Magnetotelluric two-dimensional inversion method based on DeepLabV < 3 + >

Abstract

The invention discloses a DeepLabV < 3 + >-based magnetotelluric two-dimensional inversion method, and belongs to the field of geophysical inversion methods, and the method comprises the steps: constructing a plurality of underground resistivity distribution models, and obtaining corresponding apparent resistivity and phase data through forward modeling calculation; designing a deep learning network based on a DeepLabV < 3 + > architecture, and enhancing the capability of capturing multi-scale geological features by using a cavity convolution and improved cavity spatial pyramid pooling (ASPP) module of the deep learning network; di ce Loss is adopted to replace a traditional cross entropy loss function, and the method is more suitable for the characteristics of anomaly and background information imbalance in geophysical observation data; according to the method, the network is trained and verified through the training sample set, and the optimal network parameters are obtained; the trained network is utilized to directly map the apparent resistivity and phase data of a test set into underground resistivity distribution, the end-to-end inversion process is realized, and the multi-scale feature extraction and semantic segmentation capabilities of the DeepLabV3 + network are utilized, so that compared with a traditional inversion method, the method has the advantages of being independent of an initial model, capable of obtaining a global optimal solution and the like.

Description

Technical Field

[0001] The present invention belongs to the field of geophysical inversion methods, and specifically relates to a two-dimensional magnetotelluric inversion method based on DeepLabV3+. Background Art

[0002] Magnetotellurics (MT) is a geophysical exploration technology based on natural alternating electromagnetic fields. It infers the resistivity structure of underground media by measuring the mutually orthogonal electric and magnetic field components on the surface. Its working principle is to use natural electromagnetic fields (originating from solar wind and lightning activity) with a frequency range of 0.001Hz-10kHz as the field source, and to detect electrical characteristics at different depths (tens of meters to hundreds of kilometers) through the skin effect of electromagnetic waves of different frequencies. This method has the advantages of a large detection depth range, no need for artificial sources, and sensitivity to low-resistance bodies. It is widely used in oil and gas and mineral exploration, geothermal resource evaluation, deep structure research and other fields.

[0003] Traditional magnetotelluric inversion methods are mainly linearized iterative inversion methods based on the gradient information of the objective function. This type of method approximates the nonlinear inversion problem as a linear problem and solves it by iteration. Representative methods include Occam inversion method, NLCG (nonlinear conjugate gradient) inversion method, Gauss-Newton inversion method, etc. Although linearized iterative inversion methods have been widely used, they also have some obvious limitations, such as high dependence on the initial model and prior information, and easy to fall into local optimal solutions. Another type of nonlinear inversion method that is more commonly used in low-dimensional inversion problems is the nonlinear inversion method, including simulated annealing method, genetic algorithm, ant colony algorithm, etc. They directly solve the nonlinear inverse problem. Although it is possible to obtain the global optimal solution, it usually requires high computational cost and has low practicality in high-dimensional inversion problems.

[0004] In this context, deep learning, with its excellent nonlinear mapping capabilities, provides new ideas for the intelligent inversion of MT. In the vast field of artificial intelligence and machine learning, convolutional neural network (CNN) is an important branch of deep learning research. This revolutionary neural network architecture has demonstrated remarkable technical potential in complex application scenarios such as image recognition and natural language processing with its excellent feature extraction mechanism and nonlinear conversion capabilities. CNN can automatically extract data features, reduce dependence on manual feature engineering, and demonstrate strong nonlinear modeling capabilities. Therefore, CNN has more potential than traditional methods in dealing with complex nonlinear problems. At the same time, CNN can efficiently capture the spatial features and hierarchical information of data through multi-layer convolution and pooling operations, which is highly consistent with the characteristics of MT data. Summary of the Invention

[0005] In order to overcome the above-mentioned defects, the present invention provides a two-dimensional magnetotelluric inversion method based on DeepLabV3+, which has the advantages of being independent of the initial model and being able to obtain the global optimal solution compared with traditional inversion technology.

[0006] To achieve the above-mentioned object, the present invention provides the following technical solution: a two-dimensional magnetotelluric inversion method based on DeepLabV3+, comprising constructing a two-dimensional magnetotelluric inversion training dataset, using the finite element method to forward calculate and obtain corresponding apparent resistivity and phase data; constructing a deep learning network model based on the DeepLabV3+ architecture; preprocessing the training dataset; training the model using the preprocessed training dataset; inputting measured magnetotelluric data into the trained network model for inversion, and analyzing the inversion results.

[0007] Furthermore, the step of constructing a two-dimensional magnetotelluric inversion training data set includes: constructing a variety of underground conductivity distribution models based on geological information, including a rectangular high- and low-resistance anomaly combination model, a stepped high- and low-resistance anomaly combination model, and a pyramid anomaly model; setting a range of model parameter values, and generating 20,000 groups of different geoelectric models within a regular grid by controlled step size changes and combinations of key characteristic parameters; for each group of geoelectric models, using a two-dimensional finite element method to solve Maxwell's equations and calculate the apparent resistivity and phase response at different frequency points (30 frequency points selected from the range of 0.01 Hz to 320 Hz); forming sample pairs of the geoelectric model and its corresponding apparent resistivity and phase data; and randomly dividing the sample pairs into training sets and test sets in a ratio of 9:1.

[0008] Furthermore, the step of constructing a deep learning network model based on the DeepLabV3+ architecture includes: designing a two-dimensional magnetotelluric inversion network structure based on DeepLabV3+, and the network model includes the following three main modules:

[0009] (1) Encoder module: Xception is used as the backbone network, which can be divided into three modules: EntryFlow, MiddleFlow and ExitFlow; EntryFlow is the entry part of the model, mainly used for preliminary extraction of low-level features, and contains several groups (usually 3 groups) of depth-wise separable convolution and maximum pooling operations to quickly reduce the size of feature maps and increase the number of channels; MiddleFlow is composed of multiple repeated depth-wise separable convolution stacks, mainly used to fully extract mid- and high-level semantics and feature patterns; ExitFlow is located at the end of the model, extracting higher-level abstract features, and completing the final classification prediction through global average pooling or fully connected layers;

[0010] (2) ASPP module: It contains five parallel branches: a 1×1 convolution branch, three 3×3 dilated convolution branches with different dilation rates (r=6, 12, and 18), and a global average pooling branch. The number of output channels of each branch is set to 256. The output features of each branch are concatenated in the channel dimension, compressed to 256 channels through a 1×1 convolution layer, and then processed by a batch normalization layer and a ReLU activation function;

[0011] (3) Decoder module: First, the output feature map of the ASPP module is upsampled to 1 / 4 of the original input size, and then concatenated with the low-level features output by the second stage of the encoder. Feature fusion is performed through two consecutive 3×3 convolutional layers. After being processed by the batch normalization layer and the ReLU activation function, it is upsampled again to the original input size. Finally, the number of channels is mapped to 1 through a 1×1 convolutional layer, and the final conductivity distribution prediction result is output.

[0012] Furthermore, the step of preprocessing the training data set includes: performing linear normalization processing on the apparent resistivity and phase data; and adding 5% adaptive Gaussian noise to the samples for data enhancement.

[0013] Furthermore, the step of training the model using the preprocessed training data set includes: using the Dice Loss function, whose calculation formula is:

[0014]

[0015] Among them, X is the predicted segmentation result, Y is the actual segmentation label, |X∩Y| represents the intersection of the predicted and true labels, and |X| and |Y| represent the number of pixels of the predicted and true labels, respectively.

[0016] Furthermore, the training process adopts the Adam optimizer, the initial learning rate is set to 0.001, the batch size is set to 32; the step-by-step decay learning rate adjustment strategy, the decay factor is 0.95, and the cycle is set to 100 epochs.

[0017] Furthermore, the step of inputting the measured magnetotelluric data into the trained network model for inversion includes: performing the same preprocessing operation on the measured magnetotelluric data as the training data; inputting the preprocessed data into the trained DeepLabV3+ network model; the network directly outputs the underground resistivity distribution result without iterative calculation; and performing denormalization processing on the output result to restore the true resistivity value.

[0018] Furthermore, the method also includes the steps of evaluating the quality of the inversion results: calculating the relative root mean square error (RMSE) to evaluate the overall difference between the inversion results and the true model; using actual exploration data to perform blind testing evaluation and comparing the inversion results with drilling data or other geophysical method results.

[0019] Furthermore, the method also includes the steps of model optimization and improvement: according to the evaluation results, the network structure, loss function weight or training strategy are adjusted in a targeted manner; the training samples in the low-accuracy areas are weighted to increase the attention to these areas.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] Eliminate dependence on the initial model: The method of the present invention is completely data-driven and does not require an initial model, thus avoiding the inversion failure problem caused by improper initial model selection;

[0022] Overcoming local optimal limitations: Through end-to-end mapping directly from data to model, the local optimal problem in traditional optimization methods is avoided, and the true resistivity distribution of the underground can be more accurately reflected;

[0023] Improve noise resistance: By introducing adaptive Gaussian noise into the training data for data enhancement, the noise resistance of the inversion results is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Schematic diagram of the overall process of the magnetotelluric two-dimensional inversion method based on DeepLabV3+ of the present invention;

[0025] Figure 2 Schematic diagram of the overall DeepLabV3+ network architecture used in the present invention;

[0026] Figure 3 This is a graph showing the loss of the network model training presented in the present invention;

[0027] Figure 4 A comparison diagram of the inversion results of the method of the present invention on a rectangular high- and low-resistance anomaly combination model;

[0028] Figure 5 A comparison diagram of the inversion results of the method of the present invention on a stepped high- and low-resistance anomaly combination model;

[0029] Figure 6 A comparison diagram of the inversion results of the method of the present invention on a pyramid anomaly model;

[0030] Figure 7 This is a comparison chart of the inversion performance of the method of the present invention under different noise levels. DETAILED DESCRIPTION

[0031] The technical solution of this patent is further described in detail below in conjunction with specific implementation methods.

[0032] like Figure 1 As shown, the overall implementation process of the magnetotelluric two-dimensional inversion method based on DeepLabV3+ provided by the present invention includes the following steps:

[0033] S1: Constructing a 2D magnetotelluric inversion training dataset: This training dataset is divided into three types based on the geometric anomaly morphology: a rectangular high- and low-resistance anomaly combination model, a stepped high- and low-resistance anomaly combination model, and a pyramid anomaly model. The background resistivity is set to 1000 Ω.m, and the anomaly resistivity is divided into four levels: low-resistance anomaly values ​​of 100 Ω.m and 500 Ω.m, and high-resistance anomaly values ​​of 2000 Ω.m and 5000 Ω.m.

[0034] In the parameter space construction of the geoelectric model, the present invention adopts a systematic and targeted sample generation strategy. Specifically, within a regular grid of a predefined spatial scale (length 120km, depth 50km), we generate a variety of resistivity models by controlling the step size change and combination of key characteristic parameters. These key characteristic parameters mainly include the geometric characteristics of the anomaly (such as length, thickness), spatial position and resistivity distribution. Taking into account the complexity of the parameter space, if a complete exhaustive combination method is adopted, it will inevitably lead to a sharp expansion in the number of samples, resulting in unacceptable computational overhead. To this end, the present invention proposes a data set production strategy: by designing the step size of the variable change and introducing controlled random perturbations, while ensuring computational efficiency, the diversity and statistical representativeness of the training samples are improved.

[0035] The input layer consists of two channels, specifically 20,000 sets of MT numerical simulation phase and apparent resistivity response data. During the data selection process, the response data were spatially truncated to retain only the characteristic response information of the target study area. Corresponding to the input data, the dataset labels consist of an equal number of geoelectrical models. Notably, these geoelectrical models used as labels have a wider spatial scope, encompassing not only the target study area but also the surrounding areas.

[0036] Finally, the geoelectric model and its corresponding apparent resistivity and phase data were combined into sample pairs and randomly divided into training and test sets in a ratio of 9:1, that is, 18,000 groups of samples in the training set and 2,000 groups of samples in the test set.

[0037] S2: If Figure 2 As shown in the figure, the specific implementation process of building a deep learning network model based on the DeepLabV3+ architecture is as follows:

[0038] A two-dimensional magnetotelluric inversion network structure based on DeepLabV3+ is designed. The network model mainly consists of three parts: encoder module, atrous spatial pyramid pooling (ASPP) module and decoder module.

[0039] (1) Encoder module: Xception is used as the backbone network, which can be divided into three modules: EntryFlow, MiddleFlow and ExitFlow. EntryFlow is the entry part of the model, mainly used for preliminary extraction of low-level features, and contains several groups (usually 3 groups) of depthwise separable convolution and maximum pooling operations to quickly reduce the size of the feature map and increase the number of channels; MiddleFlow is composed of multiple repeated depthwise separable convolution stacks, mainly used to fully extract mid- and high-level semantics and feature patterns; ExitFlow is located at the end of the model, extracting higher-level abstract features and completing the final classification prediction through global average pooling or fully connected layers. In addition, ExitFlow is also composed of depthwise separable convolution and layer-by-layer downsampling or feature fusion operations. Compared with the traditional Inception module, Xception no longer uses multiple convolution kernels of different sizes to operate in parallel, but decomposes the convolution into a combination of "spatial convolution separated by channel + 1×1 convolution", thereby further deepening the network depth while avoiding redundant calculations.

[0040] (2)ASPP module: The ASPP module contains 5 parallel branches:

[0041] One 1×1 convolution branch: used to capture pixel-level features. Three 3×3 dilated convolution branches with different dilation rates: the dilation rates r are set to 6, 12, and 18 respectively, used to capture contextual information of different receptive fields. One global average pooling branch: average pooling is performed on the entire feature map, then passed through a 1×1 convolution layer, and upsampled to the original feature map size to capture global contextual information. The number of output channels of each branch is set to 256. After the output features of each branch are spliced ​​in the channel dimension, the number of channels is compressed to 256 through a 1×1 convolution layer, and then processed by a batch normalization layer and a ReLU activation function.

[0042] (3) Decoder module: The specific implementation process of the decoder module is as follows: First, the output feature map of the ASPP module is upsampled to 1 / 4 of the original input size, and the bilinear interpolation method is used for upsampling. Then it is spliced ​​with the low-level features (Low-level Features) output by the second stage of the encoder. The low-level features are first reduced to 48 channels through a 1×1 convolution layer, and then spliced ​​with the upsampled high-level features. Feature fusion is performed through two consecutive 3×3 convolution layers (with 256 channels), and each convolution layer is followed by a batch normalization layer and a ReLU activation function. The fused features are upsampled again to the original input size, and the bilinear interpolation method is also used. Finally, the number of channels is mapped to 1 through a 1×1 convolution layer, and the final resistivity distribution prediction result is output.

[0043] S3: The specific implementation process of preprocessing the magnetotelluric 2D inversion training dataset is as follows:

[0044] The apparent resistivity and phase data are linearly normalized and the calculation formula is:

[0045]

[0046] Among them, X is the original data, X min and X max are the minimum and maximum values ​​of the data respectively. After normalization, the resistivity and phase data are mainly distributed in the interval [0,1].

[0047] Adaptive Gaussian noise is added to the sample for data enhancement. The calculation formula is:

[0048] x ′ =x+N(μ,σ·|x|)

[0049] x in the formula ′ Represents the new value after adding Gaussian noise, and x represents the value of the original data. N(μ,σ·|x|) represents the noise generated from the normal distribution, where: μ represents the mean of the Gaussian noise, which is taken as 0 here. σ·|x| represents the standard deviation of the noise; it is a fixed proportional factor that controls the basic intensity of the noise, which is taken as 0.05 here; |x| represents the absolute value of the original signal, ensuring that the standard deviation is proportional to the size of the original signal. The present invention multiplies the standard deviation of the noise by the absolute value of the original signal, so that the intensity of the noise can be adaptively adjusted, thereby introducing stronger noise when the signal value is large, and introducing weaker noise when the signal value is small, so as to more realistically simulate the noise impact in the actual environment.

[0050] S4: Use the DiceLoss loss function, which is calculated as follows:

[0051]

[0052] Among them, X is the predicted segmentation result, Y is the actual segmentation label, |X∩Y| represents the intersection of the predicted and true labels, and |X| and |Y| represent the number of pixels of the predicted and true labels, respectively.

[0053] The traditional cross-entropy loss function adopts an equally weighted grid point error accumulation strategy when processing data samples. This feature exposes serious limitations in scenarios with highly unbalanced sample distribution. In such cases, the sparse number of positive samples can easily be "swamped" by the errors of a large number of negative samples, resulting in a significant weakening of the recognition ability of key minority categories during the model learning process, which is ultimately reflected in the difficulty in accurately capturing the subtle features of electrical boundaries in the inversion results. In contrast, DiceLoss significantly improves this problem. Its core mechanism is that the loss function no longer relies solely on the local prediction value of the current grid point, but calculates the intersection of the predicted sample and the labeled sample, so that the error calculation has more global and correlation characteristics. This design enables the positive samples with a smaller proportion to have a more significant impact on the network weight update, thereby effectively improving the model's feature extraction and boundary recognition capabilities when processing unbalanced data sets.

[0054] S5: The specific implementation process of training the DeepLabV3+ network model using the preprocessed training dataset is as follows:

[0055] The DeepLabV3+ convolutional neural network framework of the present invention is built in the Python3.8 language environment based on the PyTorch deep learning framework, and the network model is trained using the generated training data. Pre-trained Xception weights are used for initialization; the Adam optimization algorithm is used during training, and the learning rate is 0.001; the batch size is set to 32 according to the GPU memory size; a step-by-step decay learning rate adjustment strategy is adopted, and the decay factor is 0.95; the network is trained for a total of 100 epochs, and after each epoch training is completed, the performance of the model is monitored using a validation set to prevent overfitting, and the training effect of the model is evaluated at the same time. It is worth noting that once the network training is completed, the inversion operation takes almost no time. The loss value change curve of the model training process (such as Figure 3 As shown in Figure 3), the training set loss value finally converged to 0.0345, and the validation set loss value stabilized at 0.0855.

[0056] S6: A rectangular high-low resistance combination model, a stepped high-low resistance combination model, and a pyramid model are randomly selected from the test set as examples of inversion results (see Figures 4 to 6 ), where the first example is the geoelectric model and its apparent resistivity and phase response, and the second example is the convolutional neural network inversion result and the apparent resistivity and phase response of the inversion result.

[0057] Figure 4 This figure compares the inversion results of Model I, a rectangular high- and low-resistivity combination, with the true resistivity. The background resistivity is concentrated at 1000 Ω.m, the rectangular low-resistivity anomaly has a resistance of 100 Ω.m, and the rectangular high-resistivity anomaly has a resistance of 5000 Ω.m. As can be seen from the figure, the model not only accurately locates the anomaly spatially, but also achieves highly consistent reconstruction results with the theoretical model in terms of geometry and quantitative resistivity characterization. The apparent resistivity and phase responses are also essentially the same. Based on these results, we conclude that convolutional neural networks exhibit significant nonlinear mapping capabilities in the inversion of magnetotelluric data, especially for the inversion of simple geometric rectangular high- and low-resistivity anomalies, where they can produce excellent inversion results.

[0058] Figure 5 This figure compares the inversion results of the stepped high- and low-resistivity combination model with the true resistivity. The background resistivity is 1000 Ω.m, the resistance of the stepped low-resistivity anomaly is 100 Ω.m, and the resistance of the stepped high-resistivity anomaly is 5000 Ω.m. As can be seen from the figure, the apparent resistivity response and phase response are highly similar, and the deep learning-driven convolutional neural network successfully captures and reproduces the overall morphological characteristics of the stepped anomaly. However, in terms of the precise depiction of boundary details and geometric corners, there are still noticeable deviations between the model reconstruction results and the theoretical model. These subtle differences are rooted in the limitations of magnetotelluric (MT) data acquisition: the inherent sparsity and low resolution of the data constrain the ultimate accuracy of feature reconstruction, making it difficult for the neural network to fully restore the microstructural details of the anomaly boundary.

[0059] Figure 6 The figure shows a comparison between the inversion results of the pyramid model and the true resistivity. The resistivity values ​​of the anomaly model decrease from top to bottom: 5000Ω.m, 2000Ω.m, 500Ω.m, and 100Ω.m, respectively. The background resistivity value is 1000Ω.m. As can be seen from the figure, the apparent resistivity response and phase response are essentially the same, and the boundaries between the different resistivity anomalies are well demarcated. However, continuing the previous research, the inversion results of the pyramid model again exhibit similar characteristics to the previous rectangular model: the anomaly boundary outlines show a slight geometric diffusion phenomenon. This ambiguity in boundary details causes some structural deviations between the reconstructed results and the theoretical model.

[0060] S7: Considering the impact of actual data noise on inversion, in order to further explore the potential impact of data quality on the performance of convolutional neural networks, this study introduced an adaptive Gaussian noise strategy to conduct targeted perturbation experiments on synthetic input data. First, a rectangular high-low resistance combination model was established. Secondly, a forward numerical simulation method was used to generate the apparent resistivity of the geoelectric model and its corresponding phase response data. Then, adaptive Gaussian noise was added to the model response, corresponding to 5%, 10% and 20% of each response amplitude (such as Figure 7 Finally, the noisy model response is input into the trained network to obtain the result.

[0061] Figure 7 Row (III) shows the inversion performance of the convolutional neural network under different levels of Gaussian noise contamination. This comparative analysis reveals the impact of noise on the geoelectric model inversion. Since 5% Gaussian noise was added to one-third of the sample data during training, the shape, position, and resistivity values ​​of the anomaly are largely consistent with the actual model in both the absence and addition of 5% Gaussian noise. When the noise level reaches 10%, the anomaly outline shows some distortion. When the noise level reaches 20%, the anomaly outline shows even greater deformation and begins to deviate in size, but the anomaly center position remains relatively accurate.

[0062] The preferred embodiments of the present invention are described in detail above, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A two-dimensional magnetotelluric inversion method based on DeepLabV3+, characterized in that: include: A geoelectric model construction module is used to construct a variety of underground resistivity distribution models, including a rectangular high- and low-resistance anomaly combination model, a stepped high- and low-resistance anomaly combination model, and a pyramid anomaly model; A forward calculation module, connected to the geoelectric model construction module, is used to obtain corresponding apparent resistivity and phase data of the constructed underground resistivity distribution model through finite element forward calculation; The data preprocessing module is connected to the forward calculation module and is used to normalize and enhance the acquired magnetotelluric data. The normalization process is to perform linear normalization on the apparent resistivity and phase data. The calculation formula is: Among them, X is the original data, X min and X max are the minimum and maximum values ​​of the data respectively; the data enhancement is to add adaptive Gaussian noise to the training data set, and the calculation formula is: x′=x+N(μ,σ·|x|) Where x′ represents the new value after adding Gaussian noise, and x represents the value of the original data; N(μ,σ·|x|) represents the noise generated from the normal distribution, where: μ represents the mean of the Gaussian noise, which is taken as 0; σ·|x| represents the standard deviation of the noise, σ is a fixed scaling factor, which is taken as 0.05, and |x| represents the absolute value of the original signal; A DeepLabV3+ network module, connected to the data preprocessing module, is used to build and train a deep learning network model based on the DeepLabV3+ architecture. The network model includes an Xception-based encoder module, an Atrous Spatial Pyramid Pooling (ASPP) module, and a feature decoder module. The inversion execution module is connected to the DeepLabV3+ network module and is used to input the measured magnetotelluric data preprocessed by the data preprocessing module into the trained DeepLabV3+ network model to directly output the underground resistivity distribution result.

2. The magnetotelluric two-dimensional inversion method based on DeepLab V3+ according to claim 1, characterized in that: The encoder module of the DeepLabV3+ network module uses Xception as the backbone network for feature extraction and uses void convolution instead of standard convolution.

3. The magnetotelluric two-dimensional inversion method based on DeepLab V3+ according to claim 2, characterized in that: The ASPP module of the DeepLabV3+ network module contains five parallel branches: a 1×1 convolution branch, three 3×3 dilated convolution branches with different expansion rates, and a global average pooling branch, which are used to capture multi-scale geological features. The decoder uses jump connections between low-level features and high-level features.

4. The method for two-dimensional magnetotelluric inversion based on DeepLab V3+ according to claim 3, characterized in that: The loss function used in DeepLabV3+ network module training is Dice Loss, and the calculation formula is: Among them, X is the predicted segmentation result, Y is the actual segmentation label, |X∩Y| represents the intersection of the predicted and true labels, |X| and |Y| represent the number of pixels of the predicted and true labels, respectively. The Adam optimizer is used, the initial learning rate is set to 0.001, and a step-by-step decay learning rate adjustment strategy is adopted with a decay factor of 0.

95.

5. The method for two-dimensional magnetotelluric inversion based on DeepLab V3+ according to claim 4, characterized in that: The DeepLabV3+ network module is also used to divide the training sample set, 90% of which is used for training and 10% for validation to avoid overfitting; and the performance of the trained DeepLabV3+ network model is evaluated by calculating the loss values ​​on the training set and the validation set.

Images