Deep learning based magnetic anomaly ore body positioning method

Abstract

The application relates to the technical field of deep learning, in particular to a magnetic anomaly ore body positioning method based on deep learning, which comprises the following steps: collecting multi-source geophysical data of a target area, and performing pretreatment based on the multi-source geophysical data; performing background field correction and anomaly separation based on the pretreated multi-source geophysical data, generating aerial magnetic anomaly data and ground magnetic anomaly data, and constructing a cross-height magnetic anomaly drift field based on the spatial response difference between the aerial magnetic anomaly data and the ground magnetic anomaly data; and performing enhancement processing on the ground magnetic anomaly data by using the cross-height magnetic anomaly drift field, and generating enhanced ground magnetic anomaly data. Through deep learning technology and geophysical prior knowledge, a multi-modal feature mapping model based on a U-Net architecture is constructed, and the multi-modal feature mapping model can effectively overcome the problems that the signal-to-noise ratio of traditional single-height magnetic survey data is low, deep concealed ore body signals are weak, and the deep concealed ore body signals are easily disturbed by shallow interference.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and specifically to a method for locating magnetic anomaly ore bodies based on deep learning. Background Technology

[0002] In the current field of deep resource exploration, traditional magnetic anomaly orebody location methods face two major challenges. Firstly, magnetic survey data from a single observation altitude (e.g., airborne or ground-based) often have significant limitations: airborne magnetic surveys, while covering a wide area, are affected by observation altitude, resulting in weak signals and spatial response drift in deep, concealed orebodies; while ground-based magnetic surveys, although high-resolution, are susceptible to interference from shallow non-mineralized noise and struggle to effectively separate regional and local fields. This makes it difficult for conventional processing methods to quantify and eliminate spatial response differences at different altitudes, severely impacting the accuracy of deep orebody location. Secondly, while existing deep learning-based orebody location methods have advantages in feature extraction, they are mostly used as simple data-driven black-box models, lacking effective constraints on prior knowledge of geophysical fields (such as the attenuation and shift of magnetic anomalies with altitude). This makes it difficult for existing deep learning models to adaptively focus on weak anomaly responses in deep areas when faced with low signal-to-noise ratio data. They are prone to misjudging shallow interference as mineral-induced anomalies or drowning out effective deep signals in complex background fields, thus restricting the application of intelligent mineral exploration technology in the detection of deep concealed ore bodies. Summary of the Invention

[0003] The purpose of this invention is to address the problems existing in the background technology by proposing a deep learning-based method for locating magnetic anomaly ore bodies.

[0004] The technical solution of this invention: A method for locating magnetic anomaly ore bodies based on deep learning, comprising: S1. Collect multi-source geophysical data of the target area and perform preprocessing based on the multi-source geophysical data; S2. Based on the preprocessed multi-source geophysical data, background field correction and anomaly separation are performed to generate aeromagnetic anomaly data and ground magnetic anomaly data. Based on the spatial response difference between the aeromagnetic anomaly data and the ground magnetic anomaly data, a cross-height magnetic anomaly drift field is constructed. The cross-height magnetic anomaly drift field is used to enhance the ground magnetic anomaly data to generate enhanced ground magnetic anomaly data. S3. Based on enhanced ground magnetic anomaly data, polarization processing, vertical guide enhancement, analytical signal extraction and spatial rasterization processing are performed to generate multi-channel underground anomaly feature tensors. S4. Based on the multi-channel underground anomaly feature tensor, the mapping relationship between underground magnetic anomaly features and the spatial distribution of concealed ore bodies is learned using a U-Net architecture, and an ore body probability distribution map is generated. S5. Based on the ore body probability distribution map, perform spatial probability prediction of ore bodies in the target area, and combine the connected domain analysis method to generate the location of hidden ore bodies.

[0005] As a further improvement to this technical solution, in S1, the multi-source geophysical data includes at least: airborne magnetic survey data and ground magnetic survey data.

[0006] As a further improvement to this technical solution, in step S2, airborne magnetic anomaly data and ground magnetic anomaly data are generated, and a cross-altitude magnetic anomaly drift field is constructed based on the spatial response difference between the airborne magnetic anomaly data and the ground magnetic anomaly data. The ground magnetic anomaly data is then enhanced using the cross-altitude magnetic anomaly drift field to generate enhanced ground magnetic anomaly data, including the following steps: S2.1 Read the preprocessed airborne magnetic measurement data and ground magnetic measurement data, and perform spatial registration and unified sampling interval processing on the magnetic field observation values ​​of different survey lines and different measurement points based on a unified coordinate system to generate a standardized magnetic field data matrix; S2.2. Based on the standardized magnetic field data matrix, geomagnetic diurnal variation correction and instrument drift correction are performed to obtain the corrected total magnetic field data; S2.3 Calculate the deviation of the local magnetic field in the survey area from the regional background magnetic field based on the total magnetic field data, and generate initial magnetic anomaly data; S2.4. Based on the sliding window trend surface fitting method, the initial magnetic anomaly data is separated into regional field and local anomaly field to obtain local magnetic anomaly data, including aeromagnetic anomaly data and ground magnetic anomaly data. S2.5. Construct a cross-altitude magnetic anomaly drift field based on the spatial response difference between airborne magnetic anomaly data and ground magnetic anomaly data, and use the cross-altitude magnetic anomaly drift field to enhance the ground magnetic anomaly data to generate enhanced ground magnetic anomaly data.

[0007] As a further improvement to this technical solution, in step S2.5, a cross-altitude magnetic anomaly drift field is constructed based on the spatial response difference between airborne magnetic anomaly data and ground magnetic anomaly data, and the ground magnetic anomaly data is enhanced using the cross-altitude magnetic anomaly drift field to generate enhanced ground magnetic anomaly data, including the following steps: S2.51. Extract local extreme points from the airborne magnetic anomaly data and the ground magnetic anomaly data respectively, and perform coordinate uniform projection on the extraction results to generate the magnetic anomaly amplitude, as well as the coordinates of the airborne anomaly center and the ground anomaly center. S2.52. Based on spatial distance constraints, normalized amplitude similarity constraints, and polarity consistency constraints, a matching relationship between aviation anomaly centers and ground anomaly centers is constructed to generate a set of cross-altitude corresponding anomaly pairs. S2.53. Calculate the spatial displacement relationship based on each pair of anomalies and generate a cross-height anomaly drift vector. ; S2.54, Based on cross-height anomaly drift vector Constructing a deep hidden response index And the response index for deep concealed structures. Normalization is performed to generate a normalized deep hidden response index. ; S2.55, based on deep concealed response index A weighted map of the magnetic anomaly drift field across heights is generated by mapping to a unified grid space using spatial interpolation methods. ; S2.56. Based on ground magnetic anomaly data, combined with a weighted map of the cross-height magnetic anomaly drift field. Enhanced ground magnetic anomaly data is obtained by performing enhancement processing on the target area. .

[0008] As a further improvement to this technical solution, in S2.54, based on the cross-height anomaly drift vector Constructing a deep hidden response index This includes the following steps: Calculation of span-height amplitude ratio based on magnetic anomaly amplitude Based on cross-height anomaly drift vector Calculate its drift modulus Combined with the span-height amplitude ratio With drift modulus Generate deep hidden response index .

[0009] As a further improvement to this technical solution, in step S3, based on enhanced ground magnetic anomaly data, polarization processing, vertical guide enhancement, analytical signal extraction, and spatial rasterization processing are performed to generate a multi-channel underground anomaly feature tensor, including the following steps: S3.1 Read the enhanced ground magnetic anomaly data, and construct a frequency domain polarization operator based on the geomagnetic dip angle, geomagnetic declination angle and observation latitude parameters of the survey area. Perform frequency domain polarization processing on the enhanced ground magnetic anomaly data to obtain the polarization magnetic anomaly characteristics. S3.2. Based on the fast Fourier transform, the polarization magnetic anomaly features are transformed to the wavenumber domain, and vertical gradient enhancement processing is performed to obtain vertical derivative enhancement features. S3.3 Calculate the horizontal and vertical gradients based on the polarization magnetic anomaly characteristics, and construct the analytical signal amplitude response characteristics; S3.4. Based on a unified spatial resolution, the polarization magnetic anomaly features, vertical guide augmentation data, and analytical signal amplitude response are spatially rasterized to generate a standardized set of underground anomaly feature rasteres. S3.5 For the standardized underground anomaly feature grid set, local spatial neighborhoods are extracted based on the sliding window method, and features are stacked according to the channel dimension. The polarization magnetic anomaly features, vertical guide enhancement features, and analytical signal amplitude response features are organized into a multi-channel underground anomaly feature tensor.

[0010] As a further improvement to this technical solution, in step S4, the mapping relationship between underground magnetic anomaly features and the spatial distribution of concealed ore bodies is learned using a U-Net architecture based on the multi-channel underground anomaly feature tensor, and an ore body probability distribution map is generated, including the following steps: S4.1 Read the multi-channel underground anomaly feature tensor and divide the target area into local samples according to the preset spatial window size to generate an underground anomaly feature sample set; S4.2 Construct a multimodal input feature encoding sequence based on underground anomaly feature samples from various regions, and map it to the corresponding input channel according to the feature type to generate multi-channel underground anomaly input features; S4.3 Input the multi-channel underground anomaly input features into the encoder part based on the U-Net architecture to generate underground anomaly coding features at different scales; S4.4, Weighted Map of Magnetic Anomaly Drift Field Across Heights Constructing the deep response parameter modulation matrix And modulate the matrix using deep response parameters. Spatial and channel attention enhancement is applied to the features of each layer of the encoder to generate deep anomaly-sensitive coding features; S4.5 Input the deep anomaly sensitive coding features into the decoder part based on the U-Net architecture for step-by-step decoding processing to generate underground anomaly decoding feature map; S4.6. Based on the underground anomaly decoding feature map, output the probability distribution map of the ore body corresponding to the target area through convolution mapping.

[0011] As a further improvement to this technical solution, in S4.4, a weighted map of the cross-height magnetic anomaly drift field is used. Constructing the deep response parameter modulation matrix And modulate the matrix using deep response parameters. Spatial and channel attention enhancement is applied to the features of each layer of the encoder to generate deep anomaly-sensitive coding features, including the following steps: S4.41, Weighted Map of Magnetic Anomaly Drift Field Across Heights Using a preset spatial window, perform local statistical analysis on the target area to calculate the drift response intensity, local gradient change rate and abnormal energy concentration at each spatial location, and generate a set of deep anomaly response features. S4.42. Based on the deep anomaly response feature set, normalize and map the deep anomaly response features corresponding to different spatial regions to generate a deep response parameter modulation matrix. ; S4.43, Modulation Matrix Based on Deep Response Parameters The anomalous response intensities corresponding to different spatial locations are normalized and mapped, and a spatial dynamic response weight matrix is ​​constructed. ; S4.44, Based on Spatial Dynamic Response Weight Matrix The spatial location weighting processing is applied to the underground anomaly coding features to generate weighted underground anomaly coding features; S4.45, Modulation Matrix Based on Weighted Subsurface Anomaly Coding Features and Deep Response Parameters Constructing channel dynamic adjustment coefficients ; S4.46, Based on channel dynamic adjustment coefficient Further channel-weighted enhancement processing is performed on the weighted underground anomaly coding features to generate deep anomaly sensitive coding features.

[0012] As a further improvement to this technical solution, in S4.45, the modulation matrix is ​​based on weighted underground anomaly coding features and deep response parameters. Constructing channel dynamic adjustment coefficients This includes the following steps: For each feature channel, calculate the spatial average of the weighted subsurface anomaly coding features across all spatial locations. Simultaneously, calculate the modulation matrix of the deep response parameters. The spatial mean of the entire target region is input into the learnable channel weight matrix for linear transformation, and the channel dynamic adjustment coefficients are generated by the Sigmoid function. .

[0013] As a further improvement to this technical solution, in step S5, the spatial probability prediction of ore bodies in the target area is performed based on the ore body probability distribution map, and the location of concealed ore bodies is generated by combining the connected component analysis method, including the following steps: S5.1 Calculate the probability value of the existence of ore bodies at each spatial location in the target area based on the ore body probability distribution map; S5.2. Based on the ore body probability distribution map, a set of ore body candidate areas is generated by filtering through a preset ore body probability threshold. S5.3 Construct spatial adjacency relationships based on the set of candidate ore bodies, and use the connected domain analysis method to cluster the spatially continuously distributed candidate regions to generate candidate connected regions of ore bodies; S5.4 Calculate the spatial distribution range and center coordinates of each ore body's connected domain, and generate the predicted location of the concealed ore body in the target area based on the spatial distribution results of each connected domain.

[0014] Compared with existing technologies, the technical solution of this invention has the following beneficial effects: By using deep learning technology and prior geophysical knowledge, a multimodal feature mapping model based on the U-Net architecture is constructed, which can effectively overcome the problems of low signal-to-noise ratio of traditional single-height magnetic survey data, weak signals of deep concealed ore bodies, and susceptibility to shallow interference. Furthermore, by utilizing the weight map of the cross-height magnetic anomaly drift field to enhance the spatial and channel attention of the encoder features, the network architecture can adaptively focus on the response of weak deep anomalies, thereby achieving high-precision probabilistic prediction of the spatial distribution of concealed ore bodies in complex geological backgrounds, significantly improving the intelligence level and positioning accuracy of deep resource exploration. Attached Figure Description

[0015] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] Example: Please refer to Figure 1 As shown in the figure, this embodiment provides a method for locating magnetic anomaly ore bodies based on deep learning, including the following steps: S1. Collect multi-source geophysical data of the target area and perform preprocessing based on the multi-source geophysical data (mainly including data cleaning, integration, transformation and reduction to ensure data quality and format consistency, and to prepare for subsequent analysis). In this embodiment, the multi-source geophysical data includes at least: airborne magnetic survey data and ground magnetic survey data. Airborne magnetic survey routes are designed according to the survey area and geological tasks. High-precision optically pumped magnetometers are used on UAVs or aircraft to fly along the survey routes to acquire continuous, large-scale airborne magnetic survey data. Simultaneously, ground survey points are deployed in a regular grid within the survey area, and proton precession or Overhauser magnetometers are used for point-by-point observation to collect ground magnetic survey data. Survey points are densified in key profiles or anomalous areas to improve sampling resolution. During all data acquisition processes, the spatial coordinates (latitude, longitude, and elevation), observation time, instrument parameters, and environmental information of each survey point must be recorded to ensure that subsequent multi-source data can be fused and processed under a unified spatiotemporal reference.

[0018] S2. Based on the preprocessed multi-source geophysical data, background field correction and anomaly separation are performed to generate aeromagnetic anomaly data and ground magnetic anomaly data. Based on the spatial response difference between the aeromagnetic anomaly data and the ground magnetic anomaly data, a cross-height magnetic anomaly drift field is constructed. The cross-height magnetic anomaly drift field is used to enhance the ground magnetic anomaly data to generate enhanced ground magnetic anomaly data. In this embodiment, generating aeromagnetic anomaly data and ground magnetic anomaly data, and generating enhanced ground magnetic anomaly data, includes the following steps: S2.1 Read the preprocessed airborne magnetic survey data and ground magnetic survey data, and perform spatial registration and unified sampling interval processing on the magnetic field observation values ​​of different survey lines and different survey points based on a unified coordinate system to generate a standardized magnetic field data matrix (a set of regular two-dimensional or three-dimensional magnetic field data). S2.2. Perform geomagnetic diurnal variation correction and instrument drift correction based on the standardized magnetic field data matrix (based on the reference magnetic field data from synchronous observations at fixed base stations and the instrument zero-point drift calibration parameters, perform time synchronization differential correction, low-frequency trend fitting compensation, and zero-drift elimination processing on the standardized magnetic field data matrix to suppress background interference components caused by geomagnetic field temporal variations and long-term equipment drift), and obtain the corrected total magnetic field data (this step performs the above correction processing independently on airborne magnetic measurement data and ground magnetic measurement data respectively). S2.3 Calculate the deviation of the local magnetic field in the survey area from the regional background magnetic field based on the total magnetic field data, and generate initial magnetic anomaly data. ,in, This represents the total magnetic field data. The regional background magnetic field data is generated by calculating the regional background geomagnetic field components corresponding to each spatial location based on the latitude and longitude coordinates, elevation parameters, observation time, and spherical harmonic coefficient parameters in the international geomagnetic reference field model of the target survey area, and synthesizing them to generate regional background magnetic field data. In this step, the airborne total magnetic field data and the ground total magnetic field data are calculated separately to generate the airborne initial magnetic anomaly and the ground initial magnetic anomaly. S2.4. Initial magnetic anomaly data fitted using a sliding window trend surface method. Regional fields and local anomalous fields are separated, large-scale slowly varying background magnetic fields are removed, and only anomalous responses related to local underground magnetic bodies are retained to obtain local magnetic anomaly data, including aeromagnetic anomaly data and surface magnetic anomaly data. S2.5. Based on the spatial response difference between airborne magnetic anomaly data and ground magnetic anomaly data, construct a cross-altitude magnetic anomaly drift field, and use the cross-altitude magnetic anomaly drift field to enhance the ground magnetic anomaly data to generate enhanced ground magnetic anomaly data. In this embodiment, step S2.5 mainly addresses the problem that deep concealed ore bodies in single-height magnetic anomaly data have weak signals, difficult spatial offsets that are hard to quantify, and significant differences in response at different observation heights (airborne and ground), making it difficult for conventional processing to effectively separate shallow interference from deep mineral-induced anomalies. This invention introduces the spatial response differences of cross-height magnetic anomalies as prior knowledge into the enhancement processing, and uses physical constraints at different observation scales to adaptively amplify weak deep anomalies and suppress shallow non-mineral-induced noise, thereby providing input features with higher signal-to-noise ratio and stronger ore body positioning directionality for subsequent deep learning models. The process of constructing a cross-height magnetic anomaly drift field and using this field to enhance ground magnetic anomaly data includes the following steps: S2.51. Extract local extrema from both aeromagnetic anomaly and surface magnetic anomaly data (based on sliding window neighborhood search and local gradient change analysis, automatically detect and extract local maxima and minima from both aeromagnetic and surface magnetic anomaly data, and record the polarity (positive or negative) of each extrema), and perform a unified coordinate projection on the extraction results to generate magnetic anomaly amplitudes (perform a unified coordinate projection on the extraction results, and simultaneously record the original magnetic anomaly amplitude of each extrema). and To facilitate amplitude similarity constraints in subsequent matching, the original amplitudes are normalized after taking their absolute values ​​(using min-max normalization, taking the absolute values ​​of positive and negative values ​​respectively, and then mapping the whole to the [0,1] interval) to obtain normalized amplitudes. and This process generates the final magnetic anomaly amplitude while preserving the original polarity (positive or negative) of each extreme point. This process eliminates the incomparability of amplitudes caused by differences in observation scales at different altitudes and ensures that amplitude similarity is based solely on intensity rather than algebraic values. It also generates the coordinates of the flight anomaly center. and the coordinates of the ground anomaly center (The spatial coordinates of the extreme point are the coordinates of the center of the aeronautical anomaly and the center of the ground anomaly.) , Number the value points. Indicates aviation, (representing the ground level) to eliminate the incomparability of amplitudes caused by differences in observation scales at different altitudes; S2.52. Based on spatial distance constraints, normalized amplitude similarity constraints, and polarity consistency constraints (only extreme points with the same polarity are allowed to be paired), a matching relationship between the airborne anomaly center and the ground anomaly center is constructed to generate a set of corresponding anomaly pairs across altitudes. Specifically, based on the set of local extreme points of airborne magnetic anomalies and ground magnetic anomalies, a spatial search neighborhood is first constructed with the airborne anomaly center as the benchmark. Potential ground anomaly candidate points are screened within a preset distance threshold range (the preset distance threshold ranges from 100 to 500 meters and is adaptively adjusted according to the aerial survey flight altitude, ground measurement point density, and regional geomagnetic inclination parameters in the survey area). The Euclidean spatial distance constraint between the candidate points and the airborne anomaly center is then calculated. Subsequently, the magnetic anomaly amplitudes of each candidate anomaly point are normalized, and an amplitude similarity metric (cosine similarity) is calculated based on the normalized amplitudes to filter matching candidates whose amplitude differences meet the threshold conditions. At the same time, a polarity consistency constraint is introduced to retain only extreme point pairs with consistent magnetic anomaly signs (positive or negative). Finally, anomaly point pairs that simultaneously satisfy spatial distance constraints, normalized amplitude similarity constraints, and polarity consistency constraints are jointly judged and matched to generate a set of corresponding anomaly pairs across altitudes. This ensures that the airborne and ground anomaly centers are correctly associated under the conditions of spatial proximity, similar amplitudes, and the same polarity, thus providing a reliable basis for subsequent calculations of drift vectors and deep hidden response indices. S2.53. Calculate the spatial displacement relationship based on each pair of anomalies and generate a cross-height anomaly drift vector. ( (For anomaly pair index), used to characterize the spatial migration features of homologous anomalies at different observation heights; S2.54, Based on cross-height anomaly drift vector Constructing a deep hidden response index And the response index for deep concealed structures. Normalization is performed to generate a normalized deep hidden response index. ( For all middle The minimum value, For all middle (maximum value); constructing a deep hidden response index This method integrates the spatial migration characteristics (drift modulus) of homologous anomalies from airborne altitude to ground altitude with energy attenuation or enhancement characteristics (amplitude ratio) into a comprehensive index, thereby quantitatively characterizing the probability of each anomaly originating from a deep, hidden source: the greater the drift and the more significant the enhancement of the ground-to-airborne amplitude, the higher the probability of the anomaly originating from a deep, hidden source. The larger the value, the more likely the anomaly originates from a deeper, concealed magnetic body, rather than shallow disturbances or near-surface noise; this deep concealed response index... This provides a numerical basis for subsequent spatial interpolation to generate the weight field, enabling the final enhancement field to highlight regions with typical deep-seated hidden response characteristics; where the deep-seated hidden response index... Used to quantitatively distinguish between deep mineral-induced anomalies and shallow non-mineral disturbances: The larger the value, the more significant the spatial displacement and relative energy enhancement of the anomaly from air to ground, corresponding to a hidden magnetic body deep underground. The smaller the value, the more likely it is to indicate shallow geological bodies, near-surface noise, or non-mineral anomalies. This index enables continuous quantitative separation of deep and shallow anomalies, rather than relying on absolute burial depth thresholds. It is based on a deep concealed response index. Deep bias labeling is performed on each mismatch pair: an empirical high threshold is set. With low threshold ,when The time marker is a deep mineral-induced anomaly pair, when Time markers are designated as shallow interference anomaly pairs, while those in between are considered transitional anomaly pairs; or they can be directly retained. The original values ​​are used as continuous weights, and no hard classification is performed; Furthermore, based on the cross-height anomaly drift vector Constructing a deep hidden response index This includes the following steps: Calculation of span-height amplitude ratio based on magnetic anomaly amplitude ( (This is a small constant), and the ratio reflects the relative increase in energy of the magnetic anomaly propagating from airborne altitude to ground altitude (deep, concealed sources usually cause a significant increase in ground amplitude relative to airborne amplitude), based on the cross-altitude anomaly drift vector. Calculate its drift modulus This is used to characterize the spatial migration of magnetic anomalies of the same origin at different observation heights; combined with the amplitude ratio across heights. With drift modulus Generate deep hidden response index ; S2.55, based on deep concealed response index By using spatial interpolation methods (such as weighted kernel interpolation) to map to a unified grid space, a continuous weighted map of the cross-height magnetic anomaly drift field is generated. This method is used to make the spatial continuity of the deep hidden response exponent at discrete anomaly points, forming a weighted field covering the entire area. This provides a point-by-point differentiated weighting basis for subsequent magnetic anomaly enhancement, realizing spatial reconstruction from point set to field. Specifically, it first defines a uniform grid based on the target area range (taking the smallest bounding rectangle of all known extreme points and expanding it outward by 10%), assuming the grid spacing is consistent with the resolution of the original magnetic anomaly data (e.g., 50m), generating M rows × E columns of regular grid points; for each grid point... Calculate its relationship with all known points. Euclidean distance and limit the search radius. ( (The average distance between known points) - only retain Points participate in interpolation; an inverse distance weighted kernel function is used, with weights... The power exponent , To prevent division by zero; if the number of neighboring points Then the value of that point is assigned to the entire dataset. The global mean or nearest neighbor fill (finding the nearest known point) is used; otherwise, the interpolation result is used. After calculating all grid points, the resulting two-dimensional array is normalized using the min-max method. ( For all within the entire grid range The minimum value (the minimum value of all grid points). For all within the entire grid range The maximum value (the maximum value of all grid points) is taken as the final output value. Continuous weighted field During the interpolation process, to avoid shallow outliers ( (Smaller) Overextension affects the weight of deep regions; adaptive adjustment of the search radius or weight exponent can be made: for Larger points require a smaller search radius or a higher power exponent. To maintain local focus on deep anomalies; for For smaller points, appropriately increasing the search radius or reducing the power exponent can smooth out their influence, thereby generating a continuous weight field that better reflects geological realities. ; S2.56. Based on ground magnetic anomaly data, combined with a weighted map of the cross-height magnetic anomaly drift field. Enhanced ground magnetic anomaly data is obtained by performing enhancement processing on the target area. ( This indicates enhanced magnetic anomaly data. Indicates ground magnetic anomaly data, (This represents the drift enhancement coefficient, determined through expert experience). Due to the weighted graph... In areas with high values, the probability of deep concealment is high. Multiply by Subsequently, the magnetic anomaly amplitude in the deep suspected area was enhanced, while that in the shallow interference area ( By maintaining the original amplitude (close to 0), a nonlinear adaptive processing of "deep enhancement and shallow preservation" is achieved, thereby improving the sensitivity of subsequent deep learning models to weak deep signals.

[0019] S3. Based on enhanced ground magnetic anomaly data, polarization processing, vertical guide enhancement, analytical signal extraction and spatial rasterization processing are performed to generate multi-channel underground anomaly feature tensors. In this embodiment, generating a multi-channel underground anomaly feature tensor includes the following steps: S3.1. Read the enhanced surface magnetic anomaly data, and construct a frequency-domain polar operator based on the geomagnetic dip angle, geomagnetic declination angle, and observation latitude parameters of the survey area to process the enhanced surface magnetic anomaly data. Frequency domain polarization processing is performed to correct the offset anomaly center generated under tilted magnetization conditions to be directly above the magnetic body, thus obtaining the polarization magnetic anomaly characteristics. To improve the spatial comparability of magnetic anomaly responses in different regions, the following steps are taken: Enhanced ground magnetic anomaly data is read and subjected to a two-dimensional fast Fourier transform to obtain a wavenumber domain representation. Based on the geomagnetic dip, declination, and observation latitude of the survey area, a frequency domain polarization operator is constructed. This operator is used to compensate for the displacement and distortion of magnetic anomalies in the spatial domain under tilted magnetization conditions. Its core is to introduce directional phase correction factors and amplitude gain factors to different wavenumber components, achieving an equivalent transformation of the magnetization direction from the tilted direction to the vertical direction. Subsequently, the wavenumber domain representation and the frequency domain polarization operator are multiplied point-by-point to obtain the wavenumber domain magnetic anomaly response after polarization. Finally, the wavenumber domain magnetic anomaly response is restored to the spatial domain through an inverse Fourier transform, eliminating the anomaly center shift caused by tilted magnetization, and returning the magnetic anomaly peak position to directly above the corresponding magnetic body, thus obtaining the polarized magnetic anomaly characteristics. S3.2, Based on Fast Fourier Transform, the characteristics of polarization magnetic anomalies are analyzed. Transforming to the wavenumber domain, a vertical derivative enhancement operator is constructed, and vertical gradient enhancement processing is performed on different spatial frequency components to highlight the high-frequency anomalous features corresponding to shallow boundaries and local abrupt structures, thus obtaining the vertical derivative enhancement features. Specifically, the polarization magnetic anomaly characteristics are transformed to the wavenumber domain using the Fast Fourier Transform (FFT) to obtain a wavenumber domain representation. A vertical derivative operator is then introduced into the wavenumber domain, mathematically expressed as multiplication by a wavenumber amplitude factor. (generally The enhancement order is set to 1 or 2 to selectively amplify high-frequency components. Subsequently, vertical gradient enhancement filtering is applied to different spatial frequency components, strengthening local anomalies corresponding to high-wavenumber (high-frequency) components while relatively suppressing low-wavenumber (regional background) components. This highlights the boundaries of shallow geological bodies, fault structures, and local abrupt anomalies. Finally, the enhanced wavenumber domain signal is restored to the spatial domain using inverse fast Fourier transform to obtain the vertical derivative enhancement features. ; S3.3, Based on the characteristics of polarization magnetic anomalies Calculate the horizontal gradient separately , and vertical gradient (Based on polarization magnetic anomaly characteristics) First, in the spatial domain, the horizontal gradient components are obtained by numerically differentiating them in the x and y directions using either the central difference or convolutional differential template. and The local rate of change is characterized by normalizing the difference in magnetic anomalies between adjacent grid points with the spatial step size; subsequently, the polarization magnetic anomaly characteristics are... Transforming to the wavenumber domain, and utilizing the equivalent expression of the vertical derivative in the frequency domain, the vertical response of different frequency components is enhanced by multiplying by the wavenumber modulus. Then, an inverse transform is performed back to the spatial domain to obtain the vertical gradient components. This allows for the joint formation of a three-dimensional spatial gradient response, used to characterize the spatial variation of magnetic anomalies in different directions, and the construction of analytical signal amplitude response features. ; S3.4, Characteristics of polarization magnetic anomalies based on unified spatial resolution Vertical guide augmentation data Analyzing the amplitude response of a signal Spatial rasterization processing is performed (based on a uniform spatial resolution grid set for the target survey area, to analyze polarization magnetic anomaly characteristics). Vertical guide augmentation data and analytical signal amplitude response Spatial resampling is performed by first establishing a spatial mapping relationship between the latitude and longitude coordinates of the original sampling points of each data source and the grid center position. Then, bilinear interpolation or inverse distance weighted interpolation is used to reconstruct the continuous field of irregular sampling data, mapping data with different resolutions and sampling densities to regular two-dimensional raster cells, generating a standardized raster data set with consistent cell size and spatial alignment at the same spatial scale, and generating a standardized underground anomaly feature raster set. S3.5. For the standardized underground anomaly feature grid set, local spatial neighborhoods are extracted using a sliding window method, and features are stacked according to the channel dimension to integrate polarization magnetic anomaly features. Vertical guide enhancement features Analyze the amplitude response characteristics of the signal Organization for multi-channel underground anomaly feature tensors ( Indicates the height of the spatial grid. Indicates the width of the spatial grid. Indicates the number of feature channels. (where is the real number field). In this process, normalization and outlier truncation are performed on each channel based on the multi-channel underground anomaly feature tensor, so that the feature distribution corresponding to different physical quantities satisfies the unified numerical scale constraint.

[0020] S4. Based on the multi-channel underground anomaly feature tensor, the mapping relationship between underground magnetic anomaly features and the spatial distribution of concealed ore bodies is learned using a U-Net architecture, and an ore body probability distribution map is generated. In this embodiment, generating the ore body probability distribution map includes the following steps: S4.1 Read the multi-channel underground anomaly feature tensor and perform local sample division of the target area according to the preset spatial window size (e.g., 64×64, the preset spatial window size is a fixed window scale parameter used to extract local spatial neighborhood features from uniform spatial resolution raster data, which is used to determine the spatial range covered by each sample and the corresponding number of pixels or actual geographic scale). (Based on uniform spatial raster data, the target area is gradually divided on the two-dimensional grid according to the preset window size using a sliding window or block cutting method, dividing the continuous space into a series of mutually overlapping or non-overlapping local sub-region sample sets) to generate an underground anomaly feature sample set. Among them, each underground anomaly feature sample includes at least polarization magnetic anomaly features, vertical guide number features, and analytical signal features. S4.2. Construct a multimodal input feature encoding sequence based on various underground anomaly feature samples, and map it to the corresponding input channel according to feature type to preserve the physical property differences between different magnetic anomaly responses and generate multi-channel underground anomaly input features. Specifically: First, perform type identification and normalization processing on different types of magnetic anomaly features (such as polarization magnetic anomaly features, vertical guideline enhancement features, and analytical signal amplitude response features), and maintain a one-to-one correspondence according to the spatial location of the samples; then, organize the multiple types of features at the same spatial location sequentially to construct a multimodal input feature encoding sequence; on this basis, according to the physical properties and expression dimensions of different features, map each feature component in the sequence to the corresponding input channel based on the U-Net architecture, realize the decoupled input and independent expression of different modal features in the channel dimension, and thus form a structured multi-channel underground anomaly input feature representation. S4.3. The multi-channel underground anomaly input features are input into the encoder part based on the U-Net architecture. Local texture and spatial structure features of the underground magnetic anomaly are extracted through multi-layer convolutional operations, batch normalization, and nonlinear activation operations. The receptive field is expanded layer by layer through downsampling operations to generate underground anomaly coding features at different scales. The U-Net architecture consists of an encoder and a decoder: the encoder part is composed of multi-level convolutional blocks, each containing a convolutional layer, a batch normalization layer, and a ReLU activation function, and layer-by-layer activation is achieved through max pooling or stride convolution. Downsampling is used to extract multi-scale magnetic anomaly features from local texture to global semantics. Simultaneously, cross-modal feature fusion units are introduced in each coding layer to perform spatial and channel joint coding of polarization magnetic anomaly features, vertical guide enhancement features, and analytical signal amplitude response features. The decoder part restores the spatial resolution step by step through deconvolution or upsampling operations, and splices and fuses it with the corresponding coding layer features through skip connections to preserve shallow boundaries and structural details. Finally, a spatial probability distribution map of the ore body in the target area is output through 1×1 convolution mapping to achieve pixel-level localization and probabilistic representation of the concealed ore body. S4.4, Weighted Map of Magnetic Anomaly Drift Field Across Heights Constructing the deep response parameter modulation matrix And modulate the matrix using deep response parameters. Spatial and channel attention enhancement is applied to the features of each layer of the encoder to generate deep anomaly-sensitive coding features; In this embodiment, step S4.4 mainly addresses the problem that deep concealed ore bodies have weak magnetic anomaly signals, are greatly affected by shallow interference, and exhibit spatial offsets and amplitude variations at different observation heights, making it difficult for a single data-driven deep learning network to effectively perceive and focus on the characteristics of deep mineral-induced anomalies. This method first quantitatively calculates a weighted map describing the probability of deep concealment by matching the common extreme points of airborne and ground magnetic anomalies. This is then transformed into a modulation matrix. Furthermore, spatial location weighting (enhancing deep regions with large drift and high amplitude ratio) and channel dynamic adjustment (highlighting feature channels related to deep response) are applied to the encoded features respectively. This is used to inject the differences in geophysical response across altitudes as physical priors into the deep learning model, realizing the organic integration of "data-driven" and "physical mechanism-guided" approaches, thereby significantly improving the network's signal-to-noise ratio and spatial positioning accuracy for deep weak magnetic anomalies. The process of generating deep anomaly-sensitive coding features includes the following steps: S4.41, Weighted Map of Magnetic Anomaly Drift Field Across Heights Local statistical analysis is performed on the target area using a preset spatial window to calculate the drift response intensity, local gradient change rate, and anomalous energy concentration at each spatial location, generating a set of deep anomaly response features to characterize the deep propagation characteristics of subsurface magnetic anomalies in different regions; specifically, this is based on a weighted map of the drift field of the cross-height magnetic anomaly. Using a preset spatial window (an N×N neighborhood window set based on a uniform raster resolution, for example, N is 31), a sliding local neighborhood is divided on a two-dimensional raster, and the neighborhood within each window is... Statistical analysis is performed on the selected values. First, the mean or weighted mean of the weight values ​​within the window is calculated to characterize the drift response intensity. Second, the local gradient change rate is calculated by differentiating the weights of adjacent grids within the window and combining this with spatial step size normalization to characterize the drastic changes in the drift field in the local region. Finally, the anomalous energy concentration degree is calculated based on the sum of squares or energy integral of the weight values ​​within the window to reflect the concentration of the magnetic anomaly response in that local space, thereby obtaining the drift response intensity corresponding to each spatial location (i.e., the cross-height magnetic anomaly drift field weight map). The arithmetic mean of all grid values ​​is used to characterize the average level of the deep hidden response intensity in this local region, and the local gradient change rate (i.e., the drift field weight map). The average gradient magnitude is used to characterize the severity of spatial variations in the deep response of the region (i.e., the strength of edges or abrupt changes) and the degree of anomalous energy accumulation (i.e., the drift field weight map). The sum of squares of each grid value (i.e., energy) is used to characterize the concentration or intensity clustering of the deep response in the region. S4.42. Based on the deep anomaly response feature set, normalize and map the deep anomaly response features corresponding to different spatial regions to generate a deep response parameter modulation matrix. ( For drift response intensity, For the local gradient rate of change, This indicates an abnormal energy concentration. As the weight of the drift response intensity, The weights are the local gradient rate of change. The weight of the anomalous energy concentration. ); used to weight the drift field of the magnetic anomaly across heights The statistical characteristics of the three dimensions of deep drift field response intensity, local gradient change, and energy concentration are linearly weighted and fused into a normalized spatial modulation factor. This provides a unified and quantifiable basis for controlling the deep anomaly response in subsequent steps, enabling the network to automatically expand the receptive field and enhance the feature response in deep, low-amplitude magnetic anomaly regions, while preserving local details in shallow, strong anomaly regions. This achieves adaptive and differentiated processing for magnetic anomaly regions of different depths and intensities. Essentially, it involves a weighted graph of the cross-height magnetic anomaly drift field. Further enhancements and spatial recalibration, through the spatial and channel dynamic weighting guided by the modulation matrix, enable the network architecture to actively distinguish between deep mineral-induced responses and shallow noise during the feature learning stage, thus avoiding strong shallow anomalies from drowning out weak deep signals. S4.43, Modulation Matrix Based on Deep Response Parameters The anomalous response intensities corresponding to different spatial locations are normalized and mapped, and a spatial dynamic response weight matrix is ​​constructed. It is used to characterize the importance of underground anomalies in different regions, among which, For learnable scaling parameters, For bias terms, Use the Sigmoid activation function; S4.44, Based on Spatial Dynamic Response Weight Matrix Spatial location weighting is applied to the underground anomaly coding features to generate weighted underground anomaly coding features. It is used to enhance the characteristic response intensity of deep weak magnetic anomaly regions and suppress invalid characteristic responses corresponding to background noise regions and non-mineralized anomaly regions. S4.45, Based on weighted underground anomaly coding features and deep response parameter modulation matrix Constructing channel dynamic adjustment coefficients ; Furthermore, construct the channel dynamic adjustment coefficient. This includes the following steps: For each feature channel Calculate the weighted underground anomaly coding features In all spatial locations Spatial average value (For each feature channel, the spatial average value of the weighted subsurface anomaly coding feature map at all spatial grid locations is obtained by summing the values ​​and dividing by the total number of grid cells.) Simultaneously, the modulation matrix of the deep response parameters is calculated. Spatial mean of the entire target area (modulation matrix of deep response parameters) The values ​​at all valid spatial locations within the entire target area's two-dimensional raster range are summed, then divided by the total number of raster cells to obtain the spatial average value of the channel. Input a learnable channel weight matrix Perform a linear transformation and generate channel dynamic adjustment coefficients using the Sigmoid function. ,in, for The spatial mean, A vector formed by the spatial average values ​​of each channel. Scalar coefficients This is the bias term; the spatial average value of each channel is synthesized by constructing a channel dynamic adjustment coefficient. and spatial mean Adaptively enhance the response of feature channels closely associated with deep drift fields and suppress those channels that mainly respond to shallow noise or non-mineralized anomalies. This allows the network to focus more on channel information that is truly effective for ore body location during the deep decoding stage, thereby improving the signal-to-noise ratio of deep weak magnetic anomaly signals and the accuracy of the final ore body probability prediction. S4.46, Based on channel dynamic adjustment coefficient Weighted underground anomaly coding features Further channel-weighted enhancement processing is performed to generate deep anomaly-sensitive coding features. ; S4.5. Input the deep anomaly sensitive coding features into the decoder part based on the U-Net architecture for step-by-step decoding processing. Through deconvolution upsampling and skip connection structure, the spatial resolution of the underground anomaly is restored layer by layer to generate the underground anomaly decoding feature map, which is used to characterize the comprehensive response of underground anomalies at different spatial scales. S4.6 Based on the underground anomaly decoding feature map, the probability distribution map of the ore body corresponding to the target area is output through 1×1 convolution mapping. The probability value of each spatial location is used to characterize the probability of the existence of a hidden ore body in the corresponding area.

[0021] S5. Based on the ore body probability distribution map, perform spatial probability prediction of ore bodies in the target area, and combine the connected domain analysis method to generate the location of hidden ore bodies; In this embodiment, the spatial probability prediction of ore bodies in the target area is performed, and the location of concealed ore bodies is generated by combining connected component analysis methods, including the following steps: S5.1 Calculate the probability value of ore body existence at each spatial location in the target area based on the ore body probability distribution map; specifically: based on the ore body probability distribution map output by the U-Net architecture, the decoded underground anomaly feature response is traversed spatially, and the probability response value corresponding to each grid location is output using the probability prediction layer; wherein, the convolution mapping result is converted into a probability value in the range of 0 to 1 through the Sigmoid activation function or the Softmax normalization function to characterize the probability of the existence of a hidden ore body at the corresponding spatial location; then, the probability values ​​at each location are spatially mapped and matrix-organized according to the unified spatial grid coordinates of the target area to generate the ore body existence probability distribution result corresponding to the target area; S5.2. Based on the orebody probability distribution map, high-probability anomaly regions that meet the conditions are extracted through screening using a preset orebody probability threshold, generating a set of orebody candidate regions. ,in, This represents the probability threshold of the ore body. This represents the probability distribution of the ore body. S5.3, Based on the set of candidate ore bodies Spatial adjacency relationships are constructed, and connected component analysis is used to cluster and divide spatially continuous candidate regions, generating multiple candidate connected regions for ore bodies; specifically: based on the set of candidate ore body regions... First, adjacency relationships between candidate regions are established based on spatial grid locations. Connectivity is determined for high-probability grid cells that are adjacent or continuously distributed, with adjacency relationships determined according to preset neighborhood rules. Then, connected component analysis is used to cluster candidate grids that meet spatial continuity. Interconnected candidate regions are merged and marked through a traversal search method to generate multiple independent ore body candidate connected regions. Finally, the spatial range, connectivity scale, and regional center location of each connected region are statistically analyzed to form spatial clustering results for different concealed ore bodies within the target area. S5.4 Calculate the spatial distribution range and center coordinates of each orebody's connected domain, and generate the predicted location of the concealed orebody in the target area based on the spatial distribution results of each connected domain. Specifically: For each candidate connected region of an orebody, firstly, statistically analyze the spatial coordinate information of all connected grid cells within it, and calculate the minimum outer boundary range of the corresponding connected region based on the spatial distribution range of each grid cell to characterize the spatial distribution range of the concealed orebody in the target area; then, perform weighted center calculation or geometric center calculation on all grid coordinates within the connected region to obtain the center coordinate position of the corresponding orebody candidate region; finally, combine the spatial distribution range, center coordinates, and probability response intensity of each connected region to structure and output the different orebody candidate regions, generating the predicted location of the concealed orebody in the target area.

[0022] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for locating magnetic anomaly ore bodies based on deep learning, characterized in that, include: S1. Collect multi-source geophysical data of the target area and perform preprocessing based on the multi-source geophysical data; S2. Based on the preprocessed multi-source geophysical data, background field correction and anomaly separation are performed to generate aeromagnetic anomaly data and ground magnetic anomaly data. Based on the spatial response difference between the aeromagnetic anomaly data and the ground magnetic anomaly data, a cross-height magnetic anomaly drift field is constructed. The cross-height magnetic anomaly drift field is used to enhance the ground magnetic anomaly data to generate enhanced ground magnetic anomaly data. S3. Based on enhanced ground magnetic anomaly data, polarization processing, vertical guide enhancement, analytical signal extraction and spatial rasterization processing are performed to generate multi-channel underground anomaly feature tensors. S4. Based on the multi-channel underground anomaly feature tensor, the mapping relationship between underground magnetic anomaly features and the spatial distribution of concealed ore bodies is learned using a U-Net architecture, and an ore body probability distribution map is generated. S5. Based on the ore body probability distribution map, perform spatial probability prediction of ore bodies in the target area, and combine the connected domain analysis method to generate the location of hidden ore bodies.

2. The method for locating magnetic anomaly ore bodies based on deep learning according to claim 1, characterized in that, In S1, the multi-source geophysical data includes at least: airborne magnetic survey data and ground magnetic survey data.

3. The method for locating magnetic anomaly ore bodies based on deep learning according to claim 1, characterized in that, In step S2, airborne magnetic anomaly data and ground magnetic anomaly data are generated. Based on the spatial response difference between the airborne magnetic anomaly data and the ground magnetic anomaly data, a cross-altitude magnetic anomaly drift field is constructed. The ground magnetic anomaly data is then enhanced using the cross-altitude magnetic anomaly drift field to generate enhanced ground magnetic anomaly data. This process includes the following steps: S2.1 Read the preprocessed airborne magnetic measurement data and ground magnetic measurement data, and perform spatial registration and unified sampling interval processing on the magnetic field observation values ​​of different survey lines and different measurement points based on a unified coordinate system to generate a standardized magnetic field data matrix; S2.

2. Based on the standardized magnetic field data matrix, geomagnetic diurnal variation correction and instrument drift correction are performed to obtain the corrected total magnetic field data; S2.3 Calculate the deviation of the local magnetic field in the survey area from the regional background magnetic field based on the total magnetic field data, and generate initial magnetic anomaly data; S2.

4. Based on the sliding window trend surface fitting method, the initial magnetic anomaly data is separated into regional field and local anomaly field to obtain local magnetic anomaly data, including aeromagnetic anomaly data and ground magnetic anomaly data. S2.

5. Construct a cross-altitude magnetic anomaly drift field based on the spatial response difference between airborne magnetic anomaly data and ground magnetic anomaly data, and use the cross-altitude magnetic anomaly drift field to enhance the ground magnetic anomaly data to generate enhanced ground magnetic anomaly data.

4. The method for locating magnetic anomaly ore bodies based on deep learning according to claim 3, characterized in that, In step S2.5, a cross-altitude magnetic anomaly drift field is constructed based on the spatial response difference between airborne magnetic anomaly data and ground magnetic anomaly data. This cross-altitude magnetic anomaly drift field is then used to enhance the ground magnetic anomaly data, generating enhanced ground magnetic anomaly data. This process includes the following steps: S2.

51. Extract local extreme points from the airborne magnetic anomaly data and the ground magnetic anomaly data respectively, and perform coordinate uniform projection on the extraction results to generate the magnetic anomaly amplitude, as well as the coordinates of the airborne anomaly center and the ground anomaly center. S2.

52. Based on spatial distance constraints, normalized amplitude similarity constraints, and polarity consistency constraints, a matching relationship between aviation anomaly centers and ground anomaly centers is constructed to generate a set of cross-altitude corresponding anomaly pairs. S2.

53. Calculate the spatial displacement relationship based on each pair of anomalies and generate a cross-height anomaly drift vector. ; S2.54, Based on cross-height anomaly drift vector Constructing a deep hidden response index And the response index for deep concealed structures. Normalization is performed to generate a normalized deep hidden response index. ; S2.55, based on deep concealed response index A weighted map of the magnetic anomaly drift field across heights is generated by mapping to a unified grid space using spatial interpolation methods. ; S2.

56. Based on ground magnetic anomaly data, combined with a weighted map of the cross-height magnetic anomaly drift field. Enhanced ground magnetic anomaly data is obtained by performing enhancement processing on the target area. .

5. The method for locating magnetic anomaly ore bodies based on deep learning according to claim 4, characterized in that, In S2.54, based on the cross-height anomaly drift vector Constructing a deep hidden response index This includes the following steps: Calculation of span-height amplitude ratio based on magnetic anomaly amplitude Based on cross-height anomaly drift vector Calculate its drift modulus Combined with the span-height amplitude ratio With drift modulus Generate deep hidden response index .

6. The method for locating magnetic anomaly ore bodies based on deep learning according to claim 1, characterized in that, In step S3, based on enhanced ground magnetic anomaly data, polarization processing, vertical guide enhancement, analytical signal extraction, and spatial rasterization processing are performed to generate a multi-channel underground anomaly feature tensor, including the following steps: S3.1 Read the enhanced ground magnetic anomaly data, and construct a frequency domain polarization operator based on the geomagnetic dip angle, geomagnetic declination angle and observation latitude parameters of the survey area. Perform frequency domain polarization processing on the enhanced ground magnetic anomaly data to obtain the polarization magnetic anomaly characteristics. S3.

2. Based on the fast Fourier transform, the polarization magnetic anomaly features are transformed to the wavenumber domain, and vertical gradient enhancement processing is performed to obtain vertical derivative enhancement features. S3.3 Calculate the horizontal and vertical gradients based on the polarization magnetic anomaly characteristics, and construct the analytical signal amplitude response characteristics; S3.

4. Based on a unified spatial resolution, the polarization magnetic anomaly features, vertical guide augmentation data, and analytical signal amplitude response are spatially rasterized to generate a standardized set of underground anomaly feature rasteres. S3.5 For the standardized underground anomaly feature grid set, local spatial neighborhoods are extracted based on the sliding window method, and features are stacked according to the channel dimension. The polarization magnetic anomaly features, vertical guide enhancement features, and analytical signal amplitude response features are organized into a multi-channel underground anomaly feature tensor.

7. The method for locating magnetic anomaly ore bodies based on deep learning according to claim 1, characterized in that, In step S4, the mapping relationship between underground magnetic anomaly features and the spatial distribution of concealed ore bodies is learned using a U-Net architecture based on the multi-channel underground anomaly feature tensor, and an ore body probability distribution map is generated. This includes the following steps: S4.1 Read the multi-channel underground anomaly feature tensor and divide the target area into local samples according to the preset spatial window size to generate an underground anomaly feature sample set; S4.2 Construct a multimodal input feature encoding sequence based on underground anomaly feature samples from various regions, and map it to the corresponding input channel according to the feature type to generate multi-channel underground anomaly input features; S4.3 Input the multi-channel underground anomaly input features into the encoder part based on the U-Net architecture to generate underground anomaly coding features at different scales; S4.4, Weighted Map of Magnetic Anomaly Drift Field Across Heights Constructing the deep response parameter modulation matrix And modulate the matrix using deep response parameters. Spatial and channel attention enhancement is applied to the features of each layer of the encoder to generate deep anomaly-sensitive coding features; S4.5 Input the deep anomaly sensitive coding features into the decoder part based on the U-Net architecture for step-by-step decoding processing to generate underground anomaly decoding feature map; S4.

6. Based on the underground anomaly decoding feature map, output the probability distribution map of the ore body corresponding to the target area through convolution mapping.

8. The method for locating magnetic anomaly ore bodies based on deep learning according to claim 7, characterized in that, In S4.4, the weighted map of the cross-height magnetic anomaly drift field is used. Constructing the deep response parameter modulation matrix And modulate the matrix using deep response parameters. Spatial and channel attention enhancement is applied to the features of each layer of the encoder to generate deep anomaly-sensitive coding features, including the following steps: S4.41, Weighted Map of Magnetic Anomaly Drift Field Across Heights Using a preset spatial window, perform local statistical analysis on the target area to calculate the drift response intensity, local gradient change rate and abnormal energy concentration at each spatial location, and generate a set of deep anomaly response features. S4.

42. Based on the deep anomaly response feature set, normalize and map the deep anomaly response features corresponding to different spatial regions to generate a deep response parameter modulation matrix. ; S4.43, Modulation Matrix Based on Deep Response Parameters The anomalous response intensities corresponding to different spatial locations are normalized and mapped, and a spatial dynamic response weight matrix is ​​constructed. ; S4.44, Based on Spatial Dynamic Response Weight Matrix The spatial location weighting processing is applied to the underground anomaly coding features to generate weighted underground anomaly coding features; S4.45, Modulation Matrix Based on Weighted Subsurface Anomaly Coding Features and Deep Response Parameters Constructing channel dynamic adjustment coefficients ; S4.46, Based on channel dynamic adjustment coefficient Further channel-weighted enhancement processing is performed on the weighted underground anomaly coding features to generate deep anomaly sensitive coding features.

9. The method for locating magnetic anomaly ore bodies based on deep learning according to claim 8, characterized in that, In S4.45, the modulation matrix is ​​based on weighted subsurface anomaly coding features and deep response parameters. Constructing channel dynamic adjustment coefficients This includes the following steps: For each feature channel, calculate the spatial average of the weighted subsurface anomaly coding features across all spatial locations. Simultaneously, calculate the modulation matrix of the deep response parameters. The spatial mean of the entire target region is input into the learnable channel weight matrix for linear transformation, and the channel dynamic adjustment coefficients are generated by the Sigmoid function. .

10. The method for locating magnetic anomaly ore bodies based on deep learning according to claim 1, characterized in that, In step S5, the spatial probability of ore bodies in the target area is predicted based on the ore body probability distribution map, and the location of concealed ore bodies is generated by combining the connected component analysis method, including the following steps: S5.1 Calculate the probability value of the existence of ore bodies at each spatial location in the target area based on the ore body probability distribution map; S5.

2. Based on the ore body probability distribution map, a set of ore body candidate areas is generated by filtering through a preset ore body probability threshold. S5.3 Construct spatial adjacency relationships based on the set of candidate ore bodies, and use the connected domain analysis method to cluster the spatially continuously distributed candidate regions to generate candidate connected regions of ore bodies; S5.4 Calculate the spatial distribution range and center coordinates of each ore body's connected domain, and generate the predicted location of the concealed ore body in the target area based on the spatial distribution results of each connected domain.

Images