Aviation electromagnetic rapid imaging method and device based on GAF-CNN

Abstract

The invention discloses an aviation electromagnetic rapid imaging method and device based on a GAF-CNN, and belongs to the technical field of processing and interpretation of aviation electromagnetic (AEM) data. The method comprises the following steps: acquiring aviation electromagnetic original data and carrying out data preprocessing; mapping the preprocessed data to a spatial domain based on a Gram matrix, and generating a Gramian angle summation field and a Gramian angle difference field; designing a deep convolutional neural network, wherein the network comprises four groups of convolutional layers and two full connection layers which are in communication connection in sequence; and inputting the preprocessed data, the Gramian angle summation field data and the Gramian angle difference field data into a trained deep convolutional neural network to obtain an underground resistivity model. According to the method, through innovative feature representation and deep learning network architecture, the time sequence characteristics and the space characteristics of the electromagnetic data are organically combined, and the inversion precision and the calculation efficiency are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of processing and inversion imaging of airborne electromagnetic (AEM) data, and specifically relates to an airborne electromagnetic fast imaging method and device based on Gramian angular field (GAF) and deep convolutional neural network (CNN). Background Art

[0002] Airborne electromagnetic method has become an important means in the field of geophysical exploration. Its high detection efficiency, wide spatial coverage and strong adaptability to complex terrain have made it play an important role in underground structure detection, mineral resource exploration, groundwater monitoring and environmental monitoring. However, the massive data generated by airborne electromagnetic detection and the high nonlinearity of the inversion problem have brought great challenges to achieving efficient and accurate imaging results. Traditional inversion methods have strong dependence on the initial model, huge calculation volume, serious noise interference and low optimization efficiency, which seriously restrict the further development and application of this technology, and put forward higher requirements on the effectiveness and reliability of data processing and inversion strategies.

[0003] In the past few decades, the industry has developed a variety of tools to interpret airborne electromagnetic data. Initially, people tried to quickly convert the detected electromagnetic data into resistivity parameters by using conductivity depth imaging (CDI) with the concept of pseudo-layers. Although this method can quickly obtain the resistivity distribution of underground structures, due to the lack of sufficient geological guidance, its inversion results are easily affected by geophysical noise, resulting in the appearance of misleading features.

[0004] In order to overcome the limitations of the CDI method, researchers have gradually turned to more complex optimization inversion methods. These methods are based on the construction of the objective function and use local or global optimization strategies to obtain the geoelectric model at the minimum value of the objective function through a large number of forward iterative calculations.

[0005] In recent years, with the rapid development of machine learning technology, the application of deep learning in geophysical inversion has gradually received attention. A variety of deep learning models, such as fully connected neural networks, convolutional neural networks (CNNs), and long short-term memory networks (LSTMs), have been applied to airborne electromagnetic inversion, significantly improving the robustness and adaptability of the models. However, existing methods usually rely on manual feature extraction, and have problems such as insufficient feature extraction capabilities and poor model training effects, making it difficult to give full play to the advantages of deep learning. In addition, existing deep learning strategies are mainly based on one-dimensional models and lack the ability to effectively handle high-dimensional data inversion problems, resulting in low inversion accuracy, slow calculation speed, and poor robustness to noise interference, which limits their actual application in airborne electromagnetic inversion. Summary of the invention

[0006] In order to solve the above technical problems, the present invention provides an airborne electromagnetic rapid imaging method and device based on GAF-CNN. Through innovative feature representation and deep learning network architecture, the time series characteristics and spatial characteristics of electromagnetic data are organically combined, which significantly improves the inversion accuracy and computational efficiency. Specifically, the present invention uses Gramian angular field (GAF) to convert one-dimensional time series electromagnetic data into two-dimensional images, fully retaining the temporal and spatial characteristics of the data; at the same time, combined with the powerful feature extraction capability of deep convolutional neural network (CNN), efficient and accurate imaging of resistivity distribution under complex geological conditions is achieved. It can not only effectively reduce the dependence on the initial model, but also has strong anti-noise ability, providing a new solution for airborne electromagnetic resistivity rapid imaging.

[0007] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0008] A GAF-CNN-based aerial electromagnetic rapid imaging method, the method comprising:

[0009] Step 1: Collect raw airborne electromagnetic data and perform data preprocessing;

[0010] Step 2: Map the preprocessed data to the spatial domain based on the Gram matrix to generate the Gramian angle sum field and the Gramian angle difference field;

[0011] Step 3, designing a deep convolutional neural network, the network comprising four groups of convolutional layers and two fully connected layers that are sequentially connected, the four groups of convolutional layers are used to extract multi-level features in electromagnetic signals, and the two fully connected layers are used to realize resistivity inversion;

[0012] Step 4: Input the preprocessed data, Gramian angle sum field, and Gramian angle difference field data into the trained deep convolutional neural network to obtain the underground resistivity model.

[0013] On the other hand, the present invention provides an aerial electromagnetic rapid imaging device based on GAF-CNN, comprising:

[0014] An acquisition unit, used to acquire raw airborne electromagnetic data and perform data preprocessing;

[0015] A mapping unit, used for mapping the preprocessed data to the spatial domain based on the Gram matrix to generate a Gramian angle sum field and a Gramian angle difference field;

[0016] A network unit, used for designing a deep convolutional neural network, wherein the network includes four groups of convolutional layers and two fully connected layers that are sequentially connected, wherein the four groups of convolutional layers are used for extracting multi-level features in electromagnetic signals, and the two fully connected layers are used for realizing resistivity inversion;

[0017] The output unit is used to input the preprocessed data and the Gramian angle sum field and Gramian angle difference field data into the trained deep convolutional neural network to obtain the underground resistivity model.

[0018] In a third aspect, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned GAF-CNN-based aerial electromagnetic rapid imaging method.

[0019] In a fourth aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the aforementioned GAF-CNN-based aerial electromagnetic rapid imaging method.

[0020] The beneficial effects of the present invention are:

[0021] The present invention realizes real-time resistivity inversion through a deep convolutional neural network without iterative calculation, which significantly improves the calculation efficiency and is particularly suitable for real-time inversion of massive airborne electromagnetic data.

[0022] The GAF conversion can effectively preserve the spatiotemporal characteristics of electromagnetic signals. At the same time, CNN automatically filters noise and extracts useful information through multi-layer convolution and pooling operations. It can maintain a high inversion accuracy when there is noise in the data, significantly improving the stability and reliability of the inversion results.

[0023] Through an end-to-end deep learning model, physical prior constraints are introduced into the deep learning model, combined with geoelectric physics knowledge, the potential laws in the data are automatically learned, the dependence on the initial model is reduced, various complex geological conditions are adapted, and the stability and reliability of the inversion results are improved.

[0024] Through GAF conversion, one-dimensional data is mapped to two-dimensional space, and combined with CNN's high-dimensional data processing capabilities, it is adapted to the inversion of high-dimensional electromagnetic data under complex geological conditions, significantly improving the applicability and imaging accuracy of the model.

[0025] In summary, the present invention is significantly superior to the existing technologies in terms of computational efficiency, noise resistance, feature extraction capability, initial model dependence, high-dimensional data processing capability, and combination of physics and data drive, and provides an efficient and accurate solution for the practical application of airborne electromagnetic detection technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of the GAF-CNN-based aerial electromagnetic rapid imaging method and device of the present invention;

[0027] Figure 2 This is the flow chart of electromagnetic data conversion based on Gramian angular field;

[0028] Figure 3 Schematic diagram of the deep convolutional neural network structure;

[0029] Figure 4 Generate example graphs for GASF and GADF images based on Gramian angular fields;

[0030] Figure 5 The loss function in the network training process changes with the number of iterations, where (a) is the GAF-CNN network training method and (b) is the ANN network training method;

[0031] Figure 6 The comparison diagram of the inversion results of different simulation data methods and the real model, where (a) is the noise-free simulation data condition, and (b) is the noise-containing simulation data condition;

[0032] Figure 7 Comparison of inversion results of different methods for measured data, where (a) is the Occam inversion result of measured data, (b) is the ANN inversion result of measured data, and (c) is the GAF-CNN inversion result of measured data. DETAILED DESCRIPTION

[0033] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0034] like Figure 1 As shown, the present invention provides a GAF-CNN-based airborne electromagnetic rapid imaging method, the core of which is to convert one-dimensional airborne electromagnetic data into two-dimensional images through Gramian angular field (GAF), and use deep convolutional neural network (CNN) to achieve efficient and accurate resistivity inversion. The specific steps are as follows:

[0035] Step 1: Collect raw airborne electromagnetic data and perform data preprocessing; including:

[0036] Step 1.1: Remove the background field; assume that the underground medium is a uniform half-space and calculate the background field response by forward modeling. , and from the original data Subtract the background field response , get the abnormal field response :

[0037] ,

[0038] and denote the reference uniform half-space model and the real model, respectively. represents the forward operator.

[0039] Preferably, other denoising methods, such as wavelet transform or adaptive filtering, can be used to replace the background field removal step.

[0040] Step 1.2: Data normalization;

[0041] Step 1.2.1, model parameter normalization: Since there is no linear relationship between the forward response and resistivity, the real model in the input raw data Perform logarithmic transformation and normalize it to the range [0,1]. The normalization formula is:

[0042] ,

[0043] in, For the real model, is the normalized true model, and The entire real model The maximum and minimum values ​​in space.

[0044] Step 1.2.2: Normalize the input data: Before performing the Gramian angular field (GAF) conversion, the input data abnormal field response needs to be normalized. and flight altitude Normalization is performed to ensure that the data are at the same level. The normalized formula is:

[0045] ,

[0046] in, is the normalized abnormal field response, is a one-dimensional time series, and Abnormal field response The maximum and minimum values ​​of .

[0047] Flight altitude The normalization of is done in the same way:

[0048] ,

[0049] in, is the normalized flight altitude, and Flight altitude The maximum and minimum values ​​of .

[0050] Step 2: Gramian angular field conversion: Based on the Gram matrix, the normalized abnormal field response is mapped to the spatial domain to generate the Gramian angular sum field (GASF) and the Gramian angular difference field (GADF);

[0051] Gramian angular field transformation is a technique for converting time series data into two-dimensional images. Its core idea is to use the Gram matrix to map the time series to the spatial domain, thereby retaining the temporal characteristics and local structure information of the data. Figure 2 shown.

[0052] Step 2.1, polar coordinate conversion: convert the normalized abnormal field response X into polar coordinate representation. The specific steps include:

[0053] Calculate the angle for each data point :

[0054] ,

[0055] in, is the first dimension of the time series elements;

[0056] Calculating Timestamps Radius :

[0057] ,

[0058] in, is the timestamp, is the length of the one-dimensional time series.

[0059] Step 2.2, generate Gramian angular field images: Based on polar coordinate representation, generate two two-dimensional images, Gramian angular sum field (GASF) and Gramian angular difference field (GADF), such as Figure 4 As shown. The calculation formulas of GASF and GADF are:

[0060] ,

[0061] ,

[0062] GASF and GADF images can preserve the spatiotemporal correlation of time series, which is convenient for subsequent convolutional neural network processing.

[0063] Preferably, other time series visualization methods, such as Markov transition field (MTF) or recurrence graph (RP), can be used to convert the one-dimensional time series into a two-dimensional image.

[0064] Step 3: Design a deep convolutional neural network, which includes multiple groups of convolutional layers and pooling layers that are sequentially connected; based on the image-processed signal, a deep convolutional neural network is constructed. The network has multiple convolutional layers and pooling layers, which can effectively extract deep features in electromagnetic signals and use fully connected layers to achieve the final resistivity inversion, such as Figure 3 shown.

[0065] For network input: The network input is GASF, GADF images and flight height images, and the output is the underground resistivity model.

[0066] Regarding the network structure: the network uses four groups of convolutional layers, each group of convolutional layers is followed by a maximum pooling layer, and finally the resistivity model is output through a fully connected layer. The specific structure is as follows:

[0067] The convolution kernel size of the first two groups of convolutional layers is 3×3, with a step size of 1. The first convolutional layer is used to initially extract local features of the image, such as local structure and edge information; the second convolutional layer further deepens the feature extraction and recognizes more complex image patterns, such as texture and shape. Each group of convolutional layers is followed by a 2×2 maximum pooling layer with a step size of 2. The role of the pooling layer is to reduce the size of the feature map and reduce the computational complexity, while retaining the main features and enhancing the generalization ability of the model.

[0068] The convolution kernel size of the last two groups of convolutional layers is 2×2, with a stride of 1. The third and fourth convolutional layers further extract high-level features, capture global information and abstract patterns in the image, and provide rich feature representations for subsequent fully connected layers. Each group of convolutional layers is followed by a 2×2 maximum pooling layer with a stride of 2 to further compress the feature map and retain key information.

[0069] After flattening the outputs of the last two groups of convolutional layers, they are mapped to resistivity model parameters through two fully connected layers. The first fully connected layer is used to integrate the multi-level feature information extracted by the convolutional layer, and the second fully connected layer maps the integrated multi-level feature information to the final resistivity imaging result. The output of the network is a 16×1 resistivity model parameter, corresponding to the resistivity value of the hierarchical resistivity model.

[0070] Regarding activation functions and regularization: The convolutional layer uses the ReLU activation function, and a Dropout layer is added after the two fully connected layers to prevent overfitting. At the same time, batch normalization is used to speed up training and improve model stability.

[0071] Preferably, an extensible multi-task learning framework can be used to simultaneously predict multiple parameters such as resistivity and density based on CNN, and the model efficiency can be improved by sharing feature layers; or a generative adversarial network (GAN) can be introduced to generate high-resolution imaging results using generator-discriminator adversarial training.

[0072] Step 4: Training and optimization of deep convolutional neural network: input the preprocessed data and the Gramian angle sum field and Gramian angle difference field data into the trained deep convolutional neural network to obtain the underground resistivity model.

[0073] Step 4.1, data set generation and division;

[0074] Dataset generation: A synthetic dataset was used for network training, which contains 100,000 layered resistivity models and their corresponding airborne electromagnetic responses. The resistivity value of each model was randomly generated between 1 and 1000 Ω·m, the number of layers was randomly selected between 1 and 16, and the total thickness was 600 meters. The flight altitude was randomly generated between 20 meters and 80 meters to simulate actual flight conditions.

[0075] Data enhancement: In order to improve the robustness of the model, data enhancement methods (such as rotation, translation, and noise addition) are used to expand the training set, which can effectively increase the diversity of training samples and enhance the model's adaptability to noise and complex geological conditions.

[0076] Dataset division: The dataset is divided into a training set (70%), a validation set (15%), and a test set (15%). The training set is used for model training, the validation set is used to adjust hyperparameters and prevent overfitting, and the test set is used to finally evaluate model performance.

[0077] Step 4.2: Training process;

[0078] Step 4.2.1, input training data: The input training data is image data converted by Gramian angular field (GAF), including Gramian angular sum field (GASF), Gramian angular difference field (GADF) and flight height image. These images represent electromagnetic data of different time series, and each image corresponds to a set of underground resistivity distribution.

[0079] Step 4.2.2, Neural Network Calculation of Resistivity Distribution: The convolutional neural network (CNN) passes the input image through the network for processing, and finally generates a predicted resistivity distribution map after multiple layers of convolution, pooling, activation and other operations. The network structure includes four groups of convolution layers, each of which is followed by a maximum pooling layer, and finally outputs the resistivity model parameters through two fully connected layers.

[0080] Step 4.2.3, calculate the loss function: use mean square error (MSE) as the loss function to measure the difference between the predicted resistivity distribution and the true resistivity distribution. The MSE calculation formula is as follows:

[0081] ,

[0082] in, is the normalized prediction model, Normalized true model, Indicates the sample size.

[0083] like Figure 5 As shown in (a)-(b), the loss function of GAF-CNN and ANN networks in the training phase changes with the number of iterations. The training error (dashed line) and verification error (solid line) of the two networks both decrease synchronously and gradually stabilize to the minimum value. The network learning process is stable without overfitting or underfitting. GAF-CNN reaches convergence after about 230 iterations, which is significantly faster than the 540 iterations of ANN. The convergence error (final loss value) of GAF-CNN is lower than that of the ANN algorithm, indicating that its inversion accuracy is better. In summary, GAF-CNN is significantly better than the ANN algorithm in both training efficiency (faster convergence) and inversion accuracy (lower error).

[0084] Without loss of generality, in addition to the mean square error (MSE), you can also use the cross entropy loss function, Huber loss function, etc. to adapt to different data distributions and task requirements.

[0085] Preferably, the inversion accuracy can be further improved through integrated learning strategies (such as weighted fusion of CNN+Transformer models), or the physical model can be combined to generate simulated data to enhance the training set, thus realizing physical and data-driven hybrid modeling.

[0086] Step 4.2.4, Optimization algorithm: Use the Adam optimizer for network training. The Adam optimizer combines the advantages of the momentum method and adaptive learning rate, which can accelerate convergence and improve training stability. At the same time, the learning rate is dynamically adjusted with the learning rate scheduler to further improve the convergence speed. Without loss of generality, in addition to the Adam optimizer, you can also use optimization algorithms such as stochastic gradient descent (SGD) and RMSprop.

[0087] Step 4.2.5, Back Propagation and Weight Update: After calculating the loss function, the gradient is calculated through the back propagation algorithm, and these gradients are used to update the weights and biases in the network to minimize the loss function. The back propagation process helps the network gradually learn better parameters, thereby improving the prediction accuracy.

[0088] Step 4.2.6, repeat training: Repeat steps 4.2.2 to 4.2.5 until the loss function converges or reaches the predetermined number of iterations.

[0089] Step 4.2.7, Early Stopping: During the training process, early stopping is used to prevent overfitting. When the validation set loss function stops decreasing after several consecutive rounds, training is terminated early to ensure the network has the best performance on the validation set. Without loss of generality, L2 regularization, Dropout and other methods can also be used to prevent overfitting.

[0090] Step 4.3: Training results and evaluation;

[0091] Training loss and validation loss: Monitor the changes in the loss function of the training set and validation set to ensure that the model performs well on both the training set and validation set to avoid overfitting or underfitting.

[0092] Test set evaluation: Use the test set to evaluate the final performance of the model to ensure the model's generalization ability on unseen data.

[0093] Noise robustness test: By adding different levels of Gaussian noise to the test dataset, the robustness of the model in a noisy environment is verified.

[0094] Through the above steps 4.1 to 4.3, the neural network gradually learns the mapping relationship between electromagnetic data and underground resistivity distribution, thereby providing an accurate model for subsequent rapid imaging and inversion.

[0095] In order to verify the effectiveness and robustness of the algorithm developed in this invention, simulation data is first used for testing. The generation of simulation data is based on a known geoelectric model and can provide accurate resistivity distribution as a reference to evaluate the inversion accuracy and computational efficiency of the algorithm. Figure 6 As shown in (a)-(b), the inversion results of different methods (Occam, ANN and GAF-CNN) based on simulation data are compared with the real model, where (a) is the comparison between the inversion results of different methods (Occam, ANN and GAF-CNN) and the real model under the condition of noise-free simulation data, and (b) is the comparison between the inversion results of different methods (Occam, ANN and GAF-CNN) and the real model under the condition of noisy simulation data.

[0096] A variety of typical geoelectric models are designed, including layered models, fault models, complex geological structure models, etc., covering different resistivity distribution characteristics. In this invention, only the layered model is used as an example. Based on electromagnetic field theory, the geoelectric model is forward calculated to generate corresponding airborne electromagnetic response data. Random noise (such as Gaussian noise) is added to the simulation data to simulate the noise interference in actual detection.

[0097] The noise-free and noise-containing simulation data are input into the algorithm of the present invention to generate a resistivity distribution map, and the inversion imaging results are compared with the real geoelectric model, the traditional Occam inversion method and the conventional artificial neural network (ANN) method. The inversion accuracy is evaluated by indicators such as the mean square error (MSE). The time from data input to output of the resistivity distribution map is recorded, and compared with the traditional inversion method (such as the Gauss-Newton method) and the ANN method to evaluate the improvement of the calculation efficiency. The algorithm is run with and without noise added to evaluate the stability and reliability of the inversion results of the present invention and the results of the traditional inversion method.

[0098] The simulation results show that the algorithm of the present invention can achieve high-precision resistivity inversion under various geoelectric models, and the MSE is significantly lower than that of the traditional method. Compared with the traditional method, the calculation time of the algorithm of the present invention is shortened by more than 80%, especially on large-scale data sets. When the simulation data contains noise, the algorithm of the present invention can still maintain a high inversion accuracy and show strong anti-noise ability.

[0099] On the basis of simulation data verification, the actual application effect of the algorithm of the present invention is further verified by measured data. The measured data comes from actual airborne electromagnetic detection tasks and can comprehensively evaluate the applicability of the algorithm under complex geological conditions. Figure 7 As shown in the figure, the inversion results of different methods (Occam, ANN and GAF-CNN) for measured airborne electromagnetic data are compared, where (a) is the Occam inversion result of the measured data, (b) is the ANN inversion result of the measured data, and (c) is the GAF-CNN inversion result of the measured data.

[0100] The measured data are obtained from multiple airborne electromagnetic exploration projects, covering different geological conditions (such as mineral resource areas, groundwater distribution areas, and environmental monitoring areas). This paper takes the measured data of a profile in an airborne electromagnetic exploration operation of a lead-zinc mine in a mountainous area as an example. The measured data is pre-processed by denoising and normalization to ensure that the data quality meets the algorithm input requirements.

[0101] Input the measured data into the algorithm of the present invention to generate a resistivity distribution map, and compare it with geological exploration reports, drilling data, etc. to evaluate the accuracy of the inversion results. Record the time it takes for the algorithm to process the measured data to evaluate its computational efficiency in practical applications. Run the algorithm under complex geological conditions (such as faults, folds, and uneven strata) to evaluate its adaptability and inversion effect.

[0102] Verification of measured data shows that the algorithm of the present invention can accurately reflect the underground resistivity distribution, which is highly consistent with the geological exploration results, and the inversion accuracy is significantly better than the traditional method. In practical applications, the algorithm of the present invention can complete resistivity imaging in a short time and meet the needs of real-time detection. Under complex geological conditions, the algorithm of the present invention can still provide high-precision inversion results, showing strong adaptability and robustness.

[0103] On the other hand, the present invention provides an aerial electromagnetic rapid imaging device based on GAF-CNN, wherein each unit included in the device can implement each step of the aforementioned method, specifically including:

[0104] An acquisition unit, used to acquire raw airborne electromagnetic data and perform data preprocessing;

[0105] A mapping unit, used for mapping the preprocessed data to the spatial domain based on the Gram matrix to generate a Gramian angle sum field and a Gramian angle difference field;

[0106] A network unit, used for designing a deep convolutional neural network, wherein the network includes four groups of convolutional layers and two fully connected layers that are sequentially connected, wherein the four groups of convolutional layers are used for extracting multi-level features in electromagnetic signals, and the two fully connected layers are used for realizing resistivity inversion;

[0107] The output unit is used to input the preprocessed data and the Gramian angle sum field and Gramian angle difference field data into the trained deep convolutional neural network to obtain the underground resistivity model.

[0108] In a third aspect, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned GAF-CNN-based aerial electromagnetic rapid imaging method.

[0109] In a fourth aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the aforementioned GAF-CNN-based aerial electromagnetic rapid imaging method.

[0110] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A GAF-CNN-based aeromagnetic rapid imaging method, characterized in that: The method comprises: Step 1: Collect raw airborne electromagnetic data and perform data preprocessing; Step 2: Map the preprocessed data to the spatial domain based on the Gram matrix to generate the Gramian angle sum field and the Gramian angle difference field; Step 3, designing a deep convolutional neural network, the network comprising four groups of convolutional layers and two fully connected layers that are sequentially connected, the four groups of convolutional layers are used to extract multi-level features in electromagnetic signals, and the two fully connected layers are used to realize resistivity inversion; Step 4: Input the preprocessed data, Gramian angle sum field, and Gramian angle difference field data into the trained deep convolutional neural network to obtain the underground resistivity model.

2. The GAF-CNN-based airborne electromagnetic rapid imaging method according to claim 1, characterized in that: The step 1 comprises: Step 1.1, assuming that the underground medium is a uniform half-space, calculate the background field response by forward modeling, and subtract the background field response from the original data to obtain the abnormal field response; Step 1.2: Normalize the real model, abnormal field response, and flight altitude in the original data.

3. The GAF-CNN-based airborne electromagnetic rapid imaging method according to claim 1, characterized in that: The step 2 comprises: Step 2.1, convert the normalized abnormal field response into polar coordinates to obtain the angle of each data point; Step 2.2: Based on the polar coordinate representation, generate a Gramian angle sum field and a Gramian angle difference field.

4. The GAF-CNN-based aeromagnetic rapid imaging method according to claim 3, characterized in that: The calculation formulas of the Gramian angle sum field GASF and the Gramian angle difference field GADF in step 2.2 are respectively: ， ， in, Represents the angle of each data point in the polar coordinate system, is the length of the time series.

5. The GAF-CNN-based airborne electromagnetic rapid imaging method according to claim 1, characterized in that: The structures of the four groups of convolutional layers in step 3 are: The convolution kernel size of the first two groups of convolutional layers is 3×3, and the stride is 1. The first convolutional layer is used to extract low-level features of the input image, including local structure and edge information; the second convolutional layer is used to recognize image patterns, including texture and shape; the first and second convolutional layers are followed by a 2×2 maximum pooling layer with a stride of 2; The convolution kernel size of the last two groups of convolutional layers is 2×2, and the stride is 1, which is used to capture the global information and abstract patterns in the image and construct multi-level features. The third and fourth convolutional layers are both followed by a 2×2 maximum pooling layer with a stride of 2.

6. The GAF-CNN-based airborne electromagnetic rapid imaging method according to claim 1, characterized in that: The first fully connected layer of the two fully connected layers is used to integrate the multi-level feature information extracted by the four groups of convolutional layers, and the second fully connected layer maps the integrated multi-level feature information into the final resistivity imaging result; The deep convolutional neural network adopts the ReLU activation function, and a Dropout layer is added after two fully connected layers.

7. The GAF-CNN-based airborne electromagnetic rapid imaging method according to claim 1, characterized in that: The preprocessed data in step 4 includes a flight height image; and the mean square error is used as the loss function.

8. An aerial electromagnetic rapid imaging device based on GAF-CNN, characterized in that: include: An acquisition unit, used to acquire raw airborne electromagnetic data and perform data preprocessing; A mapping unit, used for mapping the preprocessed data to the spatial domain based on the Gram matrix to generate a Gramian angle sum field and a Gramian angle difference field; A network unit, used for designing a deep convolutional neural network, wherein the network includes four groups of convolutional layers and two fully connected layers that are sequentially connected, wherein the four groups of convolutional layers are used for extracting multi-level features in electromagnetic signals, and the two fully connected layers are used for realizing resistivity inversion; The output unit is used to input the preprocessed data and the Gramian angle sum field and Gramian angle difference field data into the trained deep convolutional neural network to obtain the underground resistivity model.

9. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs; Wherein, when one or more programs are executed by the one or more processors, the one or more processors implement the GAF-CNN-based aerial electromagnetic rapid imaging method described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: Executable instructions are stored thereon, and when the instructions are executed by the processor, the processor can implement the GAF-CNN-based aerial electromagnetic rapid imaging method described in any one of claims 1 to 7.

Images