Database construction method for local geological magnetic field based on improved super learner

Abstract

The application provides a local geological magnetic field database construction method based on an improved super learner, which comprises the following steps: collecting a magnetic field distribution of a task area; fitting Gaussian semi-variogram function parameters and exponential semi-variogram function parameters, using a Kriging interpolation method to interpolate the regional magnetic field to obtain Gaussian interpolation results and exponential interpolation results; using five sub-models of machine learning to interpolate the regional magnetic field to obtain five interpolation results, using a GBM learner to adjust the interpolation result weights of each sub-model to obtain GBM weighted interpolation results of the machine learning sub-models; using the GBM learner to further adjust the weights of the Gaussian interpolation results, the exponential interpolation results and the weighted interpolation results of the machine learning sub-models finally obtained in step five to obtain final GBM learner interpolation results; and establishing a local geological magnetic field database. The technical scheme of the application can solve the technical problems of poor accuracy of the Kriging interpolation in a complex environment with strong magnetic field nonlinearity and poor generality of a single machine learning model in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of airborne magnetic exploration technology, and particularly relates to a local geological magnetic field database construction method based on an improved super learner. BACKGROUND

[0002] The geological magnetic field database is a database that includes magnetic fields caused by geological magnetic field anomalies at a specific altitude on a plane with specific latitude and longitude coordinates in a specific region. The magnetic field near the earth's surface comes from the main magnetic field (accounting for 95%, such as IGRF and WFF models), the geological magnetic field from the lithosphere (accounting for 4%), and the daily variation field from the universe and the sun (accounting for 1%). The geological magnetic field is derived from iron, cobalt, nickel, and other metal deposits in the lithosphere. Specifically, due to different rock types, different compositions, and different distributions in the lithosphere, the iron magnetic material in the lithosphere is magnetized by the geomagnetic field, which will show more significant magnetic anomalies in a small scale range. The potential magnetic source is generally iron, cobalt, nickel, and other metal deposits. Iron deposits are generally located in the upper part of the earth's crust, with a depth of tens to hundreds of meters. Cobalt and nickel deposits are often associated with iron deposits. Nickel deposits are vertically distributed at a depth of several hundred meters to one kilometer, and cobalt deposits are concentrated in copper-nickel sulfide deposits and some types of sedimentary deposits, with a vertical extension of several hundred meters. Although such minerals also appear in the deep part, the temperature and pressure conditions for mineralization are more suitable in the upper part of the earth's crust, and hydrothermal mineralization and sedimentary mineralization processes are more likely to occur in the upper part, so such interference sources are generally distributed in the upper part of the earth's crust, with a vertical distribution of several hundred meters to one kilometer.

[0003] In addition, if the region to be measured is located in a geological fault zone, the rock types in this area are discontinuous in space due to stratum faulting, and the distribution of geological magnetic sources is discontinuous, which will produce sudden geological magnetic field anomalies. In addition, the crust around the fault zone is more prone to magmatic activity and rock rupture, which further magnetizes the iron magnetic deposits. In the vicinity of a volcano, the geological magnetic field source is also affected by lava activity in the area. Plate generation or plate extinction zones are distributed at the junction of the earth's plates, and plate extinction zones cause the earth's crust to be squeezed, uplifted, or a plate to be squeezed under another plate and subducted into the mantle, while plate generation zones are located at the mid-ocean ridge, which will cause intense volcanic activity, making the distribution of the geological magnetic field in this area more complex and unpredictable.

[0004] In summary, the geological magnetic field is a kind of magnetic field which occupies a large proportion in the total magnetic field except the main magnetic field, is relatively unstable in time, and is relatively arbitrary in space. Due to the above characteristics, it is difficult to establish a model for the geological magnetic field, so it is necessary to find a suitable method to establish a database of the geological magnetic field in the region to be measured. This requires using the sample points of the geological magnetic field in the survey process as known points to perform interpolation to obtain the geological magnetic field in the region. Common interpolation methods include linear interpolation, inverse distance weighting method, etc.

[0005] The linear interpolation method adopts a linear model to extrapolate and interpolate by using the known point coordinate value. This method is logically simple, but due to the characteristics of the geological magnetic field source distribution, such as serious mutation and complex structure, the error generated by using a simple linear model for interpolation is large, especially when the survey line arrangement is sparse, and even the geological magnetic field anomaly source between the survey lines may be missed.

[0006] The inverse distance weighting method is based on the following basic assumptions: the value of the unknown point is affected more by the near point than by the far point, and the distance between the points is inversely proportional to several powers, and the power is related to the field property to be inserted. The advantage of this idea is that the idea is clear and the algorithm structure is simple, but the disadvantage is also obvious: since the geological magnetic field source often comes from the geological structure in the lithosphere, the spatial correlation of the magnetic field brought by this structure is not applied in the inverse distance weighting method, so this method is poor for global grasp, and the result is significantly affected by the uneven distribution of survey points.

[0007] The goal of the minimum curvature method is to generate the smoothest surface to meet the known survey points. This method is suitable for surface interpolation and does not depend on specific statistical models and assumptions, but this method is easy to ignore at extreme values or sharp changes, and the interpolation result is smoother than the actual one. Since the geological magnetic field source corresponds to the ore deposit which is more likely to appear near the geological special structure such as fault zone and plate boundary, the geological magnetic field is more likely to have a jump anomaly, so it is not the most suitable choice in the application scenario of the geological magnetic field.

[0008] Machine learning interpolation is a method of learning the survey sampling points to train the model and then interpolating the region to be interpolated based on the machine learning model. Compared with the interpolation methods mentioned above, the biggest advantage is that it can be well applied to complex nonlinear scenarios, such as sharp changes in small areas. However, although machine learning can well capture the complex nonlinear relationship between input features and target values, it cannot directly capture the spatial correlation of the magnetic field as the geological statistical method does, and the generalization of a single machine model is not strong enough, for example, a model has strong noise resistance but the accuracy decreases when the survey points are scarce, making it difficult to accurately interpolate in various complex environments.

[0009] Kriging method is a method based on known measuring points as samples, fitting a semi-variogram function describing the correlation of data space through the samples, which describes the relationship between the field values of two points in the region and the distance based on spatial correlation, and contains the spatial structure information of the to-be-solved field in the spatial distribution, which is the core of the Kriging method. After obtaining this function by fitting the known measuring points, the weight of each measured value can be obtained through the function, so as to complete the interpolation of the to-be-interpolated point.

[0010] Compared with other interpolation methods, the Kriging method has the following advantages in this scenario: first, based on the principle of spatial autocorrelation, the interpolation method generates more accurate predictions through the semi-variogram function to quantitatively describe the spatial correlation, especially in the case of significant spatial variation of the geological magnetic field. In contrast, the traditional interpolation method is only affected by the distance between the known measuring points and ignores the spatial correlation; second, the Kriging method can provide uncertainty estimation of each measuring point, and secondary collection can be performed in the high-uncertainty area to ensure accuracy, and other interpolation methods cannot provide uncertainty estimation; third, when the measuring points are unevenly distributed, the Kriging method can automatically adjust the interpolation weight according to the distribution of the measuring points, reducing the possibility of large interpolation error in the sparse area of the measuring points. However, the Kriging method has certain limitations in the case of complex to-be-interpolated field variation, mainly because its premise is that the to-be-interpolated quantity has stationarity in space, and the underground distribution of iron ore veins and other abrupt interference sources will destroy this premise in a small area.

[0011] In summary, the Kriging method can capture the spatial correlation of the to-be-interpolated field, but has limitations in the case of complex to-be-interpolated field with severe variation in small scale; a single machine learning method can be well applied in complex nonlinear scenarios, but cannot capture the spatial correlation of the to-be-interpolated field, and has low universality in various complex environments, so it is necessary to find a high-precision method for the interpolation of regional geological magnetic field. SUMMARY

[0012] The application provides a local geological magnetic field database building method based on an improved super learner, which can solve the technical problems of poor precision of Kriging interpolation in complex environments with strong nonlinear magnetic field and poor universality of single machine learning model in the prior art.

[0013] According to an aspect of the present application, a local geological magnetic field database building method based on an improved super learner is provided, which comprises the following steps: step one, collecting the magnetic field distribution of a task area by using a UAV as an aerial magnetic exploration platform; step two, selecting a semi-variogram function as an exponential and Gaussian function, fitting the Gaussian semi-variogram function parameters and the exponential semi-variogram function parameters by using the sampling data, and interpolating the regional magnetic field by using the Kriging interpolation method based on the Gaussian semi-variogram function and the exponential semi-variogram function to obtain the Gaussian interpolation result and the exponential interpolation result; step three, training the machine learning sub-models by using the magnetic field sampling data obtained in step one, interpolating the regional magnetic field by using the machine learning five sub-models to obtain five interpolation results, adjusting the interpolation result weights of each sub-model by using the GBM learner, and combining to obtain the GBM weighted interpolation result of the machine learning sub-models; step four, adjusting the weights of the Gaussian interpolation result, the exponential interpolation result and the machine learning sub-model weighted interpolation result finally obtained in step five by using the GBM learner, and weighting the three interpolation results by using the GBM learner to obtain the final GBM learner interpolation result; and step five, drawing a geological magnetic interference distribution map according to the final GBM learner interpolation result in step six, and building a local geological magnetic field database.

[0014] Further, in step one, the aerial magnetic exploration adopts a UAV as a detection platform, the measuring line is arranged above the to-be-detected area, and the magnetic field distribution of the region is collected by flying along the measuring line at a to-be-detected height to ensure that the measuring line is distributed north-south, and the measuring line interval is determined according to the required scale. Terrain According to B Tot = B plat&earth + B Terrain + B Sun + B grad , wherein B Tot is the total interference, B Sun is the diurnal interference, B grad is the geomagnetic gradient interference, and B plat&earth is the platform interference, the geomagnetic gradient interference.

[0015] Further, in step two, the Kriging interpolation method based on the Gaussian semi-variogram function is used to interpolate the regional magnetic field, which specifically comprises the following steps: based on the Gaussian semi-variogram function, obtaining the covariance matrix of each point on the measuring line and the covariance vector between each measuring point on the measuring line and the to-be-interpolated point; calculating and obtaining the weight vector according to the covariance matrix of each point on the measuring line and the covariance vector between each measuring point on the measuring line and the to-be-interpolated point; calculating and obtaining the to-be-interpolated point field value according to the weight vector and the vector composed of the known point values, and completing the interpolation of each point outside the measuring line according to the to-be-interpolated point field value to obtain the Gaussian interpolation result.

[0016] Further, the Kriging difference method based on the exponential semi-variation function is used to interpolate the regional magnetic field in step two, and the interpolation specifically includes: based on the exponential semi-variation function, obtaining a covariance matrix of each point on the survey line and a covariance vector between each survey point on the survey line and the point to be interpolated; obtaining a weight vector according to the covariance matrix of each point on the survey line and the covariance vector between each survey point on the survey line and the point to be interpolated; obtaining a field value of the point to be interpolated according to the weight vector and a vector composed of the known point measurements, and completing the interpolation of each point outside the survey line according to the field value of the point to be interpolated to obtain an exponential interpolation result.

[0017] Further, the field value of the point to be interpolated u can be obtained according to zλ=Z(u), wherein z is a vector composed of the known point measurements, λ is the weight vector, and Z(u) is the field value of the point to be interpolated u.

[0018] Further, the weight vector λ is obtained according to Cλ=c, wherein C is the covariance matrix based on the covariance matrix of each point on the survey line, and c is the covariance vector between each survey point on the survey line and the point to be interpolated.

[0019] Further, the five machine sub-models are a random forest model, a support vector machine (SVM), a Gaussian process regression (GPR), a K-nearest neighbor model (KNN), and a generalized additive model (GAM).

[0020] According to another aspect of the present application, there is provided a local geological magnetic field database building system based on an improved super learner, which uses the local geological magnetic field database building method based on the improved super learner as described above to build a local geological magnetic field database.

[0021] According to still another aspect of the present application, there is provided a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the local geological magnetic field database building method based on the improved super learner as described above.

[0022] According to still another aspect of the present application, there is provided a computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement the steps of the local geological magnetic field database building method based on the improved super learner as described above.

[0023] The present invention provides a method for building a local geomagnetic field database based on an improved super learner. This method adopts five machine learning models and Kriging interpolation, and innovatively introduces a GBM meta-learner that dynamically adjusts the weights of each interpolation result to integrate the various machine learning sub-models to build a super learner. This method can not only intelligently identify the weights of each sub-model in the super learner through the GBM meta-learner, but also adaptively adjust the weights of Kriging interpolation or super learner interpolation under different geomagnetic field conditions. It systematically solves the problems of poor accuracy of Kriging interpolation in complex environments with strong magnetic field nonlinearity and poor versatility of single machine learning models. The interpolation accuracy is significantly improved compared with traditional methods. Attached Figure Description

[0024] The accompanying drawings, which form part of this specification, are provided to further illustrate embodiments of the invention and, together with the textual description, explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0025] Figure 1 The general steps of a specific implementation of the present invention are shown;

[0026] Figure 2 The Kriging interpolation process is illustrated.

[0027] Figure 3 The process of building the super learner of the present invention is shown;

[0028] Figure 4 The data processing flow of the GBM meta-learner adopted in this invention is shown;

[0029] Figure 5 This diagram shows the magnetic field distribution and sampling point distribution in a region of the Yukon, Canada.

[0030] Figure 6 and Figure 7 This demonstrates the use of Gaussian and exponential semivariograms in traditional Kriging. Figure 5 Interpolation results of the magnetic field distribution in the region at the central measuring point;

[0031] Figure 8 This demonstrates the use of super learners to integrate sub-models with traditional meta-learners. Figure 5 Interpolation results of the magnetic field distribution in this region from the central measuring point;

[0032] Figure 9 This invention demonstrates the use of... Figure 5The interpolation result of the magnetic field distribution of the area by the middle measuring point. DETAILED DESCRIPTION

[0033] It should be noted that the embodiments and features of the embodiments in the present application can be combined with each other in the case of no conflict. The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. The description of the at least one exemplary embodiment is actually only illustrative, but not as any limitation on the present application and its application or use. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0034] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit the exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form, unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, there is a presence of the features, steps, operations, devices, components and / or their combinations.

[0035] Unless specifically stated otherwise, the relative arrangement of the components and steps, numerical expressions, and numerical values set forth in the various embodiments described herein are not limiting. It should be understood that the various parts shown in the drawings are not necessarily drawn to scale in proportion. The techniques, methods, and devices known to those of ordinary skill in the relevant art can not be discussed in detail, but should be considered as part of the authorized description. In all examples shown and discussed herein, any specific value should be interpreted as merely exemplary, and not as a limitation. Therefore, other examples of exemplary embodiments can have different values. It should be noted that similar reference numbers and letters represent similar items in the following drawings, so that once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0036] As Figures 1 to 4As shown, the specific embodiment according to the present application provides a local geological magnetic field database building method based on an improved super learner, which comprises the following steps: step one, collecting the magnetic field distribution of a task area by taking a UAV as an aerial magnetic exploration platform; step two, selecting a semi-variogram function as an exponential and Gaussian function, fitting Gaussian semi-variogram function parameters and exponential semi-variogram function parameters using sampling data, and obtaining Gaussian interpolation results and exponential interpolation results by using a Kriging interpolation method based on the Gaussian semi-variogram function and the exponential semi-variogram function; step three, training machine learning sub-models using the magnetic field sampling data obtained in step one, obtaining five interpolation results by using five machine learning sub-models to interpolate the regional magnetic field, adjusting the interpolation result weights of each sub-model using a GBM learner, and combining to obtain a machine learning sub-model GBM weighted interpolation result; step four, adjusting the weights of the Gaussian interpolation result, the exponential interpolation result and the machine learning sub-model weighted interpolation result finally obtained in step five using the GBM learner, and weighting the three interpolation results using the GBM learner to obtain a final GBM learner interpolation result; and step five, drawing a geological magnetic interference distribution map according to the final GBM learner interpolation result in step six, and establishing a local geological magnetic field database.

[0037] By using this configuration, a local geological magnetic field database building method based on an improved super learner is provided, which adopts five machine learning models and a Kriging interpolation method, and innovatively introduces a GBM meta-learner for dynamically adjusting the weights of each interpolation result to integrate each machine learning sub-model to establish a super learner. The method can not only intelligently identify the weights of each sub-model in the super learner through the GBM meta-learner, but also can adaptively adjust the weights of the Kriging interpolation or the super learner interpolation under different geological magnetic field conditions, systematically solving the problems of poor accuracy of the Kriging interpolation in a complex environment with strong magnetic field nonlinearity and poor generality of a single machine learning model, and greatly improving the interpolation accuracy compared with traditional methods.

[0038] The method provided by the present application comprises the following steps:

[0039] (1) Taking a UAV as an aerial magnetic exploration platform, the magnetic field distribution of a task area is collected.

[0040] (2) The semi-variogram function is adaptively determined by Kriging interpolation, and the regional magnetic field is pre-interpolated.

[0041] (3) The machine learning model is used to train the regional magnetic field, and the regional magnetic field is finely interpolated.

[0042] (4) The GBM is adopted as a meta-learner to establish a super learner, and the interpolation results are fused to obtain a super learner interpolation result.

[0043] (5) Interpolation results of super-learner are used to draw a geological magnetic interference distribution map, and a database is established.

[0044] Specifically, in step one, since the aerial magnetic exploration adopts a UAV as a detection platform, the measurement line is arranged above the to-be-detected region, and the magnetic field distribution of the region is collected by flying along the measurement line at a to-be-detected height, ensuring that the measurement line is distributed north-south, and the interval of the measurement line is determined according to the required scale; the geological magnetic field B Terrain According to B Tot =B plat&earth +B Terrain +B Sun +B grad is calculated and obtained, wherein B Tot is the total interference, B Sun is the diurnal variation interference, B grad is the geomagnetic gradient interference, and B plat&earth is the platform interference, geomagnetic gradient interference.

[0045] Further, in step two, the Kriging interpolation method based on the Gaussian semi-variation function is used to interpolate the regional magnetic field, specifically including: based on the Gaussian semi-variation function, obtaining the covariance matrix of each point on the measurement line and the covariance vector between each measurement point on the measurement line and the to-be-interpolated point; calculating and obtaining the weight vector according to the covariance matrix of each point on the measurement line and the covariance vector between each measurement point on the measurement line and the to-be-interpolated point; calculating and obtaining the field value of the to-be-interpolated point according to the vector composed of the weight vector and the known point measurement value, completing the interpolation of each point outside the measurement line according to the field value of the to-be-interpolated point, and obtaining the Gaussian interpolation result.

[0046] In step two, the Kriging interpolation method based on the exponential semi-variation function is used to interpolate the regional magnetic field, specifically including: based on the exponential semi-variation function, obtaining the covariance matrix of each point on the measurement line and the covariance vector between each measurement point on the measurement line and the to-be-interpolated point; calculating and obtaining the weight vector according to the covariance matrix of each point on the measurement line and the covariance vector between each measurement point on the measurement line and the to-be-interpolated point; calculating and obtaining the field value of the to-be-interpolated point according to the vector composed of the weight vector and the known point measurement value, completing the interpolation of each point outside the measurement line according to the field value of the to-be-interpolated point, and obtaining the exponential interpolation result.

[0047] The field value of the to-be-interpolated point u can be calculated and obtained according to zλ=Z(u), wherein z is the vector composed of the known point measurement value, λ is the weight vector, and Z(u) is the field value of the to-be-interpolated point u. The weight vector λ is calculated and obtained according to Cλ=c, wherein C is the covariance matrix based on the covariance matrix of each point on the measurement line, and c is the covariance vector between each measurement point on the measurement line and the to-be-interpolated point.

[0048] In the present application, the five machine sub-models are random forest model, support vector machine (SVM), Gaussian process regression (GPR), K-nearest neighbor model (KNN) and generalized additive model (GAM).

[0049] According to another aspect of the present application, there is provided a local geological magnetic field database building system based on an improved super learner, which builds a local geological magnetic field database using the local geological magnetic field database building method based on the improved super learner as described above.

[0050] By using such a configuration, a local geological magnetic field database building system based on an improved super learner is provided, which adopts five machine learning models and Kriging interpolation method, and innovatively introduces a GBM meta-learner for dynamically adjusting the weights of the interpolation results to integrate the machine learning sub-models to establish a super learner. The method can not only intelligently identify the weights of the sub-models in the super learner through the GBM meta-learner, but also adaptively adjust the weights of the Kriging interpolation or the super learner interpolation under different geological magnetic field conditions, systematically solving the problems of poor accuracy of Kriging interpolation in complex environments with strong magnetic field nonlinearity and poor generality of single machine learning model, and greatly improving the interpolation accuracy compared with traditional methods.

[0051] According to still another aspect of the present application, there is provided a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the local geological magnetic field database building method based on an improved super learner as described above.

[0052] According to still another aspect of the present application, there is provided a computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement the steps of the local geological magnetic field database building method based on an improved super learner as described above.

[0053] In order to have a further understanding of the present application, the following Figures 1 to 9 The local geological magnetic field database building method based on an improved super learner provided by the present application is described in detail.

[0054] As Figures 1 to 9 shown, the present application provides a method for building a database of geological magnetic field in a task area, which pre-collects magnetic field distribution samples, then performs interpolation, and finally builds a regional geological magnetic field database.

[0055] (1) An unmanned aerial vehicle is used as an aerial magnetic exploration platform to collect magnetic field distribution in the task area.

[0056] Since the aerial magnetic exploration takes the unmanned aerial vehicle as the detection platform, the detection line is arranged above the to-be-detected region, and the unmanned aerial vehicle flies along the detection line at the to-be-detected height to collect the magnetic field distribution of the region, so as to ensure that the detection line is distributed in the north-south direction, and the interval of the detection line is determined according to the required scale. Generally speaking, the interval of the detection line is one percent of the scale, for example, the interval of the detection line is 100 m for the 1:10000 scale.

[0057] (2) The adaptive Kriging interpolation determines the semi-variogram function, and pre-interpolates the magnetic field of the region.

[0058] The Kriging interpolation is a geological interpolation method based on the coordinates and measurement values of a plurality of known points in an unknown region, and grasps the spatial correlation for interpolation. The semi-variogram parameters need to be determined, so that the accuracy is higher. The semi-variogram parameters can be determined by using the maximum likelihood parameter method, and the Kriging interpolation can be completed after the semi-variogram parameters are determined.

[0059] (3) The machine learning model is adopted to train the magnetic field of the region, and the magnetic field of the region is finely interpolated.

[0060] In order to make the method adapt to various unfavorable complex environments as much as possible, such as uneven distribution of detection points, large noise, and strong nonlinear magnetic field, different machine learning models are adopted to interpolate the region to obtain the interpolation results of each machine learning. Different machine learning models have their own advantages and are suitable for different unfavorable environments.

[0061] (4) The GBM is adopted as the meta-learner to establish a super learner, and the interpolation results are fused to obtain the super learner interpolation results.

[0062] The machine learning interpolation results and the Kriging interpolation results need to be fused, and the GBM meta-learner can adaptively learn through the training set to learn how to adaptively assign weights to each interpolation method. After the GBM meta-learner is trained, the learner is used to integrate the interpolation results to obtain the super learner interpolation results.

[0063] (5) According to the improved super learner interpolation results, a geological magnetic interference distribution map is drawn, and a database is established.

[0064] In the last step, the improved super learner fuses the final interpolation results based on each machine learning sub-model and the Kriging pre-interpolation results, and draws a geological magnetic interference distribution map according to the results. In order to facilitate data access and searching, the Database Toolbox toolbox in MATLAB is adopted, and the interpolation results are stored in the database table in the form of SQL statements, so as to complete the effective management of the data.

[0065] The improved super-learner-based local geological magnetic field database construction method provided by the present application will be described in detail below in combination with specific embodiments.

[0066] (1) The unmanned aerial vehicle platform flies along the survey line over the region to be measured and collects the geological magnetic field data of the region after processing.

[0067] In the process of aerial magnetic exploration, the distribution interval of the survey line needs to be determined according to the required scale. Generally speaking, the survey line interval is one percent of the scale, for example, the survey line interval of the 1:10000 scale is 100 m. In this step, the interference from the platform has been compensated. The geomagnetic field is acquired during the platform operation due to the platform position and height, which can be compensated by a perfect geomagnetic field model, the diurnal variation field is compensated by a companion platform or a diurnal variation station, and the remaining is the geological magnetic field. Specifically, the geological magnetic field B Terrain , the total interference B Tot , the diurnal variation interference B Sun , the geomagnetic gradient interference B grad and the platform interference, geomagnetic gradient interference B plat&earth satisfy the following relationship:

[0068] B Tot =B plat&earth +B Terrain +B Sun +B grad

[0069] And the platform interference, geomagnetic main field interference B plat&earth can be solved in the first step of calibration flight and platform longitude and latitude coordinates, altitude, the geomagnetic gradient interference B grad can be solved by longitude and latitude data, the diurnal variation field B Sun is compensated by a number of platforms, and the remaining geological magnetic field B Terrain can be solved.

[0070] The geological magnetic field data obtained in this process will be used as samples for geological magnetic field interpolation to establish a geological magnetic field database.

[0071] (2) Adaptive Kriging interpolation to determine the semi-variogram function, and pre-interpolate the magnetic field of the region.

[0072] The principle of Kriging interpolation method is as follows:

[0073]

[0074] Where Z * (s0) is the value to be interpolated at the interpolation point s0, Z(s i ) is the observation value of the known measurement point s i , λ i is the measurement point si The weight of each sampling point is determined by the Kriging method, and n is the number of known sampling points on the survey line. The key of the Kriging method is how to determine the weight, which depends on the spatial correlation between the known points on the survey line and the point to be interpolated.

[0075] The semi-variogram function describes the relationship between the spatial random process and the spatial coordinates and step length. If the spatial process is stationary, its value is only related to the distance h from the sampling point, and the semi-variogram function is a unary function γ(h), which can be expressed as:

[0076]

[0077] Z(x) and Z(x+h) are the values of the random process at points x and x+h, i.e. the semi-variogram function is half of the mathematical expectation of the square of the difference between the random processes at points x and x+h. However, since the spatial distribution of the random process in the region cannot be known in advance, the semi-variogram function cannot be calculated as above, and can only be obtained by fitting the semi-variogram function from the measured values at a number of known sampling points.

[0078] The semi-variogram function is a general expression of a random process, and any random process can be expressed by this formula. In fact, since the entire random process cannot be known in advance, it can only be fitted using a suitable model based on the properties of the random process. Due to the cubic decay property of geological magnetic interference, the Gaussian model and the exponential model can better fit the geological magnetic interference process than other models. There are eight commonly used semi-variogram functions, but the Gaussian and exponential models are the most suitable for this scenario.

[0079] The commonly used semi-variogram functions include the exponential model, the spherical model, the Gaussian model, and the linear model, among which the exponential model and the Gaussian model are the models that decay exponentially with distance. First, the exponential model is:

[0080]

[0081] Then the Gaussian model is:

[0082]

[0083] where θ i is the correlation parameter of the i-th direction between the sampling point and the point to be interpolated, and d i is the distance between the sampling point and the point to be interpolated in the i-th direction.

[0084] The parameters of the semivariogram are solved using DACEToolbox, a toolbox for the Kriging method. This toolbox integrates seven semivariogram models: spherical, Gaussian, exponential, linear, generalized exponential, Matern, cubic, and linear, as well as stationary and first-order polynomial background models. The toolbox also integrates the maximum likelihood parameter (MLE) algorithm to solve for the semivariogram parameters.

[0085] Specifically, the semivariograms are selected as exponential and Gaussian functions, and the parameters of the Gaussian and exponential semivariograms are fitted using sampled data. The regional magnetic field is pre-interpolated using the Kriging interpolation method based on the Gaussian semivariogram to obtain Gaussian interpolation results. The regional magnetic field is pre-interpolated using the Kriging interpolation method based on the exponential semivariogram to obtain exponential interpolation results.

[0086] Once the semivariogram parameters are confirmed, the covariance between the measured values ​​at each point can be calculated. The covariance between any two measured values ​​satisfies the following:

[0087] Cov(Z(s i ),Z(s j ))=σ 2 -γ(h i,j )

[0088] That is, s i Point and s j The covariance of the measured values ​​at the point Cov(Z(s)) i ),Z(s j )) and the distance h between the two points i,j The relationship between them satisfies the above, Z(s) i ) and Z(s j ) is the measurement value at points i and j, σ 2 Given the total variance of the data, the covariance between points i and j can be calculated using this method. γ represents the semivariogram, which here means: s i and s j The measured values ​​Z(s) at two points i ) and Z(s j The covariance between the known data variances and the semivariogram is given by h. ij The difference between the values ​​taken at that time.

[0089] When interpolating unknown points, point i is a known point, while point j is an unknown point. In this case, σ 2 The total variance of the known data, and the distance h between the two points. i,j Since the covariance of the known point and the point to be inserted is known, it can still be calculated.

[0090] The covariance matrix is ​​determined based on the covariance matrix C of each point on the survey line and the covariance vector c between each survey point on the survey line and the point to be interpolated. The two satisfy the following equation:

[0091] Cλ=c

[0092] Where C is an n×n covariance matrix, and λ represents a vector composed of the weights of each known measurement point when interpolating the unknown point. The elements satisfy:

[0093] C i,j =Cov(Z(s) i ),Z(s j ))=σ 2 -γ(h i,j )

[0094] Where h i,j It is measuring point s i and s j The distance between them, the i-th measuring point s i Measured value Z(s) i ) and the j-th measuring point s j Measured value Z(s) j The covariance of the measured values ​​can be expressed as the total variance σ of the data. 2 The solution is obtained from the semivariogram γ.

[0095] And c is an n×1 covariance vector, and the j-th element of this vector is c j Represents the j-th known point s j The measured value Z(s) on i The covariance between the field value Z(u) at the point u to be interpolated and the field value Z(u):

[0096]

[0097] Each element of the covariance vector c can be indirectly obtained from the distance between each known measuring point and the point to be interpolated through the semivariogram and sample variance, as shown in the above formula, without needing to directly know the field value Z(u) of the point to be interpolated. Thus, the linear equation system Cλ=c can be solved, and finally the weight vector λ is obtained. Then, interpolation can be performed on the target point to complete the establishment of the spatial geological magnetic field of that point, as shown below.

[0098] zλ=Z(u)

[0099] Where z is a vector composed of known point measurements, λ is a weight vector, and Z(u) is the field value of the point u to be interpolated.

[0100] By repeatedly performing this interpolation on each point outside the survey line, the two-dimensional geomagnetic field data at that altitude is finally obtained.

[0101] Exponential semi-variogram function is suitable for continuous area with fast spatial variation of magnetic field, Gaussian semi-variogram function is suitable for continuous area with slow spatial variation of magnetic field, both of which can grasp the spatial correlation of the magnetic field.

[0102] (3) Adopting machine learning model to train the magnetic field of the region, fine interpolation of the magnetic field of the region is needed for complex and steep magnetic field, at this time, there are usually interference sources such as iron ore veins in the underground, the spatial stationarity and continuity of the magnetic field are destroyed, and the data distribution presents a complex nonlinear relationship, so a machine learning model needs to be used for processing. The role of super learner is based on ensemble learning, which comprehensively uses multiple methods to make the prediction effect more robust and accurate than each method alone. In order to achieve this purpose, first of all, a suitable sub-model is selected to obtain a set of interpolation results, and then a suitable meta-learner is selected to give different weights to the results. There are several ways to implement meta-learners: linear regression, weighted linear regression, GBM, etc. Here, through experiments, GBM as a meta-learner is more suitable.

[0103] Super learner is a special machine learning model that integrates multiple machine learning models and uses meta-learners to flexibly combine each sub-machine learning model according to specific conditions. Since super learner is used to flexibly use different machine learning models to process complex nonlinear magnetic field interpolation, it is very important to select appropriate machine learning models to overcome various adverse conditions, including sparse measurement points, different variation laws of magnetic field in each dimension, large noise, strong nonlinear magnetic field, etc.

[0104] The present application selects the following five sub-machine learning models:

[0105] Random forest model is based on the integration of multiple decision trees, which can capture the high-dimensional nonlinear relationship between input and output data.

[0106] Support vector machine (SVM) can handle high-dimensional nonlinear data by selecting appropriate kernel functions.

[0107] Gaussian process regression (GPR) is a probability-based model that can obtain predicted values and uncertainty intervals of predicted values.

[0108] K-nearest neighbor model (KNN) is a model that predicts the value of the point to be interpolated by the distance of the neighboring measurement points around the point to be interpolated.

[0109] Generalized additive model (GAM) is a model that models the additive effects of different variables.

[0110] According to the applicability of each sub-model under various adverse conditions, the meta-learner is designed to combine the interpolation results of each sub-model, and the weight of each sub-model is adaptively assigned to each point to increase the interpolation accuracy.

[0111] Specifically, the applicability of each machine learning model is as follows:

[0112] The random forest model has high robustness, but has large deviation when the measurement points are sparse or unevenly distributed, performs well in complex nonlinear magnetic field environment, and has strong noise resistance.

[0113] The support vector machine model (SVM) performs well when the measurement points are sparse or unevenly distributed, the field changes differently in multiple dimensions, and the magnetic field changes complexly and nonlinearly, but the calculation time is longer.

[0114] The Gaussian process regression (GPM) performs well in areas with sparse measurement points and variable magnetic field changes, but the calculation time is longer.

[0115] The K-nearest neighbor model (KNN) is suitable for processing smooth changes and can grasp the changes of the magnetic field in a single dimension, but performs poorly in areas with nonlinear dramatic changes in the magnetic field and areas with sparse sampling points.

[0116] The generalized additive model (GAM) is suitable for capturing the relationship between a single input variable and the output, but has limitations when dealing with the relationship between multiple input variables. It is suitable for processing changes in a single dimension in spatial data, such as considering only the changes in the magnetic field in the east-west direction.

[0117] These five machine learning models cover a variety of adverse situations where traditional interpolation models perform poorly, including situations where the magnetic field changes differently in multiple dimensions, situations where the measurement points are sparse, situations where the background noise is strong, situations where the magnetic field changes smoothly or dramatically, and situations where the magnetic field is highly nonlinear. At least one machine learning model performs well in these adverse conditions, has a wide range of applications, and has good complementary capabilities between sub-models.

[0118] The five models use grid search algorithm to optimize hyperparameters, and use five-fold cross-validation to train the model to reduce the possibility of overfitting and improve the generalization ability of the model.

[0119] (4) Use the GBM learner as a meta-learner to fuse the interpolation results to obtain the final interpolation result.

[0120] In step (3), five machine learning models with different strengths are selected. To determine how to assign appropriate weights to each interpolation result to increase interpolation accuracy at different locations, a meta-learner is needed to integrate each interpolation result to generate the final interpolation result.

[0121] The gradient boosting machine (GBM) is selected as the meta-learner, which takes the decision tree as the basic unit, can integrate each machine learning model with different applicability to establish a super learner, and effectively process the complex nonlinear changes of the spatial magnetic field. The super learner integrates the interpolation results of each machine learning model, trains each measurement point as a training set, learns the characteristics of the region, and learns how to give different weights to each machine sub-learning model according to the specific interpolation point, and adaptively combines the results of each sub-model to intelligently generate a strong prediction result. In short, it is to adaptively determine how to assign weights to each sub-model at the current point, and adaptively assign weights to each point to obtain more accurate interpolation results. The LSBoost algorithm is used in the GBM training process to optimize the hyperparameters, and five-fold cross-validation is used to reduce the possibility of overfitting. The trained GBM inputs the features at any interpolation point, adaptively assigns weights to the prediction results of each sub-model, and the weight of each sub-machine learning model is dynamically adjusted according to the different positions. Since the geological magnetic field involved in the present application is smooth in a large range, but when it is near the geological disturbance source such as iron ore vein, the magnetic field is no longer smooth and is replaced by a complex nonlinear relationship, so the weight of the sub-model suitable for the local area is adaptively and dynamically adjusted near each interpolation point, which significantly improves the interpolation accuracy.

[0122] The method is innovative in that the dynamic weighting mechanism of the GBM meta-learner can flexibly adjust the weights of each sub-model at each spatial point, thereby achieving excellent performance in both smooth regions with gentle magnetic field changes and non-smooth regions with dramatic magnetic field changes. Compared with traditional single machine learning interpolation methods and Kriging interpolation methods, the present application has higher interpolation stability and accuracy in complex geological environments.

[0123] (5) Draw a geological magnetic disturbance distribution map according to the interpolation results of the super learner, and establish a database.

[0124] The present application uses MATLAB to draw a heat map, and the interpolation results are stored automatically using the Database Toolbox of MATLAB. Specifically, after the interpolation is completed, the encapsulated function in the Database Toolbox is used to insert the pre-established database table in the form of an SQL statement. This process is fully automated and also facilitates future comparison and expansion of interpolation results with current data.

[0125] Effect verification

[0126] In order to practically verify that the present application has performance improvement compared with the traditional method, experimental verification is carried out. The experimental results show that the interpolation method of the present application has 10% to 30% error reduction in average relative error and 5% to 10% error reduction in root mean square error compared with the traditional method, which fully proves the superiority and universality of the present application under complex magnetic field conditions.

[0127] The experimental data set selects the land magnetic map distribution published by the Canadian Department of Land Information, which is located in a mountainous area near Yukon, the vegetation is virgin forest, and there are underground deposits in the area, and the magnetic environment is relatively complex. A rectangular area with an east-west length of 8km, a north-south length of 7km, and an area of 56km 2 is selected as the experimental site.

[0128] The magnetic field distribution in this rectangular area has the following characteristics: it contains a "plain" area with a flat magnetic field distribution, and there are several local magnetic anomalies in the "plain" area. It also includes "hilly" areas where the magnetic field changes dramatically with space, the spatial stability is destroyed, and the spatial complexity presents a complex nonlinear change with space.

[0129] When taking measurement points in this area, the measurement point distribution simulates the requirement of south-north distribution of the specification of aerial magnetic exploration, and the east-west measurement line interval is 600m, and the measurement point interval on each measurement line is 150m. There are a total of 12 measurement lines in the entire area, and 44 measurement points are arranged on each measurement line, which uniformly and comprehensively covers the rectangular area, as shown in Figure 5 .

[0130] Each model will use these measurement points to interpolate the area, and compare with the data published by the Canadian government to judge the performance of each model.

[0131] The following models are used for interpolation in this area:

[0132] 1. Ordinary Kriging Gaussian semi-variogram interpolation (referred to as KG model) ;

[0133] 2. Ordinary Kriging exponential semi-variogram interpolation (referred to as KE model) ;

[0134] 3. Linear element learner combination of random forest, SVM, GPR, GAM, KNN (referred to as SL-L5 model) ;

[0135] 4. GBM element learner combination of random forest, SVM, GPR, GAM, KNN (referred to as SL-GBM5 model) ;

[0136] 5. Linear element learner, combination of SL-GBM5 model and KE, KG model (referred to as KEG-SL-GBM5-L model) ;

[0137] 6. Take the GBM meta-learner, combine the SL-GBM5 model and the KE, KG model and use the GBM as a meta-learner of the super learner model (referred to as KEG-SL-GBM5-GBM model).

[0138] For the pros and cons of the interpolation results, the following three indicators are taken:

[0139] 1. Maximum absolute error (ME): compare the interpolation results with the official geological magnetic map published by Canada, and the maximum absolute value error is the maximum absolute error;

[0140] 2. Mean absolute error (MAE): compare the interpolation results with the official geological magnetic map published by Canada, and the average absolute error MAE is obtained by averaging the absolute value of each point error;

[0141] 3. Root mean square error (RMSE): compare the interpolation results with the official geological magnetic map published by Canada, and the average value of the square of the sum of each point error is obtained.

[0142] Since the interpolation results are prone to distortion at the edge when lacking measurement points, all the interpolation results are clipped by 10% of the size of the edge, and the central and eastern 7.2km, southern and northern 5.6km area is taken, and the official geological magnetic map is also clipped, and the two are aligned for comparison, avoiding the edge error caused by the distortion of the interpolation at the edge to affect the evaluation of the interpolation quality.

[0143] The interpolation error of the six methods is shown in Table 1.

[0144] The KEG-SL-GBM5-GBM model obtained by combining the Kriging Gaussian and exponential semi-variation function with the super learner using the GBM meta-learner weighting, and the traditional Kriging interpolation and the super learner interpolation alone:

[0145] The ME is reduced by 16.9% compared with the traditional Kriging Gaussian interpolation, and is equivalent to the Kriging exponential interpolation, the super learner model using linear meta-learner or GBM meta-learner interpolation;

[0146] The MAE is reduced by 30.3% compared with the traditional Kriging Gaussian or exponential interpolation, is reduced by 21.5% compared with the super learner interpolation using linear meta-learner alone, and is reduced by 10.8% compared with the super learner interpolation using GBM meta-learner alone;

[0147] The RMSE is reduced by 19.2% compared to traditional Kriging Gaussian interpolation, 12.5% ​​compared to traditional Kriging exponential interpolation, 14.8% lower than super learner interpolation using only the linear meta-learner, and 8.9% lower than super learner interpolation using only the GBM meta-learner.

[0148] Table 1

[0149] Method class KE KG SL-L5 ME 169.42 214.39 175.70 MAE 22.83 22.14 20.37 RMSE 31.90 34.87 32.76 Method class SL-GBM5 KEG-SL-GBM5-L KEG-SL-GBM5-GBM ME 203.77 174.90 180.71 MAE 16.92 19.46 15.94 RMSE 30.62 29.78 27.90

[0150] In summary, this invention provides a method for building a local geomagnetic field database based on an improved super learner. This method adopts five machine learning models and the Kriging interpolation method, and innovatively introduces a GBM meta-learner that dynamically adjusts the weights of each interpolation result to integrate the various machine learning sub-models to build a super learner. This method can not only intelligently identify the weights of each sub-model in the super learner through the GBM meta-learner, but also adaptively adjust the weights of Kriging interpolation or super learner interpolation under different geomagnetic field conditions. It systematically solves the problems of poor accuracy of Kriging interpolation in complex environments with strong magnetic field nonlinearity and poor versatility of single machine learning models. The interpolation accuracy is significantly improved compared with traditional methods.

[0151] For ease of description, spatial relative terms such as "above," "on top of," "on the upper surface of," "above," etc., are used herein to describe the spatial positional relationship of a device or feature as shown in the figures to other devices or features. It should be understood that spatial relative terms are intended to encompass different orientations in use or operation beyond the orientation of the device as described in the figures. For example, if the device in the figures were inverted, a device described as "above" or "on top of" other devices or structures would subsequently be positioned as "below" or "under" other devices or structures. Thus, the exemplary term "above" can include both "above" and "below." The device may also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatial relative descriptions used herein will be interpreted accordingly.

[0152] Furthermore, it should be noted that the use of terms such as "first" and "second" to define components is merely for the purpose of distinguishing the corresponding components. Unless otherwise stated, the above terms have no special meaning and therefore should not be construed as limiting the scope of protection of this invention.

[0153] The above merely provides the preferred embodiments of the present application, and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. made within the principles and technical scope of the present application shall fall into the scope of the present application.

Claims

1. A local geological magnetic field database building method based on an improved super learner, characterized in that, The improved super learner-based local geological magnetic field database building method comprises the following steps: Step 1: collecting magnetic field distribution of a task area by using a UAV as an aerial magnetic exploration platform; Step 2: selecting a semi-variogram function as an exponential function and a Gaussian function, fitting Gaussian semi-variogram function parameters and exponential semi-variogram function parameters by using sampling data, interpolating the regional magnetic field by using a Kriging interpolation method based on the Gaussian semi-variogram function and the exponential semi-variogram function to obtain Gaussian interpolation results and exponential interpolation results; Step 3: training machine learning sub-models by using the magnetic field sampling data obtained in step 1, interpolating the regional magnetic field by using five machine learning sub-models to obtain five interpolation results, adjusting the interpolation result weights of each sub-model by using a GBM learner, and combining to obtain a GBM weighted interpolation result of the machine learning sub-models; Step 4: adjusting the weights of the Gaussian interpolation results, the exponential interpolation results and the machine learning sub-model weighted interpolation results obtained in step 3 by using the GBM learner, and weighting the three interpolation results by using the GBM learner to obtain a final GBM learner interpolation result; Step 5: drawing a geological magnetic interference distribution map according to the final GBM learner interpolation result in step 4, and building a local geological magnetic field database.

2. The improved super learner based local geologic magnetic field database building method according to claim 1, wherein, In step one, because airborne magnetic surveys use drones as the detection platform, survey lines are deployed and fly in a loop along the survey lines at the target altitude to collect the magnetic field distribution of the area, ensuring that the survey lines are distributed north-south, and the spacing of the survey lines is determined according to the required scale; the geological magnetic field... according to Calculation and acquisition, where, For total interference, For solar variation interference, Due to geomagnetic gradient interference, This includes platform interference and geomagnetic gradient interference. 3.The improved super learner based local geologic magnetic field database building method according to claim 2, characterized in that, In step 2, the Kriging interpolation method based on the Gaussian semi-variogram function comprises the following steps: Based on the Gaussian semi-variogram function, the covariance matrix of each point on the survey line and the covariance vector between each point on the survey line and the point to be interpolated are obtained; The weight vector is calculated according to the covariance matrix of each point on the survey line and the covariance vector between each point on the survey line and the point to be interpolated; The field value of the point to be interpolated is calculated according to the weight vector and the vector composed of the known point values, the interpolation of each point outside the survey line is completed according to the field value of the point to be interpolated, and the Gaussian interpolation result is obtained. 4.The improved super learner based local geologic magnetic field database building method according to claim 2, wherein, In step 2, the Kriging interpolation method based on the exponential semi-variogram function comprises the following steps: Based on the exponential semi-variogram function, the covariance matrix of each point on the survey line and the covariance vector between each point on the survey line and the point to be interpolated are obtained; The weight vector is calculated according to the covariance matrix of each point on the survey line and the covariance vector between each point on the survey line and the point to be interpolated; The field value of the point to be interpolated is calculated according to the weight vector and the vector composed of the known point values, the interpolation of each point outside the survey line is completed according to the field value of the point to be interpolated, and the exponential interpolation result is obtained.

5. The improved super learner based local geologic magnetic field database building method according to claim 3 or 4, characterized in that, The field value of the point to be interpolated u can be calculated according to The field value of the point to be interpolated u can be calculated according to The field value of the point to be interpolated u can be calculated according to The field value of the point to be interpolated u can be calculated according to The field value of the point to be interpolated u can be calculated according to The field value of the point to 6.The improved super learner based local geologic magnetic field database building method according to claim 5, wherein, The weight vector λ is determined according to The calculation is obtained, wherein C is the covariance matrix based on the covariance matrix of each point on the survey line, and c is the covariance vector between each survey point on the survey line and the point to be interpolated.

7. The improved super learner based local geologic magnetic field database building method according to claim 6, wherein, The five machine sub-models are a random forest model, a support vector machine (SVM), a Gaussian process regression (GPR), a K nearest neighbor model (KNN) and a generalized additive model (GAM).

8. A local geomagnetic field database building system based on an improved super learner, characterized in that, The improved super learner-based local geological magnetic field database building system uses the improved super learner-based local geological magnetic field database building method according to any one of claims 1 to 7 to build a local geological magnetic field database.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the steps of the improved super learner-based local geological magnetic field database building method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program, when executed by a processor, implements the steps of the improved super learner based local geologic magnetic field database building method according to any one of claims 1 to 7.

Images