Two-dimensional image-based airborne electromagnetic deep learning rapid inversion method and system

Abstract

The application discloses a two-dimensional image-based airborne electromagnetic deep learning rapid inversion method and system, relates to the technical field of airborne electromagnetism, and comprises the following steps: acquiring one-dimensional electromagnetic response sequences corresponding to multiple continuous measuring points on a measuring line, longitudinally stacking and splicing the one-dimensional electromagnetic response sequences on the same measuring line according to the spatial order of the measuring points along the measuring line to obtain a two-dimensional original feature matrix, performing sample processing on the two-dimensional original feature matrix to obtain a two-dimensional image data set, inputting the two-dimensional image data set into a pre-established end-to-end two-dimensional convolutional coding network model, and outputting a resistivity two-dimensional image as an airborne electromagnetic deep learning rapid inversion result, so that the overall stratum is layered, the high and low resistance lateral extension mode is highly consistent, and the imaging integrity of the inversion profile is improved.

Description

Technical Field

[0001] This invention relates to the field of aerospace electromagnetic technology, specifically to a rapid inversion method and system for aerospace electromagnetic deep learning based on two-dimensional images. Background Technology

[0002] Airborne electromagnetic inversion is a core method for mineral exploration and underground structural detection. Traditional iterative inversion algorithms such as Gauss-Newton require repeated electromagnetic field forward modeling and parameter correction, resulting in a huge amount of iterative computation. Processing a single survey line can take several days, making it difficult to meet the needs of rapid interpretation in the field. In recent years, deep learning, with its advantage of forward inference without iterative computation, has increased the inversion speed by hundreds of times compared to traditional methods, becoming the mainstream research direction for accelerating the industry.

[0003] Existing airborne electromagnetic deep learning-based rapid inversion methods all use one-dimensional electromagnetic data from a single measurement point as independent samples. The prediction processes for each measurement point are independent of each other, which leads to the current technology severing the stratigraphic spatial relationship between adjacent measurement points on the survey line. As a result, it is impossible to learn the naturally occurring lateral continuity features of underground strata along the survey line, ultimately resulting in poor imaging integrity of the inversion profile. Summary of the Invention

[0004] To address the shortcomings mentioned in the background section, the present invention aims to provide a rapid inversion method and system for airborne electromagnetic deep learning based on two-dimensional images.

[0005] Firstly, the objective of this invention can be achieved through the following technical solution: a fast inversion method for airborne electromagnetic deep learning based on two-dimensional images, the method comprising the following steps: A one-dimensional electromagnetic response sequence corresponding to multiple consecutive measuring points on the measuring line is obtained. The one-dimensional electromagnetic response sequences on the same measuring line are stacked and spliced ​​vertically according to the spatial order of the measuring points along the measuring line to obtain a two-dimensional original feature matrix. The horizontal direction of the two-dimensional original feature matrix corresponds to the measuring points in the measuring line space. The original two-dimensional feature matrix is ​​sampled to obtain a two-dimensional image dataset. The two-dimensional image dataset is then input into a pre-established end-to-end two-dimensional convolutional coding network model, and the output is a two-dimensional resistivity image, which serves as the result of rapid inversion of aerospace electromagnetic deep learning. The end-to-end two-dimensional convolutional coding network model is used to extract the lateral continuous features of the two-dimensional image dataset.

[0006] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the processing of the end-to-end two-dimensional convolutional coding network model, comprising the following steps: By extracting shallow electromagnetic spatial features from a two-dimensional image dataset, and fusing multi-scale information from the transverse and longitudinal directions of the stratigraphy through multi-scale residual void convolution, the high and low resistivity anomaly regions of the two-dimensional image dataset are highlighted through CBAM adaptive weighting. Finally, a continuous and smooth two-dimensional resistivity profile is output as a two-dimensional resistivity image.

[0007] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the end-to-end two-dimensional convolutional coding network model includes a basic shallow feature extraction module, a three-level residual dilated convolutional feature fusion module, a CBAM channel-space dual attention module, and an output mapping module.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the basic shallow feature extraction module is formed by sequentially stacking a Conv2d two-dimensional convolutional layer, a BN2d batch normalization layer, and a LeakyReLU activation function; the two-dimensional image is convolved by the Conv2d two-dimensional convolutional layer; the convolution output is fed into the BN2d batch normalization layer for feature standardization; the standardized convolutional features are nonlinearly mapped by the LeakyReLU activation function to finally obtain the electromagnetic shallow spatial features.

[0009] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the three-level residual hole convolution feature fusion module adopts the internal structure of ResBlock residual units, which is composed of three layers of differentiated custom residual units layer1, layer2, and layer3 connected in series. By inputting electromagnetic shallow spatial features into layer1, the first feature map is obtained by extracting the stratigraphic details of the local area of ​​the measuring point in the lateral measuring line direction and the longitudinal depth direction. Layer2 captures the large-scale continuous stratigraphic structure along the measuring line as the second feature map. The first feature map and the second feature map are input into layer3. Layer3 fuses the first feature map and the second feature map to output a multi-scale fused feature map.

[0010] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the CBAM channel spatial dual attention module includes a channel attention CA submodule and a spatial attention SA submodule. The channel attention CA submodule performs global average pooling to compress the spatial dimension of the multi-scale fused feature map. The spatial attention SA submodule performs global average pooling and global max pooling on the compressed feature map. Two sets of spatial features are spliced ​​along the channel dimension to highlight the high and low resistance abnormal regions of the two-dimensional image dataset, thereby obtaining the attention-enhanced feature map. The attention-enhanced feature map is compressed by a two-layer convolutional cascade structure in the output mapping module, resulting in a continuous and smooth two-dimensional resistivity profile.

[0011] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the loss function used for training the end-to-end two-dimensional convolutional coding network model is obtained by summing the mean squared error loss and the total variational regularization loss.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: the process of sample processing of the two-dimensional original feature matrix, including: The original two-dimensional feature matrix is ​​segmented into blocks using an overlapping sliding window method, where the horizontal step size of the window is set to half the width of the window to obtain subsample images; The subsample images are filled with missing and outlier values ​​in both directions. Then, the numerical range is compressed by logarithmic transformation. After global normalization, the normalized sample images are converted to multi-channel tensor format, and finally, a two-dimensional image dataset is output.

[0013] Secondly, in order to achieve the above objectives, this invention discloses a rapid inversion system for airborne electromagnetic systems based on two-dimensional images using deep learning, comprising: The data acquisition module is used to acquire one-dimensional electromagnetic response sequences corresponding to multiple consecutive measuring points on the measuring line. The one-dimensional electromagnetic response sequences on the same measuring line are vertically stacked and spliced ​​according to the spatial order of the measuring points along the measuring line to obtain a two-dimensional original feature matrix. The horizontal direction of the two-dimensional original feature matrix corresponds to the measuring points in the measuring line space. The inversion output module is used to process the original two-dimensional feature matrix to obtain a two-dimensional image dataset. The two-dimensional image dataset is then input into a pre-established end-to-end two-dimensional convolutional coding network model, and the output is a two-dimensional resistivity image, which serves as the fast inversion result of deep learning for airborne electromagnetics. The end-to-end two-dimensional convolutional coding network model is used to extract the lateral continuous features of the two-dimensional image dataset.

[0014] The beneficial effects of this invention are: This invention models a two-dimensional image by stitching together one-dimensional measurement point data of a survey line. It relies on the residual dilated convolution and CBAM attention structure of an end-to-end two-dimensional convolutional coding network, combined with MSE+TV composite loss to constrain profile smoothing, thereby achieving a high degree of consistency between the overall stratigraphic layering and the lateral extension morphology of high and low resistivity, and improving the imaging integrity of the inverted profile. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the electromagnetic data dimension conversion of the present invention; Figure 3 This is a schematic diagram of the end-to-end two-dimensional convolutional coding network structure of the present invention; Figure 4 This is a schematic diagram of the inversion results of the present invention; Figure 5 This is a schematic diagram of the loss curve during the training process of this invention; Figure 6 This is a schematic diagram of the system structure 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: like Figure 1 As shown, a fast inversion method for airborne electromagnetic systems based on two-dimensional images using deep learning includes the following steps: S101: Obtain the one-dimensional electromagnetic response sequence corresponding to multiple consecutive measuring points on the measuring line, and stack and splice the one-dimensional electromagnetic response sequence on the same measuring line longitudinally according to the spatial order of the measuring points along the measuring line to obtain a two-dimensional original feature matrix. The two-dimensional original feature matrix corresponds to the measuring points in the measuring line space laterally. like Figure 2 As shown, firstly, one-dimensional electromagnetic response sequences corresponding to several continuous measuring points are collected along the airborne electromagnetic survey line, with each measuring point corresponding to a set of one-dimensional observation data; then, all the one-dimensional sequences of measuring points arranged in sequence on the same survey line are stacked vertically according to the spatial order of the measuring points to form a two-dimensional original feature matrix. The matrix corresponds horizontally to the measuring points in the survey line space and vertically to the multi-channel observation data inside the measuring points, thereby converting the scattered one-dimensional measuring point data into parameters that have both horizontal and vertical spatial characteristics. The process of sample processing on the original two-dimensional feature matrix includes: The original two-dimensional feature matrix is ​​segmented into blocks using an overlapping sliding window method, where the horizontal step size of the window is set to half the window width to obtain sub-sample images. Batch sub-image samples are generated by cutting adjacent windows with 50% data overlap, which can both expand the total number of samples and avoid the breakage of lateral stratigraphic information caused by window segmentation.

[0018] Standardized preprocessing operations are performed uniformly on all sub-image samples: first, missing and abnormal values ​​in the image are filled and repaired bidirectionally; then, the numerical span is compressed by logarithmic transformation to reduce the training interference caused by extreme values; then, global normalization is performed based on the overall extreme values ​​of all samples to uniformly map the values ​​to a fixed interval; finally, it is converted into a standard multi-channel tensor format suitable for convolutional networks to obtain a two-dimensional image dataset that can be directly used for model training, and finally output as a two-dimensional image dataset.

[0019] S102: Perform sample processing on the original two-dimensional feature matrix to obtain a two-dimensional image dataset. Input the two-dimensional image dataset into a pre-established end-to-end two-dimensional convolutional coding network model to output a two-dimensional resistivity image, which serves as the result of rapid inversion of aerospace electromagnetic deep learning. The end-to-end two-dimensional convolutional coding network model is used to extract the lateral continuous features of the two-dimensional image dataset.

[0020] Among them, such as Figure 3 As shown, the end-to-end two-dimensional convolutional coding network model is built on PyTorch. The input is a preprocessed single-channel aero-electromagnetic two-dimensional image tensor [1, PatchWidth, DepthNum] (PatchWidth is the width of the training image, and DepthNum is the number of sampling points in the depth direction). The output is a single-channel underground normalized resistivity two-dimensional image with the same size as the input.

[0021] The network consists of four core sequential units: a basic shallow feature extraction module, a three-level residual dilated convolution feature fusion module, a CBAM channel-spatial dual attention module, and an output mapping module. The overall data flow logic is as follows: first, shallow electromagnetic spatial features are extracted; then, multi-scale information from the transverse and longitudinal dimensions of the geological survey line is fused through multi-scale residual dilated convolution; high and low resistivity anomalies are highlighted using CBAM adaptive weighting; finally, a continuous and smooth two-dimensional resistivity profile is obtained. The model employs all convolutions combined with batch normalization and LeakyReLU activation, along with residual skip connections to mitigate deep gradient vanishing, adapting to the learning of continuous transverse features in two-dimensional survey line images.

[0022] All feature processing in the model of this invention is based on three basic operators: two-dimensional convolution Conv2d, batch normalization BN2d, and LeakyReLU activation function. The functions of these three types of operators will be explained first: (1) Conv2d two-dimensional convolutional layer The algorithm primarily employs a 3×3 standard convolution kernel, supplemented by 1×1 dimensionality-transformed convolution and dilated convolution with a dilation factor. Convolution extracts local spatial features of the image through a sliding window, with padding parameters maintaining consistent input and output feature map sizes. Dilated convolution expands the receptive field by setting a dilation factor, capturing long-distance lateral stratigraphic relationships without increasing the number of kernel parameters. The 1×1 convolution is used only for channel upscaling, downscaling, and dimension matching, without altering the spatial dimensions.

[0023] (2) BN2d Batch Normalization Layer Placed after each convolution operation, it standardizes all feature channels within a single batch, unifies the feature distribution range of each layer, eliminates training oscillations caused by data scale fluctuations, accelerates model convergence speed, and also has a certain ability to resist noise and outliers.

[0024] (3) LeakyReLU activation function As a nonlinear activation layer, a fixed negative half-axis slope of 0.2 is set. Unlike the standard ReLU which directly truncates negative values, this operator preserves a weak gradient for negative electromagnetic signals, avoiding permanent inactivation of neurons in shallow and deep networks, fully preserving the low-amplitude electromagnetic features corresponding to weak underground anomalies, and improving the robustness of the model in complex strata.

[0025] The basic shallow feature extraction module is composed of Conv2d two-dimensional convolutional layers, BN2d batch normalization layers, and LeakyReLU activation functions stacked in sequence. The two-dimensional image is convolved by the Conv2d two-dimensional convolutional layer, and the convolution output is fed into the BN2d batch normalization layer for feature standardization. The standardized convolutional features are then nonlinearly mapped by the LeakyReLU activation function to finally obtain the electromagnetic shallow spatial features.

[0026] Specifically, the input is a single-channel preprocessed two-dimensional electromagnetic image. The first layer of Conv2d is configured with 1 input channel, 32 output channels, a 3×3 convolution kernel, and padding=1 to ensure the image size remains unchanged. The convolution output is fed into BN2d for feature standardization. Finally, a nonlinear mapping is introduced through LeakyReLU. The function of this module is to upscale the original single-channel electromagnetic image to 32-channel shallow texture features, extract the basic spatial variation patterns of the measurement points, and provide a stable input for subsequent multi-scale feature fusion.

[0027] The three-level residual void convolution feature fusion module adopts the internal structure of ResBlock residual units, which is composed of three layers of differentiated custom residual units: layer1, layer2, and layer3 connected in series. By inputting the shallow electromagnetic spatial features into layer1, the module extracts the stratigraphic details of the local area of ​​the measuring point in the lateral measuring line direction and the vertical depth direction as the first feature map and inputs it into layer2. Layer2 captures the large-scale continuous stratigraphic structure along the measuring line as the second feature map. The first and second feature maps are input into layer3. Layer3 fuses the first and second feature maps to output a multi-scale fused feature map.

[0028] Specifically, the entire module consists of three layers of differentiated custom residual units: layer1, layer2, and layer3 connected in series. First, the internal structure of the general ResBlock residual unit will be explained, and then the differentiated design of the three layers will be introduced separately.

[0029] General ResBlock Residual Cell Internal Structure The main convolutional layer of a unit follows a sequence of dilated Conv2d → BN2d → LeakyReLU → standard Conv2d → BN2d. The shortcut branch uses a 1×1 Conv2d for dimensional alignment if the input and output channels of a unit are inconsistent; channel matching is performed using an identity mapping. The final output is the pixel-by-pixel sum of the main convolutional layer output and the shortcut branch input feature map, followed by LeakyReLU activation. These shortcut connections allow shallow features to be directly transferred to deeper layers, mitigating the vanishing gradient problem during deep network training.

[0030] Differentiated configuration of three-layer residual units Layer 1: ResBlock (32, 64, dilation=1), with a dilation factor of 1, i.e., ordinary convolution, extracts only local small stratigraphic details at the measurement point; Layer 2: ResBlock (64, 128, dilation=2), with a dilated factor of 2, greatly expands the receptive field and captures large-scale continuous stratigraphic structures along the measurement line; Layer 3: ResBlock (128, 64, dilation=1), aggregates the multi-scale features of the first two layers, unifies the channel dimension to 64, and outputs to the attention module.

[0031] The CBAM channel-space dual attention module includes a channel attention CA submodule and a spatial attention SA submodule. The channel attention CA submodule performs global average pooling to compress the spatial dimension of the multi-scale fused feature map. The spatial attention SA submodule performs global average pooling and global max pooling on the compressed feature map and splices two sets of spatial features along the channel dimension to highlight the high and low resistance anomaly regions of the two-dimensional image dataset and obtain the attention-enhanced feature map. Specifically, the channel attention CA submodule first performs global average pooling to compress the spatial dimension of the input 64-channel features; then two layers of 1×1 Conv2d are used to complete the channel dimensionality reduction and dimensionality increase transformations, with ReLU activation in between; finally, Sigmoid generates channel weights in the 0~1 interval, and multiplies the weights with the original features channel by channel, automatically suppressing invalid channels without geological information and amplifying the key feature channels corresponding to high and low resistivity anomalies.

[0032] Then, based on the spatial attention SA submodule, global average pooling and global max pooling are performed on the feature map respectively, and two sets of spatial features are spliced ​​along the channel dimension; 7×7 Conv2d is used to fuse the two types of pooling information, and sigmoid outputs pixel-level spatial weights; the weights are multiplied pixel by pixel with the channel-enhanced features to strengthen the feature response of the transverse continuous anomaly and the longitudinal stratigraphic interface area of ​​the survey line.

[0033] The attention-enhanced feature map is compressed by a two-layer convolutional cascade structure in the output mapping module, resulting in a continuous and smooth two-dimensional resistivity profile.

[0034] The output mapping module has a two-layer convolutional cascade structure as follows: Conv2d (64,32,3,padding=1) + LeakyReLU + Conv2d (32,1,1). The first 3×3 convolution compresses the attention-enhanced 64-channel high-dimensional features to 32 channels, and LeakyReLU preserves the nonlinear stratigraphic information. The second 1×1 convolution completes the channel compression, and finally outputs a single-channel, normalized two-dimensional resistivity profile with the same size as the input, fully preserving the continuous lateral stratigraphic distribution characteristics of the survey line.

[0035] The loss function used for training the end-to-end two-dimensional convolutional coding network model is obtained by summing the mean squared error loss and the total variational regularization loss.

[0036] The training input is preprocessed two-dimensional airborne electromagnetic response data (i.e., observations), and the output is the subsurface resistivity distribution at the corresponding location, which serves as the label. Both are subjected to missing value imputation, logarithmic transformation to compress the dynamic range, global Min-Max normalization to the [0,1] range, and are then segmented into fixed-size sub-images using a sliding window (50% overlap), with a window size of 64×100 (horizontal sampling points × vertical layer number). Finally, they are converted into PyTorch tensors with a shape of [BatchSize, Channels, Height, Width], where the number of channels is 1 (single-channel image), and the batch size (BatchSize) is fixed at 4 during training. The optimizer is Adam, the loss function is mean squared error, and the total number of training epochs is 1000. Figure 5 The loss curve during the training process is shown.

[0037] Specifically, the design process of the loss function includes: in This represents the total loss of the model in a single iteration, used for backpropagation to update all weight parameters of the network. This represents the mean square error main loss, which measures the global numerical deviation between the predicted resistivity and the true resistivity. It represents the total variational canonical loss, which constrains the lateral and longitudinal smoothness of the prediction profile and suppresses resistivity jumps between measurement points. This represents the TV loss weighting coefficient, which is fixed in this method to balance fitting accuracy and profile continuity.

[0038] (1) Mean square error loss Partial Formulas in, This represents the total number of pixels within a single batch (batch size × image width × image height). This represents the normalized logarithmic resistivity prediction value of the i-th pixel output by the network; This represents the true label of the normalized logarithmic resistivity corresponding to the i-th pixel.

[0039] (2) Total variational regularized loss Partial Formulas Where B represents the training batch size; H represents the number of pixels in the vertical (depth) direction of a single image; and W represents the number of pixels in the horizontal (survey lines and points) direction of a single image. This represents the predicted resistivity of the b-th image, the h-th depth layer, and the w-th measurement point location; the first summation represents the pixel gradient penalty in the vertical depth direction, constraining the smoothness of the strata; the second summation represents the pixel gradient penalty in the horizontal measurement line direction, ensuring the continuity of the strata at adjacent measurement points.

[0040] Specifically, the present invention will be further illustrated below through embodiments, such as... Figure 4 As shown, the actual resistivity profile on the left and the density map predicted by the model on the right are highly consistent in terms of overall stratigraphic stratification and the lateral extension of high and low resistivity. The boundaries of the 80m high resistivity layer, the 280m low resistivity strip, and the 320m lower stratigraphic layer are completely restored, without the lateral resistivity jumps and stratigraphic fractures common in traditional one-dimensional single-point measurement schemes.

[0041] Example 2: To achieve the above objective, such as Figure 5 As shown, based on Embodiment 1, this invention discloses a fast inversion system for airborne electromagnetic systems based on two-dimensional images using deep learning, comprising: Data acquisition module 11 is used to acquire one-dimensional electromagnetic response sequences corresponding to multiple consecutive measuring points on the measuring line, and to vertically stack and splice the one-dimensional electromagnetic response sequences on the same measuring line according to the spatial order of the measuring points along the measuring line to obtain a two-dimensional original feature matrix. The horizontal direction of the two-dimensional original feature matrix corresponds to the measuring points in the measuring line space. The inversion output module 12 is used to process the two-dimensional original feature matrix to obtain a two-dimensional image dataset. The two-dimensional image dataset is input into a pre-established end-to-end two-dimensional convolutional coding network model, and the output is a two-dimensional resistivity image, which serves as the fast inversion result of deep learning for airborne electromagnetic systems. The end-to-end two-dimensional convolutional coding network model is used to extract the lateral continuous features of the two-dimensional image dataset.

[0042] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0043] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0044] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0045] The foregoing has shown and described the basic principles, main features, and advantages of this disclosure. Those skilled in the art should understand that this disclosure is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this disclosure. Various changes and modifications can be made to this disclosure without departing from its spirit and scope, and all such changes and modifications fall within the scope of this disclosure as claimed.

Claims

1. A fast inversion method for airborne electromagnetic systems based on two-dimensional images using deep learning, characterized in that, The method includes the following steps: A one-dimensional electromagnetic response sequence corresponding to multiple consecutive measuring points on the measuring line is obtained. The one-dimensional electromagnetic response sequences on the same measuring line are stacked and spliced ​​vertically according to the spatial order of the measuring points along the measuring line to obtain a two-dimensional original feature matrix. The horizontal direction of the two-dimensional original feature matrix corresponds to the measuring points in the measuring line space. The original two-dimensional feature matrix is ​​sampled to obtain a two-dimensional image dataset. The two-dimensional image dataset is then input into a pre-established end-to-end two-dimensional convolutional coding network model, and the output is a two-dimensional resistivity image, which serves as the result of rapid inversion of aerospace electromagnetic deep learning. The end-to-end two-dimensional convolutional coding network model is used to extract the lateral continuous features of the two-dimensional image dataset.

2. The fast inversion method for airborne electromagnetic deep learning based on two-dimensional images according to claim 1, characterized in that, The processing procedure of the end-to-end two-dimensional convolutional coding network model includes the following steps: By extracting shallow electromagnetic spatial features from a two-dimensional image dataset, and fusing multi-scale information from the transverse and longitudinal directions of the stratigraphy through multi-scale residual void convolution, the high and low resistivity anomaly regions of the two-dimensional image dataset are highlighted through CBAM adaptive weighting. Finally, a continuous and smooth two-dimensional resistivity profile is output as a two-dimensional resistivity image.

3. The fast inversion method for airborne electromagnetic deep learning based on two-dimensional images according to claim 2, characterized in that, The end-to-end two-dimensional convolutional coding network model includes a basic shallow feature extraction module, a three-level residual dilated convolutional feature fusion module, a CBAM channel-space dual attention module, and an output mapping module.

4. The fast inversion method for airborne electromagnetic deep learning based on two-dimensional images according to claim 3, characterized in that, The basic shallow feature extraction module is composed of a Conv2d two-dimensional convolutional layer, a BN2d batch normalization layer, and a LeakyReLU activation function stacked in sequence. The two-dimensional image is convolved by the Conv2d two-dimensional convolutional layer, and the convolution output is fed into the BN2d batch normalization layer for feature standardization. The standardized convolutional features are then nonlinearly mapped by the LeakyReLU activation function to finally obtain the electromagnetic shallow spatial features.

5. The fast inversion method for airborne electromagnetic deep learning based on two-dimensional images according to claim 4, characterized in that, The three-level residual void convolution feature fusion module adopts the internal structure of ResBlock residual units, which is composed of three layers of differentiated custom residual units: layer1, layer2, and layer3 connected in series. By inputting the electromagnetic shallow spatial features into layer1, the module extracts the stratigraphic details of the local area of ​​the measuring point in the lateral measuring line direction and the longitudinal depth direction as the first feature map and inputs it into layer2. Layer2 captures the large-scale continuous stratigraphic structure along the measuring line as the second feature map. The first and second feature maps are input into layer3. Layer3 fuses the first and second feature maps to output a multi-scale fused feature map.

6. The fast inversion method for airborne electromagnetic deep learning based on two-dimensional images according to claim 5, characterized in that, The CBAM channel spatial dual attention module includes a channel attention CA submodule and a spatial attention SA submodule. The channel attention CA submodule performs global average pooling to compress the spatial dimension of the multi-scale fused feature map. The spatial attention SA submodule performs global average pooling and global max pooling on the compressed feature map. Two sets of spatial features are spliced ​​along the channel dimension to highlight the high and low resistance abnormal regions of the two-dimensional image dataset and obtain the attention-enhanced feature map. The attention-enhanced feature map is compressed by a two-layer convolutional cascade structure in the output mapping module, resulting in a continuous and smooth two-dimensional resistivity profile.

7. The fast inversion method for airborne electromagnetic deep learning based on two-dimensional images according to claim 6, characterized in that, The training loss function of the end-to-end two-dimensional convolutional coding network model is obtained by summing the mean squared error loss and the total variational regularization loss.

8. The fast inversion method for airborne electromagnetic deep learning based on two-dimensional images according to claim 1, characterized in that, The process of sample processing on the two-dimensional original feature matrix includes: The original two-dimensional feature matrix is ​​segmented into blocks using an overlapping sliding window method, where the horizontal step size of the window is set to half the width of the window to obtain subsample images; The subsample images are filled with missing and outlier values ​​in both directions. Then, the numerical range is compressed by logarithmic transformation. After global normalization, the normalized sample images are converted to multi-channel tensor format, and finally, a two-dimensional image dataset is output.

9. A rapid inversion system for airborne electromagnetic systems based on two-dimensional images using deep learning, comprising the rapid inversion method for airborne electromagnetic systems based on two-dimensional images as described in any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire one-dimensional electromagnetic response sequences corresponding to multiple consecutive measuring points on the measuring line. The one-dimensional electromagnetic response sequences on the same measuring line are vertically stacked and spliced ​​according to the spatial order of the measuring points along the measuring line to obtain a two-dimensional original feature matrix. The horizontal direction of the two-dimensional original feature matrix corresponds to the measuring points in the measuring line space. The inversion output module is used to process the original two-dimensional feature matrix to obtain a two-dimensional image dataset. The two-dimensional image dataset is then input into a pre-established end-to-end two-dimensional convolutional coding network model, and the output is a two-dimensional resistivity image, which serves as the fast inversion result of deep learning for airborne electromagnetics. The end-to-end two-dimensional convolutional coding network model is used to extract the lateral continuous features of the two-dimensional image dataset.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, The memory stores a computer program that can run on a processor. When the processor loads and executes the computer program, it employs the rapid inversion method for airborne electromagnetic deep learning based on two-dimensional images, as described in any one of claims 1 to 8.

Images