Method and apparatus for using transfer learning to locate igneous carbonate rock-type rare earth deposits

Abstract

The present application proposes a method and apparatus for transfer learning-based localization of igneous carbonate-hosted rare earth mineralization. The technology relates to electromagnetic exploration and includes obtaining exploration data for electric, magnetic, seismic, and gravity methods in the target area. Transfer learning-based localization is applied to the electric, magnetic, seismic, and gravity exploration data to determine the cross-sectional map corresponding to the anomalous position of igneous carbonate-hosted rare earth mineralization in the target area. Feature decomposition dimensionality reduction, feature enhancement, and weighted fusion processing are applied to the cross-sectional map, followed by segmentation of the igneous carbonate-hosted rare earth mineralization geological body in the fused image to obtain the spatial distribution of the detection target. The integrated analysis of multi-source geophysical data using images improves the accuracy of identification and localization of igneous carbonate-hosted rare earth mineralization.

Description

FIELD OF THE INVENTION

The present invention relates to the technical field of geophysical exploration of igneous carbonate rock-type rare earth deposits. More specifically, it pertains to a method and apparatus for using transfer learning to locate igneous carbonate rock-type rare earth deposits.

BACKGROUND OF THE INVENTION

Rare earths, also known as “industrial vitamins,” have excellent physical properties such as light, electricity, and magnetism. They are widely used in cutting-edge technology fields such as national defense, aerospace, low-carbon energy, and special materials. Rare earths are of great strategic significance to national security, the national economy, and high-tech development, and are strategic mineral resources that countries around the world are competing for.

The morphology of igneous carbonate rock-type rare earth deposits is complex. As a typical mantle-derived magmatic rock, it is formed under extensional tectonic background and is mainly related to continental rifts, mantle plumes, or large-scale strike-slip faults. The composition of carbonate rocks varies greatly. In addition to typical calcite carbonate rocks, dolomite carbonate rocks, and dolomite-calcite carbonate rocks, there are also various carbonate rocks with rich magnetite and fluorite, as well as rich biotite.

Due to the complex rock composition and uneven spatial distribution of ore-bearing carbonate rocks, and the varied tectonic shapes of ore-bearing carbonate rocks and surrounding rock slate/sandstone affected by tectonic movement, with steep dip angles and rapid lateral lithological changes, effective geophysical exploration methods and refined resolution of geophysical data processing inversion are the key issues for the distribution characteristics of carbonate rock deposits.

INVENTION SUBJECT MATTER

The present invention provides a method and device for transfer learning-based positioning of rare earth minerals of igneous carbonate rocks, which comprehensively analyzes multi-source geophysical data using images as carriers, and improves the accuracy of identification and positioning of rare earth minerals of igneous carbonate rocks.

To achieve the above-mentioned object, the technical solution adopted by the present invention includes:

A transfer learning-based positioning method for rare earth minerals of igneous carbonate rocks, including:

Obtaining electric, magnetic, seismic, and gravity exploration data of the target area to be estimated; performing transfer learning-based positioning on the electric, magnetic, seismic, and gravity exploration data to determine the sectional images corresponding to the abnormal positions of rare earth minerals of igneous carbonate rocks in the target area;

Performing feature decomposition and dimensionality reduction, feature enhancement, and weighted fusion processing on the sectional images;

Segmenting the occurrences of rare earth mineralized geological bodies in the fused images to obtain the spatial occurrence morphology of the detection target.

Preferably, performing transfer learning-based positioning on the electric, magnetic, seismic, and gravity exploration data includes:

Analyzing the electric, magnetic, seismic, and gravity exploration data to obtain the data volumes corresponding to abnormal electrical resistivity, magnetization, velocity, and density of rare earth minerals of igneous carbonate rocks; and positioning the data volumes corresponding to abnormal electrical resistivity, magnetization, velocity, and density.

Preferably, the sectional images corresponding to the abnormal positions of the data anomalies include: digital sectional images of abnormal electrical resistivity, magnetization, velocity, and density.

Preferably, determining the sectional images corresponding to the abnormal positions of rare earth minerals of igneous carbonate rocks in the target area includes:

Generating digital sectional images of abnormal electrical resistivity, magnetization, velocity, and density of the data volumes corresponding to the abnormal positions of rare earth minerals of igneous carbonate rocks, respectively, at the same depth;

Configuring the digital sectional images of abnormal electrical resistivity, magnetization, velocity, and density in a common coordinate system.

Preferably, performing image feature decomposition on the sectional images includes:

Using a transfer learning-based downsampling network to downsample the input signals and output the downsampling results, i.e., performing initial feature extraction on digital sectional images of abnormal electrical resistivity, magnetization, velocity, and density, respectively, to obtain downsampling features at different scales for each image.

Preferably, performing feature enhancement on the sectional images includes:

Using Spearman correlation to calculate the correlation coefficients between the initial features at different scales for each image;

Removing irrelevant initial features based on the comparison results between the correlation coefficients and the coefficient threshold to obtain enhanced features.

Preferably, performing weighted fusion processing on the enhanced sectional images includes:

Calculating the initial weight matrices corresponding to each sectional image using local nuclear norm and average operations;

Performing interpolation operation on the initial weight matrices using the soft-max operation and bicubic interpolation to obtain the weight coefficients corresponding to each sectional image;

Using the local energy map algorithm to maintain the structural and detail information of each sectional image;

Using the similarity measurement as the fusion decision value and comparing the fusion results calculated by the local energy with the weighted average fusion rule.

Preferably, using Spearman correlation to calculate the correlation coefficients between the initial features at different scales for each image:

The Spearman correlation coefficient r_s^k is used to represent the correlation between the initial features F_i,1:Mi^k of the i-th layer convolution module outputting M channels.

$r_s^k=1-\frac{6\sum _{i=1}^M{d}^2}{M\left(M^2-1\right)}$

where d is the rank difference between F_i,p^k and F_i,q^kp,q∈[1,2,L,M_i}; k represents the image category number; the minimum number of features is selected with r_s^k as the selected features after feature selection, represented by F_i,1:C^k.

Preferably, the ZCA algorithm is utilized for redundancy reduction on the secondary image features, to achieve enhancement of section image features, including:

calculating the covariance matrix Cov_i,m^k of X=(F_i,1^k,F_i,2^k,F_i,C^k) and perform singular value decomposition on the covariance matrix Cov_i,m^k as follows:

${\mathrm{Cov}}_{i,m}^k=F_{i,m}^k\times {\left(F_{i,m}^k\right)}^T,m\in \left\{1,2,L,C\right\}$ ${\mathrm{Cov}}_{i,m}^k=U\sum V^T$

using the following equation to calculate the secondary extracted feature

${\stackrel{)}{F}}_{i,1;C}^k:{\stackrel{)}{F}}_{i,1;C}^k=U\sqrt{{\left(\sum +\epsilon I\right)}^{-1}}{U}^T{F}_{i,C}^k$

where U and V are both unitary matrices satisfying U^TU=I,V^TV=I; Σ represents a diagonal matrix, with all elements except the main diagonal being 0, and each element on the main diagonal is called a singular value. I represents an identity matrix, and ε represents a minimum value for eliminating ill-conditioned matrices.

performing weighted fusion processing on the enhanced sectional images, comprising: calculating the initial weight matrix S_i,*^k of the image using local nuclear norm and average operation according to the following formula:

$S_{i,*}^k\left(x,y\right)=\frac{\sum _{p=x-t}^{p=x+t}\sum _{q=y-t}^{q=y+t}{{\stackrel{)}{F}}_{i,1;C}^k\left(p,q\right)}^{*}}{\left(2t+1\right)·\left(2t+1\right)}$

where p,q=∈{1,2,L,M_i}, x, and y represent the pixel coordinates, t represents the step size for iterating over all pixels, and k represents the image category number. The following equation is used to perform interpolation of the weight matrix using soft-max operation and bicubic interpolation to obtain weight coefficients ω_i^k:

${\omega }_i^k\left(x,y\right)=\frac{S_{i,*}^k\left(x,y\right)}{S_{i,*}^1\left(x,y\right)+S_{i,*}^2\left(x,y\right)},k\in \left\{1,2\right\}$

Considering the specificity and locality of underground targets, Local Energy Map is used to preserve the structural and detailed information of the image:

${\mathrm{LE}}_k\left(x,y\right)=\sum _m\sum _{n}L\left\{I_k\right\}{\left(i+m,y+n\right)}^2$

where LE_k(x,y) represents the total energy map function, L{Ik} represents the local energy map function, m×n represents the size of the neighbourhood, and the similarity measurement between I₁ and I₂ is calculated as the fusion decision value using the following equation:

$M\left(x,y\right)=\frac{2\prod _{k=1,2}\left(\sum _m\sum _{n}L\left\{I_k\right\}\left(x+m,y+n\right)\right)}{\sum _{k=1,2}{\mathrm{LE}}_k\left(x,y\right)}$

Let M(x,y)=∈[−1,1] be the similarity measure between I₁ and I₂, where a value closer to 1 indicates a stronger similarity between the images. Set a threshold value T, if M(x,y)<T, then use the weighted averaging fusion rule to calculate the fusion result:

$F_{M<T}\left(x,y\right)={\omega }_i^1\left(x,y\right)I_1\left(x,y\right)+{\omega }_i^2\left(x,y\right)I_2\left(x,y\right)$

If M(x,y)≥T, then the fusion result is calculated by comparing the local energy:

$F_{M\ge T}\left(x,y\right)=\{\begin{array}{cc}L\left\{I_1\right\}\left(x,y\right),& \mathrm{if}{\mathrm{LE}}_1\left(x,y\right)\ge {\mathrm{LE}}_2\left(x,y\right)\\ L\left\{I_2\right\}\left(x,y\right),& \mathrm{if}{\mathrm{LE}}_1\left(x,y\right)<{\mathrm{LE}}_2\left(x,y\right)\end{array}$

The fusion strategy is represented as:

$F\left(x,y\right)=\{\begin{array}{cc}{F}_{M<T}\left(x,y\right),& \mathrm{if}M\left(x,y\right)<T\\ F_{M\ge T}\left(x,y\right),& \mathrm{if}M\left(x,y\right)\ge T\end{array}$

In the second aspect, the present invention also provides a transfer learning positioning device for igneous carbonate rock-derived rare earth ore, comprising:

An image acquisition module, set to acquire electrical, magnetic, seismic, and gravity exploration data of the area to be evaluated, and perform transfer learning positioning on the electrical, magnetic, seismic, and gravity exploration data to determine the cross-sectional view corresponding to the abnormal position of igneous carbonate rock-derived rare earth ore data in the area to be evaluated;

An image processing module, set to perform feature decomposition and dimensionality reduction, feature enhancement, and weighted fusion processing on the respective cross-sectional images;

A storage positioning module, set to segment the geological bodies of igneous carbonate rock-derived rare earth ore mineralization in the fusion image, and obtain the spatial storage morphology of the detection target.

Compared with the prior art, the igneous carbonate rock-derived rare earth ore transfer learning positioning method and device provided by the present invention comprehensively analyze multi-source geophysical data through images, and accurately identify and locate the target body under big data conditions. The accuracy is ensured while improving the efficiency of comprehensive interpretation.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic flowchart of the igneous carbonate rock-derived rare earth ore transfer learning positioning method of an embodiment of the present invention;

FIG. 2 is a schematic flowchart of feature extraction of a single input image according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of feature extraction, feature dimensionality reduction, and feature enhancement from a single SI image according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of the establishment process of the weight matrix according to an embodiment of the present invention;

FIG. 5 is a schematic diagram of transfer learning according to an embodiment of the present invention;

FIG. 6 is a schematic diagram of the cross-sectional view of an embodiment of the present invention, wherein FIG. 6 (a) is a resistivity image, FIG. 6 (b) is a density image, FIG. 6 (c) is a magnetization image, and FIG. 6 (d) is a velocity image;

FIG. 7 is a schematic diagram of the transfer learning result of an embodiment of the present invention.

DETAILED DESCRIPTION

In order to make the purpose, technical scheme, and beneficial effects of the present invention clearer and more comprehensible, the embodiments of the present invention will be described below in conjunction with the accompanying drawings. It should be noted that, in the absence of conflict, the features in the embodiments of the present application can be combined with each other in any manner.

As shown in FIG. 1, the embodiment of the present invention provides an igneous carbonate rock-derived rare earth ore transfer learning positioning method, including:

S101, acquiring electrical, magnetic, seismic, and gravity exploration data of the area to be evaluated, and performing transfer learning positioning on the electrical, magnetic, seismic, and gravity exploration data to determine the cross-sectional view corresponding to the abnormal position of igneous carbonate rock-derived rare earth ore data in the area to be evaluated;

S102, performing feature decomposition and dimensionality reduction, feature enhancement, and weighted fusion processing on the respective cross-sectional images;

S103, segmenting the geological bodies of igneous carbonate rock-derived rare earth ore mineralization in the fusion image, and obtaining the spatial storage morphology of the detection target.

In the embodiments of the present invention, transfer learning-based positioning of electrical, magnetic, seismic, and gravity exploration data includes:

Analyzing the electrical, magnetic, seismic, and gravity exploration data to obtain the data volumes corresponding to the electrical resistivity anomalies, magnetization anomalies, velocity anomalies, and density anomalies of igneous carbonate-hosted rare earth ore genesis; and positioning the data volumes corresponding to the electrical resistivity anomalies, magnetization anomalies, velocity anomalies, and density anomalies.

In the embodiments of the present invention, the sectional images corresponding to the data anomalies include: a digital sectional image of electrical resistivity anomalies, a digital sectional image of magnetization anomalies, a digital sectional image of velocity anomalies, and a digital sectional image of density anomalies.

In the embodiments of the present invention, electrical, magnetic, seismic, and gravity exploration data are obtained in the target area, and then transfer learning-based positioning is performed based on the data.

In the embodiments of the present invention, determining the sectional images corresponding to the data anomaly locations of igneous carbonate-hosted rare earth ore genesis in the target area includes:

Generating digital sectional images of electrical resistivity anomalies, magnetization anomalies, velocity anomalies, and density anomalies with the same depth, respectively;

Configuring the data of the digital sectional images of electrical resistivity anomalies, magnetization anomalies, velocity anomalies, and density anomalies on a common coordinate system.

In the embodiments of the present invention, feature decomposition and dimension reduction, feature enhancement, and weighted fusion processing of the sectional images can be understood as image feature decomposition, image feature dimension reduction, and image feature enhancement processing of the sectional images, respectively, and the processed sectional images are weighted and fused.

In the embodiments of the present invention, feature decomposition and dimension reduction of the sectional images include: using a transfer learning-based downsampling network to downsample the input signal, outputting the downsampling result, i.e., performing initial feature extraction on the digital sectional images of electrical resistivity anomalies, magnetization anomalies, velocity anomalies, and density anomalies, obtaining the downsampling features at different scales for each image.

In the embodiments of the present invention, by analyzing the electrical, magnetic, seismic, and gravity exploration data obtained in the target area, the data volumes corresponding to the electrical resistivity anomalies, magnetization anomalies, velocity anomalies, and density anomalies of igneous carbonate-hosted rare earth ore genesis are obtained, and then transfer learning-based positioning is performed on the electrical, magnetic, seismic, and gravity exploration data volumes to determine the location of the igneous carbonate-hosted rare earth ore genesis in the target area through multi-source information fusion technology.

In the embodiments of the present invention, image feature decomposition of the sectional images includes:

Using the transfer learning-based network ResNet50 to perform initial feature extraction on the digital sectional images of electrical resistivity anomalies, magnetization anomalies, velocity anomalies, and density anomalies, and obtaining M initial features for each image.

As shown in FIG. 2, an example is given to illustrate the number of image features M extracted by the fixed network ResNet50 of transfer learning at different convolutional layers i (i=1,2,3,4,5) for a given input image. The i-th convolutional module outputs M channels of features denoted by F_i,1:Mi^k, where i∈{1,2,3,4,5}, and k represents the category number of the input image. For example, F_i,1:Mi^k represents 1024 channels of features extracted by ResNet50's fixed network at the 4th convolutional module for input image I1, which can also be understood as 1024 features extracted from the image.

In the embodiments of the present invention, the feature enhancement processing of the sectional image includes:

For each image, calculating the correlation coefficients between the initial features at different scales using Spearman correlation;

Removing the irrelevant initial features based on the comparison results between the correlation coefficients and the threshold, and obtaining the enhanced features.

In the embodiments of the present invention, the image feature dimension reduction and enhancement processing of the sectional image include:

For the M initial features of each image, calculating the correlation coefficients between the initial features using Spearman correlation;

Removing the irrelevant initial features based on the comparison results between the correlation coefficients and the threshold, and obtaining the secondary image features after dimensionality reduction;

Using the ZCA algorithm to perform redundancy reduction on the secondary image features to achieve feature enhancement of the sectional image.

As shown in FIG. 3, the embodiments of the present invention provide a process diagram of single SI image feature extraction, feature dimension reduction, and feature enhancement. (a) The source image SI is a contour map of resistivity; (b) The input image I after preprocessing; (c) A series of initial features extracted from I (using the ResNet50 fixed network), here showing the features with 1024 channels output from the 4th convolution layer (i=4). (d) The correlation coefficients between the initial features calculated using Spearman correlation. (e) The secondary features after dimensionality reduction. The irrelevant features are removed based on the correlation coefficients, where the embodiments of the present invention consider the features with correlation coefficients<0.3 as irrelevant. (f) An example of a feature in (e), here showing the 525th feature in the 4th convolution module. (g) The feature after enhancement, where the ZCA algorithm is used for redundancy reduction to achieve feature enhancement.

In the embodiments of the present invention, the enhanced sectional images are weighted and fused, including:

Calculating the initial weight matrix for each sectional image using local Nuclear Norm and average operation;

Interpolating the initial weight matrix using soft-max operation and bicubic interpolation to obtain the weight coefficients for each sectional image;

Using the Local energy map algorithm to preserve the structural and detail information of each sectional image;

Using similarity measurement as the decision value for fusion, comparing the fusion results calculated by weighted averaging with local energy calculation.

The embodiments of the present invention also include calculating the correlation coefficients between the initial features at different scales for each image using Spearman correlation:

The Spearman correlation coefficient r_s^k is used to represent the correlation between the initial features F_i,1:Mi^k of the i-th layer convolution module outputting M channels.

$r_s^k=1-\frac{6\sum _{i=1}^M{d}^2}{M\left(M^2-1\right)}$

where d is the rank difference between F_i,p^k and F_i,q^k, p,q∈{1,2,L,M_i}; k represents the image category number; the minimum number of features is selected with r_s^k as the selected features after feature selection, represented by F_i,1:C^k.

In the present embodiment of the invention, four source images SI (SI1, SI2, SI3, and SI4) are used, which are digitized sections of anomalies in resistivity, magnetization, velocity, and density, respectively. After preprocessing, four input images I (I1, I2, I3, and I4) are obtained. Then, the ResNet50 transfer learning network is used to extract initial features from the four input images respectively.

In this embodiment of the invention, ZCA algorithm is used to remove redundancy from the secondary image features, achieving enhancement of the section image features:

calculating the covariance matrix Cov_i,m^k of X=(F_i,1^k,F_i,2^k,L,F_i,C^k) and perform singular value decomposition on the covariance matrix Cov_i,m^k as follows:

${\mathrm{Cov}}_{i,m}^k=F_{i,m}^k\times {\left(F_{i,m}^k\right)}^T,m\in \left\{1,2,L,C\right\}$ ${\mathrm{Cov}}_{i,m}^k=U\Sigma V^T$

using the following equation to calculate the secondary extracted feature

${\stackrel{)}{F}}_{i,1:C_{:}}^k{\stackrel{)}{F}}_{i,1:C}^k=U\sqrt{{\left(\Sigma +\epsilon I\right)}^{-1}}{U}^T{F}_{i,C}^k$

where U and V are both unitary matrices satisfying U^TU=I,V^TV=I; Σ represents a diagonal, matrix, with all elements except the main diagonal being 0, and each element on the main diagonal is called a singular value. I represents an identity matrix, and ε represents a minimum value for eliminating ill-conditioned matrices.

1) Image Feature Dimensionality Reduction

After ResNet50 network extracts features, the features are high-dimensional, and redundant features inevitably exist. In order to improve the efficiency of the algorithm, the embodiments of the present invention propose to use Spearman correlation coefficient (Gibbs 2015) to reduce the dimensionality of the initial features and remove correlated features.

The redundant features in the initial features F_i,1:Mi^k output by the third layer convolution module are removed by using the Spearman correlation coefficient, i.e., feature dimensionality reduction. It should be noted that the dimensionality reduction here is a process of feature selection, the number of features will decrease, but each feature matrix will not undergo any changes (FIG. 3 (c)-(e)). The Spearman correlation coefficient of the initial features F_i,1:Mi^k output by the third layer module is represented by r_s^k:

$r_s^k=1-\frac{6\sum _{i=1}^M{d}^2}{M\left(M^2-1\right)}$

where d is the rank difference between F_i,p^k and F_i,q^k,p,q∈{1,2,L,M_i}. When |r_s|→1, F_i,p^k and F_i,q^k are more correlated; when |r_s|→0,F_i,p^k and F_i,q^k are considered unrelated. The r_s is used to select the unrelated features, and at this point, multiple combinations will be obtained. The embodiments of the present invention select the group with the smallest number of features as the feature selected after feature selection, which is represented by F_i,1:C^k.

2) Image Feature Enhancement

Due to the ambiguity of single features, it can affect the image fusion effect. Therefore, this patent uses zero-phase component analysis (ZCA) for feature enhancement (Whitening). ZCA is commonly used to process image data, especially when using convolutional neural networks. After the image is whitened by ZCA, it will rotate back to the original pixel space, and the neighboring features of the image can still maintain the original feature structure after being whitened by ZCA. Therefore, the whitened image and the original image will be as similar as possible, and the convolutional neural network needs to use the information between pixels (Ren et al. 2021; Sheng et al. 2018). The following describes the operation process of the ZCA algorithm. First, calculate the covariance matrix Cov_i,m^k of X=(F_i,1^k,F_i,2^k,L,F_i,C^k) according to equation (2) and perform singular value decomposition on Cov_i,m^k:

${\mathrm{Cov}}_{i,m}^k=F_{i,m}^k\times {\left(F_{i,m}^k\right)}^T,m\in \left\{1,2,L,C\right\}$ ${\mathrm{Cov}}_{i,m}^k=U\Sigma V^T$

Then, in the embodiment of the present invention, the second extracted feature

${\stackrel{)}{F}}_{i,1:C}^k$

(FIG. 2 (e)˜(g)) is obtained using equation (3).

${\stackrel{)}{F}}_{i,1:C}^k=U\sqrt{{\left(\Sigma +\epsilon I\right)}^{-1}}{U}^T{F}_{i,C}^k$

Where U and V are unitary matrices, and U^TU=I,V^TV=I satisfies; Σ represents a diagonal matrix, except that all elements except the main diagonal are 0, and each element on the main diagonal is called a singular value. I represents the identity matrix, and ε represents a small value avoiding bad matrix inversion to eliminate ill-conditioned matrices.

In the embodiment of the present invention, the enhanced section image is subjected to weighted fusion processing, including:

The initial weight matrix S_i,*^k of the image is calculated using local nuclear norm and average operation by the following equation:

$S_{i,*}^k\left(x,y\right)=\frac{\sum _{p=x}^{p=x+t}\sum _{q=y-t}^{q=y+t}{{\stackrel{)}{F}}_{i,1:C}^k\left(p,q\right)}^{*}}{\left(2t+1\right)·\left(2t+1\right)}$

Where p,q∈{1,2,L,M_i}, x,y represent the coordinates of the pixel, t represents the step size when calculating the sum of each pixel, and k represents the image category number.

The weight coefficient matrix ω_i^k is obtained by interpolating the weight matrix using soft-max operation and bilinear interpolation with the following equation:

${\omega }_i^k\left(x,y\right)=\frac{S_{i,*}^k\left(x,y\right)}{S_{i,*}^1\left(x,y\right)+S_{i,*}^2\left(x,y\right)},k\in \left\{1,2\right\}$

Consider the specificity and local characteristics of underground targets and use the local energy map to preserve the structural and detailed information of the image:

$L{E}_k\left(x,y\right)=\sum _m\sum _{n}L\left\{I_k\right\}{\left(x+m,y+n\right)}^2$

where LE_k(x,y) represents the total energy map function, L{Ik} represents the local energy map function, m×n represents the size of the neighborhood, and the similarity measurement between I₁ and I₂ is used as the fusion decision value calculated by the following equation:

$M\left(x,y\right)=\frac{2\prod _{k=1,2}\left(\sum _m\sum _{n}L\left\{I_k\right\}\left(x+m,y+n\right)\right)}{\sum _{k=1,2}L{E}_k\left(x,y\right)},$

where the closer M(x,y)=∈[−1,1] is to 1, the stronger the similarity between the images.

If M(x,y)<T, the fusion result is calculated using the weighted average fusion rule:

$F_{M<T}\left(x,y\right)={\omega }_i^1\left(x,y\right)I_1\left(x,y\right)+{\omega }_i^2\left(x,y\right)I_2\left(x,y\right)$

Otherwise, the fusion result is calculated by comparing the local energy:

$F_{M\ge T}\left(x,y\right)=\{\begin{array}{cc}L\left\{I_1\right\}\left(x,y\right),& \mathrm{if}{\mathrm{LE}}_1\left(x,y\right)\ge L{E}_2\left(x,y\right)\\ L\left\{I_2\right\}\left(x,y\right),& \mathrm{if}{\mathrm{LE}}_1\left(x,y\right)<L{E}_2\left(x,y\right)\end{array}$

The fusion strategy is represented by:

$F\left(x,y\right)=\{\begin{array}{cc}{F}_{M<T}\left(x,y\right),& \mathrm{if}M\left(x,y\right)<T\\ F_{M\ge T}\left(x,y\right),& \mathrm{if}M\left(x,y\right)\ge T\end{array}$

The process of establishing the weight matrix is shown in (a) the final feature of I1, (b) the initial weight matrix S_i,*¹, (c) the final weight matrix of I1, (d) the final feature of I2, (e) the initial weight matrix S_i,*², and (f) the final weight matrix of I2.

After the image feature extraction process is completed, the next step is to fuse multiple sources of geophysical images. In this patent, the extracted features are considered as the fusion weights to achieve adaptive image fusion. For a series of image features related to underground structures, there must be correlation between the rows of the matrix. Generally, the matrix is low-rank, so this invention uses nuclear norm to solve the problem of low-rank reconstruction and obtain the initial weight matrix of the image. In this implementation, local nuclear norm and average operation are used to calculate the initial weight matrix S_i,*^k of the image, as shown in equation (4):

$\begin{array}{cc}{S}_{i,*}^k\left(x,y\right)=\frac{\sum _{p=x-t}^{p=x+t}\sum _{q=y-t}^{q=y+t}{{\stackrel{)}{F}}_{i,1:C}^k\left(p,q\right)}^{*}}{\left(2t+1\right)·\left(2t+1\right)}& \left(4\right)\end{array}$

As shown in FIG. 4, the window size in this patent is set to 2. At this point, the initial weight matrix of the image is obtained. However, there is a difference in size between the initial weight matrix obtained after the source image is subjected to convolution and downsampling and the size of the original image. Therefore, this patent uses soft-max operation and bilinear interpolation to interpolate the weight matrix to obtain the final weight coefficient matrix as shown in equation (5):

$\begin{array}{cc}{\omega }_i^k\left(x,y\right)=\frac{S_{i,*}^k\left(x,y\right)}{S_{i,*}^1\left(x,y\right)+S_{i,*}^2\left(x,y\right)},k\in \left\{1,2\right\}& \left(5\right)\end{array}$

Considering the specificity and local characteristics of underground targets, we propose to use the Local Energy Map (LE) (Qiu et al. 2016) to preserve the structural and detailed information of the image. The definition of LE is shown in Equation (6):

$\begin{array}{cc}{\mathrm{LE}}_k\left(x,y\right)=\sum _m\sum _{n}L\left\{I_k\right\}{\left(x+m,y+n\right)}^2& \left(6\right)\end{array}$

Where m×n represents the size of the neighborhood. The similarity measurement between I₁ and I₂ is used as the decision value for fusion and calculated using Equation (7):

$\begin{array}{cc}M\left(x,y\right)=\frac{2\prod _{k=1,2}\left(\sum _m\sum _{n}L\left\{I_k\right\}\left(x+m,y+n\right)\right)}{\sum _{k=1,2}{\mathrm{LE}}_k\left(x,y\right)}& \left(7\right)\end{array}$

If M(x,y)∈[−1,1] tends towards 1, it indicates a strong similarity between the images. A threshold value T is set, and if M(x,y)<T is greater than T, the fusion result is calculated using the weighted average fusion rule, as shown in Equation (8):

$\begin{array}{cc}{F}_{M<T}\left(x,y\right)={\omega }_i^1\left(x,y\right)I_1\left(x,y\right)+{\omega }_i^2\left(x,y\right)I_2\left(x,y\right)& \left(8\right)\end{array}$

If M(x,y)≥T is less than or equal to T, the fusion result is calculated by comparing the local energy using Equation (9):

$\begin{array}{cc}{F}_{M\ge T}\left(x,y\right)=\{\begin{array}{cc}L\left\{I_1\right\}\left(x,y\right),& \mathrm{if}{\mathrm{LE}}_1\left(x,y\right)\ge {\mathrm{LE}}_2\left(x,y\right)\\ L\left\{I_2\right\}\left(x,y\right),& \mathrm{if}{\mathrm{LE}}_1\left(x,y\right)<{\mathrm{LE}}_2\left(x,y\right)\end{array}& \left(9\right)\end{array}$

Therefore, the fusion strategy can be shown in Equation (10):

$\begin{array}{cc}F\left(x,y\right)=\{\begin{array}{cc}{F}_{M<T}\left(x,y\right),& \mathrm{if}M\left(x,y\right)<T\\ F_{M\ge T}\left(x,y\right),& \mathrm{if}M\left(x,y\right)\ge T\end{array}& \left(10\right)\end{array}$ $\mathrm{or}F\left(x,y\right)=\{\begin{array}{cc}{\omega }_i^1\left(x,y\right)I_1\left(x,y\right)+{\omega }_i^2\left(x,y\right)I_2\left(x,y\right),& \mathrm{if}M\left(x,y\right)<T\\ \frac{1}{2}\left(I_1\left(x,y\right)+I_2\left(x,y\right)\right),& \mathrm{if}M\left(x,y\right)\ge T\end{array}$

In this patent, the value of T=0.9 is set to 9. The final fusion result was determined by comparing the fusion results at different T values with prior information.

In the present embodiment of the invention, the abnormal resistivity data body, abnormal magnetization data body, abnormal velocity data body, and abnormal density data body are respectively used to generate digital cut-out maps with the same depth of abnormal resistivity, abnormal magnetization, abnormal velocity, and abnormal density. The data of the digital cut-out maps of abnormal resistivity, abnormal magnetization, abnormal velocity, and abnormal density are configured on a common coordinate system to form a first registration color image and a second registration color image. The first and second registration color images are pre-processed to form a first source image and a second source image. The four images are decomposed into low-frequency sub-band images and high-frequency sub-band images. The two low-frequency sub-band images are fused, and the two high-frequency sub-band images are fused. The fused low-frequency sub-band image and the fused high-frequency sub-band image are reconstructed. The geological bodies in the fused image are segmented using a segmentation method, and the spatial morphology of the detection target is obtained. The present invention identifies geological targets using images as carriers to comprehensively analyze multi-source geophysical data, realizes accurate identification and positioning of target bodies under big data conditions, and can improve the efficiency of comprehensive interpretation while ensuring accuracy.

The present embodiment of the invention also provides a positioning device for the migration learning of igneous carbonate rock-origin rare earth minerals, including:

An image acquisition module is configured to obtain electrical, magnetic, seismic, and gravity exploration data of the area to be estimated and to perform migration learning positioning on the electrical, magnetic, seismic, and gravity exploration data to determine the cut-out map corresponding to the abnormal location of the igneous carbonate rock-origin rare earth mineral data in the area to be estimated.

An image processing module is configured to perform feature decomposition and dimension reduction, feature enhancement, and weighted fusion processing on the cut-out maps, respectively.

A storage and positioning module is configured to segment the igneous carbonate rock-origin rare earth mineralized geological bodies in the fused image and obtain the spatial morphology of the detection target.

Implementation Example

The test was conducted using the method and device for transfer learning-based rare earth mineral localization described in the embodiments of the present invention. Bayan Obo is the largest rare earth mine in the world, which has experienced compression tectonic effects, including the development of conglomerate pancake structures, mylonitization, and folding, transforming the horizontal stratigraphic structure into steep structural bedding with an E-W trend. The compressed structures were invaded by carbonate rocks, indicating that they formed before the formation of carbonate rocks and provided a favorable pathway for magma upwelling.

The main body of the carbonate rock is distributed parallel to the east-west direction, and the veins are distributed in the NE direction, forming a characteristic of a serial arrangement, indicating that the carbonate rocks intruded in a strike-slip environment. The sponge meteorite structure shows that the magnetite is magma iron, and the spatial distribution of the mineral-containing dolomite is unclear due to the unclear intrusion center of the magma.

The bedding and fold axis orientations in the Bayan Obo mining area are 120-100°, 50-70°, and E-W direction. Shear zones are developed in carbonate rocks, including eyeball-shaped structures with NW and NE directions, and conjugate folds and shear zones indicate the compression in the north-south direction. The structural style after the intrusion of carbonate rocks is conjugate, and there is no significant displacement between mining areas, indicating near-in-situ tectonics.

The formation and mineralization processes of the Bayan Obo deposit were divided into four mineralization periods (Middle Paleozoic and Paleozoic). The mineralization process is divided from early to late into magmatic period (coarse-grained dolomite), magmatic-hydrothermal period (fine-grained dolomite), hydrothermal period (impregnation, banded and massive mineralization), and early Paleozoic superimposed alteration period (vein mineralization). The formation and mineralization process of the Bayan Obo deposit is characterized by the interaction of melt-fluid derived from carbonatite magma with the carbonate rock body and surrounding rocks. It has been confirmed that the rare earth resource amount is 5570 T with an average grade of 5.6%.

The ore body is a thick and large lens body produced in the contact zone. However, the mineral composition is complex, and the spatial output is irregular, making it difficult to locate the ore body in space. In this implementation example, the transfer learning modeling was performed using the measured electromagnetic detection, magnetic, and gravity exploration data of the Bayan Obo rare earth deposit. The data set includes electrical exploration data, magnetic exploration data, seismic data, and gravity exploration data.

Transfer learning refers to training a basic network with strong feature extraction ability on a large data set and then transferring the network to a task of multi-source geophysical image fusion. Transfer learning aims to solve the problem of lack of label data or difficult access to label data in the target field using existing experience.

In the target evaluation area, electrical, magnetic, seismic, and gravity exploration data were obtained, and the obtained electrical exploration data, magnetic exploration data, seismic exploration data, and gravity exploration data were processed to obtain FIG. 6 (a) resistivity image, FIG. 6 (b) density image, FIG. 6 (c) magnetic susceptibility image, and FIG. 6 (d) velocity image. The images in FIG. 6 reflect inconsistent positions of underground target bodies, mainly due to the multi-solution nature of a single geophysical image.

Based on the results of rock physical property tests in the area, including minerals, dolomite has the characteristics of high resistance, high magnetic susceptibility, high density, and high velocity. The fusion coefficient in the fusion results represents the distribution range of dolomite, with high coefficient values indicating the distribution of dolomite. According to the drilling results, when the fusion coefficient is greater than 0.5, it represents the dolomite, as shown in the green, yellow, and red areas in the figure. This result (FIG. 7) is consistent with the range of dolomite revealed by the six boreholes in the profile.

Although the embodiment of the present invention is described above, it is for the purpose of facilitating understanding of the technical solution of the present invention and is not intended to limit the present invention. Any modifications and changes in form and detail that fall within the scope of the core technical solutions disclosed in the present invention are within the knowledge of those skilled in the art of the relevant technology, but the scope of protection defined by the appended claims shall still prevail.

Claims

1. A transfer learning positioning method for igneous carbonate-type rare earth mineralization, comprising: acquiring electrical, magnetic, seismic, and gravity exploration data for the target area; performing transfer learning positioning on the electrical, magnetic, seismic, and gravity exploration data to determine the cross-sectional image corresponding to the abnormal position of the data for the igneous carbonate-type rare earth mineralization in the target area;performing feature decomposition and dimension reduction, feature enhancement, and weighted fusion processing on the cross-sectional image, respectively;segmenting the spatial occurrence form of the igneous carbonate-type rare earth mineralized geological body in the fused image to obtain the spatial occurrence form of the detection target.

2. The method according to claim 1, wherein the transfer learning positioning of the electrical, magnetic, seismic, and gravity exploration data comprises: analyzing the electrical, magnetic, seismic, and gravity exploration data to obtain data volumes corresponding to electrical resistivity anomalies, magnetization anomalies, velocity anomalies, and density anomalies for igneous carbonate-type rare earth mineralization; performing positioning on the data volumes corresponding to the electrical resistivity anomalies, magnetization anomalies, velocity anomalies, and density anomalies, respectively; the cross-sectional image corresponding to the data anomaly position comprising: a digital cross-sectional image of the electrical resistivity anomaly, a digital cross-sectional image of the magnetization anomaly, a digital cross-sectional image of the velocity anomaly, and a digital cross-sectional image of the density anomaly.

3. The method according to claim 2, wherein determining the cross-sectional image corresponding to the abnormal position of the data for the igneous carbonate-type rare earth mineralization in the target area comprises: generating digital cross-sectional images of electrical resistivity anomaly, magnetization anomaly, velocity anomaly, and density anomaly at the same depth for the data volumes corresponding to electrical resistivity anomalies, magnetization anomalies, velocity anomalies, and density anomalies, respectively;configuring the digital cross-sectional images of the electrical resistivity anomaly, magnetization anomaly, velocity anomaly, and density anomaly in a common coordinate system.

4. The method according to claim 1, wherein performing feature decomposition and dimension reduction on the cross-sectional image comprises: using a transfer learning selection network to perform stepwise dimension reduction on the input signal and output the reduced feature, i.e., performing initial feature extraction on the digital cross-sectional images of the electrical resistivity anomaly, magnetization anomaly, velocity anomaly, and density anomaly, respectively, to obtain reduced features at different scales for each image.

5. The method according to claim 4, wherein performing feature enhancement on the cross-sectional image comprises: calculating the initial feature correlation coefficient between different scales of reduced features using Spearman correlation for each image;removing irrelevant initial features based on the comparison result between the correlation coefficient and the coefficient threshold to obtain enhanced features.

6. The method according to claim 5, wherein performing weighted fusion processing on the enhanced cross-sectional image comprises: calculating the initial weight matrix corresponding to each cross-sectional image using the local nuclear norm and average operation;performing interpolation operation using the soft-max operation and bicubic interpolation on the initial weight matrix to obtain weight coefficients corresponding to each cross-sectional image;maintaining the structural information and detail information of each cross-sectional image using the local energy map algorithm;using similarity measurement as the fusion decision value and using the weighted averaging fusion rule to compare the fusion results calculated by the local energy calculation.

7. Method as claimed in claim 5, characterized in that for each image at different scales, the initial feature correlation calculated using Spearman correlation includes: The Spearman correlation coefficient of the initial features of M output channels of the i-th layer convolution module is used to represent:The Spearman correlation coefficient rsk is used to represent the correlation between the initial features Fi,1:Mik of the i-th layer convolution module outputting M channels.

8. The method as described in claim 6, characterized by: performing weighted fusion processing on the enhanced sectional images, comprising: calculating the initial weight matrix Si,*k of the image using local nuclear norm and average operation according to the following formula:

9. A transfer learning-based localization device for igneous carbonatite-related rare earth mineralization, comprising: an image acquisition module configured to acquire electrical, magnetic, seismic, and gravity exploration data of the target area to be estimated; and to perform transfer learning-based localization on the acquired data to determine the section corresponding to the anomalous location of igneous carbonatite-related rare earth mineralization in the target area;an image processing module configured to perform feature decomposition and dimension reduction, feature enhancement, and weighted fusion processing on the determined section;and a localization module configured to segment the spatial occurrence of the igneous carbonatite-related rare earth mineralization in the fusion image to obtain the spatial occurrence morphology of the detection target.