CN-121981891-A - Physical constraint generation countermeasure network super-resolution reconstruction method and system for geomagnetic grid data

CN121981891ACN 121981891 ACN121981891 ACN 121981891ACN-121981891-A

Abstract

The invention discloses a physical constraint generation countermeasure network super-resolution reconstruction method and a physical constraint generation countermeasure network super-resolution reconstruction system for geomagnetic grid data, which relate to the technical field of geophysical measurement and artificial intelligence and comprise the following steps of S1, data preprocessing, namely loading low-resolution geomagnetic field data and high-resolution geomagnetic field data, carrying out geographic area screening, invalid data processing and gridding processing, and constructing a training data set and a test data set; the method comprises the steps of S2, constructing a network model, S3, designing a loss function, namely, constructing a comprehensive loss function by combining content loss, space continuity loss, edge retention loss, counterloss and feature matching loss, and S4, training the model. The invention reconstructs the low-resolution geomagnetic anomaly map into the high-resolution grid under the condition of not increasing extra measurement cost, thereby obviously improving the resolution and quality of geomagnetic field data and providing more accurate data support for related research and application.

Inventors

XU WEI
XIA LINGLONG
YIN JIANING
QIN FANGJUN
HAN YU
ZHU TIANGAO
CHEN JIAWEI

Assignees

中国人民解放军海军工程大学

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (8)

1. The physical constraint generation countermeasure network super-resolution reconstruction method for geomagnetic mesh data is characterized by comprising the following steps of: S1, data preprocessing, namely loading low-resolution geomagnetic field data and high-resolution geomagnetic field data, and carrying out geographic area screening, invalid data processing and gridding processing to construct a training data set and a testing data set; S2, constructing a network model, namely designing a generator network and a lightweight class discriminator network; s3, designing a loss function, namely constructing a comprehensive loss function by combining content loss, space continuity loss, edge retention loss, counterloss and feature matching loss; S4, model training, namely optimizing a training process by using a built comprehensive loss function training generator and a discriminator and adopting a gradient punishment and learning rate scheduling strategy; s5, super-resolution reconstruction, namely converting the low-resolution geomagnetic field data into high-resolution geomagnetic field data by using a trained generator.
2. The method of claim 1, wherein the preprocessing of data comprises extracting data blocks using a high overlap rate block extraction technique to reduce reconstruction artifacts, fusing overlapping blocks using a gaussian weight mask to smooth boundaries, and processing boundary regions using a gradient fill method to avoid edge effects.
3. The method for generating a super-resolution reconstruction of an antagonizing network of claim 1 wherein the generator network comprises an initial feature extraction layer, an enhanced residual block, pixelShuffle up-sampling modules, and an anti-artifact smoothing layer.
4. The method for reconstructing the super-resolution of the antagonism network by generating physical constraints for geomagnetic grid data according to claim 1, wherein the enhanced residual block comprises batch normalization and Dropout, the feature extraction capability and the generalization capability are improved, and the PixelShuffle up-sampling module is used for reducing checkerboard artifacts caused by conventional transpose convolution.
5. The method for reconstructing the super-resolution of the antagonism network generated by physical constraint of geomagnetic grid data according to claim 1, wherein the discriminator network adopts a lightweight design, comprises a multi-layer convolution feature extraction and classification head, and focuses on local texture detail recognition.
6. The method for reconstructing the super-resolution of the countermeasure network by physical constraint generation for geomagnetic grid data according to claim 1, wherein the content loss is to ensure that the generated super-resolution geomagnetic field data accords with a physical rule while keeping high precision, a multi-objective loss function comprehensively considering numerical precision, countermeasure learning and physical constraint is designed, and model performance is improved by multi-dimensional constraint; Content loss, wherein the content loss is used for guaranteeing the numerical consistency between a super-resolution result and high-resolution reference data, and adopts Root Mean Square Error (RMSE) as a content loss function, compared with Mean Square Error (MSE), the RMSE is more sensitive to abnormal values, and can effectively improve the overall numerical precision, and the method is defined as follows: wherein y represents a super-resolution result generated by the model, y represents high-resolution reference data, and N is the number of data samples; Combat losses-combat losses are built based on a generated combat network (GAN) framework, and the combat losses are defined as a capability measure of the generator to misjudge the generated result by the discriminator through combat training of the discriminator and the generator, and are expressed as follows: Wherein, the A representation discriminator network for discriminating the generated result from the true high resolution data, the counterloss weight being set to a smaller value in order to ensure that the content fidelity is prioritized over the countergeneration effect; The calculation mode of the feature matching loss is aimed at guiding a generator to learn the feature representation of a discriminator in different levels, enabling the generated data to be more similar to the distribution rule of a real geomagnetic abnormal field from a low-level texture to a high-level structure through the difference of the feature space between a matched generated result and real data in the discriminator, extracting feature graphs of a plurality of key middle layers in a discriminator network, respectively calculating the L1 loss of the generated result and the real data in the feature space of each level, and taking the average value of the loss of all levels, wherein the definition is as follows: G (x) represents the super-resolution output result of the generator for the low-resolution input x, y represents real high-resolution geomagnetic data, dk (z) represents the characteristic output of the kth layer of the discriminator network, nk represents the total element number of the characteristic diagram of the kth layer of the discriminator; The spatial continuity loss is realized by calculating gradient differences in horizontal, vertical and diagonal directions, and is defined as follows: Wherein gh and gv respectively represent average values of gradients in horizontal and vertical directions, gd1 and gd2 respectively represent average values of gradients in two diagonal directions, and the average value is obtained by calculating absolute difference values of adjacent pixels and taking the average value; Edge preservation loss, namely ensuring that the model better retains edge details in geomagnetic field data while enhancing spatial continuity, calculating the gradient map difference between a generated result and reference data based on a Sobel operator, and measuring gradient consistency by adopting L1 loss, wherein the definition is as follows: Wherein ∇ x and ∇ y represent horizontal and vertical gradients calculated based on Sobel operator, respectively; the comprehensive loss function is configured through carefully designed weights, the loss function is combined into a final objective function, and numerical precision and physical constraint are balanced: The weight parameters are set to content loss weight λc=80.0, counterloss weight λa=0.001, feature matching loss weight λf=0.01, spatial continuity loss weight λcont=0.01, and edge preserving loss weight λe=30.0.
7. The method for reconstructing physical constraint generation of geomagnetic mesh data according to claim 1, wherein the model training includes improving training stability by using a label smoothing technique, dynamically adjusting a learning rate by using a cosine annealing learning rate scheduler, and preventing gradient explosion by using gradient clipping.
8. The physical constraint generation countermeasure network super-resolution reconstruction system for geomagnetic grid data is characterized by comprising a data processing module and a model training module box resolution conversion module, wherein the data processing module comprises data input, data preprocessing and data model construction, and the resolution conversion module is used for converting the ground resolution into high resolution.

Description

Physical constraint generation countermeasure network super-resolution reconstruction method and system for geomagnetic grid data Technical Field The invention relates to the technical field of geophysical measurement and artificial intelligence, in particular to a physical constraint generation countermeasure network super-resolution reconstruction method and system for geomagnetic grid data. Background In the prior art, a high-resolution geomagnetic anomaly reference map for geophysical exploration, mineral resource investigation, geological structure research and the like is mainly obtained by spatially interpolating satellite height measurement inversion, ship-borne and aviation magnetic measurement data. However, the space between the satellite magnetic measurement data grids is generally larger than 4 km, a large number of blank grids exist in 30 km shallow land frames and island regions, hundred-meter-level data can be acquired by shipborne or aviation magnetic measurement, but the data distribution is sparse due to limitation of the space between the measuring lines, the operation cost and the meteorological conditions, and the research requirement is not satisfied. The traditional spatial interpolation methods such as inverse distance weighting, common kriging, bicubic interpolation and the like are limited by spatial autocorrelation assumption, fine magnetic anomalies with the wavelength smaller than 2 times of grid spacing cannot be recovered, stretching artifacts along the direction of a measuring line are easy to generate for sparse measuring line data, SRCNN, SRGAN and other super-resolution networks are widely applied in the field of natural images, but original frames of the super-resolution networks only depend on pixel level loss, lack of geophysical priori, and do not meet the non-divergence condition of the V.B=0 when being directly migrated to geomagnetic anomaly map reconstruction, and meanwhile weak amplitude anomalies are easy to be excessively smoothed, so that the reliability of geomagnetic field data research application is reduced. Therefore, the invention introduces no-divergence projection and multi-objective physical constraint loss on the basis of SRGAN, and provides a geomagnetic anomaly map super-resolution reconstruction method which takes numerical accuracy and physical consistency into consideration. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides a physical constraint generation countermeasure network super-resolution reconstruction method and system for geomagnetic grid data. In order to achieve the above purpose, the present invention adopts the following technical scheme: A physical constraint generation countermeasure network super-resolution reconstruction method for geomagnetic mesh data comprises the following steps: S1, data preprocessing, namely loading low-resolution geomagnetic field data and high-resolution geomagnetic field data, and carrying out geographic area screening, invalid data processing and gridding processing to construct a training data set and a testing data set; s2, constructing a network model, namely designing a generator network and a lightweight class discriminator network; s3, designing a loss function, namely constructing a comprehensive loss function by combining content loss, space continuity loss, edge retention loss, counterloss and feature matching loss; S4, model training, namely optimizing a training process by using a built comprehensive loss function training generator and a discriminator and adopting a gradient punishment and learning rate scheduling strategy; s5, super-resolution reconstruction, namely converting the low-resolution geomagnetic field data into high-resolution geomagnetic field data by using a trained generator. Preferably, the data preprocessing includes extracting data blocks using a high overlap rate block extraction technique to reduce reconstruction artifacts, fusing overlapping blocks using a Gaussian weight mask, smoothing boundaries, and processing boundary regions using a gradient fill method to avoid edge effects. Preferably, the generator network comprises an initial feature extraction layer, an enhanced residual block, pixelShuffle up-sampling modules, and an anti-artifact smoothing layer. Preferably, the enhanced residual block comprises batch normalization and Dropout, the feature extraction capability and the generalization capability are improved, and the PixelShuffle up-sampling module is used for reducing checkerboard artifacts caused by traditional transpose convolution. Preferably, the arbiter network employs a lightweight design, including multi-layer convolution feature extraction and classification heads, focusing on local texture detail recognition. Preferably, the content loss designs a multi-objective loss function comprehensively considering numerical precision and resisting learning and physical constraint in order to ensure that the gene