A method for inverting seafloor topography using machine learning based on terrain unit partitioning

Abstract

The present invention discloses a method for inverting seabed topography using machine learning based on terrain unit zoning, belonging to the field of geophysical technology. The method is used for inverting seabed topography, comprising: forming a feature data set required by a machine learning model using gravity anomalies, shortwave gravity anomalies, vertical gravity gradient anomalies, and residual vertical gravity gradient anomalies, wherein the feature data set constitutes a training set, and a gridded feature data set constitutes a prediction set; performing an overall correlation analysis on the training set, adjusting various parameters of the model, inputting all training sets into the machine learning model for training, and then inputting the prediction set into the trained machine learning model to obtain the water depth value of the entire area, and fusing the zoning prediction results to obtain the final water depth of the entire area. The present invention utilizes the powerful nonlinear mapping capability of machine learning to solve the problem that traditional inversion methods ignore gravity information and the effects of nonlinear terms of seabed topography, thereby improving the inversion accuracy of areas with large terrain fluctuations and rapid changes.

Description

Technical Field

[0001] The invention discloses a method for inverting seabed topography using machine learning based on terrain unit partitioning, and belongs to the field of geophysical technology. Background Art

[0002] Traditional seabed topography surveys primarily rely on shipborne sonar bathymetry to acquire seabed topography data, which suffers from low efficiency and high costs. Most traditional research methods only consider the linear mapping relationship between gravity anomalies and seabed topography, ignoring the impact of nonlinear terms on the inversion results, which limits the accuracy of the inversion results. Furthermore, vertical gravity gradient anomalies, as an extension of gravity anomaly data, are more sensitive to the high-frequency components of seabed topography than gravity itself. Existing technologies utilize vertical gravity gradient anomalies, take into account the second-order effects of Park theory, and employ simulated annealing to invert seabed topography, improving accuracy by 22%. This method considers the role of nonlinear terms and confirms that the use of vertical gravity gradient anomalies can reflect more detailed topographic information and improve inversion accuracy. In seabed topography inversion, the influence of high-order nonlinear terms between the seabed gravity field and topography cannot be ignored.

[0003] Machine learning has powerful nonlinear mapping capabilities and shows great potential in the field of earth science. However, the relationship between different seabed topography and gravity fields is uncertain, especially in areas with large terrain fluctuations, such as ridges and trenches. The causes of different terrains are complex, and the relationship between them and gravity is even more complex. Using only a single model is not enough to reflect this complex relationship, especially when there is less training data, which will lead to reduced inversion accuracy. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for inverting seabed topography using machine learning based on terrain unit partitioning, so as to solve the problem of low accuracy of seabed topography inversion using machine learning under complex terrain in the prior art.

[0005] A method for inverting seafloor topography using machine learning based on topographic unit partitioning, comprising:

[0006] S1. Separate shortwave gravity anomalies from gravity anomalies;

[0007] S2. Separate the residual vertical gravity gradient anomaly from the vertical gravity gradient anomaly;

[0008] S3. Gravity anomalies, shortwave gravity anomalies, vertical gravity gradient anomalies, and residual vertical gravity gradient anomalies are combined into a feature dataset required for a machine learning model. The feature dataset constitutes a training set, and the gridded feature dataset constitutes a prediction set. The machine learning model uses a BP neural network model.

[0009] S4. Perform a correlation analysis on the training set and the water depth data, using the Pearson correlation coefficient to measure the correlation between the training set and the water depth data. Based on the correlation between the training set and the water depth data, the weight of the training set in the model training is determined. All training sets are input into the machine learning model for training, and the prediction set is then input into the trained machine learning model to obtain the water depth value for the entire area.

[0010] S5. Grid the water depth values ​​of the entire area to obtain a seabed topography model of the entire area;

[0011] The topography and landforms of the area are determined by the seabed topography model of the entire area. The topography of the study area is divided into topographic units according to the three topographic features of ridges, trenches, and basins. The ship-surveyed water depth data and gravity data of each sub-area are processed by steps S1, S2, and S3 respectively to obtain the characteristic data sets of each sub-area. The characteristic data sets of the ship-surveyed water depth control point positions are used as partition training sets. The partition training sets are input into the machine learning model for training, and then the partition prediction sets are input into the machine learning model to obtain the partition prediction results. The partition prediction results are fused to obtain the final water depth of the entire area.

[0012] S1 includes:

[0013] Measured shortwave gravity Δg at a single beam point res Use control point j n The water depth is calculated using the Bouguer plate formula:

[0014]

[0015] Where, Indicates that at control point j n The short-wave component on the surface of the ocean; G is the gravitational constant; Δρ is the optimal density difference constant between seawater and the oceanic crust; D represents the reference water depth, which is the maximum water depth of the shipborne bathymetric data; is the water depth value of the control point;

[0016] The iterative method is used to calculate the correlation and root mean square error between the predicted water depth and the measured water depth corresponding to different density difference constants. The density difference constant with the minimum root mean square error and the maximum correlation coefficient is the optimal value.

[0017] S2 includes:

[0018] Considering gravity data and water depth as two different signals, the coherence between them is:

[0019]

[0020] in, is the cross-spectral coherence function; G(k) and H(k) represent the Fourier transform of gravity signal and terrain signal respectively; G *(k), H * (k) represents the complex conjugate of G(k) and B(k), respectively;

[0021] The detrended water depth data and vertical gravity gradient anomaly were analyzed for coherence. Linear regression is performed on the data within the band range greater than 0.5, and the band vertical gravity gradient anomaly-band water depth proportional factor is obtained. The obtained proportional factor is multiplied by the ship-measured water depth to obtain the reference vertical gravity gradient anomaly. The reference vertical gravity gradient anomaly is subtracted from the vertical gravity gradient anomaly to obtain the residual vertical gravity gradient anomaly of the sounding point, and the vertical gravity gradient anomaly field of the study area is obtained by interpolation.

[0022] After obtaining the shortwave gravity anomaly, the longwave gravity anomaly of each point is obtained by subtracting the shortwave gravity anomaly from the measured single-beam point gravity anomaly. The longwave gravity anomaly is gridded, and then the gridded longwave gravity anomaly is subtracted from the gridded gravity anomaly to obtain the gridded shortwave gravity anomaly.

[0023] Compared with the existing technology, the present invention has the following beneficial effects: based on the machine learning method, the present invention proposes a method for inverting the seabed topography according to the terrain unit by combining gravity anomalies and vertical gravity gradient anomalies, and uses the powerful nonlinear mapping ability of machine learning to solve the problem that the traditional inversion method ignores gravity information and the effects of nonlinear terms of seabed topography, thereby improving the inversion accuracy in areas with large terrain fluctuations and rapid changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a technical flow chart of the present invention;

[0025] Figure 2 It is the BP neural network structure diagram used in the present invention;

[0026] Figure 3 This is the flow chart for obtaining feature datasets. DETAILED DESCRIPTION

[0027] 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.

[0028] A method for inverting seafloor topography using machine learning based on topographic unit partitioning, comprising:

[0029] S1. Separate shortwave gravity anomalies from gravity anomalies;

[0030] S2. Separate the residual vertical gravity gradient anomaly from the vertical gravity gradient anomaly;

[0031] S3. Gravity anomalies, shortwave gravity anomalies, vertical gravity gradient anomalies, and residual vertical gravity gradient anomalies are combined into a feature dataset required for a machine learning model. The feature dataset constitutes a training set, and the gridded feature dataset constitutes a prediction set. The machine learning model uses a BP neural network model.

[0032] S4. Perform a correlation analysis on the training set and the water depth data, using the Pearson correlation coefficient to measure the correlation between the training set and the water depth data. Based on the correlation between the training set and the water depth data, the weight of the training set in the model training is determined. All training sets are input into the machine learning model for training, and the prediction set is then input into the trained machine learning model to obtain the water depth value for the entire area.

[0033] S5. Grid the water depth values ​​of the entire area to obtain a seabed topography model of the entire area;

[0034] The topography and landforms of the area are determined by the seabed topography model of the entire area. The topography of the study area is divided into topographic units according to the three topographic features of ridges, trenches, and basins. The ship-surveyed water depth data and gravity data of each sub-area are processed by steps S1, S2, and S3 respectively to obtain the characteristic data sets of each sub-area. The characteristic data sets of the ship-surveyed water depth control point positions are used as partition training sets, and the partition training sets are input into the machine learning model for training. The partition prediction sets are then input into the machine learning model to obtain partition prediction results, and the partition prediction results are fused to obtain the final water depth of the entire area.

[0035] S1 includes:

[0036] Measured shortwave gravity Δg at a single beam point res Use control point j n The water depth is calculated using the Bouguer plate formula:

[0037]

[0038] Where, Indicates that at control point j n The short-wave component on the surface of the ocean; G is the gravitational constant; Δρ is the optimal density difference constant between seawater and the oceanic crust; D represents the reference water depth, which is the maximum water depth of the shipborne bathymetric data; is the water depth value of the control point;

[0039] The iterative method is used to calculate the correlation and root mean square error between the predicted water depth and the measured water depth corresponding to different density difference constants. The density difference constant with the minimum root mean square error and the maximum correlation coefficient is the optimal value.

[0040] S2 includes:

[0041] Considering gravity data and water depth as two different signals, the coherence between them is:

[0042]

[0043] in, is the cross-spectral coherence function; G(k) and H(k) represent the Fourier transform of gravity signal and terrain signal respectively; G * (k), H * (k) represents the complex conjugate of G(k) and B(k), respectively;

[0044] The detrended water depth data and vertical gravity gradient anomaly were analyzed for coherence. Linear regression is performed on the data within the band range greater than 0.5, and the band vertical gravity gradient anomaly-band water depth proportional factor is obtained. The obtained proportional factor is multiplied by the ship-measured water depth to obtain the reference vertical gravity gradient anomaly. The reference vertical gravity gradient anomaly is subtracted from the vertical gravity gradient anomaly to obtain the residual vertical gravity gradient anomaly of the sounding point, and the vertical gravity gradient anomaly field of the study area is obtained by interpolation.

[0045] After obtaining the shortwave gravity anomaly, the longwave gravity anomaly of each point is obtained by subtracting the shortwave gravity anomaly from the measured single-beam point gravity anomaly. The longwave gravity anomaly is gridded, and then the gridded longwave gravity anomaly is subtracted from the gridded gravity anomaly to obtain the gridded shortwave gravity anomaly.

[0046] The technical process of the present invention is as follows Figure 1 First, the gravity anomaly, vertical gravity gradient anomaly and ship-measured water depth data are taken along the same grid points and placed in the same document to form a data set. The data set is then divided into a training set and a prediction set. The training set undergoes an overall correlation analysis and is imported into the neural network model for training. After the training is completed, the training set is combined to form a pre-estimated water depth model. At the same time, the training set is partitioned by terrain units to form multiple training sets, which are subjected to partition correlation analysis and then imported into the neural network model for training. After multi-model training, a multi-region water depth model is formed, which is then fused to form a full-region seabed terrain model. The neural network model in the embodiment is specifically a BP neural network, that is, the machine learning model, such as Figure 2 , four types of data are input to the input layer, processed by the hidden layer, and form the output layer, which finally outputs the predicted water depth. The process of obtaining the feature data set is as follows Figure 3Gravity anomalies and ship-surveyed bathymetry data were applied to the Bouguer plate formula to form shortwave gravity anomalies at the ship-surveyed points. Detrending, coherence analysis, and linear regression were performed on the ship-surveyed bathymetry data and vertical gravity gradient anomalies to obtain a scaling factor. This factor was multiplied by the ship-surveyed bathymetry data to obtain a reference vertical gravity gradient anomaly at the ship-surveyed points. This factor was then subtracted from the vertical gravity gradient anomaly to obtain a residual vertical gravity gradient anomaly at the ship-surveyed points. This residual gravity gradient anomaly was then fused with the shortwave gravity anomaly at the ship-surveyed points to form a characteristic dataset. The error characteristics of the inversion results of the seafloor topography model obtained by traditional gravity geological inversion were statistically analyzed with those of the unpartitioned BP model and the partitioned BP-S model. The error statistics for the three model inversion results and 14,000 single-beam checkpoints are shown in Table 1.

[0047] Table 1 Statistics of internal check errors of different models;

[0048] Model Root mean square error / m Average relative error BP neural network 89 2.07％ Method of the present invention 47 1.45％ Gravity geological method 91 2.82％ ; It can be seen from Table 1 that the method of the present invention has advantages in both root mean square error and mean relative error.

[0049] 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 inverting seabed topography using machine learning based on terrain unit partitioning, characterized in that: include: S1. Separate shortwave gravity anomalies from gravity anomalies; S2. Separate the residual vertical gravity gradient anomaly from the vertical gravity gradient anomaly; S3. Gravity anomalies, shortwave gravity anomalies, vertical gravity gradient anomalies, and residual vertical gravity gradient anomalies are combined into a feature dataset required for a machine learning model. The feature dataset constitutes a training set, and the gridded feature dataset constitutes a prediction set. The machine learning model uses a BP neural network model. S4. Perform a correlation analysis on the training set and the water depth data, using the Pearson correlation coefficient to measure the correlation between the training set and the water depth data. Based on the correlation between the training set and the water depth data, the weight of the training set in the model training is determined. All training sets are input into the machine learning model for training, and the prediction set is then input into the trained machine learning model to obtain the water depth value for the entire area. S5. Grid the water depth values ​​of the entire area to obtain a seabed topography model of the entire area; The topography of the area is determined by the seabed topography model of the entire area. The terrain of the study area is divided into topographic units according to the three topographic features of ridges, trenches, and basins. The ship-surveyed water depth data and gravity data of each sub-area are processed in steps S1, S2, and S3 respectively to obtain the characteristic data set of each sub-area. The characteristic data set of the ship-surveyed water depth control point position is used as the partition training set. The partition training set is input into the machine learning model for training. The partition prediction set is then input into the machine learning model to obtain the partition prediction result. The partition prediction results are then integrated to obtain the final water depth of the entire area. S1 includes: Measured shortwave gravity Δg at a single beam point res Use control point j n The water depth is calculated using the Bouguer plate formula: Where, Indicates that at control point j n The short-wave component on the surface of the ocean; G is the gravitational constant; Δρ is the optimal density difference constant between seawater and the oceanic crust; D represents the reference water depth, which is the maximum water depth of the shipborne bathymetric data; is the water depth value of the control point; The iterative method is used to calculate the correlation between the predicted water depth and the measured water depth corresponding to different density difference constants and the root mean square error. The density difference constant with the minimum root mean square error and the maximum correlation coefficient is the optimal value. S2 includes: Considering gravity data and water depth as two different signals, the coherence between them is: in, is the cross-spectral coherence function; G(k) and H(k) represent the Fourier transform of gravity signal and terrain signal respectively; G * (k), H * (k) represents the complex conjugate of G(k) and H(k), respectively; The detrended water depth data and vertical gravity gradient anomaly were analyzed for coherence. Linear regression is performed on the data within the band range greater than 0.5, and the band vertical gravity gradient anomaly-band water depth proportional factor is obtained. The obtained proportional factor is multiplied by the ship-measured water depth to obtain the reference vertical gravity gradient anomaly. The reference vertical gravity gradient anomaly is subtracted from the vertical gravity gradient anomaly to obtain the residual vertical gravity gradient anomaly of the sounding point, and the vertical gravity gradient anomaly field of the study area is obtained by interpolation.

2. The method for inverting seabed topography using machine learning based on terrain unit partitioning according to claim 1, characterized in that: After obtaining the shortwave gravity anomaly, the longwave gravity anomaly of each point is obtained by subtracting the shortwave gravity anomaly from the measured single-beam point gravity anomaly. The longwave gravity anomaly is gridded, and then the gridded longwave gravity anomaly is subtracted from the gridded gravity anomaly to obtain the gridded shortwave gravity anomaly.

Images