A method and system for generating seafloor terrain model

Abstract

The invention discloses a method and system for generating a seabed terrain model. The invention adopts an improved CNN network to generate a mid-band seabed terrain model, does not ignore the influence of nonlinear terms and high-order terms, reduces the influence of uneven distribution and sparse distribution of mid-band data measured by ship, greatly improves the precision and accuracy of the seabed terrain model, and is a method for generating seabed terrain that is easy to operate and has universal adaptability.

Description

Technical Field

[0001] The invention relates to a method and system for generating a seabed terrain model, and belongs to the field of marine application research. Background Art

[0002] Currently, the proportional factor method (SAS method) is used to invert the mid-band seabed topography. However, since the SAS method ignores the influence of nonlinear terms and high-order terms on data accuracy, the accuracy of the inverted mid-band seabed topography is not high, which further leads to the low accuracy of the entire seabed topography model (the model mainly includes the long-band seabed topography model and the mid-band seabed topography model, and the ship-measured sea depth and gravity anomaly in the low band have a low correlation and are therefore negligible). Summary of the invention

[0003] The present invention provides a method and system for generating a seabed terrain model, which solves the problems disclosed in the background technology.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A method for generating a seabed terrain model, comprising:

[0006] Obtain long-wave seabed topography model based on ship-measured sea depth data;

[0007] Obtain the gravity anomaly data of the mid-band, input the gravity anomaly data of the mid-band into the pre-trained CNN network, and obtain the mid-band seafloor topography model; wherein, the CNN network is trained by the stochastic gradient descent method with momentum, and the L2 regularization factor is used in the stochastic gradient descent method with momentum to attenuate the weight value; the CNN network uses the ELU function as the activation function;

[0008] The long-wave band seafloor topography model and the medium-wave band seafloor topography model are superimposed to obtain the seafloor topography model.

[0009] Based on the ship-measured sea depth data, the long-wave seabed topography model is obtained, including:

[0010] Grid processing is performed on the ship-measured sea depth data to generate a grid sea depth model;

[0011] The grid ocean depth model is transformed from the spatial domain to the frequency domain, and the transformed grid ocean depth model is high-pass filtered to obtain a long-wave seabed topography model.

[0012] The samples for training the CNN network include medium-band gravity anomaly data and medium-band ship-surveyed sea depth data. The medium-band gravity anomaly data are extracted from the gravity anomaly data model through a bandpass filter, and the medium-band ship-surveyed sea depth data are extracted from the ship-surveyed residual sea depth data. The ship-surveyed residual sea depth data is the ship-surveyed sea depth data minus the long-band sea depth data.

[0013] The objective function of the stochastic gradient descent method with momentum is:

[0014]

[0015] Among them, α is the trade-off norm penalty term, Ω is the standard objective function, θ is the starting point of gradient descent, X is the input variable, y is the target value corresponding to the input variable, and J is the function about θ. is the objective function.

[0016] Before the data is input into the CNN network, the data is normalized using min-max.

[0017] A seabed terrain model generation system, comprising:

[0018] Long-wave model module: obtains long-wave seabed topography model based on ship-measured sea depth data;

[0019] Mid-band model module: obtain mid-band gravity anomaly data, input the mid-band gravity anomaly data into the pre-trained CNN network, and obtain the mid-band seafloor topography model; the CNN network is trained using the stochastic gradient descent method with momentum, and the L2 regularization factor is used in the stochastic gradient descent method with momentum to attenuate the weight value; the CNN network uses the ELU function as the activation function;

[0020] Overlay module: Overlay the long-wave band seafloor topography model and the medium-wave band seafloor topography model to obtain the seafloor topography model.

[0021] The long-waveband model module includes:

[0022] Gridding module: Gridding the ship-measured sea depth data to generate a gridded sea depth model;

[0023] Model acquisition module: transform the grid ocean depth model from the spatial domain to the frequency domain, perform high-pass filtering on the transformed grid ocean depth model, and obtain the long-wave seabed topography model.

[0024] The samples for training the CNN network include medium-band gravity anomaly data and medium-band ship-surveyed sea depth data. The medium-band gravity anomaly data are extracted from the gravity anomaly data model through a bandpass filter, and the medium-band ship-surveyed sea depth data are extracted from the ship-surveyed residual sea depth data. The ship-surveyed residual sea depth data is the ship-surveyed sea depth data minus the long-band sea depth data.

[0025] It also includes a standardization processing module, which performs min-max standardization on the data before inputting it into the CNN network.

[0026] A computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, cause the computing device to perform a method for generating a seafloor terrain model.

[0027] The beneficial effects achieved by the present invention are as follows: the present invention adopts an improved CNN network to generate a mid-band seabed terrain model, does not ignore the influence of nonlinear terms and higher-order terms, reduces the influence of uneven distribution and sparse distribution of mid-band data measured by ship, greatly improves the precision and accuracy of the seabed terrain model, and is a seabed terrain generation method that is easy to operate and has universal adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of the method of the present invention;

[0029] Figure 2 is a detailed flow chart of the method of the present invention;

[0030] Figure 3 is a schematic diagram of residual distribution;

[0031] FIG4( a ) is a mid-band seafloor topography model obtained by using the method of the present invention;

[0032] Figure 4(b) shows the mid-band seafloor topography model obtained using the scale factor method;

[0033] FIG5( a ) is a statistical diagram of the proportion of sea depth difference in the model of the present invention;

[0034] Figure 5(b) is a statistical diagram of the proportion of sea depth difference in the V19.1 model;

[0035] Figure 5(c) is a statistical diagram of the proportion of sea depth difference in the GEBCO model;

[0036] Figure 5(d) is a statistical diagram of the proportion of sea depth difference in the ETOPO1 model;

[0037] FIG6( a ) is a schematic diagram showing the error distribution between the model of the present invention and the ship survey grid model;

[0038] Figure 6(b) is a schematic diagram of the error distribution between the V19.1 model and the ship-surveyed grid model;

[0039] Figure 6(c) is a schematic diagram of the error distribution between the GEBCO model and the ship-measured grid model;

[0040] Figure 6(d) is a schematic diagram of the error distribution between the ETOPO1 model and the ship-measured grid model. DETAILED DESCRIPTION

[0041] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0042] like Figure 1 As shown, a method for generating a seabed terrain model comprises the following steps:

[0043] Step 1, obtaining a long-wave seabed topography model based on ship-measured sea depth data;

[0044] Step 2, obtaining the gravity anomaly data of the mid-band, inputting the gravity anomaly data of the mid-band into the pre-trained CNN network, and obtaining the mid-band seafloor topography model; wherein the CNN network is trained by the stochastic gradient descent method with momentum, and the L2 regularization factor is used in the stochastic gradient descent method with momentum to attenuate the weight value; the CNN network uses the ELU function as the activation function;

[0045] Step 3: superimpose the long-wave band seabed topography model and the medium-wave band seabed topography model to obtain a seabed topography model.

[0046] The above method uses an improved CNN network to generate a mid-band seabed terrain model. It does not ignore the influence of nonlinear terms and higher-order terms, and reduces the impact of uneven distribution and sparse distribution of ship-measured mid-band data, greatly improving the precision and accuracy of the seabed terrain model. It is a simple-to-operate and universally adaptable seabed terrain generation method.

[0047] The above-mentioned ship-surveyed depth data are depth data measured over a long period of time in the study area. Since the ship-surveyed depth data are data obtained from measurements at an early period, there are large errors in the ship-surveyed depth data, which affects the data accuracy. Therefore, it is necessary to preprocess them, mainly to eliminate data points with excessive gross errors and mean errors higher than the threshold.

[0048] The ship-measured bathymetric data usually include control point data and check point data, wherein the data used to generate the model is the control point data, and the check point data is used to check the generated model.

[0049] like Figure 2 , the control point data in the ship-measured sea depth data are gridded to generate a grid sea depth model with a resolution of 1×1, and the grid sea depth model is transformed from the spatial domain to the frequency domain.

[0050] The conversion from spatial domain to frequency domain mainly uses Fourier transform to convert data into frequency domain. The conversion formula can be expressed as:

[0051] U(k x ,k y ,0)=F[u(x,y,0)]

[0052] This formula represents the Fourier transform of u(x,y,0), k x , k y are the wave numbers in the x and y directions respectively;

[0053] Design the following high-pass filter:

[0054] W1(k)=1-exp[-2(πks) 2 ]

[0055] Where W1(k) is a high-pass filter, s is the Wiener filter parameter, k is the circular frequency,

[0056] The converted grid ocean depth model is subjected to high-pass filtering, that is, the mid-band and low-band are filtered out, and a long-band seafloor topography model is generated based on the retained long-band data (the seafloor topography model is generated through the ocean depth).

[0057] According to the correlation analysis of ship-measured sea depth data and gravity anomaly data, the two have a strong correlation reaction in the mid-band, and the low-band is weak and does not participate in the reaction, so a band-pass filter is needed to extract the mid-band gravity anomaly data from the gravity anomaly data model; the band-pass filter is designed as follows:

[0058] W2(k)=[1+Ak 4 exp(4πkd)] -1

[0059] W(k)=W1(k)+W2(k)

[0060] Where W(k) is a bandpass filter, W2(k) is a low-pass filter, A is a constant, and d is the average sea depth.

[0061] In order to generate the mid-band seafloor topography model and not ignore the influence of nonlinear terms and higher-order terms in the generation process, a neural network, specifically a CNN network, can be used. However, due to the limitations of ship measurement data, the traditional CNN network has limitations on it, so the CNN network needs to be improved, as follows:

[0062] 1. Using CNN to estimate the seabed topography, the maximum and minimum values ​​can reach several thousand meters in height difference, and gravity anomaly and sea depth are two different dimensions, which are affected by the dimensions of the indicators;

[0063] Therefore, the data input into the improved CNN network is subjected to min-max standardization processing, and both the sample data and the mid-band gravity anomaly data obtained when using the network need to be standardized; the formula for min-max standardization processing is: f = (X′-min) / (max-min), where X′ is the data to be changed; min is the minimum value of the group of data; max is the maximum value of the group of data.

[0064] 2. Traditional CNN networks generally use the ReLU function as the activation function. The advantage of this function is that it overcomes the cause of gradient disappearance and speeds up training;

[0065] The ReLU function is:

[0066] ReLU=max(0,x)

[0067] Where x is the value of the input signal.

[0068] In the improved CNN network, the ELU function is used to replace the ReLU function as the activation function. Compared with the ReLU function, it has a certain output when a negative number is input, and the output has a certain anti-interference ability, eliminating the necrosis situation; its function form is as follows:

[0069]

[0070] Among them, λ is a setting parameter, usually 0.01.

[0071] 3. CNN networks are generally trained using the ordinary gradient descent method. The main purpose of the ordinary gradient descent method is to find the minimum value of the objective function through iteration, or converge to the minimum value, and gradually adjust the input weights and biases between layers through training errors;

[0072] In the improved CNN network, the stochastic gradient descent method with momentum is used for training. The exponential weighted average method is used in the gradient calculation, so that there is a correlation between each gradient calculation, which offsets the situation that the gradient swings up and down, while the true descent direction (towards the right) can be well maintained, making the convergence optimization faster. The specific mathematical iteration formula is as follows:

[0073]

[0074] Among them, J(θ 0 ) is θ 0 The function of θ 0 is the independent variable, θ 1 is the updated value, is the gradient and η is the step size.

[0075] If v t-1represents the average of the first t-1 numbers, n t represents the tth number, then the exponentially weighted average of t numbers is defined as:

[0076] v t =αv t-1 +βn t

[0077] Among them, α is the weighted average coefficient, and β=1-α.

[0078] Let the momentum be v, and its momentum is updated according to the above formula:

[0079]

[0080] θ″=θ′+v

[0081] Among them, ∈ is the learning rate, v′ is the previous momentum, θ′ is the initial parameter, and θ″ is the updated parameter.

[0082] In order to prevent the effect of overfitting, the traditional stochastic gradient descent method with momentum is improved. The L2 regularization factor is used to decay the weight value in the improved stochastic gradient descent method with momentum. The principle is to add a parameter norm to the objective function to limit its learning ability. The regularized objective function is as follows:

[0083]

[0084] Among them, α is the trade-off norm penalty term, Ω is the standard objective function, θ is the starting point of gradient descent, X is the input variable, y is the target value corresponding to the input variable, and J is the function about θ. is the objective function.

[0085] The improved CNN network is trained using samples, specifically, the training data is imported, the number of iterations and the learning rate are set, multiple cycles are performed, the optimal parameters are selected, the test data is used for prediction, and optimization is performed according to the set conditions; wherein, the samples for training the CNN network include medium-band gravity anomaly data and medium-band ship-surveyed sea depth data, wherein the medium-band gravity anomaly data is extracted from the gravity anomaly data model through a bandpass filter, and the medium-band ship-surveyed sea depth data is extracted from the ship-surveyed residual sea depth data, and the ship-surveyed residual sea depth data is the ship-surveyed sea depth data minus the long-band sea depth data.

[0086] The trained CNN network can directly input the medium-band gravity anomaly data to obtain the medium-band seabed topography model. Finally, the long-band seabed topography model and the medium-band seabed topography model are superimposed to obtain the seabed topography model.

[0087] In order to verify the above method, take the southern part of the Hawaiian Emperor Seamount Chain, along the seamount chain trend at 40°~45°N, 168°~172°E as an example:

[0088] The gravity anomaly data of the mid-band is input into the improved CNN network to obtain the mid-band seafloor topography model, and its residual is calculated. The improved CNN network method (B-CNN), the same scale factor method and the traditional CNN method are compared, and the values ​​are shown in Table 1 and Figure 3 shown.

[0089] Table 1 Numerical statistics of residual analysis / m

[0090]

[0091]

[0092] From Table 1, we can see that the residual sum of squares, RMSE and MAE have been significantly reduced, and Figure 3 It shows that the residual distribution of the improved CNN for terrain estimation is significantly better than the least squares, and the residual is significantly reduced.

[0093] From Figure 4(a) to (b), we can see that the improved CNN estimation method uses nonlinear processing methods to display the terrain and landforms more specifically. In seamounts, sea valleys and continuous terrain areas, the estimated terrain is detailed and the terrain display is more comprehensive.

[0094] The superimposed seabed terrain model was compared and analyzed with the internationally used high-precision terrain model. The data of different seabed terrain models are shown in Table 2.

[0095] Table 2 Numerical statistics of different seabed topography models / m

[0096]

[0097] There is no significant difference in average water depth between the terrain model constructed by the present invention and the high-precision terrain model, the correlation coefficient reaches 0.98, the root mean square RMS and standard deviation STD are minimal, and its stability from the standard deviation evaluation data is higher than that of the existing terrain model.

[0098] In order to further analyze the model accuracy, the check point data was used for verification. Specifically, the corresponding sea depth value was obtained by interpolating each model at the check point, and the accuracy was compared by subtracting it from the ship-measured check point, as shown in Table 3.

[0099] Table 3 Difference statistics of different seabed topography models at checkpoints / m

[0100]

[0101] At the check points, the average sea depth values ​​of the check points of each model are close to each other, and the terrain model constructed by the present invention has a comparable accuracy to the high-precision terrain model.

[0102] 5(a) to (d) show the depth difference distribution histograms of different models. The depth difference of the model of the present invention accounts for 43.15% within the range of 50m, and the depth difference within the range of 200m reaches 78.81%, which is higher than that of other terrain models.

[0103] Table 4 Relative error statistics of different seabed terrain models / m

[0104]

[0105] Figures 6(a) to (d) show the relative error distribution of each model and the NGDC grid model (ship survey grid model), and the numerical statistical results are given in Table 4. The maximum relative error of the model constructed by the present invention is 1.0202, which is lower than that of the other three terrain models, and the standard deviation STD is the smallest, reaching 0.0913.

[0106] Verification has shown that the correlation between the seabed topography constructed by the present invention and the internationally widely used high-precision terrain model is above 0.98. At the same time, its relative error distribution is significantly reduced and the accuracy is significantly improved. It has met the needs of constructing a high-precision terrain model and has high precision and accuracy.

[0107] Based on the same technical solution, the present invention also discloses a software system of the above method, a seabed terrain model generation system, comprising:

[0108] Long-wave model module: Obtain long-wave seabed topography model based on ship-measured sea depth data.

[0109] The long-waveband model module includes:

[0110] Gridding module: Gridding the ship-measured sea depth data to generate a gridded sea depth model;

[0111] Model acquisition module: transform the grid ocean depth model from the spatial domain to the frequency domain, perform high-pass filtering on the transformed grid ocean depth model, and obtain the long-wave seabed topography model.

[0112] Standardization processing module: Before the data is input into the CNN network, the data is subjected to min-max standardization processing.

[0113] Mid-band model module: obtain mid-band gravity anomaly data, input the mid-band gravity anomaly data processed by the standardization processing module into the pre-trained CNN network, and obtain the mid-band seabed topography model; the CNN network is trained using the stochastic gradient descent method with momentum, and the L2 regularization factor is used in the stochastic gradient descent method with momentum to attenuate the weight value; the CNN network uses the ELU function as the activation function; the training data is also processed by the standardization processing module.

[0114] The samples for training the CNN network include medium-band gravity anomaly data and medium-band ship-surveyed sea depth data. The medium-band gravity anomaly data are extracted from the gravity anomaly data model through a bandpass filter, and the medium-band ship-surveyed sea depth data are extracted from the ship-surveyed residual sea depth data. The ship-surveyed residual sea depth data is the ship-surveyed sea depth data minus the long-band sea depth data.

[0115] Overlay module: Overlay the long-wave band seafloor topography model and the medium-wave band seafloor topography model to obtain the seafloor topography model.

[0116] Based on the same technical solution, the present invention also discloses a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by a computing device, the computing device executes a method for generating a seabed terrain model.

[0117] Based on the same technical solution, the present invention also discloses a computing device, including one or more processors, one or more memories and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing a method for generating a seabed terrain model.

[0118] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0119] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0120] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0121] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0122] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are included in the scope of the claims of the present invention to be approved.

Claims

1. A method for generating a seabed terrain model, characterized in that: include: Obtain long-wave seabed topography model based on ship-measured sea depth data; Obtain the gravity anomaly data of the medium band, input the gravity anomaly data of the medium band into the pre-trained CNN network, and obtain the mid-band seabed topography model; wherein, the CNN network is trained by the stochastic gradient descent method with momentum, and the L2 regularization factor is used in the stochastic gradient descent method with momentum to attenuate the weight value; the CNN network uses the ELU function as the activation function; the samples for training the CNN network include the mid-band gravity anomaly data and the mid-band ship-surveyed sea depth data, the mid-band gravity anomaly data is extracted from the gravity anomaly data model through a bandpass filter, the mid-band ship-surveyed sea depth data is extracted from the ship-surveyed residual sea depth data, and the ship-surveyed residual sea depth data is the ship-surveyed sea depth data minus the long-band sea depth data; The long-wave band seafloor topography model and the medium-wave band seafloor topography model are superimposed to obtain the seafloor topography model.

2. A method for generating a seabed terrain model according to claim 1, characterized in that: Based on the ship-measured sea depth data, the long-wave seabed topography model is obtained, including: Grid processing is performed on the ship-measured sea depth data to generate a grid sea depth model; The grid ocean depth model is transformed from the spatial domain to the frequency domain, and the transformed grid ocean depth model is high-pass filtered to obtain a long-wave seabed topography model.

3. The method for generating a seabed terrain model according to claim 1, characterized in that: The objective function of the stochastic gradient descent method with momentum is: ； in, α To weigh the norm penalty term, is the standard objective function, θ is the starting point of gradient descent, X is the input variable, y is the target value corresponding to the input variable, J About θ The function of is the objective function.

4. The method for generating a seabed terrain model according to claim 1, characterized in that: Before the data is input into the CNN network, the data is normalized using min-max.

5. A seabed terrain model generation system, characterized in that: include: Long-wave model module: obtains long-wave seabed topography model based on ship-measured sea depth data; Medium-band model module: obtain the gravity anomaly data of the medium-band, input the gravity anomaly data of the medium-band into the pre-trained CNN network, and obtain the medium-band seabed topography model; wherein, the CNN network is trained by the stochastic gradient descent method with momentum, and the L2 regularization factor is used in the stochastic gradient descent method with momentum to attenuate the weight value; the CNN network uses the ELU function as the activation function; the samples for training the CNN network include the medium-band gravity anomaly data and the medium-band ship-surveyed sea depth data, the medium-band gravity anomaly data is extracted from the gravity anomaly data model through a bandpass filter, and the medium-band ship-surveyed sea depth data is extracted from the ship-surveyed residual sea depth data, and the ship-surveyed residual sea depth data is the ship-surveyed sea depth data minus the long-band sea depth data; Overlay module: Overlay the long-wave band seafloor topography model and the medium-wave band seafloor topography model to obtain the seafloor topography model.

6. A seabed terrain model generation system according to claim 5, characterized in that: The long-waveband model module includes: Gridding module: Gridding the ship-measured sea depth data to generate a gridded sea depth model; Model acquisition module: transform the grid ocean depth model from the spatial domain to the frequency domain, perform high-pass filtering on the transformed grid ocean depth model, and obtain the long-wave seabed topography model.

7. A seabed terrain model generation system according to claim 5, characterized in that: It also includes a standardization processing module, which performs min-max standardization on the data before inputting it into the CNN network.

8. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions which, when executed by a computing device, cause the computing device to perform any one of the methods according to claims 1 to 4.

Images