A method for magnetotelluric structure constrained gravity inversion based on semantic segmentation
By combining semantic segmentation and structure-guided matrix constraints with magnetotelluric data, the gravity inversion process is optimized, overcoming the shortcomings of gravity exploration methods in vertical resolution and achieving efficient and accurate identification and inversion of anomalous areas.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JILIN UNIVERSITY
- Filing Date
- 2023-07-11
- Publication Date
- 2026-05-29
AI Technical Summary
Traditional gravity exploration methods have weak vertical resolution, making it difficult to distinguish between deep and shallow anomalies. Joint inversion calculations are costly and complex, and clustering algorithm inversion results are susceptible to human interference. Effectively incorporating high-resolution structural information has become a challenge.
A magnetotelluric structure-constrained gravity inversion method based on semantic segmentation is adopted. The magnetotelluric inversion model is identified by a semantic segmentation fully convolutional neural network, a structure guidance matrix is constructed, and iterative optimization is performed by combining the Gauss-Newton-conjugate gradient method. Gravity inversion is performed only in the target anomaly area.
This improves the vertical resolution of gravity inversion, obtains clearer information on anomalous boundaries, reduces the ambiguity of inversion, and enhances the accuracy and efficiency of inversion results.
Smart Images
Figure CN116894382B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of two-dimensional gravity data inversion processing technology, and in particular proposes a high-precision method for magnetotelluric structure-constrained gravity inversion based on semantic segmentation. Background Technology
[0002] Gravity exploration is an important branch of geophysical exploration methods. It has mature applications in groundwater and geothermal resource detection, geological structure interpretation, and the exploration of deposits such as oil, coal, copper, iron, and tungsten, characterized by high efficiency and low cost. However, its vertical resolution is relatively weak, making it difficult to distinguish between deep and shallow anomalies at the same location. Compared to gravity exploration, electromagnetic exploration, rich in frequency information, and high-resolution seismic exploration offer superior vertical resolution, providing gravity methods with higher-resolution vertical information for inversion. Traditional joint inversion methods are computationally expensive and complex due to the need to simultaneously calculate multiple physical parameters, and are also affected by inconsistent convergence speeds of different methods. Furthermore, joint inversion based on clustering algorithms requires empirical adjustment of the number of cluster cores and classification thresholds, leading to human interference in the accuracy of the inversion results. Therefore, effectively incorporating high-resolution structural information into gravity inversion calculations remains a significant challenge. Summary of the Invention
[0003] The purpose of this invention is to provide a method for magnetotelluric structure-constrained gravity inversion based on semantic segmentation, thereby improving the longitudinal resolution of gravity inversion.
[0004] like Figure 1 As shown, the present invention includes the following steps:
[0005] 1) Input the actual observation data of magnetotellurics, and set the initial resistivity model, reference model, and initial regularization factor of magnetotellurics method;
[0006] 2) Establish the objective function Φ1(m1) for the inversion of two-dimensional magnetotelluric data;
[0007] 3) The objective function for inverting two-dimensional magnetotelluric data is iteratively optimized using the OCCAM method to obtain the magnetotelluric inversion resistivity model.
[0008] 4) Construct an image recognition training set composed of magnetotelluric inversion models, and establish a semantic segmentation fully convolutional neural network model;
[0009] 5) The magnetotelluric inversion model is identified and extracted by a semantic segmentation fully convolutional neural network, the target anomaly area is segmented from the background area and transformed into a structure guidance matrix composed of 0 and 1.
[0010] 6) Input the actual observed gravity data, and set the initial density model, reference model, and initial regularization factor for gravity;
[0011] 7) Establish the objective function Φ2(m2) for two-dimensional gravity data inversion based on the structural guidance matrix;
[0012] 8) The inversion range is constrained by the structure guiding matrix, and the objective function of the two-dimensional gravity data inversion is iteratively optimized by the Gauss-Newton-conjugate gradient method to obtain the density model.
[0013] In step 1), the magnetotelluric observation data, the initial resistivity model and the reference model are input, and the regularization factor is initialized. The magnetotelluric observation data is obtained by calculating the forward response of the established resistivity theoretical model.
[0014] In step 2), the objective function for the inversion of two-dimensional magnetotelluric data is constructed, and its expression is as follows:
[0015] Φ1(m1)=‖W d1 (d1-f(m1))‖ 2 +α1·||W m1 (m1-m 1ref )|| 2 (1)
[0016] Where f(m1) is the forward modeling response of the magnetotelluric data; d1 is the actual observed magnetotelluric data; W d1 W is the diagonal inverse matrix of the magnetotelluric data noise error; m1 m is the smoothing matrix for the resistivity model. 1ref α1 is the resistivity reference model; α1 is the regularization factor for magnetotellurics.
[0017] In step 3), the objective function for the inversion of two-dimensional magnetotelluric data is iteratively optimized using the OCCAM method. The expression for its resistivity model is as follows:
[0018]
[0019] Among them, A 1k Let m be the Jacobian matrix of magnetotellurics. 1k These are the resistivity model parameters for the k-th iteration.
[0020] Compared to gravity, data acquired by the electromagnetic method contains rich frequency information and has a higher vertical resolution. In order to make full use of various geophysical parameter information and effectively combine the advantages of various geophysical methods, the magnetotelluric data inversion model is selected as the structural image guide to improve the resolution of gravity inversion.
[0021] In step 4), an image recognition training set composed of magnetotelluric inversion models is constructed, and a semantic segmentation fully convolutional neural network model is established. The training set requires constructing a large number of resistivity models obtained from magnetotelluric OCCAM inversion and performing anomaly labeling. The specific composition and principle of the fully convolutional neural network (FCN) are shown below:
[0022] The data in each convolutional network layer is a three-dimensional array of h×w×d, where h is the height, w is the width, and d is the number of features or channels. The input image consists of h pixels high, w pixels wide, and d color channels. First, the input image is processed through the first convolutional layer and max pooling layer, reducing its size to half. Then, after the second convolutional layer and max pooling layer, the image size is reduced to one-quarter. Next, after the third, fourth, and fifth convolutional layers and max pooling layers, the image size is reduced to one-eighth, one-sixth, and one-third of the original size, respectively. Following the sixth and seventh convolutional layers and max pooling layers, the image size remains at one-third of the original size. Finally, the eighth layer produces a pixel-level feature map with a size one-third of the original image. This feature map is then upsampled and enlarged by a factor of two to restore it to one-sixth of the original image size before being fused with the feature map from the fourth layer. The fused feature map is then upsampled and enlarged by a factor of two to restore it to one-eighth of the original image size before being fused with the feature map from the third layer. Finally, the fused feature map is upsampled and enlarged by a factor of eight to restore it to the original image size, thus obtaining the predicted and recognized feature map.
[0023] In step 5), the semantic segmentation fully convolutional neural network segments the target region and background region of the magnetotelluric inversion model to obtain the structure guidance matrix. Since the background is black and the number of three channels is 0, we can take advantage of the fact that not all three channels are 0, assign a value of 1 to the pixel and assign a value of 0 to the other pixels, and output it in matrix form, saving it as a txt file, thus obtaining the magnetotelluric inversion structure guidance matrix R.
[0024] In step 6), the gravity observation data, the initial density model and the reference model are input, and the regularization factor is initialized. The gravity observation data is obtained by calculating the forward response of the established density theory model.
[0025] In step 7), the objective function for two-dimensional gravity data inversion based on the structural guidance matrix is established, and its expression is as follows:
[0026] Φ2(m2)=‖w d2 (d2-f(m2))‖ 2 +α2·||W m2 (m2-m 2ref )|| 2 (3)
[0027] Constraint condition: m2=R·m2 (4)
[0028] Where f(m2) is the forward response; d2 is the observed data; W d2 W is the diagonal inverse matrix of data noise error; m2 m is the model smoothing matrix; 2ref α is the reference model; α2 is the regularization factor.
[0029] In step 8), under the constraint of the structural guiding matrix, the objective function for the two-dimensional gravity data inversion is iteratively optimized using the Gauss-Newton-conjugate gradient method. The expression for the (k+1)th density model is as follows:
[0030]
[0031] m 2k+1 =R·m 2k+1 (6)
[0032] Among them, A 2k Let m be the Jacobian matrix of the gravity method. 2k These are the density model parameters for the k-th iteration.
[0033] Since the structure guidance matrix consists of 0s and 1s, the constrained gravity model is only inverted within the target region, while the background region values remain unchanged. This effectively reduces the dimensionality of the inversion solution and lowers the ambiguity of the inversion. Compared to single gravity inversion, the magnetotelluric structure-constrained gravity inversion method based on semantic segmentation can effectively improve the vertical resolution of gravity inversion and obtain clearer anomaly boundary information.
[0034] The beneficial effects of this invention are:
[0035] 1. This invention applies the semantic segmentation fully convolutional network method to the identification and extraction of magnetotelluric inversion models, which can perform feature identification of inverted resistivity models in large batches, quickly and accurately, distinguish target anomaly areas from background areas, and thus effectively obtain the structural information of resistivity models.
[0036] 2. The structural guidance matrix of this invention consists of only two elements, 0 and 1. Therefore, the density model calculated with the addition of the structural guidance matrix will have sparsity, and the model boundary will be more clearly characterized. In particular, it avoids the problem of anomalies that are difficult to distinguish due to smoothing constraints in the vertical direction.
[0037] 3. The magnetotelluric structure-constrained gravity inversion method based on semantic segmentation of this invention performs inversion calculations only in the target anomaly region, while the background region remains unchanged. This effectively reduces the scope of the inversion calculation and achieves the goal of reducing the ambiguity of the inversion. Compared with the inversion of a single gravity method and the joint inversion based on the cross-gradient method, this method more efficiently combines multiple physical property parameter information, thus making the inversion results more accurate and reliable. Attached Figure Description
[0038] Figure 1 The flowchart shows the magnetotelluric structure-constrained gravity inversion method based on semantic segmentation.
[0039] Figure 2 A schematic diagram of the theoretical model of resistivity and density;
[0040] Figure 3 This is a schematic diagram of image recognition based on magnetotelluric inversion results.
[0041] Figure 4 Schematic diagrams of density models obtained by different inversion methods. Detailed Implementation
[0042] like Figure 1 As shown, a magnetotelluric structure-constrained gravity inversion method based on semantic segmentation includes the following steps:
[0043] 1) Input the actual observation data of magnetotellurics, and set the initial resistivity model, reference model, and initial regularization factor of magnetotellurics method;
[0044] 2) Establish the objective function Φ1(m1) for the inversion of two-dimensional magnetotelluric data;
[0045] 3) The objective function for inverting two-dimensional magnetotelluric data is iteratively optimized using the OCCAM method to obtain the magnetotelluric inversion resistivity model.
[0046] 4) Construct an image recognition training set composed of magnetotelluric inversion models, and establish a semantic segmentation fully convolutional neural network model;
[0047] 5) The magnetotelluric inversion model is identified and extracted by a semantic segmentation fully convolutional neural network, the target anomaly area is segmented from the background area and transformed into a structure guidance matrix composed of 0 and 1.
[0048] 6) Input the actual observed gravity data, and set the initial density model, reference model, and initial regularization factor for gravity;
[0049] 7) Establish the objective function Φ2(m2) for two-dimensional gravity data inversion based on the structural guidance matrix;
[0050] 8) The inversion range is constrained by the structure guiding matrix, and the objective function of the two-dimensional gravity data inversion is iteratively optimized by the Gauss-Newton-conjugate gradient method to obtain the density model.
[0051] In step 1), the magnetotelluric observation data, the initial resistivity model and the reference model are input, and the regularization factor is initialized. The magnetotelluric observation data is obtained by calculating the forward response of the established resistivity theoretical model.
[0052] In step 2), the objective function for the inversion of two-dimensional magnetotelluric data is constructed, and its expression is as follows:
[0053] Φ1(m1)=‖W d1 (d1-f(m1))‖ 2 +α1·||W m1 (m1-m 1ref )|| 2 (1)
[0054] Where f(m1) is the forward modeling response of the magnetotelluric data; d1 is the actual observed magnetotelluric data; W d1 W is the diagonal inverse matrix of the magnetotelluric data noise error; m1 m is the smoothing matrix for the resistivity model. 1ref α1 is the resistivity reference model; α1 is the regularization factor for magnetotellurics.
[0055] In step 3), the objective function for the inversion of two-dimensional magnetotelluric data is iteratively optimized using the OCCAM method. The expression for its resistivity model is as follows:
[0056]
[0057] Among them, A 1k Let m be the Jacobian matrix of magnetotellurics. 1k These are the resistivity model parameters for the k-th iteration.
[0058] Compared to gravity, data acquired by the electromagnetic method contains rich frequency information and has a higher vertical resolution. In order to make full use of various geophysical parameter information and effectively combine the advantages of various geophysical methods, the magnetotelluric data inversion model is selected as the structural image guide to improve the resolution of gravity inversion.
[0059] In step 4), an image recognition training set composed of magnetotelluric inversion models is constructed, and a semantic segmentation fully convolutional neural network model is established. The training set requires constructing a large number of resistivity models obtained from magnetotelluric OCCAM inversion and performing anomaly labeling. The specific composition and principle of the fully convolutional neural network (FCN) are shown below:
[0060] The data in each convolutional network layer is a three-dimensional array of h×w×d, where h is the height, w is the width, and d is the number of features or channels. The input image consists of h pixels high, w pixels wide, and d color channels. First, the input image is processed through the first convolutional layer and max pooling layer, reducing its size to half. Then, after the second convolutional layer and max pooling layer, the image size is reduced to one-quarter. Next, after the third, fourth, and fifth convolutional layers and max pooling layers, the image size is reduced to one-eighth, one-sixth, and one-third of the original size, respectively. Following the sixth and seventh convolutional layers and max pooling layers, the image size remains at one-third of the original size. Finally, the eighth layer produces a pixel-level feature map with a size one-third of the original image. This feature map is then upsampled and enlarged by a factor of two to restore it to one-sixth of the original image size before being fused with the feature map from the fourth layer. The fused feature map is then upsampled and enlarged by a factor of two to restore it to one-eighth of the original image size before being fused with the feature map from the third layer. Finally, the fused feature map is upsampled and enlarged by a factor of eight to restore it to the original image size, thus obtaining the predicted and recognized feature map.
[0061] In step 5), the semantic segmentation fully convolutional neural network segments the target region and background region of the magnetotelluric inversion model to obtain the structure guidance matrix. Since the background is black and the number of three channels is 0, we can take advantage of the fact that not all three channels are 0, assign a value of 1 to the pixel and assign a value of 0 to the other pixels, and output it in matrix form, saving it as a txt file, thus obtaining the magnetotelluric inversion structure guidance matrix R.
[0062] In step 6), the gravity observation data, the initial density model and the reference model are input, and the regularization factor is initialized. The gravity observation data is obtained by calculating the forward response of the established density theory model.
[0063] In step 7), the objective function for two-dimensional gravity data inversion based on the structural guidance matrix is established, and its expression is as follows:
[0064] Φ2(m2)=‖W d2 (d2-f(m2))‖ 2 +α2·||W m2 (m2-m 2ref )|| 2 (3)
[0065] Constraint condition: m2=R·m2 (4)
[0066] Where f(m2) is the forward response; d2 is the observed data; W d2 W is the diagonal inverse matrix of data noise error; m2 m is the model smoothing matrix; 2refα is the reference model; α2 is the regularization factor.
[0067] In step 8), under the constraint of the structural guiding matrix, the objective function for the two-dimensional gravity data inversion is iteratively optimized using the Gauss-Newton-conjugate gradient method. The expression for the (k+1)th density model is as follows:
[0068]
[0069] m 2k+1 =R·m 2k+1 (6)
[0070] Among them, A 2k Let m be the Jacobian matrix of the gravity method. 2k These are the density model parameters for the k-th iteration.
[0071] Since the structure guidance matrix consists of 0s and 1s, the constrained gravity model is only inverted within the target region, while the background region values remain unchanged. This effectively reduces the dimensionality of the inversion solution and lowers the ambiguity of the inversion. Compared to single gravity inversion, the magnetotelluric structure-constrained gravity inversion method based on semantic segmentation can effectively improve the vertical resolution of gravity inversion and obtain clearer anomaly boundary information.
[0072] The following section verifies the magnetotelluric structure-constrained gravity inversion method based on semantic segmentation provided by this invention.
[0073] The effectiveness of the algorithm is demonstrated by performing single-property inversion, joint inversion, and constrained inversion based on semantic segmentation and image recognition on the theoretical model. A set of complex combined models, such as... Figure 2 As shown, the resistivity model is as follows: Figure 2 As shown in a, the density model is as follows: Figure 2 As shown in b. The survey area is 6 km wide and 3 km deep, and the density in the background field is assumed to be 0 g / cm³. 3 Resistivity 10 2 The anomaly has a density of Ω·m, consists of two rectangular anomalies, and has a near-surface model density of 0.5 g / cm³. 3 The resistivity is 4 Ω·m, and the density of the model near the Earth's surface is 1 g / cm³. 3The resistivity is 0.5 Ω·m. The magnetotelluric method uses 29 observation points spaced 200m apart, and selects 10 frequencies ranging from 1 to 1000 Hz for calculation. Regarding grid division, the magnetotelluric method divides the area 0-8km wide and 0-2km deep into a regular 79×20 grid, with each grid size being 100m×100m, and adds an extension region around it to make the overall grid size 89×27. The gravity method also divides the area 0-8km wide and 0-2km deep into a regular 79×20 grid, with each grid size being 100m×100m, and adds an extension region around it to make the overall grid size 85×20.
[0074] We first perform forward modeling on the magnetotelluric theoretical model to obtain the anomalous response, then perform two-dimensional inversion to obtain the magnetotelluric inverted resistivity model. This model is then input into a semantic segmentation fully convolutional neural network for recognition, semantically segmenting the target and background regions, and finally obtaining the structure guidance matrix. The recognition results are as follows: Figure 3 As shown in the figure. Next, a forward model of the gravity theory is performed to obtain the anomalous response. Then, the structural guidance matrix is imported during the gravity inversion calculation to obtain the gravity inversion model under structural constraints. We compared and tested a single gravity inversion method, a joint inversion method based on cross-gradients, and a gravity inversion method constrained by semantic segmentation. The results are shown in the figure. Figure 4 As shown.
[0075] As can be seen from the images, single gravity inversion cannot distinguish the two anomalies vertically; it can only invert the overall outline, and the vertical accuracy is not high. The joint inversion method improves the vertical resolution, but the inversion results of the anomaly boundary are relatively blurry, and the ability to depict the details of the anomaly is slightly lacking. In contrast, the inversion based on semantic segmentation constraints has a significant improvement in vertical resolution. The overall anomaly shape is very similar, the outline is clear, and the position and size are closer to the real model.
Claims
1. A magnetotelluric structure-constrained gravity inversion method based on semantic segmentation, characterized in that: Includes the following steps: 1) Input the actual observation data of magnetotellurics, set the initial resistivity model and reference model, and initialize the regularization factor; 2) Establish the objective function for 2D magnetotelluric data inversion ; 3) The objective function for inverting two-dimensional magnetotelluric data is iteratively optimized using the OCCAM method to obtain the magnetotelluric inversion resistivity model; 4) Construct an image recognition training set composed of magnetotelluric inversion models, and establish a semantic segmentation fully convolutional neural network model; 5) The magnetotelluric inversion model is identified and extracted by a semantic segmentation fully convolutional neural network, the target anomaly area is segmented from the background area and transformed into a structure guidance matrix composed of 0 and 1; 6) Input actual gravity observation data, set the initial density model and reference model, and initialize the regularization factor; 7) Establish the objective function for two-dimensional gravity data inversion based on the structural guidance matrix. ; 8) The inversion range is constrained by the structure guiding matrix, and the objective function of the two-dimensional gravity data inversion is iteratively optimized by the Gauss-Newton method to obtain the density model; In step 2), the objective function for the inversion of two-dimensional magnetotelluric data is constructed, and its expression is as follows: (1) Among them, Forward modeling response of magnetotelluric data; These are actual observational data of magnetotelluric data; This is the diagonal inverse matrix of noise error for magnetotelluric data. This is the smoothing matrix for the resistivity model; For resistivity reference model; This is the regularization factor for the magnetotelluric method; in step 3), the objective function for the inversion of two-dimensional magnetotelluric data is iteratively optimized using the OCCAM method, and its resistivity model expression is as follows: in, The Jacobian matrix of magnetotelluric data. For the first k The resistivity model parameters of the next iteration; compared with gravity, the data collected by the electromagnetic method contains rich frequency information and has a higher longitudinal resolution. In order to make full use of various geophysical parameter information and effectively combine the advantages of various geophysical methods, the magnetotelluric data inversion model is selected as the structural image guide to improve the resolution of gravity inversion. In step 5), the semantic segmentation fully convolutional neural network segments the target area and background area of the magnetotelluric inversion model to obtain the structure guidance matrix. Since the background is black and the number of three channels is 0, we can use the fact that not all three channels are 0 to assign a value of 1 to the pixel and assign a value of 0 to the other pixels, and output it in the form of a matrix and save it as a txt file, so as to obtain the magnetotelluric inversion structure guidance matrix R. In step 7), the objective function for two-dimensional gravity data inversion based on the structural guidance matrix is established, and its expression is as follows: in, Forward response; For observational data; This is the diagonal inverse matrix of the data noise error; The model smoothing matrix; For reference model; As a regularization factor; In step 8), under the constraint of the structural guiding matrix, the objective function for the inversion of two-dimensional gravity data is iteratively optimized using the Gauss-Newton method to obtain the first... k The expression for the +1 density model is as follows: in, Let Jacobian matrix be the value of the gravity method. For the first k Density model parameters for the next iteration.
2. The magnetotelluric structure-constrained gravity inversion method based on semantic segmentation according to claim 1, characterized in that: In step 4), an image recognition training set composed of magnetotelluric inversion models is constructed, and a semantic segmentation fully convolutional neural network model is established. The training set needs to construct a large number of resistivity models obtained from magnetotelluric OCCAM inversion and label anomalies. The specific composition and principle of the fully convolutional neural network are as follows: The data in each convolutional network layer is a three-dimensional array of h × w × d, where h is the height, w is the width, and d is the number of features or channels. The input image consists of h pixels high, w pixels wide, and d color channels. First, the input image passes through the first convolutional layer and max-pooling layer, reducing its size to half. Then, after the second convolutional layer and max-pooling layer, the image size is reduced to one-quarter. Next, after the third, fourth, and fifth convolutional layers and max-pooling layers, the image size is reduced to 1 / 8, 1 / 16, and 1 / 32 of its original size, respectively. Finally, after the sixth and... After seven convolutional and max-pooling layers, the image size remains at 1 / 32 of the original size. The final eighth layer yields a pixel-level feature map with a size of 1 / 32 of the original image size. This feature map is then upsampled and enlarged by a factor of two to restore it to 1 / 16 of the original image size before being fused with the feature map from the fourth layer. The fused feature map is then upsampled and enlarged by a factor of two to restore it to 1 / 8 of the original image size before being fused with the feature map from the third layer. Finally, the fused feature map is upsampled and enlarged by a factor of eight to restore it to the original image size, thus obtaining the predicted and recognized feature map.