A first arrival picking method based on a deep residual semantic segmentation network

Abstract

The application designs a first arrival picking method based on a deep residual semantic segmentation network, which comprises the following steps: S1, constructing a first arrival training set, a verification set and a test set; S2, constructing a deep residual semantic segmentation network structure; S3, training the deep residual semantic segmentation network structure constructed in step S2 according to the training set, the verification set and the test set in step S1, and obtaining a first arrival picking network model; and S4, picking the first arrival of all actual seismic data according to the first arrival picking network model obtained in step S3. The application adopts an asymmetric encoder-decoder structure, uses sub-pixel convolution to realize high-resolution reconstruction of feature maps, and improves the precision and efficiency of first arrival automatic picking through the skip connection between the encoder and the decoder. Finally, the first arrival picking model is obtained by training the seismic data through the deep residual semantic segmentation network structure, thereby achieving the technical effects of short training time and high first arrival picking accuracy.

Description

Technical Field

[0001] This invention belongs to the field of seismic data processing, specifically relating to an initial arrival picking method based on a deep residual semantic segmentation network. Background Technology

[0002] First-arrival acquisition is a fundamental task in seismic exploration. The quality of the first-arrival acquisition results directly affects the effectiveness of subsequent refraction wave static correction, vertical seismic profile interpretation, and seismic tomography. With advancements in seismic acquisition technology and increasingly complex exploration terrain, the volume of acquired seismic data is enormous, and the signal-to-noise ratio (SNR) is low. Traditional automatic first-arrival acquisition methods are not accurate or precise enough for low SNR data, requiring extensive manual adjustments, which is time-consuming, labor-intensive, and inefficient, failing to meet industrial standards for efficiency and accuracy.

[0003] With the rapid development of deep learning technology and its widespread application in the field of seismic exploration, first arrival picking can be viewed as a binary classification problem in principle. High-precision first arrival picking methods can be implemented by combining image semantic segmentation technology in computer vision to classify each signal moment of seismic data and obtain the first arrival position.

[0004] Chinese patent CN108072896B discloses a fully automatic seismic wave first arrival picking algorithm. It obtains reference first arrivals from seismic data obtained through constant-speed scanning shot gathers, then calculates the first arrival using an improved Coppens algorithm. By analyzing the difference in energy ratios between accurately picked and inaccurate first arrivals, quantifying the trend of energy ratio changes, and calculating the reliability factor of the first arrival, accurate first arrival data can be obtained for subsequent processes by setting a reliability threshold constraint. This method not only picks the first arrival but also automatically judges its accuracy, effectively improving the efficiency of first arrival picking. Through picking seismic data from different work areas, the results show that the algorithm can quickly and accurately pick first arrivals, greatly reducing the time spent on manual interaction, effectively improving the efficiency of first arrival picking, and saving a significant amount of manpower.

[0005] Chinese patent CN111626355A discloses a method for first arrival picking of seismic data based on a Unet++ convolutional neural network, comprising: acquiring seismic data and the first arrival time of the seismic data; automatically generating seismic data labels based on the first arrival time; extracting a training set, a validation set, and a test set from the acquired seismic data; modifying the structure and adjusting the parameters of the Unet++ model according to the specific characteristics of the seismic data to establish a deep learning network model suitable for seismic data; training the improved Unet++ convolutional neural network model based on the seismic data training and validation sets to obtain a first arrival picking model; and using the trained first arrival picking model to pick the first arrival of data in the test set to obtain the first arrival picking result. This method can achieve fast and accurate first arrival picking of seismic data and has good noise resistance.

[0006] However, most neural networks currently used for first-arrival picking are encoder-decoder structures improved based on FCN, U-Net, etc., and still suffer from defects such as accuracy errors and low efficiency in first-arrival picking.

[0007] However, the U-Net network was proposed to solve the problem of medical image segmentation. Medical images are small in size, clear in content, and simple to classify, while seismic data are large in size, noisy, and complex in content. Therefore, directly applying U-Net to first arrival picking would yield poor results. Traditional first arrival picking methods mainly include manual picking and automatic picking. Manual picking relies on staff to manually mark the first arrival position based on their experience. This method is highly subjective, inconsistent, and inefficient. Automatic picking algorithms construct equations based on different characteristics of the waveform in terms of energy, frequency, etc., for identification. Common methods include: energy ratio method, correlation method, image processing method, neural network method, etc. Most of these methods are based on single-channel data picking and lack consideration of the correlation between multi-channel data. Therefore, they are not effective for first arrival picking of seismic data with low signal-to-noise ratio.

[0008] Therefore, for massive amounts of medium and low signal-to-noise ratio data, there is an urgent need to find a method for automatic first arrival picking that meets both accuracy and efficiency standards. Summary of the Invention

[0009] To address the aforementioned issues, this paper proposes a first-arrival picking method based on a deep residual semantic segmentation network to improve the accuracy and efficiency of first-arrival picking for seismic data.

[0010] To achieve the above effects, this invention designs an initial arrival picking method based on a deep residual semantic segmentation network.

[0011] A first-arrival picking method based on a deep residual semantic segmentation network includes:

[0012] Step S1: Construct the initial training set, validation set, and test set;

[0013] Step S2: Construct a deep residual semantic segmentation network structure;

[0014] Step S3: Train the deep residual semantic segmentation network structure constructed in step S2 based on the training set, validation set and test set in step S1 to obtain the initial picking network model.

[0015] Step S4: Perform initial arrival picking on all actual seismic data based on the initial arrival picking network model trained in step S3.

[0016] Preferably, in step 1, the method for constructing the initial training set, validation set, and test set includes:

[0017] Step S11: Crop the obtained seismic data and normalize the seismic data;

[0018] Step S12: Classify the label data into 0 and 1 categories;

[0019] Step S13: Finally, randomly divide the earthquake data and label data into training set, validation set and test set.

[0020] Preferably, the size of both the seismic data and the label data is 2000×256, where 2000 represents the number of sampling points and 256 represents the number of seismic traces.

[0021] Preferably, the ratio of the earthquake data and label data to the training set, validation set, and test set is 6:2:2.

[0022] Preferably, in step S2, the deep residual semantic segmentation network structure adopts an asymmetric encoding and decoding structure; the encoder and decoder use skip connections to superimpose feature map dimensions.

[0023] Preferably, the encoder includes a residual module; the residual module consists of 13 convolutional layers with a kernel size of 3*3.

[0024] Preferably, the decoder uses subpixel convolutional layers for super-resolution reconstruction of the feature map.

[0025] Preferably, the operation method of the deep residual semantic segmentation network structure is as follows:

[0026] Step S21: First, perform a convolution process on the input data using a large-size convolution kernel;

[0027] Step S22: Then, the feature map obtained after max pooling is fed into the encoder, and the encoder performs deep convolution three times to learn deep feature semantic information.

[0028] Step S23: Next, the feature map processed by the encoder is sent to the decoder to perform high-resolution reconstruction of the feature map and restore the size of the feature map.

[0029] Step S24: Finally, the feature map obtained from each upsampling is fused with the feature map saved by the corresponding layer encoder to learn multi-scale feature information by combining low-level and high-level features.

[0030] Preferably, in step S3, the method for training the initial arrival picking network model includes:

[0031] Step S31: Input the training set into the deep residual semantic segmentation network for training;

[0032] Step S32: Based on the set training parameters, obtain the training model of the network;

[0033] Step S33: Predict the segmentation effect of the trained model on the test set to obtain the corresponding segmentation result;

[0034] Step S34: Finally, the segmented Mask results are superimposed on the input data to obtain the final segmentation effect.

[0035] Preferably, in step S4, the specific method for performing first arrival picking on all actual seismic data based on the first arrival picking network model trained in step S3 is to extract the first arrival time from other data in the seismic data that have not participated in the training using the trained deep residual semantic segmentation network model.

[0036] The advantages and effects of this application are as follows:

[0037] 1. This invention designs an end-to-end deep learning network structure based on deep residuals, uses sub-pixel convolution to achieve high-resolution reconstruction of feature maps, and uses skip connections between the encoder and decoder to fuse multi-scale feature information, thereby improving the accuracy and efficiency of automatic first arrival picking. Using this network structure to train seismic data, a first arrival picking model is obtained, which has a short training time and high accuracy in picking first arrivals.

[0038] 2. The deep residual semantic segmentation network proposed in this application adopts an asymmetric encoder-decoder structure, increasing the proportion of the encoder. To improve the learning ability of the network, the encoder part uses a deep residual module, enabling the network to learn more feature information in the seismic data. The decoder part uses sub-pixel convolutional layers to reconstruct the feature map at high resolution, achieving accurate localization of the initial arrival.

[0039] 3. The deep residual network structure designed in this application is based on U-Net and integrates a deep residual module and a sub-pixel convolutional layer. The encoder of the deep residual network uses the deep residual module to perform downsampling operations. The increase in the number of network layers enables the network to have a stronger learning ability, thereby enabling it to learn more feature information in seismic data and find the initial arrival position.

[0040] 4. The decoder of the deep residual network designed in this application uses sub-pixel convolutional layers for upsampling operations to reconstruct the feature map at high resolution, which can achieve accurate localization. This avoids the technical problem of U-Net's decoder using deconvolution for upsampling operations, which results in low resolution of the recovered feature map and poor accuracy of the network segmentation results.

[0041] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0042] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0044] Figure 1 This is a flowchart of the first arrival picking method based on a deep residual semantic segmentation network according to an embodiment of the present invention;

[0045] Figure 2 This is a diagram of the U-Net network structure provided in an embodiment of the present invention;

[0046] Figure 3 This is a diagram of the deep residual semantic segmentation network structure used in this invention;

[0047] Figure 4 This refers to the residual module in the deep residual semantic segmentation network structure used in this invention;

[0048] Figure 5 This is the sub-pixel convolutional layer used in this invention;

[0049] Figure 6 This is a diagram showing the initial pickup model recognition results of the present invention;

[0050] Figure 7 This is a diagram showing the initial and final picking effect of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0052] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0053] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0054] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0055] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0056] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0057] Example 1

[0058] This embodiment mainly introduces the most basic design method of an initial arrival picking method based on a deep residual semantic segmentation network, which includes the following steps: Please refer to the following for details. Figure 1 , Figure 1 This is a flowchart of the first arrival picking method based on a deep residual semantic segmentation network according to an embodiment of the present invention.

[0059] Step S1: Construct earthquake data training set, validation set and test set; preprocess the acquired earthquake data, automatically generate label data according to the marked first arrival time, and then divide the earthquake data and label data into training set, validation set and test set according to the ratio;

[0060] Step S2: Design a deep residual semantic segmentation network structure. The deep residual semantic segmentation network structure is an asymmetric encoder-decoder structure with a large encoder ratio and a small decoder ratio. The encoder and decoder in the deep residual semantic segmentation network structure adopt a skip connection.

[0061] Step S3: Train the deep residual semantic segmentation network constructed in step S2 based on the training set, validation set and test set of step S1.

[0062] Step S4: Based on the network model trained in step S3, perform initial picking on all actual seismic data.

[0063] Furthermore, the deep residual semantic segmentation network encoder uses a residual module; the residual module consists of 13 convolutional layers with a kernel size of 3*3.

[0064] Furthermore, the deep residual semantic segmentation network decoder uses subpixel convolutional layers for super-resolution reconstruction of the feature map.

[0065] Furthermore, the deep residual semantic segmentation network encoder and decoder use skip connections to stack feature map dimensions.

[0066] Further, the construction process of the training sample set in step S1 is as follows: given the first arrival time of each trace of the original seismic data, the original seismic data is cropped and labels are automatically generated; the original seismic data is cropped in units of 256 traces, and the final sample set of seismic data and label data has an average size of 2000×256, where 2000 represents the number of sampling points and 256 represents the number of seismic traces; the label data is set to 0 and 1 for classification, with 0 before the first arrival time and 1 after the first arrival time; finally, the seismic data and label data are randomly divided into training set, validation set and test set.

[0067] This invention designs an end-to-end deep learning network structure based on deep residuals, uses subpixel convolution to achieve high-resolution reconstruction of feature maps, and uses skip connections between the encoder and decoder to fuse multi-scale feature information, thereby improving the accuracy and efficiency of initial automatic picking.

[0068] Example 2

[0069] Based on the above embodiment 1, this embodiment mainly introduces the optimal design of an initial arrival picking method based on a deep residual semantic segmentation network.

[0070] 1. Construct the initial training set, validation set, and test set:

[0071] The obtained seismic data was cropped into units of 256 channels and then normalized (as shown in Equation 1).

[0072]

[0073] Where: x' x,j x represents the value of the i-th sample point in the j-th channel after normalization; i,j x represents the value of the i-th sample point in the j-th channel without normalization; min x represents the minimum value in the entire dataset; max This represents the maximum value in the entire dataset.

[0074] Then, corresponding seismic data labels are created based on the marked first arrival times.

[0075] The labels were created as follows: pixel values ​​before the first arrival time were set to 0, and pixel values ​​after the first arrival time were set to 1; the size of both the seismic data and the label data in the sample set was 2000×256, where 2000 represents the number of sampling points and 256 represents the number of seismic traces; finally, the data were shuffled and divided into training set, validation set, and test set in a 6:2:2 ratio.

[0076] 2. Design the deep residual semantic segmentation network structure:

[0077] This invention is based on, for example Figure 2The U-Net network structure shown is used to design a deep residual semantic segmentation network. For example... Figure 3 The deep residual semantic segmentation network shown adopts an asymmetric encoding and decoding structure, increases the proportion of the encoder, enables the network to learn richer semantic information, and improves network performance; the decoder uses deep residual module groups to learn features, and the decoder uses sub-pixel convolutional layers to recover high-resolution feature maps.

[0078] First, a large-size convolutional kernel is used to perform a convolution process on the input data to increase the receptive field and reduce the network training parameters. Then, the feature map obtained after max pooling is fed into the encoder. The encoder is used three times as follows: Figure 4 The residual modules shown perform deep convolution to learn deep feature semantic information. The number of residual modules used in the deep convolution stacking are 3, 4, and 6 respectively. Convolution controls the stride instead of pooling to downsample and reduce the feature map size. To prevent overfitting due to excessive network complexity, batch normalization is performed after each convolution operation, which also speeds up network training. Compared to the ResNet34 network structure, this encoder reduces the number of network parameters while ensuring effective extraction of high-level semantic information, thus improving model training efficiency. The encoder parameters are shown in Table 1. The encoder saves the feature map after each downsampling for fusion with the corresponding decoder feature map. The feature map processed by the encoder is fed into the decoder, which consists of convolutional layers and sub-pixel convolutional layers, as shown in Table 1. Figure 5 The sub-pixel convolutional layers shown achieve high-resolution reconstruction of the feature maps, restoring their original size. The feature maps obtained from each upsampling are then fused with the feature maps saved by the corresponding encoder layer, combining low-level and high-level features to learn multi-scale feature information, enabling the network to accurately locate initial boundaries. The decoder's structural parameters are shown in Table 2. The feature map output by the decoder undergoes two convolutions to obtain the segmentation result. Except for the last convolution, all other convolution operations use ReLU activation, and the final convolution uses the Sigmoid activation function to calculate the probability. The output prediction result has the same size as the data input into the network.

[0079]

[0080] Table 1 Encoder Parameters

[0081] Layername Layers Decoder_1 Concat,3*3,512;PixelShuffle(2),128 Decoder_2 Concat,3*3,256;PixelShuffle(2),64 Decoder_3 Concat,3*3,256;PixelShuffle(2),64

[0082] Table 2 Decoder Parameters

[0083] 3. Training the initial arrival picking network model:

[0084] The ultimate goal of first-arrival picking is to provide information on whether each point in the seismic spatial data volume is at its first arrival or not, which falls under the category of a binary classification problem. The loss function used is a cross-entropy loss function combining the LogSoftmax function and the negative log-likelihood loss function (as shown in Equation 2). First, the network output is processed by LogSoftmax, and then the negative log-likelihood loss function is used to calculate the degree of loss between the predicted value and the label value.

[0085]

[0086] Where x is the predicted result vector, class is the label value (0 or 1), and N is the dimension of vector x. class and x j These are the class-th and j-th values ​​in vector x, respectively.

[0087] Specifically, the training set is input into a deep residual semantic segmentation network for training, and a training model of the network is obtained based on the set training parameters (learning rate, number of iterations, etc.). Then, the segmentation effect of the training model is predicted on the test set to obtain the corresponding segmentation result. Finally, the segmented mask result is superimposed on the input data to obtain the final segmentation effect.

[0088] This invention uses MIoU (Mean Intersection over Union), the most commonly used metric in semantic segmentation, to evaluate the model's performance. MIoU is the ratio of the intersection to the union of the true and predicted values, as shown in equation (3):

[0089]

[0090] Where: p ij This represents the number of labels with value i that are predicted to be j; p ji This represents the number of labels with value j that are predicted to be i; p ii k+1 represents the number of correct predictions; k+1 represents the total number of categories.

[0091] After training, the training accuracy and training time of the traditional U-Net network, the network with only residual modules added to the traditional U-Net network, and the improved algorithm of this invention were evaluated. The evaluation results are shown in Table 3.

[0092]

[0093] Table 3 Training accuracy and time for U-Net, RNet, and RSNet

[0094] As shown in the table above, the traditional U-Net achieves an accuracy of 75.80% on the validation set and takes 13.4 hours. This invention, based on the traditional U-Net network, combines the advantages of residual learning and sub-pixel convolutional layers to redesign a deep residual semantic segmentation network that achieves an accuracy of 99.96% on the validation set and takes 4.5 hours. Compared with the traditional U-Net network, both the accuracy and efficiency of initial pickup are significantly improved.

[0095] 4. Use the model to pick the first arrival and extract the time:

[0096] First arrival times were extracted from other data in the untrained seismic data using a pre-trained deep residual semantic segmentation network model. The first arrival picking and segmentation results are as follows: Figure 6 As shown, the left image is the input seismic data, and the right image is the segmentation result using the first-arrival model. The results demonstrate that the first-arrival picking model trained using a deep residual semantic segmentation network performs well in picking up seismic wave data, accurately depicting the arrival times of the seismic waves. The program then processes the automatically identified first-arrival images to obtain the first-arrival times of the seismic data and compares them with the original seismic data. Figure 7 As shown in the figure. The results of the examples of this invention demonstrate that the method proposed in this invention is efficient and accurate.

[0097] The deep residual semantic segmentation network proposed in this invention employs an asymmetric Encoder-Decoder structure, increasing the proportion of the Encoder. To enhance the network's learning ability, the Encoder part uses a deep residual module, enabling the network to learn more feature information from seismic data; the Decoder part uses sub-pixel convolutional layers to perform high-resolution reconstruction of the feature map, achieving accurate first-arrival localization. Using this network structure to train a first-arrival picking model on seismic data results in short training time and high accuracy in picking first arrivals.

[0098] The network structure designed in this invention is based on U-Net and integrates deep residual modules and subpixel convolutional layers.

[0099] Comparison of the deep residual network of this invention with U-Net:

[0100] U-Net's encoder uses convolution and pooling for downsampling, which fails to learn enough feature information from seismic data, causing the network to fail to converge properly. In contrast, the encoder of the deep residual network uses deep residual modules for downsampling, increasing the number of network layers and giving the network a stronger learning ability. It can learn more feature information from seismic data and find the location of the first arrival.

[0101] U-Net's decoder uses deconvolution for upsampling, resulting in low-resolution reconstructed feature maps and poor network segmentation accuracy. In contrast, the decoder of deep residual networks uses subpixel convolutional layers for upsampling, enabling high-resolution reconstruction of feature maps and achieving accurate localization.

[0102] The above description is merely a preferred embodiment of the present invention and does not limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter alterations to these embodiments within the spirit and principles of the present invention, achieved through conventional substitutions or by achieving the same function without departing from the principles and spirit of the present invention, fall within the scope of protection of the present invention.

Claims

1. A first-arrival picking method based on a deep residual semantic segmentation network, characterized in that, It includes: Step S1: Construct the initial training set, validation set, and test set in a ratio of 6:2:2; Step S2: Construct a deep residual semantic segmentation network structure, adopting an asymmetric encoder-decoder structure, in which the encoder accounts for a large proportion and the decoder accounts for a small proportion; the encoder and decoder use skip connections to stack feature map dimensions; the encoder includes a residual module; the convolutional kernel size of the convolutional layer in the residual module is 3*3, and the decoder uses sub-pixel convolutional layers to perform super-resolution reconstruction of the feature map; Step S3: Train the deep residual semantic segmentation network structure constructed in step S2 based on the training set, validation set and test set in step S1 to obtain the initial picking network model. The method for training the initial picking network model includes: step S31, inputting the training set into the deep residual semantic segmentation network for training; step S32, obtaining the training model of the network based on the set training parameters; step S33, predicting the segmentation effect of the training model on the test set to obtain the corresponding segmentation result; step S34, finally superimposing the segmented Mask result on the input data to obtain the final segmentation effect. Step S4: Perform first arrival picking on all actual seismic data based on the first arrival picking network model trained in step S3. The operation method of the deep residual semantic segmentation network structure is as follows: Step S21: First, perform a convolution process on the input data using a large-size convolution kernel; Step S22: Then, the feature map obtained after max pooling is fed into the encoder. The encoder performs deep convolution three times to learn deep feature semantic information. The number of residual modules used in the deep convolution stacking are 3, 4 and 6 respectively. The stride is controlled by convolution to replace pooling for downsampling to reduce the size of the feature map. Batch normalization is performed after each convolution operation. Step S23: Next, the feature map processed by the encoder is sent to the decoder to perform high-resolution reconstruction of the feature map and restore the size of the feature map. Step S24: Finally, the feature map obtained from each upsampling is fused with the feature map saved by the corresponding layer encoder to learn multi-scale feature information by combining low-level and high-level features.

2. The first arrival picking method based on a deep residual semantic segmentation network according to claim 1, characterized in that, In step S1, the method for constructing the initial training set, validation set, and test set includes: Step S11: Crop the obtained seismic data and normalize the seismic data; Step S12: Classify the label data into 0 and 1 categories; Step S13: Finally, randomly divide the earthquake data and label data into training set, validation set and test set; In step S12, the value is set to 0 before the initial arrival time and to 1 after the initial arrival time.

3. The first arrival picking method based on a deep residual semantic segmentation network according to claim 2, characterized in that, The size of both the seismic data and the label data is 2000×256, where 2000 represents the number of sampling points and 256 represents the number of seismic traces.

4. The first arrival picking method based on a deep residual semantic segmentation network according to any one of claims 1 or 3, characterized in that, In step S4, the specific method for picking the first arrival of all actual seismic data based on the first arrival picking network model trained in step S3 is to extract the first arrival time of other data in the seismic data that have not participated in the training using the trained deep residual semantic segmentation network model.

Images