A measured gravity and magnetic data denoising method based on prior knowledge deep learning

Abstract

The application provides a gravity and magnetic data denoising method based on prior knowledge deep learning, which comprises the following steps: constructing a data set fusing prior knowledge, performing artificial filtering processing on gravity and magnetic data to obtain a denoising result, slicing the result to form a label, simulating noise data, superimposing the noise data on the label to form input data of a deep learning network model corresponding to the label one by one, building the deep learning network model, constructing a loss function fusing prior knowledge to optimize the prediction result of the deep learning, training the deep learning network model and optimizing the hyperparameters, inputting gravity and magnetic data of a prediction area into the trained deep learning network model to obtain a prediction result of the deep learning and perform optimization processing on the prediction result to obtain a final gravity and magnetic data denoising result. The application can effectively process noisy gravity and magnetic data obtained through an aerial, shipborne or satellite platform, and remove a certain amount of noise caused by platform motion, instrument observation or surrounding environment interference.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of gravity and magnetic exploration, and particularly relates to a gravity and magnetic data denoising method based on prior knowledge deep learning. BACKGROUND

[0002] In the technical field of gravity and magnetic exploration, the gravity and magnetic data obtained through aerial, shipborne and satellite platforms are all disturbed by a certain amount of noise signals, even covering useful information, due to platform motion, instrument observation or surrounding environmental interference, which seriously affects the subsequent processing and interpretation application of the gravity and magnetic data. The manual filtering method such as Gaussian filtering method and the selection of filtering parameters have certain experience, and the processing process is also relatively complex. The existing deep learning method lacks the introduction of prior knowledge, and in the construction process of the data set, the label data is mostly obtained by simple structure of the underground geological body forward simulation, so that the simulated label data is relatively simple and cannot truly simulate the gravity and magnetic data. SUMMARY

[0003] In view of the technical problems in the background art, the application provides a gravity and magnetic data denoising method based on prior knowledge deep learning, which has reasonable concept and can effectively process the noisy gravity and magnetic data obtained through aerial, shipborne and satellite platforms, remove a certain amount of noise signals generated due to platform motion, instrument observation or surrounding environmental interference, and even cover useful information, so as to provide high-quality basic data for the subsequent processing and interpretation application of the gravity and magnetic data.

[0004] To solve the above technical problems, the application provides a gravity and magnetic data denoising method based on prior knowledge deep learning, which mainly includes the following steps:

[0005] (1) Training stage

[0006] (1.1) Construct a data set fused with prior knowledge, on the basis of analyzing the noise characteristics and composition of the gravity and magnetic data, perform artificial filtering processing on the gravity and magnetic data to obtain a denoising result, slice the result to form a label, simulate noise data according to the noise characteristics of the gravity and magnetic data, superimpose the noise data on the label to form input data of a deep learning network model corresponding to the label one by one;

[0007] (1.2) According to the gravity and magnetic data denoising requirement, build a deep learning network model;

[0008] (1.3) Construct a loss function fused with prior knowledge, the harmonic field is a feature of the gravity field and has smooth characteristics, add a Laplace constraint in the loss function to optimize the prediction result of the deep learning;

[0009] (1.4) training and hyperparameter tuning of the deep learning network model, stopping training when a set number of iterations or a loss function reaches a threshold, and saving the deep learning network model;

[0010] (2) prediction phase

[0011] (2.1) inputting the gravity and magnetic data of the prediction area into the trained deep learning network model to obtain a deep learning prediction result;

[0012] (2.2) performing fine optimization processing on the deep learning prediction result to obtain a final gravity and magnetic data denoising result.

[0013] The gravity and magnetic data denoising method based on prior knowledge deep learning, wherein the specific process of step (1.1) of constructing a data set that fuses prior knowledge is: starting from original satellite gravity data containing noise, smooth denoising results are obtained by performing omnidirectional filtering on the noisy gravity data; then sample label data is generated by cropping according to a set window size and a moving step; and then simulated noise data is added to the corresponding sample label according to the noise characteristics of gravity, thereby forming a large number of sample sets containing sample data and label data.

[0014] The gravity and magnetic data denoising method based on prior knowledge deep learning, wherein step (1.1) is to select original satellite gravity data containing noise in the area with a longitude of 45°-90° and a latitude of -45°-0°, perform omnidirectional filtering on the data with a grid spacing of 0.075° to obtain smooth denoising results, crop the denoising grid using a moving window, select a window size of 160x160 for the sample size, and a moving step of 4 times the grid spacing to obtain 12100 sample label data; and for each sample label data, add noise data simulated by the following formula (7):

[0015]

[0016] In the above formula (7), A is the simulated strip noise amplitude, κ is the wavelength, α is the direction angle, a is the wavelength difference, R0 is the Gaussian white noise, x is the horizontal coordinate of the noise data calculation point, and y is the vertical coordinate of the noise data calculation point.

[0017] The gravity and magnetic data denoising method based on prior knowledge deep learning, wherein step (1.2) is to first construct a mapping relationship between noisy data and noise-free data by learning a large number of training data for original satellite gravity data denoising, and this mapping relationship can be expressed as:

[0018] d2=Net(d1,θ) (1);

[0019] In equation (1) above, d1 and d2 are the input and output data of the deep learning network, respectively, and θ is the parameter of the deep learning network.

[0020] The proposed method for denoising gravity and magnetic data based on prior knowledge deep learning comprises: a deep learning network model with a 6-layer structure, consisting of 25 two-dimensional convolutional layers, 5 downsampling layers, 5 upsampling layers, and 5 skip connections; the input data d1 and output data d2 of the deep learning network model have a resolution of 160×160, and the resolutions of the data in layers 2 to 6 are 80×80, 40×40, 20×20, 10×10, and 5×5, respectively. Conversely, the number of channels increases by a factor of 2 from 64 in layer 1 to 2048 in layer 6; and each of the two-dimensional convolutional layers and downsampling layers consists of one two-dimensional convolution operator, one batch normalization operator, and one Leaky-ReLU activation function, while the upsampling layer is implemented through two-dimensional transposed convolution.

[0021] The gravity and magnetism data denoising method based on prior knowledge deep learning, wherein the specific process of training the deep learning network model in step (1.4) is as follows: during network training, the input data is from... Figure 1 The data enters the network from the top left corner and extracts features at the current layer scale through two consecutive convolutional operations in the first layer of the deep learning network. Then, downsampling is performed to halve the resolution of the data while retaining the features obtained from the previous layer. This process is repeated until the last layer, continuously increasing the receptive field to gradually extract features from local to global information in the input data. Then, upsampling is performed on the right side to restore the resolution of the data layer by layer, and information fusion with the same-resolution features on the left side is achieved through skip connections. Finally, five two-dimensional convolutional layers are used at the output of the network to obtain the denoised result of the noisy data.

[0022] The gravity and magnetism data denoising method based on prior knowledge deep learning, wherein the specific process of step (1.3) is as follows:

[0023] First, the Dice function is introduced into gravity data denoising to characterize the similarity or overlap between the denoised data and the original data. When the Dice function is 1, it indicates that the denoised data and the original data completely overlap. Therefore, the Dice loss function is constructed as follows:

[0024]

[0025] In equation (2) above, d represents the predicted value and the theoretical value of the denoised gravity data, respectively; N represents the number of training samples; and M represents the number of grid points per sample. i Let j be a variable indicating the change in the number of training samples, and j be a variable indicating the change in the number of grid points per sample.

[0026] The Laplace operator is introduced again to constrain the smoothness of the prediction result of the deep learning network model; the Laplace operator can be expressed as the gravity data denoising prediction value The sum of the two-dimensional partial derivatives in the x direction and the y direction on the two-dimensional grid is respectively:

[0027]

[0028] The Laplace loss can be expressed in the following discrete form:

[0029]

[0030] Therefore, the total loss function, which is the sum of the above formula (2) and formula (4), can be expressed as:

[0031]

[0032] In the above formula (5), λ is a weight coefficient, used to adjust the proportion between the Dice loss and the Laplace loss.

[0033] The gravity and magnetic data denoising method based on prior knowledge deep learning, wherein the fine optimization processing mode of the prediction result of the deep learning in the step (2.2) is: the prediction result of the deep learning is low-pass filtered by introducing a Gaussian filter, thereby obtaining the final gravity data denoising result, wherein the Gaussian kernel used is:

[0034]

[0035] By adopting the above technical solution, the present application has the following beneficial effects:

[0036] The gravity and magnetic data denoising method based on prior knowledge deep learning has reasonable conception, can be widely applied in the field of gravity and magnetic exploration technology, can effectively process the noisy gravity and magnetic data obtained through an aerial, shipborne or satellite platform, remove noise signals of a certain amount, even useful information, generated due to platform motion, instrument observation or surrounding environmental interference, and thus provide high-quality basic data for subsequent processing and interpretation of gravity and magnetic data.

[0037] Compared with the artificial filtering method, the present application does not need to perform complex filtering parameter selection, and can avoid the human error caused by artificial selection of filtering parameters in the artificial filtering method; compared with the existing deep learning method, the present application has the advantages of introducing prior knowledge from two aspects, one is that the artificial filtering processing result is used as a label, noise data is simulated according to the noise characteristics of the gravity and magnetic data, and is superimposed on the label to form the input data of the deep learning network model corresponding to the label; the other is that a loss function fusing prior knowledge is constructed, according to the smooth characteristics of the gravity field, a Laplace constraint is added in the loss function to optimize the prediction result of the deep learning. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed in the following description of the specific embodiments or the prior art will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0039] Figure 1 The workflow diagram of the present application based on prior knowledge deep learning gravity and magnetic data denoising method;

[0040] Figure 2 The sample set label data clipping schematic diagram involved in the present application based on prior knowledge deep learning gravity and magnetic data denoising method;

[0041] Figure 3 The sample and label example diagram involved in the present application based on prior knowledge deep learning gravity and magnetic data denoising method ((a, b), (c, d), (a, b), (e, f) are the sample and label of serial numbers 11, 200, 2100 respectively);

[0042] Figure 4 The gravity data denoising deep learning network model schematic diagram involved in the present application based on prior knowledge deep learning gravity and magnetic data denoising method;

[0043] Figure 5 The training set and validation set loss curve diagram involved in the present application based on prior knowledge deep learning gravity and magnetic data denoising method;

[0044] Figure 6 The verification data example diagram involved in the present application based on prior knowledge deep learning gravity and magnetic data denoising method ((a) verification data (b) verification data denoising theoretical value (c) verification data noise);

[0045] Figure 7The verification data denoising result graph (a) U-net denoising result (b) U-net removed noise (c) U-net+GF denoising result (d) U-net+GF removed noise) involved in the prior knowledge deep learning-based gravity and magnetic data denoising method of the application;

[0046] Figure 8 The test area lunar Bouguer gravity data graph involved in the prior knowledge deep learning-based gravity and magnetic data denoising method of the application;

[0047] Figure 9 The real data denoising effect comparison graph (a) ODF denoising result (b) ODF removed noise (c) the denoising result of the method of the application (d) the removed noise of the method of the application) involved in the prior knowledge deep learning-based gravity and magnetic data denoising method of the application. DETAILED DESCRIPTION

[0048] The technical solutions of the application will be described below in a clear and complete manner with reference to the drawings. Obviously, the described embodiments are only some of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0049] The application will be further explained and described below with reference to specific embodiments.

[0050] As shown in the figure, the prior knowledge deep learning-based gravity and magnetic data denoising method provided by the embodiment has the following specific process: Figure 1 S100, training phase

[0051] S101, first, construct a data set fused with prior knowledge, on the basis of analyzing the noise characteristics and composition of gravity and magnetic data, perform artificial filter processing on the gravity and magnetic data by using a filter method such as Gaussian filter, obtain a denoising result based on the artificial filter method, slice the result according to a certain step length to form a large number of labels, simulate noise data according to the noise characteristics of the gravity and magnetic data, superimpose the noise data on the labels to form input data of a deep learning network model corresponding to the labels one by one;

[0052] S102, build a deep learning network model according to the gravity and magnetic data denoising requirement;

[0053] S103, construct a loss function fused with prior knowledge, the harmonic field is a feature of the gravity field and has a smooth characteristic, and a Laplace constraint is added to the loss function;

[0054]

[0055] ​S104, training and hyperparameter tuning are performed on the deep learning network model, and the training is stopped when the set iteration number is met or the loss function reaches the threshold value, and the deep learning network model is saved;

[0056] S200, prediction stage

[0057] First, the gravity and magnetic data of the prediction area are input into the trained deep learning network model to obtain the deep learning prediction result; then, the deep learning prediction result is finely optimized by a filtering method such as Gaussian filtering to obtain the final gravity and magnetic data denoising result.

[0058] The present application adopts a supervised deep learning method, and the data processing workflow is as shown in Figure 1

[0059] In the training data set stage, first, noise-free gravity and magnetic data are constructed as output labels, and these labels are the results of denoising by an artificial filtering method, that is, the labels contain prior knowledge of useful signals of the gravity and magnetic field; then, the sum of the labels and the gravity and magnetic noise generated by simulation is used as the input data of the deep neural network, that is, the input data contains the experience of artificial noise recognition; on this basis, the deep learning network model is trained by fine-tuning the hyperparameters to obtain the optimal neural network parameters. In the prediction stage, the trained network is used to predict the noisy gravity and magnetic data to obtain the denoised gravity and magnetic data, and according to the smooth property of the gravity and magnetic field as a harmonic field, the deep learning prediction result is optimized by an artificial filtering method to obtain the final result. The key to the above process is to develop a sufficient number of noisy gravity and magnetic data sets with high accuracy to fully and objectively reflect the complexity of the gravity and magnetic noise and the characteristics of the gravity and magnetic field.

[0060] In this embodiment, the noisy lunar gravity data denoising is taken as an example, and the specific process of constructing the data set fused with prior knowledge in the above step S101 is as follows:

[0061] ​Deep learning is a technology that simulates the learning process of human brain by building and training neural network model, in which the construction of data set plays a crucial role. The construction of deep learning data set in the present application is different from the previous scholars who start from the theoretical model of underground geological body and obtain the smooth denoising result by filtering the noisy satellite gravity data in all directions in a certain area, and then cut a large number of sample label data according to a certain window size and moving step, and then add simulated noise data to the corresponding sample label according to the noise characteristics of the moon gravity, so as to form a large number of sample sets containing sample data and label data. The advantage of the above data set construction is that the deep learning network model can not only learn the artificial processing experience, but also learn more rich noise characteristics without regional limitation, so in the prediction stage of other areas, the denoising result can be close to or even better than the artificial denoising result.

[0062] Specifically, the present application selects the original satellite gravity data with noise in the region of longitude 45°-90° and latitude-45°-0°, the grid spacing is 0.075°, and the smooth denoising result (as shown in Figure 2 ) is obtained by filtering in all directions. The denoising grid is cut by using a moving window, the window size is selected as the sample size (160x160), and the moving step is 4 times the grid spacing, so as to obtain 12100 sample label data. For each sample label data, noise data simulated by the following method is added:

[0063]

[0064] Wherein A is the simulated strip noise amplitude, κ is the wavelength, α is the direction angle, a is the wavelength difference, R0 is the Gaussian white noise, x is the horizontal coordinate of the noise data calculation point, and y is the vertical coordinate of the noise data calculation point.

[0065] In the present application, A=1, κ is randomly selected from the integers 3-6, a is randomly selected from the integers 1-8, and α=0 0 , 30 0 , 60 0 , 90 0 , 120 0 , 150 0 indicates 6 direction strip noise, and R0 is Gaussian white noise with standard deviation 1 and mean 0. The sum of the sample label data and the simulated noise data is obtained, and 12100 samples and corresponding labels are obtained, the first 10100 are selected as the training set, and the last 2000 are selected as the verification set. Figure 3 The serial numbers of the samples and labels are 11, 200 and 2100, the left side is the sample data with noise, and the right side is the label data without noise.

[0066] The step S102 is to solve the denoising problem of the original satellite gravity data by using the deep learning technology, and the core is to construct a complex mapping relationship between the noisy data and the noise-free data by learning a large amount of training data, and the mapping relationship can be expressed as:

[0067] d2=Net(d1,θ) (1);

[0068] Wherein d1 and d2 are input data and output data of the deep learning network model respectively, and θ is a parameter of the deep learning network model.

[0069] As shown in Figure 4 , the application designs a deep learning network model based on U-Net architecture for denoising the original satellite gravity data. The deep learning network model designed by the application has a 6-layer structure, which is composed of 25 two-dimensional convolution layers (kernel size is 3*3, step is 1), 5 down-sampling layers (kernel size is 2*2, step is 2), 5 up-sampling layers (kernel size is 2*2, step is 2) and 5 skip connections. The red number represents the grid size of the data, and the black number represents the channel number. The resolution of the input data and the output data is 160*160, and the resolution of the data of the 2nd to 6th layers is 80*80, 40*40, 20*20, 10*10 and 5*5 respectively. The channel number is increased from 64 in the 1st layer to 2048 in the 6th layer in the form of 2 times. Among them, the two-dimensional convolution layer and the down-sampling layer are composed of one two-dimensional convolution operator, one batch normalization operator and one Leaky-ReLU activation function, and the up-sampling layer is realized by two-dimensional transpose convolution.

[0070] The specific process of training the deep learning network model in the above step S104 is as follows:

[0071] During the network training, the input data enters the network from the upper left corner of Figure 1 , and the feature information of the current scale is extracted through the first layer of continuous two convolution operations. Then, the down-sampling operation is performed to halve the resolution of the data and retain the feature information obtained in the upper layer. The above operation is repeated until the last layer, in which the receptive field is continuously increased, and the feature extraction from local information to global information of the input data is gradually realized. Then, the up-sampling operation on the right side is entered, the resolution of the data is recovered layer by layer, and the information fusion with the same resolution feature on the left side is realized through the skip connection. Finally, 5 two-dimensional convolution layers are used at the output end of the network to obtain the denoising result of the noisy data.

[0072] The specific process of the above step S103 is as follows:

[0073] In deep learning, the loss function plays a crucial role; it is a non-negative real-number function that measures the difference between the model's predicted output and the true target value. The Dice function focuses on the overlap between the predicted result and the true label, which can better preserve the details and structure of the image. This invention introduces it into gravity data denoising to characterize the similarity or overlap between the denoised data and the original data. When the Dice function is 1, it indicates that the denoised data and the original data completely overlap. Therefore, the Dice loss function is constructed as follows:

[0074]

[0075] in d represents the predicted value and the theoretical value of gravity data after denoising, respectively; N represents the number of training samples; M represents the number of grid points per sample; i represents the variable indicating the change in the number of training samples; and j represents the variable indicating the change in the number of grid points per sample.

[0076] The Dice loss function can constrain the similarity between denoised data and original data, but it struggles to achieve a smooth transition between adjacent data. However, Bouguer gravity anomalies caused by underground geological bodies exhibit smoothness on a plane. Therefore, the Laplacian operator is introduced to constrain the smoothness of the prediction results. The Laplacian operator can be represented as the denoised prediction value of gravity data. The sum of the two-dimensional partial derivatives with respect to the x and y directions on the two-dimensional grid:

[0077]

[0078] The Laplace loss can then be expressed in the following discrete form:

[0079]

[0080] Therefore, the overall loss function can be expressed as:

[0081]

[0082] Where λ is the weighting coefficient, used to adjust the ratio between Dice loss and Laplace loss.

[0083] The validation data denoising results show that the denoised prediction values ​​of gravity data constrained by Dice loss and Laplace loss are... A small amount of high-frequency noise still exists. Therefore, a Gaussian filter is introduced to perform low-pass filtering on the deep learning prediction results, thereby obtaining the final denoising result of the gravity data. The Gaussian kernel used is:

[0084]

[0085] Comparative analysis of results:

[0086] After setting up the deep neural network architecture and preparing the training and validation sets, training of the neural network can begin. After multiple tests and adjustments to key hyperparameters, the final settings were a learning rate of 0.01, a batch size of 60, a dropout rate of 0.2, and a loss function threshold of 0.0001. This invention uses the Adam algorithm to optimize the deep neural network parameters. The computer configuration was an AMD Ryzen 9 7945HX with Radeon Graphics 2.50GHz, 16.0GB of RAM, and PyTorch version 2.1.2. The loss curves for the training and validation sets are shown below. Figure 5 As shown, in the initial stage of training, the losses of both the training set and the validation set are relatively large, but as the number of iterations increases, the loss curves decay rapidly, and eventually the two loss curves tend to stabilize and approach zero, without overfitting or underfitting.

[0087] Denoising results of model data:

[0088] Figure 6 Here is an example of model data, where 'a' represents the model data, 'b' is the theoretical value of the denoised model data (label), and 'c' represents the noise contained in the model data (the difference between 'a' and 'b'). The U-Net neural network built using this invention can be used to learn... Figure 7 The denoising result of a Figure 7 b represents the noise removed by the U-Net neural network, compared to... Figure 6 b and Figure 6 c. Most of the noise has been removed, but a small amount of high-frequency interference still remains. The results after Gaussian filtering optimization are shown below. Figure 7 As shown in Figure c, the comparison shows that the denoising method combining deep learning and manual filtering proposed in this invention can achieve satisfactory denoising results.

[0089] Real data denoising results:

[0090] Based on the satisfactory denoising results obtained from the theoretical data, original noisy satellite gravity data within the longitude range of 6°–18° and latitude range of 0°–12° were selected. Figure 8 The test was conducted with a grid spacing of 0.075°. A comparison of the denoising results of the proposed denoising method and the ODF method is shown below. Figure 9 As shown, where Figure 9 9a represents the denoising result of the ODF method, 9b represents the noise removed by the ODF method, 9c represents the denoising result of the method of this invention, 9d represents the noise removed by the method of this invention, and 9e represents the curve comparison of the two methods on the AB cross-section. From the planar and cross-sectional results, both methods effectively remove the interference of strip noise, but the method of this invention better preserves some local anomaly information, such as... Figure 8 As shown in the red box.

[0091] The present application concept is reasonable, can effectively process the noisy gravity and magnetic data obtained by aviation, shipborne and satellite platform, remove a certain amount of noise signals caused by platform motion, instrument observation or surrounding environment interference, etc., even cover the useful information, thereby providing high-quality basic data for subsequent processing and interpretation application of gravity and magnetic data.

[0092] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not limited to them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for removing noise from gravity and magnetic data based on deep learning with prior knowledge, characterized in that, The method mainly comprises the following steps: (1) training phase (1.1) constructing a dataset fusing prior knowledge, on the basis of analyzing noise characteristics and composition of gravity and magnetic data, performing artificial filtering processing on the gravity and magnetic data to obtain a denoising result, slicing the result to form a label, simulating noise data according to noise characteristics of the gravity and magnetic data, superimposing the noise data on the label to form input data of the deep learning network model corresponding to the label one by one; (1.2) according to the denoising requirement of the gravity and magnetic data, building a deep learning network model; (1.3) constructing a loss function fusing prior knowledge, the harmonic field is a feature of the gravity field and has a smooth characteristic, a Laplace constraint is added to the loss function to optimize the prediction result of the deep learning, and the specific process is as follows: firstly, the Dice function is introduced into the gravity data denoising to represent the similarity or overlap degree between the denoised data and the original data, when the Dice function is 1, it indicates that the denoised data and the original data are completely overlapped, therefore, the Dice loss function is constructed as follows: （2）； In the above formula (2), , is the gravity data denoising prediction value, is the gravity data denoising theoretical value, is the number of training samples, is the number of grid points of each sample, is a variable indicating the change in the number of training samples, is a variable indicating the change in the number of grid points of each sample; then, the Laplace operator is introduced to constrain the smoothness of the prediction result of the deep learning network model; The Laplacian can be expressed as the gravity data denoising prediction value The two-dimensional grid on the respective Direction and The sum of the two-dimensional partial derivatives of the direction （3）； the Laplace loss can be expressed in the following discrete form: （4）； therefore, the total loss function, that is, the sum of the above formula (2) and formula (4), can be expressed as: （5）； In the above formula (5), is a weight coefficient, used to adjust the proportion between the Dice loss and the Laplacian loss. (1.4) training and hyperparameter optimization of the deep learning network model, stopping training when the set iteration number is reached or the loss function reaches a threshold value, and saving the deep learning network model; (2) prediction phase (2.1) inputting the gravity and magnetic data of a prediction area into the trained deep learning network model to obtain a prediction result of the deep learning; (2.2) performing fine optimization processing on the prediction result of the deep learning to obtain a final gravity and magnetic data denoising result.

2. The deep learning based on prior knowledge method for denoising gravity and magnetic data according to claim 1, wherein The specific process of step (1.1) constructing a dataset fusing prior knowledge is as follows: starting from the original satellite gravity data containing noise, the smooth denoising result is obtained by performing full-direction filtering on the noisy gravity data; then, sample label data is generated by cropping according to the set window size and moving step; then, a large number of sample sets containing sample data and label data are formed by adding simulated noise data to the corresponding sample label according to the noise characteristics of the gravity.

3. The deep learning based on priori knowledge method for denoising gravity and magnetic data according to claim 2, wherein, In step (1.1), the original satellite gravity data containing noise in the region with a longitude of 45°-90° and a latitude of -45°-0° is selected, the grid spacing is 0.075°, the smooth denoising result is obtained by performing full-direction filtering, the denoising grid is cropped by using a moving window, the window size is selected as 160x160 of the sample size, the moving step is 4 times the grid spacing, and 12100 sample label data are obtained; for each sample label data, noise data simulated by the following formula (7) is added: （7）； In the above equation (7), is the amplitude of the simulated strip noise, is the wavelength, is the direction angle, is the wavelength difference, is the Gaussian white noise, is the abscissa of the noise data calculation point, is the ordinate of the noise data calculation point.

4. The deep learning based on priori knowledge method for denoising gravity and magnetic data according to claim 1, wherein, In step (1.2), the mapping relationship between the noisy data and the noise-free data is first constructed by learning a large number of training data for denoising of the original satellite gravity data, and the mapping relationship can be expressed as: （1）； In the above formula (1), , are input data and output data of the deep learning network, respectively, is a parameter of the deep learning network.

5. The deep learning based on priori knowledge method for denoising gravity and magnetic data according to claim 4, wherein: The deep learning network model has a 6-layer structure, consisting of 25 two-dimensional convolutional layers, 5 downsampling layers, 5 upsampling layers, and 5 skip connections; the input data of the deep learning network model... and output data The resolution of all layers is 160×160. The resolutions of the data in layers 2 to 6 are 80×80, 40×40, 20×20, 10×10, and 5×5, respectively. Conversely, the number of channels increases by a factor of 2 from 64 in layer 1 to 2048 in layer 6. Furthermore, the two-dimensional convolutional layer and the downsampling layer are each composed of one two-dimensional convolution operator, one batch normalization operator, and one Leaky-ReLU activation function. The upsampling layer is implemented through two-dimensional transposed convolution.

6. The deep learning based on priori knowledge method for denoising gravity and magnetic data according to claim 1, wherein: The specific process of training the deep learning network model in the step (1.4) is as follows: when the network is trained, the input data enters the network from the upper left corner of figure 1, and the feature information of the scale of the current layer is extracted through the first two convolution operations of the deep learning network; then, the data resolution is halved through downsampling operation, and the feature information obtained in the upper layer is retained; the above operation is repeated until the last layer, in the process, the receptive field is continuously increased, and the feature extraction from local information to global information of the input data is gradually realized; then, the upsampling operation on the right side is entered, the data resolution is recovered layer by layer, and the information fusion with the left side features of the same resolution is realized through the jump connection; finally, five two-dimensional convolution layers are used at the output end of the network to obtain the denoising result of the noise-containing data.

7. The deep learning based on priori knowledge method for denoising gravity and magnetic data according to claim 1, wherein, The specific process of training the deep learning network model in the step (1.4) is as follows: when the network is trained, the input data enters the network from the upper left corner of figure 1, and the feature information of the scale of the current layer is extracted through the first two convolution operations of the deep learning network; then, the data resolution is halved through downsampling operation, and the feature information obtained in the upper layer is retained; the above operation is repeated until the last layer, in the process, the receptive field is continuously increased, and the feature extraction from local information to global information of the input data is gradually realized; then, the upsampling operation on the right side is entered, the data resolution is recovered layer by layer, and the information fusion with the left side features of the same resolution is realized through the jump connection; finally, five two-dimensional convolution layers are used at the output end of the network to obtain the denoising result of the noise-containing data. （6）。

Images