CN-121999234-A - Surface wave detection dispersion curve segmentation method, medium, equipment and product

CN121999234ACN 121999234 ACN121999234 ACN 121999234ACN-121999234-A

Abstract

The invention discloses a method, medium, equipment and product for segmenting a surface wave detection dispersion curve, and relates to the field of seismic surface wave detection data processing, wherein the method comprises the steps of obtaining a surface wave dispersion energy diagram and a true value dispersion curve label, and constructing a training set and a verification set; the method comprises the steps of carrying out improved construction on a MiTUNet framework to construct a dispersion curve segmentation model, inputting two shallow convolution detail branches and a four-stage MiT main coding branch at the coding end, adjusting the channel number to be aligned with a channel corresponding to a decoding end feature to be connected in a jumping manner after coding features are selected according to the size and need to be subjected to attention gating, carrying out normalization weighted fusion on the aligned coding features and the decoding end feature corresponding to the decoding end feature to be connected in the jumping manner, training and verifying the model by using a training set and a verification set, and inputting a dispersion energy diagram to be segmented into the trained model to obtain a dispersion curve segmentation result. The model of the invention can meet the requirement of the division task of the dispersion energy diagram.

Inventors

CHEN SONG
LIU FEI
WU JIANQIANG
LIU LEI

Assignees

中国地质调查局武汉地质调查中心(中南地质科技创新中心)

Dates

Publication Date: 20260508
Application Date: 20260410

Claims (10)

1. The surface wave detection dispersion curve segmentation method is characterized by comprising the following steps of: s1, acquiring a surface wave dispersion energy diagram and a true value dispersion curve label, and constructing a training set and a verification set; S2, improving MiTUNet frames to construct a dispersion curve segmentation model AGSE-MiTUNet, inputting a model at the encoding end, inputting two shallow convolution detail branches and a four-stage MiT main encoding branch at the same time, wherein the two shallow convolution detail branches respectively output a full resolution characteristic E0 and a half resolution characteristic E1, and the four-stage MiT main encoding branch sequentially outputs multi-scale characteristics E2, E3 and E4 and bottleneck characteristics B; The jump connection is as follows: The jump connection features E0-E2 are firstly subjected to attention gating screening features, and after the channel number of the screening features is aligned with the channel corresponding to the decoding end feature to be jump connected, aligned jump connection features E0 '-E2' are obtained; after the channel number of the jump connection features E3 and E4 is adjusted to be aligned with the channel of the decoding end feature corresponding to the jump connection to be obtained, aligned jump connection features E3 'to E4'; the aligned jump connection characteristics and the decoding end characteristics corresponding to the jump connection to be subjected to normalization weighted fusion; And S3, training and verifying the dispersion curve segmentation model by using a training set and a verification set, and inputting the dispersion energy diagram to be segmented into the trained model to obtain a dispersion curve segmentation result.
2. The method for segmenting the surface wave detection dispersion curve according to claim 1, wherein the surface wave dispersion energy map includes dispersion energy information and grid coordinate coding information, and the grid coordinate coding information is normalized frequency coordinates and normalized speed coordinates corresponding to each pixel position on the target dispersion grid.
3. The method of claim 1, wherein the main code branch of the dispersion curve segmentation model is composed of four stage feature extraction modules, each stage feature extraction module comprises an overlapped block embedded layer and a transform coding layer, model input features are input to a first stage feature extraction module, the input features of each stage feature extraction module are input to the transform coding layer after convolution projection is performed on the overlapped block embedded layer of the current stage, the transform coding layer outputs the features of the stage and is used as the input of the next stage feature extraction module, the transform coding layer is formed by stacking a plurality of transform blocks, and the overlapped block embedded layer of the first stage adopts convolution kernel as Convolution projection with step length of 4, and overlapping block embedding layer in subsequent stage adopts convolution kernel as Convolution projection with step size of 2.
4. The method for partitioning a surface wave detection dispersion curve according to claim 1, wherein, The decoder of each stage comprises an up-sampling main branch, a fusion unit, a residual refinement block and a stage tail attention module, wherein the up-sampling main branch is used for up-sampling and channel projection of the input features of the decoding stage, the output features of the up-sampling main branch and the corresponding aligned jump connection features are subjected to normalized weighted fusion, the fusion features are sequentially input into the residual refinement block and the decoding stage tail attention module, the features of the decoding stage are output, and the output features of the current decoding stage are used as the input features of the next decoding stage; The residual refinement block consists of a main branch and a shortcut branch, wherein the main branch comprises a first 3 multiplied by 3 convolution, batch normalization, reLU, a second 3 multiplied by 3 convolution and batch normalization which are sequentially connected in series, and the shortcut branch carries out identity mapping on the input of the residual refinement block; The main upsampling leg of the first, second and third decoding stages comprises a series of nearest neighbor upsampling and a1 x 1 convolution; the main upsampling leg of the fourth and fifth decoding stages includes tandem nearest neighbor upsampling, 3 x 3 convolution, and 1 x 1 convolution; The output characteristics of the fifth decoding stage are subjected to a convolution refinement layer to obtain refinement characteristics, and then a figure logits is partitioned by a 3×3 convolution partition head output dispersion curve, wherein the convolution refinement layer is composed of 3×3 convolution, batch normalization and a ReLU activation function.
5. The method for partitioning a surface wave detection dispersion curve according to claim 4, wherein, The attention modules of the first and second stage decoders employ a channel extrusion excitation mechanism, and the attention modules of the third, fourth and fifth stage decoders employ a spatial-channel joint extrusion excitation mechanism.
6. The method for partitioning a surface wave detection dispersion curve according to claim 1, wherein, The input feature map is firstly serialized, then the attention branch and the mixed feedforward branch are sequentially input, and finally the output feature of the transducer block is obtained through map reconstruction; let the input feature map of the transducer block be The characteristic sequence after x serialization is Then during the training phase, the attention branch and the hybrid feed forward branch respectively satisfy: Wherein, the Representing the attention branching output characteristics, The representation space is reduced from the point of attention, Representing the normalization of the first layer, The operation of the method is shown at DropPath, Representing the hybrid feed-forward branch output characteristics, A hybrid feed forward operation is shown and, Representing the normalization of the second layer, Representing the output characteristics of the transducer block, Representing a feature reconstruction operation; in the reasoning phase DropPath degenerates into an identity mapping, the attention branch and the mixed feed forward branch satisfy respectively: The hybrid feed forward operation includes a first linear transformation, a channel-by-channel convolution, GELU activation functions, and a second linear transformation in series.
7. The method for partitioning a surface wave detection dispersion curve according to claim 1, wherein, The model training adopts a binary Tversky loss function with effective area mask constraint, and the logarithmic probability map of the model output is set as The corresponding predictive probability map is The label is The effective area mask is If no effective area mask is provided, fetch Then the predictions and labels after mask constraint are respectively as follows And The training loss function is defined as: Wherein, the Representation of The function of the function is that, Representing an element-by-element multiplication, Pixel statistics representing curves correctly predicted within the active area, Background pixel statistics representing mispredicted curves within the active area, Curve pixel statistics representing the mispredicted background within the active area, And Representing the height and width of the dispersion grid respectively, And The weight coefficients of the false detection item and the omission item are respectively represented, Representing locations in a dispersion grid Prediction after the mask constraint is applied, Representing locations in a dispersion grid Labeling after mask constraint is processed.
8. A computer-readable storage medium, in which a computer program is stored, characterized in that the method according to any one of claims 1 to 7 is implemented when the computer program is executed by a processor.
9. An electronic device comprising a processor and a memory, the processor being interconnected with the memory, wherein the memory is configured to store a computer program comprising computer readable instructions, the processor being configured to invoke the computer readable instructions to perform the method of any of claims 1-7.
10. A computer program product comprising computer programs/instructions which, when executed by a processor, implement the method of any of claims 1 to 7.

Description

Surface wave detection dispersion curve segmentation method, medium, equipment and product Technical Field The invention relates to the technical field of seismic surface wave detection data processing, in particular to a surface wave detection dispersion curve segmentation method, medium, equipment and a product. Background The goal of the seismic surface wave dispersion energy map segmentation is to reconstruct an elongated and continuous dispersion curve over a frequency-velocity grid. The targets are generally distributed in narrow bands along the energy peak ridges, which requires both accurate positioning of local energy peak positions and maintenance of geometric continuity of the curve at start and stop endpoints, across-frequency extension processes and among different modes. Therefore, the segmentation result of the dispersion energy map is sensitive to pixel-level positioning errors, endpoint offsets, local breaks and background texture misjudgment, and any fine deviation can further influence the accuracy of subsequent dispersion parameter extraction and formation medium interpretation. Existing MiTUNet and related code-decode segmentation networks typically employ a multi-stage main coding branch to extract multi-scale features and recover spatial resolution stage by a decoder. The structure represented by the MiT main encoder is embedded by overlapping blocks in the encoding process to gradually reduce the spatial resolution, and then the high-level semantic information is extracted by a transducer block. Such a structure is beneficial to enhancing global context modeling capability, but as spatial resolution decreases step by step, shallow high resolution details and precise geometric position information may be continuously attenuated. For the long, narrow-band and boundary-sensitive targets such as the dispersion curve, the deep low-resolution semantic features are only relied on for step-by-step up-sampling recovery, so that curve endpoint deviation, local fracture and detail loss are easily caused. In the decoding stage, the existing method generally introduces the coding end features into the decoding layer of the corresponding scale through jump connection so as to supplement the space detail information. If only direct splicing, simple addition or unified compression is adopted for fusion, the differences among different scale features in terms of semantic strength, spatial granularity and channel distribution are difficult to effectively coordinate. In particular, in the high resolution reconstruction stage, background texture and noise components in shallow features may be introduced into the decoding backbone along with the jump connection, and form mismatch with deep semantic features, thereby causing background texture leakage, pseudo-curve activation, and curve purity degradation. In addition, the dispersive energy map segmentation not only requires deep semantic support, but also requires sufficient shallow positioning evidence to be preserved in the end recovery process. If the decoder relies on conventional progressive upsampling only, but lacks a screening mechanism for high-resolution jump features, a hierarchical attention recalibration mechanism for different scale features, and a specialized refinement path for end reconstruction, it is difficult to simultaneously compromise the semantic aggregation capability of the low-resolution stage and the geometric reconstruction accuracy of the high-resolution stage. Especially for the targets with high continuity requirement and strong background interference of the dispersion curve, how to enhance shallow detail retention capacity, promote jump connection fusion quality, inhibit high-resolution background noise interference and improve continuity and boundary consistency of terminal output under MiTUNet frames constitutes a key technical problem in the tasks. Disclosure of Invention Aiming at the defects of the existing seismic surface wave dispersion energy map segmentation method in the aspects of detail preservation, geometric structure reconstruction and multi-scale feature fusion, the invention provides a surface wave detection dispersion curve segmentation method, which comprises the following steps: s1, acquiring a surface wave dispersion energy diagram and a true value dispersion curve label, and constructing a training set and a verification set; S2, improving MiTUNet frames to construct a dispersion curve segmentation model AGSE-MiTUNet, inputting a model at the encoding end, inputting two shallow convolution detail branches and a four-stage MiT main encoding branch at the same time, wherein the two shallow convolution detail branches respectively output a full resolution characteristic E0 and a half resolution characteristic E1, and the four-stage MiT main encoding branch sequentially outputs multi-scale characteristics E2, E3 and E4 and bottleneck characteristics B; The jump connection is as follows: The jump connect