A method for classifying shallow seabed sediments using airborne radar.

Abstract

The present invention discloses a method for classifying shallow seabed seismicity using airborne radar bathymetry, which belongs to the field of marine surveying and mapping technology. The method first extracts multivariate waveforms and terrain features of airborne radar bathymetry based on the "three-segment" waveform decomposition and quadratic surface LM fitting algorithm; then constructs a dual-branch CNN dynamic transfer learning classification model that takes both intensity and point cloud features into consideration, and effectively improves the overall stability of the network structure by performing dual training on the intensity branch and the point cloud branch; finally, the fully connected layer parameters of the point cloud branch are adaptively adjusted by the dynamic control network to achieve parameter weight migration, thereby enhancing the generalization and robustness of the overall dual-branch CNN dynamic transfer learning classification model. Through this method, the present invention realizes the construction of a classification model for shallow seabed seismicity using airborne radar bathymetry, which will provide important technical support for the urgent needs of my country's ocean, surveying and mapping, transportation, navy and other fields for high-precision bottom seismic classification in shallow waters.

Description

Technical Field

[0001] The invention belongs to the technical field of ocean surveying and mapping, and in particular relates to a method for classifying shallow seabed geophysical information using airborne radar bathymetry. Background Art

[0002] Airborne radar bathymetry technology has the characteristics of high measurement accuracy, high measurement point density, high work efficiency, strong mobility, and measurement continuity. It is particularly suitable for rapid detection of complex terrains such as shallow water areas and areas near islands and reefs, and can achieve seamless splicing of the above-water and underwater terrain of the coastline. The airborne radar bathymetry system uses a 532nm green laser with strong water penetration ability as a means of seabed detection. By obtaining the time difference between the reflected laser from the sea surface and the seabed reaching the receiver, various corrections (system calibration, attitude correction, refraction correction, and tide level correction, etc.) can be made to calculate the three-dimensional coordinates of the seabed laser point. Airborne radar bathymetry technology can obtain rich original waveform information and seabed terrain information, providing important data support for the classification of shallow seabed bottom sediments.

[0003] The seabed sediment type is an important marine environmental parameter and the basis for conducting seabed scientific research. The classification of seabed sediments has important scientific and practical significance for marine engineering construction, marine environmental monitoring, seabed scientific research, marine resource exploitation, and marine fishery development. Shallow sea areas around coastlines, islands and reefs are the sea areas with the most frequent human activities. The types of seabed sediments are very complex. The complex shallow seabed sediments play a vital role in the construction of near-shore ports, cross-sea bridges, development of offshore wind power platforms, matching field positioning, low-frequency long-range sonar range prediction, submarine sinking and concealed seabed selection, underwater targets and underwater combat environment detection.

[0004] At present, the main methods for shallow seabed sediment classification and detection include in-situ grab sampling and optical remote sensing detection. Due to the various limitations of field grab sampling, there are very few sample data with actual labels. Small sample training sets are an important factor restricting the development of seabed sediment classification. Traditional sediment classification methods often have low generalization and robustness problems for small sample training sets. To address this problem, a shallow seabed sediment classification method based on airborne radar bathymetry is proposed. The original waveform and terrain feature information of airborne radar bathymetry are combined with a dual-branch CNN (Convolutional Neural Networks) transfer learning classification method to achieve reasonable classification of shallow seabed sediments. Summary of the invention

[0005] In view of the above-mentioned technical problems existing in the prior art, the present invention proposes a method for classifying shallow seabed seismicity using airborne radar bathymetry. The method has a reasonable design and solves the problems of low generalization and robustness of the prior art for small sample training sets. By combining multivariate features with a dual-branch CNN dynamic transfer learning classification model, accurate classification of shallow seabed seismicity is achieved, providing important technical support for the urgent needs of my country's ocean, surveying and mapping, transportation, navy and other fields for high-precision seismic classification in shallow waters.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A method for classifying shallow seabed geophysical materials by airborne radar bathymetry comprises the following steps:

[0008] Step 1: Extract the multivariate waveform and terrain features of airborne radar bathymetry based on the “three-segment” waveform decomposition and quadratic surface LM (Levenberg-Marquardt) fitting algorithm;

[0009] Step 2: Build a dual-branch CNN dynamic transfer learning classification model that takes both intensity and point cloud features into consideration, and improve the overall stability of the network structure by dual training the intensity branch and the point cloud branch;

[0010] Step 3: Adaptively adjust the fully connected layer parameters of the point cloud branch through the dynamic control network to achieve parameter weight migration and enhance the generalization and robustness of the overall two-branch CNN dynamic transfer learning classification model.

[0011] Preferably, in step 1, the method for extracting the multivariate waveform and terrain features of airborne radar sounding is as follows:

[0012] Step 1.1: Extract the multivariate waveform features of airborne radar depth sounding based on the “three-segment” waveform decomposition method;

[0013] Step 1.2: Extract the multivariate terrain features of airborne radar bathymetry based on the quadratic surface LM fitting algorithm.

[0014] Preferably, the "three-segment" waveform includes Gaussian function fitting water surface echo, double exponential function fitting water body echo and Weber function fitting bottom echo; the multivariate waveform characteristics of airborne radar depth sounding include amplitude, wave width, half-wave width, waveform area, skewness, kurtosis, backscattering cross-section coefficient, residual amplitude and residual wave width; the multivariate terrain characteristics of airborne radar depth sounding include slope, Gaussian curvature, roughness, water depth standard deviation, elevation entropy, undulation, mean deviation, coefficient of variation, and concavity coefficient.

[0015] Preferably, in step 2, the method for constructing the dual-branch CNN dynamic transfer learning classification model is as follows:

[0016] Step 2.1: The airborne radar sounding amplitude information is the intensity information, and the airborne radar sounding intensity image is obtained by interpolating the extracted amplitude features;

[0017] Step 2.2: Segment the airborne radar bathymetric intensity image according to the resolution of the seabed point cloud. The segmented image unit represents only one type of bottom sediment (for mixed bottom sediments, the type of bottom sediment with the largest proportion is selected). Intensity branch classification is achieved by distinguishing the color and texture features of the image unit.

[0018] Step 2.3: Transfer the training parameters of the intensity branch to the point cloud branch, use the extracted multivariate waveform and terrain feature parameters as the input matrix of the point cloud branch, and complete the point cloud branch training based on the autonomously constructed CNN network model and transfer training parameters;

[0019] Step 2.4: To ensure that the difference between the intensity branch and the point cloud branch is getting smaller and smaller and the two branches are close to each other, the MK-MMD probability adaptation method is used to measure the distance between the two branch distributions; assuming that the intensity branch and the point cloud branch have Y t A and Y t F , using the MK-MMD probability adaptation method to measure Y t A With Y t F The difference between distance(Y t A ,Y t F ), and add the difference to the loss function of the two-branch CNN transfer learning model. The loss function is shown in formula (1):

[0020]

[0021] In the formula, z i (θ) is the point cloud branch sample x i The predicted label of ; where i = 1, 2, ..., N; θ = (θ 1 ,θ 2 ) is the parameter of the dynamic control network in the point cloud branch; y i is the label of the point cloud branch sample; λ is the regularization coefficient; P = (p 1 ,…,p m ) is the weighted parameter predicted by the dynamic regulatory network, ||·|| 1 Indicates L 1 norm;

[0022] Step 2.5: L 1The norm can make the weight matrix of the dual-branch CNN transfer learning model more sparse and alleviate the problem of model overfitting. The target loss function is optimized as shown in formula (2):

[0023]

[0024] in, L 2 =||P|| 1 ;

[0025] Step 2.6: Use the equivalent form with differentiability shown in formula (3) to express L 1 Norm:

[0026]

[0027] In the formula, ε is a small positive constant.

[0028] Preferably, in step 3, the training strategy of the dynamic control network is as follows:

[0029] Step 3.1: Get the weighted vector through the dynamic control network, transfer the vector to the point cloud branch, and dynamically control the connection parameters between the hidden layer and the output layer in the network. The update rule of is shown in formula (4):

[0030]

[0031] In the formula, is the connection parameter between the hidden layer and the output layer, m represents the number of iterations, τ M is the learning rate, Loss is the target loss function;

[0032] Step 3.2: Dynamically adjust the connection parameters between the network input layer and the hidden layer The update rule of is shown in formula (5):

[0033]

[0034] In the formula, is the connection parameter between the input layer and the hidden layer, m represents the number of iterations, τ M is the learning rate, for The first derivative of is the activation function, is the input of the jth neuron in the hidden layer, is the output of the i-th neuron in the input layer, n is the number of neurons in the last fully connected layer, is the connection parameter between the hidden layer and the output layer, To dynamically regulate the local gradient of the output layer in the network;

[0035] Step 3.3: Fix the dynamic control network and train the point cloud branch; then fix the point cloud branch, train the dynamic control network, and complete the adaptive dynamic control of the training parameters through continuous cross-iteration.

[0036] Beneficial technical effects brought by the present invention:

[0037] The present invention proposes a method for classifying shallow seabed sediments using airborne radar bathymetry. Compared with the prior art, the present invention extracts multivariate waveforms and terrain features of airborne radar bathymetry based on a "three-segment" waveform decomposition and a quadratic surface LM fitting algorithm; then a two-branch CNN transfer learning classification model that takes into account both intensity branches and point cloud branches is constructed to improve the overall stability of the network structure; the parameters of the fully connected layer are adaptively adjusted by dynamically regulating the network to enhance the generalization and robustness of the sediment classification model, thereby effectively improving the accuracy and computational efficiency of the traditional sediment classification model. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 The present invention is a flow chart of the method for classifying shallow seabed geophysical materials using airborne radar bathymetry.

[0039] Figure 2 Schematic diagram of the dual-branch CNN dynamic transfer learning classification model constructed in the present invention.

[0040] Figure 3 Schematic diagram of adaptively adjusting the fully connected layer of the dynamic control network in the present invention. DETAILED DESCRIPTION

[0041] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0042] The present invention provides a method for classifying shallow seabed geophysical materials by airborne radar sounding, the process of which is as follows: Figure 1 As shown, the following steps are included:

[0043] Step 1: Extraction of multivariate waveform and terrain features of airborne radar sounding.

[0044] According to the existing airborne radar sounding waveform decomposition algorithm, the "three-segment" waveform decomposition method is used to extract multivariate waveform features. The highly stable and efficient LM algorithm is used to establish a quadratic surface fitting model of seabed terrain points to extract multivariate terrain features.

[0045] In a further embodiment, step 1 specifically includes the following steps:

[0046] Step 1.1: Extract the multivariate waveform characteristics (amplitude, wave width, half-width, waveform area, skewness, kurtosis, backscatter cross-sectional coefficient, residual amplitude, residual wave width) of airborne radar bathymetry based on the "three-stage" (Gaussian function fits water surface echo, double exponential function fits water body echo, and Weber function fits bottom echo) waveform decomposition method;

[0047] Step 1.2: Extract the multivariate terrain features (slope, Gaussian curvature, roughness, standard deviation of water depth, elevation entropy, undulation, mean deviation, coefficient of variation, and concavity coefficient) of airborne radar bathymetry based on the quadratic surface LM fitting algorithm.

[0048] In specific implementation, the multivariate waveforms and terrain features of airborne radar bathymetry can provide an important parameter basis for the shallow seabed seismic classification model and serve as input data to improve classification accuracy.

[0049] Step 2: Build a two-branch CNN dynamic transfer learning classification model.

[0050] Excellent transfer learning classifier models can make machine learning systems more reliable and robust. Aiming at the problem of small sample data sets encountered in shallow seabed sediment classification, a two-branch CNN dynamic transfer learning sediment classification method is proposed. By constructing a two-branch CNN dynamic transfer learning classification model that takes both intensity features and point cloud features into consideration, the intensity branch and the point cloud branch are dually trained, which effectively improves the overall stability and generalization ability of the classification model network structure. Figure 2 .

[0051] In a further embodiment, step 2 specifically includes the following steps:

[0052] Step 2.1: The airborne radar sounding amplitude information is the intensity information, and the airborne radar sounding intensity image is obtained by interpolating the extracted amplitude features;

[0053] Step 2.2: Segment the airborne radar bathymetric intensity image according to the resolution of the seabed point cloud. The segmented image unit represents only one type of bottom sediment (for mixed bottom sediments, the type of bottom sediment with the largest proportion is selected). Intensity branch classification is achieved by distinguishing the color and texture features of the image unit.

[0054] Step 2.3: Transfer the training parameters of the intensity branch to the point cloud branch, use the extracted multivariate waveform and terrain feature parameters as the input matrix of the point cloud branch, and complete the point cloud branch training based on the autonomously constructed CNN network model and transfer training parameters;

[0055] Step 2.4: To ensure that the difference between the intensity branch and the point cloud branch is getting smaller and smaller and the two branches are close to each other, the MK-MMD probability adaptation method is used to measure the distance between the two branch distributions; assuming that the intensity branch and the point cloud branch have Y t A and Y t F , using the MK-MMD probability adaptation method to measure Y t A With Y t F The difference between distance(Y t A ,Y t F ), and add the difference to the loss function of the two-branch CNN transfer learning model. The loss function can be expressed as formula (1):

[0056]

[0057] In the formula, z i (θ) is the point cloud branch sample x i (i=1,2,…,N) predicted labels; θ=(θ 1 ,θ 2 ) is the parameter of the dynamic control network in the point cloud branch; y i is the label of the point cloud branch sample; λ is the regularization coefficient; P = (p 1 ,…,p m ) is the weighted parameter predicted by the dynamic regulatory network, ||·|| 1 Indicates L 1 Norm.

[0058] Step 2.5: L 1 The norm can make the weight matrix of the two-branch CNN transfer learning model more sparse, alleviating the problem of model overfitting. Furthermore, the target loss function can be optimized as follows:

[0059]

[0060] in, L 2 =||P|| 1 .

[0061] Step 2.6: Since L 1 The norm is not differentiable at zero, and it is difficult to apply to the gradient descent training method, so the following differentiable equivalent form is used to represent L 1 Norm:

[0062]

[0063] In the formula, ε is a small positive constant.

[0064] In the specific implementation, the overall classification accuracy and Kappa coefficient of the test set samples were high, the obtained bottom sediment classification results were highly consistent with the distribution of ground objects in high-definition aerial digital photos, and the overall prediction results were good.

[0065] Step 3: Dynamically regulate the network to adaptively adjust the fully connected layer of the point cloud branch to enhance the generalization and robustness of the overall bottom classification model.

[0066] In the point cloud branch module, multivariate waveforms and terrain features are used as data input, and the dynamic control network is used to adaptively adjust the fully connected layer parameters of the point cloud branch. By cross-iteratively training the dynamic control network and the point cloud branch, the training parameter weight migration is achieved to enhance the generalization and robustness of the overall bottom classification model. Figure 3 , where p 1 、p 2 、p 3 、p 4 、p i and p N is the training parameter, w 1 、w 2 、w 3 、w 4 、w i and w N is the training weight.

[0067] In a further embodiment, step 3 specifically includes the following steps:

[0068] Step 3.1: Get the weighted vector through the dynamic control network, transfer the vector to the point cloud branch, and dynamically control the connection parameters between the hidden layer and the output layer in the network. The update rule can be expressed as formula (4):

[0069]

[0070] In the formula, is the connection parameter between the hidden layer and the output layer, m represents the number of iterations, τ M is the learning rate, and Loss is the target loss function.

[0071] Step 3.2: Dynamically adjust the connection parameters between the network input layer and the hidden layer The update rule can be expressed as formula (5):

[0072]

[0073] In the formula, is the connection parameter between the input layer and the hidden layer, m represents the number of iterations, τ M is the learning rate, for The first derivative of is the activation function, is the input of the jth neuron in the hidden layer, is the output of the i-th neuron in the input layer, n is the number of neurons in the last fully connected layer, is the connection parameter between the hidden layer and the output layer, To dynamically regulate the local gradient of the output layer in the network.

[0074] Step 3.3: Fix the dynamic control network and train the point cloud branch; then fix the point cloud branch, train the dynamic control network, and complete the adaptive dynamic control of the training parameters through continuous cross-iteration.

[0075] In specific implementation, the output of the adaptive dynamic control network can provide weight prediction values ​​of training parameters for the point cloud branch, thereby improving the generalization ability for different training samples and reflecting the superior overall stability and robustness of the sediment classification model.

[0076] In summary, the present invention provides a method for classifying shallow seabed seismology using airborne radar bathymetry, the method comprising: extracting multivariate waveforms and terrain features based on waveform decomposition and quadratic surface LM fitting algorithm; constructing a dual-branch CNN dynamic transfer learning classification model, and effectively improving the overall stability of the network structure through dual training of the intensity branch and the point cloud branch; further adaptively adjusting the fully connected layer parameters of the point cloud branch through the dynamic control network to achieve parameter weight migration, thereby enhancing the generalization and robustness of the overall bottom classification model.

[0077] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for classifying shallow seabed geophysical materials by airborne radar sounding, characterized in that: The following steps are involved: Step 1: Extract the multivariate waveform and terrain features of airborne radar bathymetry based on the "three-segment" waveform decomposition and quadratic surface LM fitting algorithm; Step 2: Build a dual-branch CNN dynamic transfer learning classification model that takes both intensity and point cloud features into consideration, and improve the overall stability of the network structure by dual training the intensity branch and the point cloud branch; Step 2.1: The airborne radar sounding amplitude information is the intensity information, and the airborne radar sounding intensity image is obtained by interpolating the extracted amplitude features; Step 2.2: Segment the airborne radar bathymetric intensity image according to the resolution of the seabed point cloud. The segmented image units only represent one type of bottom material. Intensity branch classification is achieved by distinguishing the color and texture features of the image units. Step 2.3: Transfer the training parameters of the intensity branch to the point cloud branch, use the extracted multivariate waveform and terrain feature parameters as the input matrix of the point cloud branch, and complete the point cloud branch training based on the autonomously constructed CNN network model and transfer training parameters; Step 2.4: To ensure that the difference between the intensity branch and the point cloud branch is getting smaller and smaller and the two branches are close to each other, the MK-MMD probability adaptation method is used to measure the distance between the two branch distributions; assuming that the intensity branch and the point cloud branch have Y t A and Y t F , using the MK-MMD probability adaptation method to measure Y t A With Y t F The difference between distance(Y t A ,Y t F ), and add the difference to the loss function of the two-branch CNN transfer learning model. The loss function is shown in formula (1): In the formula, z i (θ) is the predicted label of the point cloud branch sample xi; where i = 1, 2, ..., N; θ = (θ1, θ2) is the parameter of the dynamic control network in the point cloud branch; y i is the label of the point cloud branch sample; λ is the regularization coefficient; P = (p1,…,p m ) is the weighted parameter predicted by the dynamic regulatory network, ·1 represents the L1 norm; Step 2.5: The L1 norm can make the weight matrix of the dual-branch CNN transfer learning model more sparse, alleviating the overfitting problem of the model; optimize the target loss function, as shown in formula (2): in, L2=P1; Step 2.6: Use the differentiable equivalent form shown in formula (3) to express the L1 norm: In the formula, ε is a very small positive constant; Step 3: Adaptively adjust the fully connected layer parameters of the point cloud branch through the dynamic control network to achieve parameter weight migration and enhance the generalization and robustness of the overall two-branch CNN dynamic transfer learning classification model.

2. The method for classifying shallow seabed geophysical materials by airborne radar sounding according to claim 1, characterized in that: In step 1, the method for extracting the multivariate waveform and terrain features of airborne radar sounding is as follows: Step 1.1: Extract the multivariate waveform features of airborne radar depth sounding based on the "three-segment" waveform decomposition method; Step 1.2: Extract the multivariate terrain features of airborne radar bathymetry based on the quadratic surface LM fitting algorithm.

3. The method for classifying shallow seabed geophysical materials by airborne radar sounding according to claim 2, characterized in that: The "three-segment" waveform includes Gaussian function fitting water surface echo, double exponential function fitting water body echo and Weber function fitting bottom echo; the multivariate waveform characteristics of airborne radar bathymetry include amplitude, wave width, half-wave width, waveform area, skewness, kurtosis, backscattering cross-section coefficient, residual amplitude and residual wave width; the multivariate terrain characteristics of airborne radar bathymetry include slope, Gaussian curvature, roughness, water depth standard deviation, elevation entropy, undulation, mean deviation, coefficient of variation and concavity coefficient.

4. The method for classifying shallow seabed geophysical materials by airborne radar sounding according to claim 1, characterized in that: In step 3, the training strategy of the dynamic control network is as follows: Step 3.1: Get the weighted vector through the dynamic control network, transfer the vector to the point cloud branch, and dynamically control the connection parameters between the hidden layer and the output layer in the network. The update rule of is shown in formula (4): In the formula, is the connection parameter between the hidden layer and the output layer, m represents the number of iterations, τ M is the learning rate, Loss is the target loss function; Step 3.2: Dynamically adjust the connection parameters between the network input layer and the hidden layer The update rule of is shown in formula (5): In the formula, is the connection parameter between the input layer and the hidden layer, m represents the number of iterations, τ M is the learning rate, for The first derivative of is the activation function, is the input of the jth neuron in the hidden layer, is the output of the i-th neuron in the input layer, n is the number of neurons in the last fully connected layer, is the connection parameter between the hidden layer and the output layer, To dynamically regulate the local gradient of the output layer in the network; Step 3.3: Fix the dynamic control network and train the point cloud branch; then fix the point cloud branch, train the dynamic control network, and complete the adaptive dynamic control of the training parameters through continuous cross-iteration.

Images