing seabed topography inversion technology has a greater impact on nonlinear errors in areas with large seabed topography fluctuations, resulting in lower accuracy of the inverted seabed topography, the present invention proposes a seabed topography inversion method and system, which can effectively reduce the impact of nonlinear errors in areas with large seabed topography fluctuations, thereby improving the accuracy of the inverted seabed topography and reducing the requirements for computing performance to improve computing efficiency.

[0006] To achieve the purpose of the present invention, the present invention adopts the following technical solutions:

[0007] A method for inverting seabed topography, the method comprising the following steps:

[0008] Acquiring ocean gravity data and preprocessing the ocean gravity data;

[0009] Construct a seabed topography inversion model, and divide the preprocessed ocean gravity data into a training set and a verification set. Use the training set to train the seabed topography inversion model, and use the verification set to verify the trained seabed topography inversion model. After verification, a trained seabed topography inversion model is obtained.

[0010] The preprocessed ocean gravity prediction data is input into the trained seabed topography inversion model for seabed topography inversion, and the inverted seabed topography results are output.

[0011] In the above technical scheme, by preprocessing the collected ocean gravity data, the practicability and reliability of the ocean gravity data can be effectively improved; the constructed seabed topography inversion model can obtain a seabed topography inversion model that effectively grasps the changes in ocean gravity data after seabed topography inversion training of the training set and verification of the verification set, so that seabed topography inversion can be performed through the seabed topography inversion model, which can effectively reduce the influence of nonlinear term errors in areas with large seabed topography fluctuations, thereby improving the accuracy of the seabed topography obtained by inversion, and reducing the requirements for computing performance to improve computing efficiency.

[0012] Furthermore, the process of preprocessing the ocean gravity data includes:

[0013] The acquired ocean gravity data include gravity anomaly data, vertical gravity gradient data and vertical deviation data;

[0014] The gravity anomaly data is gridded, and the gravity geological method is used to calculate the short-wave gravity anomaly data after gridding in the gravity anomaly data. The expression is:

[0015]

[0016] The SAS method is used to perform bandpass filtering on the vertical gravity gradient data in the ocean gravity data. The expression is:

[0017]

[0018] Normalize the shortwave gravity anomaly data, vertical gravity gradient data and vertical deviation data;

[0019] Among them, SAS method represents Smith And Sandwell method, i is the position of the grid point, j is the position of the control point, E(j) is the water depth at the control point, D represents the reference water depth, Δρ is the density difference between the seafloor crust and seawater, and G is the gravitational constant. obs (i) is the gridded sea surface gravity anomaly, g long (i) is the long-wave gravity anomaly of the control point grid, g short (i) is the gridded short-wave gravity anomaly. k represents the wave number, W lowpass represents Gaussian low-pass filtering, W bandpass represents a bandpass filter composed of a Gaussian high-pass filter and a Wiener low-pass filter, z0 represents the average water depth, and λ1 and λ2 represent the cutoff wavelength.

[0020] In the above technical scheme, the gravity anomaly data in the acquired ocean gravity data is gridded, which can effectively simplify the process of extracting short-wave gravity anomaly data, thereby reducing the requirements for computer performance; the vertical gravity gradient data in the ocean gravity data is bandpass filtered by the SAS method, which can effectively reduce the random noise and low-frequency system errors existing in the original vertical gravity gradient data; the short-wave gravity anomaly data, vertical gravity gradient data and vertical deviation data are normalized, which can effectively improve the sensitivity of the model to the data during the model training process, so as to improve the performance of the trained model, thereby improving the accuracy and inversion efficiency of the seabed topography obtained by inversion, and reducing the requirements for computing performance to improve the computing efficiency. The acquired ocean gravity data includes gravity anomaly data, vertical gravity gradient data and vertical deviation data, which can also effectively solve the problem of information omissions in a single data source.

[0021] Furthermore, the process of constructing the seafloor topography inversion model includes:

[0022] A back propagation neural network is used to construct a seabed topography inversion model, wherein the seabed topography inversion model includes a plurality of input layers, a plurality of hidden layers and an output layer;

[0023] The constructed seafloor topography inversion model expression is:

[0024]

[0025] Among them, (ψ,λ) represents the longitude and latitude of the sea depth at the prediction point, represents the weights and biases between hidden layers, b (3) represents the weight and bias between the hidden layer and the output layer, n1 and n2 represent the number of nodes in the first and second layers of the hidden layer, f represents the tansig activation function, g represents the purelin activation function, and h represents the output of the first hidden layer.

[0026] In the above technical scheme, a back-propagation neural network is used to construct a seabed topography inversion model, which can effectively restore the nonlinear information of the seabed topography ignored by traditional inversion methods. Several input layers, several hidden layers and output layers are set for the constructed seabed topography inversion model, which can improve the accuracy of neural network training in the model and effectively reduce the cost of model training.

[0027] Furthermore, the process of using the training set to perform seabed topography inversion training on the seabed topography inversion model includes:

[0028] Setting several rounds of iterative training to perform seabed topography inversion training on the seabed topography inversion model;

[0029] According to the ocean gravity data in the training set, the weights and thresholds between the input layer, hidden layer and output layer are initialized, and the particle swarm algorithm is used to iteratively optimize the initialized weights and thresholds, and the optimized weights and thresholds are output;

[0030] Set the loss function to optimize the weights and thresholds after each round of iterative optimization, and use the validation set to verify the parameters after each round of iterative optimization. When the set iterative training round ends or the loss function converges, the optimal weights and thresholds are obtained;

[0031] The obtained optimal weights and optimal thresholds are set as new parameters between the input layer, hidden layer and output layer in the seabed topography inversion model to obtain a trained seabed topography inversion model.

[0032] Furthermore, the process of iteratively optimizing the initialized weights and thresholds using the particle swarm algorithm includes:

[0033] According to the training set data received by the input layer, the particle population is set and the parameters of the particle population are initialized;

[0034] Calculate the spatial dimension of particles based on the particle population after parameter initialization;

[0035] Calculate the fitness of particles according to the particle population after parameter initialization, and extract the individual best fitness eigenvalue and the global best fitness eigenvalue;

[0036] According to the calculated particle fitness, the best position of the individual particle is calculated, and according to the best position of the individual particle, the global best position of the particle group is calculated;

[0037] According to the best position of the individual particle and the global best position of the particle group, the position and speed of the particle are updated, and the fitness of the particle after the updated position and speed is calculated to confirm whether the fitness of the particle after the updated position and speed is better than the global best fitness. If it is better, the fitness of the current particle is taken as the global best fitness and iterates until the maximum number of updates. If it is not better, the global best fitness is not changed and iterates until the maximum number of updates.

[0038] Among them, the parameters of particle population include quantity, speed, position and update times.

[0039] Furthermore, the expression for calculating the spatial dimension of particles based on the initialized particle population is:

[0040] C = I × H + H × O + H + O;

[0041] Among them, C represents the spatial dimension of the particle, I represents the number of input layer nodes, H represents the number of hidden layer nodes, and O represents the number of output layer nodes.

[0042] Furthermore, the expression for calculating the fitness of particles based on the initialized particle population is:

[0043]

[0044] Among them, K represents the Kth particle, E(K) represents the fitness of the individual particle, and y n Indicates the actual sea depth, d n represents the expected ocean depth value, and O represents the number of output layer elements.

[0045] Furthermore, according to the calculated particle fitness, the best individual position of each particle in the past generations is calculated as:

[0046] T i =(t i1 ,t i2, …,t iC ),i∈[1,2,…,N];

[0047] Among them, i is the number index of particles, N is the number of particles, C is the particle dimension, T i is the optimal position of the ith particle. Based on the optimal position of the individual particle, the global optimal position of the particle swarm is calculated, which is recorded as:

[0048] Z z =(Z z1 ,Z z2 ,…,Z zC )

[0049] Among them, Z z is the global optimal position, and C is the particle dimension.

[0050] Furthermore, the expression for updating the particle position is:

[0051] j i+1 =ωj i +s1r1(T i -l i )+s2r2(Z z -l i );

[0052] The expression for updating the particle's velocity is:

[0053] l i+1 = l i +j i+1 ;

[0054] Among them, ω represents the inertia weight, s1 and s2 represent the learning factor of a single particle and the learning factor of the particle group, and r1 and r2 represent random numbers in the range of [0,1].

[0055] In the above technical solution, the particle swarm algorithm is used to iteratively optimize the initialized weights and thresholds, and then optimize the parameters of the seabed topography inversion model, which can solve the problem that the neural network falls into the local optimal solution when inverting the sea depth. In the particle swarm algorithm, the parallel computing characteristics of each particle help to reduce the calculation time of the program, improve the computing efficiency while ensuring the accuracy of the results, and finally obtain a seabed topography model with higher accuracy than most current research, thereby improving the accuracy and inversion efficiency of the inverted seabed topography.

[0056] A seabed topography inversion system, the system comprising:

[0057] A data acquisition module, used to acquire ocean gravity data;

[0058] A preprocessing module, used for preprocessing the ocean gravity data;

[0059] Model building module, used to build seabed topography inversion model;

[0060] The preprocessing module is also used to divide the preprocessed ocean gravity data into a training set and a verification set;

[0061] A model training module is used to perform seabed topography inversion training on a seabed topography inversion model using a training set;

[0062] The model verification module is used to verify the trained seabed topography inversion model using the verification set, and obtain the trained seabed topography inversion model after the verification passes;

[0063] The seabed topography inversion module is used to input the preprocessed ocean gravity prediction data into the trained seabed topography inversion model for seabed topography inversion and output the inverted seabed topography results.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] The present invention proposes a method and system for inverting seabed topography. By preprocessing the collected ocean gravity data, the practicability and reliability of the ocean gravity data can be effectively improved. After the constructed seabed topography inversion model is trained by seabed topography inversion of a training set and verified by a verification set, a seabed topography inversion model that effectively grasps the changes in ocean gravity data can be obtained, so that seabed topography inversion can be performed through the seabed topography inversion model, and the influence of nonlinear term errors in areas with large seabed topography fluctuations can be effectively reduced, thereby improving the accuracy of the seabed topography obtained by inversion, and reducing the requirements for computing performance, so as to improve computing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1A flowchart of the steps of a method for inverting seabed topography provided in an embodiment of the present application;

[0067] Figure 2 A schematic diagram of the network structure of the seabed topography inversion model provided in the embodiment of the present application;

[0068] Figure 3 A flowchart of the steps of iteratively optimizing a seabed topography inversion model using a particle swarm algorithm provided in an embodiment of the present application;

[0069] Figure 4 A schematic diagram of ship test points and control points in the experiment provided in the embodiment of the present application;

[0070] Figure 5 A schematic diagram of four types of ocean gravity field data in the experiment provided in the embodiment of the present application;

[0071] Figure 6 The sea depth map obtained by inverting the seabed topography inversion model provided in the embodiment of the present application and the regression map of the ship-measured sea depth and the model;

[0072] Figure 7 A schematic diagram of an experiment comparing the sea depth obtained by inverting the seabed topography inversion model provided in an embodiment of the present application with the existing sea depth model;

[0073] Figure 8 A histogram of differences between different models and check points in the experiment provided in the embodiment of the present application;

[0074] Fig. 9 A distribution diagram of the proportion of the absolute difference between each model and the check point within 200m in different water depth intervals in the experiment provided in the embodiment of the present application and a histogram of the absolute value difference of each model;

[0075] Fig.10 The power spectral density and power spectral density difference diagram of each model in the experiment provided in the embodiment of the present application;

[0076] Fig.11 A schematic diagram of the structure of a seabed topography inversion system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0077] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Preferred embodiments of the present invention are provided in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0078] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0079] Embodiment 1:

[0080] This embodiment provides a method for inverting seabed topography. Figure 1 , the method comprises the following steps:

[0081] Step S1: Acquire ocean gravity data and pre-process the ocean gravity data;

[0082] Step S2: construct a seabed topography inversion model, and divide the preprocessed ocean gravity data into a training set and a verification set, use the training set to perform seabed topography inversion training on the seabed topography inversion model, and use the verification set to verify the trained seabed topography inversion model, and obtain a trained seabed topography inversion model after verification;

[0083] Step S3: input the preprocessed ocean gravity prediction data into the trained seabed topography inversion model to perform seabed topography inversion, and output the inverted seabed topography results.

[0084] In step S1, the process of preprocessing the ocean gravity data includes:

[0085] The acquired ocean gravity data include gravity anomaly data, vertical gravity gradient data and vertical deviation data;

[0086] The gravity anomaly data is gridded, and the gravity geological method is used to calculate the short-wave gravity anomaly data after gridding in the gravity anomaly data. The expression is:

[0087]

[0088] The SAS method is used to perform bandpass filtering on the vertical gravity gradient data in the ocean gravity data. The expression is:

[0089]

[0090] Normalize the shortwave gravity anomaly data, vertical gravity gradient data and vertical deviation data;

[0091] Among them, SAS method represents Smith And Sandwell method, i is the position of the grid point, j is the position of the control point, E(j) is the water depth at the control point, D represents the reference water depth, Δρ is the density difference between the seafloor crust and seawater, and G is the gravitational constant. obs (i) is the gridded sea surface gravity anomaly, g long (i) is the long-wave gravity anomaly of the control point grid, g short (i) is the gridded short-wave gravity anomaly. k represents the wave number, W lowpass represents Gaussian low-pass filtering, W bandpass represents a bandpass filter composed of a Gaussian high-pass filter and a Wiener low-pass filter, z0 represents the average water depth, and λ1 and λ2 represent the cutoff wavelength.

[0092] Specifically, GGM calculates the shortwave gravity anomaly:

[0093] The sea surface gravity anomaly is composed of the short-wave gravity anomaly caused by the seabed topography and the long-wave gravity anomaly caused by the deep density unevenness. Its expression is:

[0094] g obs (i) = g short (i)+g long (i)(1)

[0095] Where i is the location of the gravity anomaly grid point, g obs (i) is the sea surface gravity anomaly at the grid point, g short (i) is the shortwave gravity anomaly at the grid point, g long (i) is the long-wave gravity anomaly at the grid point. The short-wave gravity anomaly at the control point is g short (j) can be calculated from the water depth at the control point using the Bouguer plate formula, expressed as:

[0096] g short (j) = 2πGΔρ(E(j)-D)(2)

[0097] Where j is the position of the control point, E(j) is the water depth at the control point, D represents the reference water depth, Δρ is the density difference between the ocean crust and seawater, and G is the gravitational constant. After gridding, the long-wave gravity anomaly g of the grid point is obtained. long (i) According to formula (1), the gridded short-wave gravity anomaly g is obtained short (i).

[0098] SAS method to deal with vertical gradient:

[0099] Since there are random noise and low-frequency systematic errors in the original vertical gravity gradient data, it is also necessary to perform bandpass filtering on the vertical gravity gradient. The filter used is the bandpass filter used in the SAS method, as shown in formula (3):

[0100]

[0101] Where k is the wave number, W lowpass is a Gaussian low-pass filter, W bandpass It is a bandpass filter composed of Gaussian high-pass filter and Wiener low-pass filter, z0 is the average water depth, λ1 and λ2 are the cutoff wavelengths.

[0102] It can be understood that gridding the gravity anomaly data in the acquired ocean gravity data can effectively simplify the process of extracting short-wave gravity anomaly data, thereby reducing the requirements for computer performance; using the SAS method to perform bandpass filtering on the vertical gravity gradient data in the ocean gravity data can effectively reduce the random noise and low-frequency system errors existing in the original vertical gravity gradient data; normalizing the short-wave gravity anomaly data, vertical gravity gradient data and vertical deviation data can effectively improve the sensitivity of the model to the data during the model training process, so as to improve the performance of the trained model, thereby improving the accuracy and inversion efficiency of the seabed topography obtained by inversion, and reducing the requirements for computing performance to improve computing efficiency. The acquired ocean gravity data includes gravity anomaly data, vertical gravity gradient data and vertical deviation data, which can also effectively solve the problem of information omissions in a single data source.

[0103] In step S2, the process of constructing the seafloor topography inversion model includes:

[0104] A back propagation neural network is used to construct a seabed topography inversion model, wherein the seabed topography inversion model includes a plurality of input layers, a plurality of hidden layers and an output layer;

[0105] The constructed seafloor topography inversion model expression is:

[0106]

[0107] Among them, (ψ,λ) represents the longitude and latitude of the sea depth at the prediction point, represents the weights and biases between hidden layers, b (3) represents the weight and bias between the hidden layer and the output layer, n1 and n2 represent the number of nodes in the first and second layers of the hidden layer, f represents the tansig activation function, g represents the purelin activation function, and h represents the output of the first hidden layer.

[0108] Specifically, the BPNN (seabed topography inversion model constructed using back propagation neural network) designed by the present invention is as follows: Figure 2 As shown, it is composed of an input layer, two hidden layers and an output layer. Among them, the input data are longitude (LON), latitude (LAT), shortwave gravity anomaly (GA), vertical gravity gradient (VGG), vertical deviation east component (EVD) and vertical deviation north component (NVD), and the input layer is designed to be 6 nodes. The output data is the sea depth value, and the output layer is designed to be 1 node. The number of nodes in the hidden layer is related to the training accuracy and training cost of the neural network model. Increasing the number of neurons in the hidden layer can improve the accuracy of neural network training. The present invention determines that the number of hidden layers is two layers, the first layer has 4 nodes, the second layer has 7 nodes, and a total of 11 nodes.

[0109] After normalization, the gravity data enters the input layer, is passed to the next layer through the connection weights and thresholds, and is passed to the output layer layer by layer. Since the "tansig" activation function has good nonlinear conduction ability and can construct a nonlinear relationship between gravity data and seabed topography, "tansig" will be selected as the activation function of the input layer and hidden layer of the neural network, while the output layer activation function is the "purelin" function with good flexibility. The sea depth expression of the whole process is:

[0110]

[0111] Where (ψ,λ) is the longitude and latitude of the sea depth at the prediction point, are the weights and biases between hidden layers, is the weight and bias between the hidden layer and the output layer, n1, n2 are the number of nodes in the first and second layers of the hidden layer, f, g are the "tansig" and "purelin" activation functions. h is the output of the first hidden layer, the expression is:

[0112]

[0113] In the formula, is the weight and bias from the input layer to the hidden layer, where i is the node index of the first hidden layer, j is the index of the input layer node, and f is the activation function "tansig". lon is longitude, lat is latitude, ga is the shortwave gravity anomaly extracted by GGM, vgg is the vertical gradient processed by SAS method, evd, nvd are the east and north components of the vertical deviation.

[0114] It can be understood that the use of back-propagation neural network to construct a seabed topography inversion model can effectively restore the nonlinear information of the seabed topography that is ignored by traditional inversion methods. Setting 1 input layer, 2 hidden layers and 1 output layer for the constructed seabed topography inversion model can improve the accuracy of neural network training in the model and effectively reduce the cost of model training.

[0115] In step S2, see Figure 3 The process of using the training set to train the submarine topography inversion model includes:

[0116] S21: setting several rounds of iterative training to perform seabed topography inversion training on the seabed topography inversion model;

[0117] S22: Initialize the weights and thresholds between the input layer, hidden layer and output layer according to the ocean gravity data in the training set, and iteratively optimize the initialized weights and thresholds using the particle swarm algorithm, and output the optimized weights and thresholds;

[0118] S23: Setting a loss function to optimize the weights and thresholds after each round of iterative optimization, and using a validation set to verify the parameters after each round of iterative optimization, and obtaining the optimal weights and thresholds when the set iterative training round ends or the loss function converges;

[0119] S24: The obtained optimal weight and optimal threshold are set as new parameters between the input layer, the hidden layer and the output layer in the seabed topography inversion model to obtain a trained seabed topography inversion model.

[0120] In step S22, see Figure 3 ,The process of iteratively optimizing the initialized weights and thresholds using the particle swarm algorithm includes:

[0121] S221: according to the training set data received by the input layer, a particle population is set and the parameters of the particle population are initialized;

[0122] S222: Calculate the spatial dimension of particles according to the particle population after parameter initialization;

[0123] S223: Calculate the fitness of the particles according to the particle population after parameter initialization, and extract the individual best fitness eigenvalue and the global best fitness eigenvalue;

[0124] S224: Calculate the best position of the individual particle according to the calculated particle fitness, and calculate the global best position of the particle group according to the best position of the individual particle;

[0125] S225: according to the best position of the individual particle and the best position of the particle group, the position and speed of the particle are updated, and the fitness of the particle after the updated position and speed is calculated to confirm whether the fitness of the particle after the updated position and speed is better than the global best fitness. If it is better, the fitness of the current particle is taken as the global best fitness and iterates until the maximum number of updates. If it is not better, the global best fitness is not changed and iterates until the maximum number of updates.

[0126] Among them, the parameters of particle population include quantity, speed, position and update times.

[0127] In step S222, the expression for calculating the spatial dimension of the particle according to the initialized particle population is:

[0128] C = I × H + H × O + H + O;

[0129] Among them, C represents the spatial dimension of the particle, I represents the number of input layer nodes, H represents the number of hidden layer nodes, and O represents the number of output layer nodes.

[0130] Specifically, determine the number and dimensions of particles:

[0131] The number of particles represents the size of the population, and the size of the number affects the complexity of the algorithm. Therefore, the population size should not be too large. Based on existing experience and combined with the input data, the population size is set to 10. In the C-dimensional space, there is a set of N particles. This set exchanges and updates information through a series of specific rules and mechanisms to achieve the optimization goal of the algorithm. The dimension C can be determined according to formula (6):

[0132] C=I×H+H×O+H+O(6)

[0133] In formula (6), I is the number of input layer nodes, H is the number of hidden layer nodes, and O is the number of output layer nodes. The dimension of the particle in the present invention is 71.

[0134] In step S223, the expression for calculating the fitness of the particles according to the initialized particle population is:

[0135]

[0136] Among them, K represents the Kth particle, E(K) represents the fitness of the individual particle, and y n Indicates the actual sea depth, d n represents the expected ocean depth value, and O represents the number of output layer elements.

[0137] Specifically, initialize the particle position and velocity range:

[0138] In space, the position of the i-th particle is expressed as follows, setting the maximum position Lmax and maximum speed J max Limit the particles to avoid blind searching in space. Particle position L i Usually in [-L max ,L max ] range. max If the value is too large, the particle unit will move in larger steps, which may miss the best position. If the value is too small, the algorithm will converge too slowly. i Usually in [-J max ,J max ] range, the position and speed of the present invention are both in the range of [-1,1].

[0139] Calculate the particle fitness value:

[0140] In the iterative process, the position of the particle is predicted by the neural network. The smaller the error between the predicted value and the actual value, the greater the fitness. The initial fitness value is calculated using formula (7) as the fitness function.

[0141]

[0142] In formula (7), k is a particle, E(K) represents the fitness of the individual, and y n Indicates the actual sea depth, d n represents the expected ocean depth value, and O represents the number of output layer elements.

[0143] In step S224, the best individual position of each particle in the past generations is calculated according to the calculated particle fitness, and the expression is:

[0144] T i =(t i1 ,t i2, …,t iC ),i∈[1,2,…,N];

[0145] Among them, i is the number index of particles, N is the number of particles, C is the particle dimension, T i is the optimal position of the ith particle. Based on the optimal position of the individual particle, the global optimal position of the particle swarm is calculated, and the expression is:

[0146] Z z =(Z z1 ,Z z2 ,…,Z zC )

[0147] Among them, Z z is the global optimal position, and C is the particle dimension.

[0148] Furthermore, the expression for updating the particle position is:

[0149] j i+1 =ωj i +s1r1(T i -l i )+s2r2(Z z -l i );

[0150] The expression for updating the particle's velocity is:

[0151] l i+1 = l i +j i+1 ;

[0152] Among them, ω represents the inertia weight, s1 and s2 represent the learning factor of a single particle and the learning factor of the particle group, and r1 and r2 represent random numbers in the range of [0,1].

[0153] Specifically, update the best individual particle position and the best global position of the particle swarm:

[0154] For each particle, the individual best position is updated according to the currently calculated fitness value, and the optimal position is found in the iterative process, that is, the individual extreme value, which is recorded as T i In the overall operation of the particle swarm, each particle will share and interact with information. The best global position reached by all particles in the search process is the global extreme value, which is recorded as Z. z .

[0155] T i =(t i1 ,t i2, …,t iC ),i∈[1,2,…,N](8)

[0156] Z z =(Z z1 ,Z z2 ,…,Z zC )(9)

[0157] Update particle positions and velocities:

[0158] Each particle in the particle swarm is affected by the individual extreme value and the global extreme value during its movement. By comprehensively analyzing these two types of information and continuously iterating, the particle updates its moving speed and position, as shown in formulas (10) and (11).

[0159] j i+1 =ωj i +s1r1(T i -l i )+s2r2(Z z -l i ) (10)

[0160] l i+1 = l i +j i+1 (11)

[0161] In formulas (10) and (11), ω is the inertia weight, which is used to balance the influence of two adjacent motion speeds. In the present invention, the value is 0.5. s1 and s2 are the learning factors of a single particle and a particle swarm, respectively. In the present invention, the values ​​are both 2. r1 and r2 are random numbers in the range of [0,1], which are used to increase the randomness of particle motion. After optimization by the particle swarm algorithm, the optimal weights and thresholds are assigned to the BP neural network for training, as shown in Figure 2 shown.

[0162] In the above technical solution, the particle swarm algorithm is used to iteratively optimize the initialized weights and thresholds, and then optimize the parameters of the seabed topography inversion model, which can solve the problem that the neural network falls into the local optimal solution when inverting the sea depth. In the particle swarm algorithm, the parallel computing characteristics of each particle help to reduce the calculation time of the program, improve the computing efficiency while ensuring the accuracy of the results, and finally obtain a seabed topography model with higher accuracy than most current research, thereby improving the accuracy and inversion efficiency of the inverted seabed topography.

[0163] In this embodiment, by preprocessing the collected ocean gravity data, the practicability and reliability of the ocean gravity data can be effectively improved; the constructed seabed topography inversion model can obtain a seabed topography inversion model that effectively grasps the changes in the ocean gravity data after seabed topography inversion training of the training set and verification of the verification set, so that seabed topography inversion can be performed through the seabed topography inversion model, and the influence of nonlinear term errors in areas with large seabed topography fluctuations can be effectively reduced, thereby improving the accuracy of the seabed topography obtained by inversion, and reducing the requirements for computing performance to improve computing efficiency.

[0164] Embodiment 2:

[0165] This embodiment is based on the inversion method of seabed topography described in Example 1, and provides corresponding experimental data for further explanation, as follows:

[0166] Taking the South China Sea as an example, the longitude and latitude range is (112°~119° east longitude, 12°~20° north latitude). This implementation used ship-measured water depth data, a priori bathymetric data and altimetry gravity data.

[0167] 1) Priori bathymetric data: This embodiment uses the SRTM15+V2.5.5 model released by the Scripps Institution of Oceanography (SIO), with a model resolution of 15 seconds, for gross error elimination and accuracy assessment of ship survey data. In terms of evaluating the accuracy of the inversion results, the ETOPO1 model and the SIO topo_23.1 model are used. The ETOPO1 model was released in August 2008, with a spatial resolution of 1′. The model integrates a large number of seabed topography-related models and measured ship survey point data, and includes depth data for both land and sea depth. The SIO topo_23.1 model integrates satellite altimetry data and single-beam and multi-beam ship survey data, with a spatial resolution of 1′. The model can cover areas with sparse data in the previous model. The two models are mainly obtained by inversion by the SAS method, which inevitably ignores the nonlinear terms of the seabed topography. Therefore, they are used to compare with the method of the present invention to verify the ability of the method of the present invention to restore the nonlinear terms of the seabed topography.

[0168] 2) Ship-surveyed bathymetric data: There are a total of 136,028 bathymetric points in the study area in the ship-surveyed bathymetric data. Since the NCEI ship-surveyed bathymetric data has a large time span and uneven observation quality, there are many gross error data. This embodiment is based on the SRTM15+V2.5.5 model, interpolates the model water depth to the ship-surveyed points, and eliminates data points where the difference between the two is greater than twice the mean error. A total of 4,531 gross error points were eliminated, and 131,497 data points were retained, with a data elimination rate of 3.33%. This embodiment uses the bathymetric points on the "SO49" and "V3604" routes as verification points (see Figure 4 Red survey line, spatial distribution of ship-measured water depth points, where red lines represent check points, black lines represent control points, the background water depth model is SRTM15+V2.5.5), and the rest of the data are used as control points.

[0169] 3) Altimetry gravity data:

[0170] The gravity data uses the latest version of the gravity model V32.1 released by the Scripps Institution of Oceanography. The SIO model re-tracks the waveforms of Jason-1 / 2, Envisat, Altika and Cryosat-LRM altimetry data, and reduces the wavelength of the smoothing filter of the surface wave height from 45km to 30km to obtain more accurate altimetry information. The final inversion product is as follows: Figure 5 As shown (a. gravity anomaly data grav_32.1; b. vertical gradient data curv_32.1; c. north component of vertical deviation north_32.1; d. east component of vertical deviation east_32.1).

[0171] 4) Model evaluation: Gravity data is input into the neural network and the prediction results obtained after training are as follows: Figure 6 (a), the predicted water depth range is -5200m~0m. The terrain in the Zhongsha Islands and the northwest continental shelf varies greatly, while in the southern basin, except for the central seamount chain, the terrain is relatively flat. The predicted sea depth is verified using the checkpoints on the above-mentioned "SO49" and "V3604" routes, and the predicted sea depth is interpolated to the checkpoints. There are 637 worse points with a mean error of more than three times, which are mainly distributed near the Zhongsha Islands and the central seamount chain; regression analysis is performed on the ship-measured sea depth and the predicted sea depth, such as Figure 6 (b) shows that the scale factor of the regression equation is 0.996, indicating that the ship-measured sea depth is consistent with the predicted sea depth. 2 It shows that 98.9% of the water depth differences can be explained in the neural network fitting relationship.

[0172] 5) Differences in existing models: The existing bathymetric models ETOPO1 and SIO topo_23.1 were used for lateral assessment. Both ETOPO1 and SIO models show that there are a large number of seamounts in the lower right corner of the area ( Figure 7 The PSO-BP model also reflects this detail. However, the number of seamounts depicted by the three models in this area is different, so further exploration combined with subsequent ship survey data is needed to determine the specific fluctuations in this area. By analyzing the differences with existing bathymetric models (such as Figure 7 c and 7d), it can be found that the inversion model has the same trend as the ETOPO1 and SIO models as a whole, and the differences with the two models are reflected in the geomorphic details such as seamounts, trenches, and submarine basins. The most obvious differences mainly appear in the Zhongsha Islands and Xisha Islands. It is speculated that the main reason is the quality of the shipborne bathymetric data used in the model, the data processing method and the fusion technology.

[0173] To further determine the differences in numerical statistics among the models, each model was interpolated to the checkpoint for comparison. The results are shown in Table 1. Compared with the first two models, the PSO-BP model has the smallest average difference with the checkpoint, indicating the smallest system deviation. At the same time, the accuracy of this model is also the highest, with a root mean square value of 126.4m. It is worth mentioning that compared with the SIO model, its relative error is relatively large, which is mainly due to the large error in shallow water areas (such as Figure 6 This difference leads to improved accuracy compared to other models, but the relative error is relatively large.

[0174] Table 1: Statistical results of the differences between different models and checkpoints

[0175]

[0176] like Figure 8 , showing the depth difference histogram of ETOPO1, SIO, and PSO-BP models, divided by 50m steps. In the range of -50 to 0m, the difference ratio of the PSO-BP model is higher than that of the ETOPO1 model. Since the SIO model integrates high-precision ship depth measurement, it has a higher difference ratio in this part than the PSO-BP model. However, when the normal distribution curve is drawn, it is found that the PSO-BP model has a sharper "peak" and a longer "tail". 97.82% of the depth difference values ​​are concentrated in the range of ±400m, which is higher than 95.55% of the ETOPO1 model and 97.35% of the SIO model. This proves that the accuracy of the inversion model is better than the previous two in terms of overall effect.

[0177] To further study the performance of the model in different water depth ranges, the study area was divided into five intervals with a span of 1000m. The proportion of the absolute difference between each model and the check point within different water depth ranges was counted and found to be less than 200m. Fig. 9 As shown in a. In the range of ≤1000m and >4000m, 93% and 97% of the water depth differences of the PSO-BP model are less than 200m; except for the range of ≤1000m, which is slightly worse than the SIO model, the PSO-BP model has good performance in other water depth ranges, especially in deep-sea topography, which has a good improvement compared with the two models, indicating that the particle swarm algorithm may be beneficial to restore deep-sea topography. Fig. 9 b is the histogram of the absolute value differences of each model. In terms of the number of check points with a difference of less than 200 m, the PSO-BP model is slightly better than the other models with a proportion of 91.78%, which is closest to the ship measurement data.

[0178] In the frequency domain, in order to examine the performance of each model at different wavelengths, the radial power spectrum density diagram of each model was calculated (see Fig.10 a) At the same wavelength, the higher the power spectrum density, the higher the energy, which can provide more seabed topography details. Fig.10 It can be seen from Figure 1 that in the wavelength range of 20 to 100 km, the power spectral density of the PSO-BP model is higher than that of the other two models, and it performs better in recovering the details of the seabed topography. In the 10 to 20 km band, the power spectral density of the SIO model is higher than that of the PSO-BP model, mainly because the SIO model incorporates high-precision ship survey data.

[0179] In this embodiment, due to the problem of random initialization, the BP neural network (seabed topography inversion model) is optimized by introducing a particle swarm algorithm and applied to the depth of the South China Sea to test its feasibility. The input of the neural network includes multiple gravity field data obtained from satellite altimetry, including gravity anomalies, vertical gradients and vertical deviations. The results show that the PSO-BP neural network significantly improves the efficiency of the network in extracting feature information. The inversion model is very consistent with the globally recognized seabed topography model, proving that the method proposed in the present invention can accurately represent the topography. Compared with the existing sea depth model, since the ETOPO1 model and the SIO model are mainly inverted by the SAS method, the nonlinear terms of the seabed topography are inevitably ignored. The inversion accuracy of this method is 41.1 meters and 20.8 meters higher than that of the ETOPO1 model and the SIO model, respectively, and the nonlinear terms of the seabed topography are restored to a certain extent. In the analysis of the spatial domain, the PSO-BP model (trained seabed topography inversion model) shows the ability to recover deep terrain. Even in areas where ship survey data are scarce, it can also show good robustness. In the frequency domain analysis, the PSO-BP model has a better ability to reflect details, especially in the seafloor topography band that is highly correlated with gravity.

[0180] Embodiment three:

[0181] This embodiment provides a seabed topography inversion system, see Fig.11 , the system comprising:

[0182] A data acquisition module, used to acquire ocean gravity data;

[0183] A preprocessing module, used for preprocessing the ocean gravity data;

[0184] Model building module, used to build seabed topography inversion model;

[0185] The preprocessing module is also used to divide the preprocessed ocean gravity data into a training set and a verification set;

[0186] A model training module is used to perform seabed topography inversion training on a seabed topography inversion model using a training set;

[0187] The model verification module is used to verify the trained seabed topography inversion model using the verification set, and obtain the trained seabed topography inversion model after the verification passes;

[0188] The seabed topography inversion module is used to input the preprocessed ocean gravity prediction data into the trained seabed topography inversion model for seabed topography inversion and output the inverted seabed topography results.

[0189] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for inverting seabed topography, characterized in that: The method comprises the following steps: Acquiring ocean gravity data and preprocessing the ocean gravity data; Construct a seabed topography inversion model, and divide the preprocessed ocean gravity data into a training set and a verification set. Use the training set to train the seabed topography inversion model, and use the verification set to verify the trained seabed topography inversion model. After verification, a trained seabed topography inversion model is obtained. The preprocessed ocean gravity prediction data is input into the trained seabed topography inversion model for seabed topography inversion, and the inverted seabed topography results are output.

2. The method for inverting seabed topography according to claim 1, characterized in that: The process of preprocessing the ocean gravity data includes: The acquired ocean gravity data include gravity anomaly data, vertical gravity gradient data and vertical deviation data; The gravity anomaly data is gridded, and the short-wave gravity anomaly data after gridding in the gravity anomaly data is calculated using the gravity geological method. The expression is: The SAS method is used to perform bandpass filtering on the vertical gravity gradient data in the ocean gravity data. The expression is: Normalize shortwave gravity anomaly data, vertical gravity gradient data and vertical deviation data; Wherein, SAS method stands for Smith and Sandwell method, i is the position of the grid point, j is the position of the control point, E(j) is the water depth at the control point, D represents the reference water depth, Δρ is the density difference between the seafloor crust and seawater, G is the gravitational constant, and g obs (i) is the gridded sea surface gravity anomaly, g long (i) is the long-wave gravity anomaly of the control point grid, g short (i) is the gridded short-wave gravity anomaly, k represents the wave number, W lowpass represents Gaussian low-pass filtering, W bandpass represents a bandpass filter composed of a Gaussian high-pass filter and a Wiener low-pass filter, z0 represents the average water depth, and λ1 and λ2 represent the cutoff wavelength.

3. The method for inverting seabed topography according to claim 2, characterized in that: The process of constructing the seafloor topography inversion model includes: A back propagation neural network is used to construct a seabed topography inversion model, wherein the seabed topography inversion model includes a plurality of input layers, a plurality of hidden layers and an output layer; The constructed seafloor topography inversion model expression is: Among them, (ψ,λ) represents the longitude and latitude of the sea depth at the prediction point, represents the weights and biases between hidden layers, b (3) represents the weight and bias between the hidden layer and the output layer, n1 and n2 represent the number of nodes in the first and second layers of the hidden layer, f represents the tansig activation function, g represents the purelin activation function, and h represents the output of the first hidden layer.

4. The method for inverting seabed topography according to claim 3, characterized in that: The process of using the training set to train the seabed topography inversion model includes: Setting several rounds of iterative training to perform seabed topography inversion training on the seabed topography inversion model; According to the ocean gravity data in the training set, the weights and thresholds between the input layer, hidden layer and output layer are initialized, and the particle swarm algorithm is used to iteratively optimize the initialized weights and thresholds, and the optimized weights and thresholds are output; Set the loss function to optimize the weights and thresholds after each round of iterative optimization, and use the validation set to verify the parameters after each round of iterative optimization. When the set iterative training round ends or the loss function converges, the optimal weights and thresholds are obtained; The obtained optimal weights and optimal thresholds are set as new parameters between the input layer, hidden layer and output layer in the seabed topography inversion model to obtain a trained seabed topography inversion model.

5. The method for inverting seabed topography according to claim 4, characterized in that: The process of iteratively optimizing the initialized weights and thresholds using the particle swarm algorithm includes: According to the training set data received by the input layer, the particle population is set and the parameters of the particle population are initialized; Calculate the spatial dimension of particles based on the particle population initialized with parameters; Calculate the fitness of particles according to the particle population after parameter initialization, and extract the individual best fitness eigenvalue and the global best fitness eigenvalue; According to the calculated particle fitness, the best position of the individual particle is calculated, and according to the best position of the individual particle, the global best position of the particle group is calculated; According to the individual best position of the particle and the global best position of the particle group, the position and speed of the particle are updated, and the fitness of the particle after the updated position and speed is calculated to confirm whether the fitness of the particle after the updated position and speed is better than the global best fitness. If better, the fitness of the current particle is taken as the global best fitness and iterates until the maximum number of updates. If not, the global best fitness is not changed and iterates until the maximum number of updates.

6. The method for inverting seabed topography according to claim 5, characterized in that: The expression for calculating the spatial dimension of particles based on the initialized particle population is: C = I × H + H × O + H + O; Among them, C represents the spatial dimension of the particle, I represents the number of input layer nodes, H represents the number of hidden layer nodes, and O represents the number of output layer nodes.

7. The method for inverting seabed topography according to claim 5, characterized in that: The expression for calculating the fitness of particles based on the initialized particle population is: Among them, K represents the Kth particle, E(K) represents the fitness of the individual particle, and y n Indicates the actual sea depth, d n represents the expected sea depth value, and O represents the number of output layer elements.

8. The method for inverting seabed topography according to claim 6, characterized in that: According to the calculated particle fitness, the best individual position of each particle in the past generations is calculated. The expression is: T i =(t i1 ,t i2 ,...,t iC ),i∈[1,2,...,N]; Among them, i is the number index of particles, N is the number of particles, C is the particle dimension, T i is the optimal position of the ith particle. According to the best position of individual particles, the global best position of the particle group is calculated, and the expression is: WITH z =(From z1 ,WITH z2 ,...,WITH zC ) Among them, Z z is the global optimal position, and C is the particle dimension.

9. The method for inverting seabed topography according to claim 8, characterized in that: The expression for updating the particle's position is: j i+1 =ωj i +s1r1(T i -l i )+s2r2(Z z -l i ); The expression for updating the particle's velocity is: l i+1 ＝l i +j i+1 ； Among them, ω represents the inertia weight, s1 and s2 represent the learning factor of a single particle and the learning factor of the particle group, and r1 and r2 represent random numbers in the range of [0, 1].

10. A seabed topography inversion system, characterized in that: The system comprises: A data acquisition module, used to acquire ocean gravity data; A preprocessing module, used for preprocessing the ocean gravity data; Model building module, used to build seabed topography inversion model; The preprocessing module is also used to divide the preprocessed ocean gravity data into a training set and a verification set; A model training module is used to perform seabed topography inversion training on a seabed topography inversion model using a training set; The model verification module is used to verify the trained seabed topography inversion model using the verification set, and obtain the trained seabed topography inversion model after the verification passes; The seabed topography inversion module is used to input the preprocessed ocean gravity prediction data into the trained seabed topography inversion model for seabed topography inversion and output the inverted seabed topography results.

Images