A multi-source, multi-physical-field gravity data super-resolution reconstruction method

Abstract

The application relates to the technical field of marine geophysical exploration and mapping, and provides a multi-source and multi-physical-field gravity data super-resolution reconstruction method.The method comprises the following steps: acquiring preprocessed low-resolution satellite altimetry gravity data and preprocessed high-resolution shipborne gravity data, and constructing a training data set; constructing a dual regression deep learning model; training the dual regression deep learning model by using the training data set, so as to obtain a trained dual regression deep learning model; inputting preprocessed low-resolution satellite gravity data of a target area to be reconstructed into the trained dual regression deep learning model, and acquiring super-resolution reconstruction gravity data of the target area to be reconstructed.Through dual network closed-loop constraint and adaptive degradation modeling, the application can significantly reduce the dependence on high-resolution shipborne gravity data, and improve the reconstruction accuracy and generalization ability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of marine geophysical exploration and mapping, and particularly relates to a multi-source and multi-physical-field gravity data super-resolution reconstruction method. BACKGROUND

[0002] High-resolution and high-precision marine gravity field data is of great significance in marine resource exploration, seabed topography mapping and marine scientific research. At present, marine gravity data mainly relies on two observation methods: one is satellite altimetry technology, which has global coverage and is suitable for data acquisition in large areas of the sea; the other is shipborne gravity measurement, which can provide high spatial resolution data and is suitable for fine measurement in local areas. However, the spatial resolution of satellite altimetry gravity data is limited, which cannot meet the high-precision research needs of local complex structures or small-scale features. Although shipborne measurement has high resolution, it is limited by high cost, long measurement period and complex geographical environment, and its spatial coverage is usually sparse and discontinuous. Therefore, how to fuse the two complementary data sources, taking into account high resolution and wide coverage, to construct high-precision and high-resolution marine gravity field is one of the key challenges in current marine geophysical research.

[0003] Traditional data fusion methods often face problems such as complex modeling, large computational overhead, difficulty in effectively handling boundary effects and complex terrain interference when dealing with multi-source gravity data, resulting in low accuracy of the fusion results and difficulty in extending to large areas of the sea. In addition, such methods often rely on intensive prior knowledge and manual parameter tuning in practical applications, making it difficult to adapt to large-scale data-driven research needs. In recent years, with the rapid development of deep learning technology, image super-resolution reconstruction methods have shown strong data reconstruction capabilities in many fields, providing a new technical path for high-resolution and high-precision reconstruction of marine gravity fields. However, marine gravity data has significantly different spatial distribution patterns and physical characteristics from natural images, and directly applying traditional super-resolution network models cannot achieve ideal results. On the one hand, existing methods mostly ignore the physical consistency between geophysical laws and data, which can easily lead to results deviating from reality. On the other hand, training deep models usually relies on large amounts of high-resolution labeled samples, while high-resolution shipborne gravity data is costly and limited in quantity, further exacerbating the difficulty of model training. Moreover, the low-resolution satellite altimetry gravity map formed by low-resolution satellite altimetry gravity data not only has a large grid spacing, but also has inaccurate pixel values due to the influence of satellite acquisition methods. Therefore, from low-resolution to high-resolution gravity data, not only the pixel values of the missing grids need to be predicted, but also the pixel values of the existing grids in the low-resolution map need to be corrected. Correspondingly, the process from high-resolution to low-resolution gravity data is not a simple downsampling process, but also needs to consider the noise, filtering and other effects contained in the low-resolution map.

[0004] Therefore, how to fully tap the potential information of a large amount of unlabeled satellite gravity data under the condition of limited high-resolution ship-borne data, and construct a multi-source gravity data fusion and reconstruction framework with semi-supervised ability and capable of simultaneously capturing physical consistency and statistical characteristics is a core problem to be solved at present. SUMMARY

[0005] In view of the problems in the prior art, the present application provides a multi-source and multi-physical-field gravity data super-resolution reconstruction method to solve the problems in the prior art.

[0006] The present application provides a multi-source and multi-physical-field marine gravity data super-resolution reconstruction method, comprising:

[0007] Step S101: obtaining gravity data, the gravity data comprising low-resolution satellite altimetry gravity data and high-resolution ship-borne gravity data;

[0008] Step S102: preprocessing the low-resolution satellite altimetry gravity data and the high-resolution ship-borne gravity data, obtaining preprocessed low-resolution satellite altimetry gravity data and preprocessed high-resolution ship-borne gravity data, and constructing a paired sample data set and an unpaired sample data set according to the preprocessed low-resolution satellite altimetry gravity data and the preprocessed high-resolution ship-borne gravity data; the paired sample data set and the unpaired sample data set constitute a training data set;

[0009] Step S103: constructing a dual regression deep learning model, the dual regression deep learning model comprising a primary regression network P and a dual regression network D;

[0010] Step S104: training the dual regression deep learning model using the training data set to obtain a trained dual regression deep learning model;

[0011] Step S105: inputting preprocessed low-resolution satellite gravity data of a target area to be reconstructed into the primary regression network P in the trained dual regression deep learning model to obtain super-resolution reconstruction gravity data of the target area.

[0012] Further, the paired sample data set is composed of preprocessed low-resolution satellite altimetry gravity data and preprocessed high-resolution ship-borne gravity data The unpaired sample data set is composed of preprocessed low-resolution satellite altimetry gravity data ;

[0013] The low-resolution satellite altimetry gravity data in the training data set is The preprocessed low-resolution satellite altimetry gravity data in the paired sample dataset The preprocessed low-resolution satellite altimetry gravity data in the unpaired sample dataset .

[0014] Further, the preprocessing comprises data cleaning, quality control, outlier detection and rejection.

[0015] Further, the quality control comprises first-stage quality control and second-stage quality control on the high-resolution shipborne gravity data.

[0016] Further, the first-stage quality control comprises:

[0017] calculating the local mean μ and the standard deviation σ of the high-resolution shipborne gravity data within a spatial window, and rejecting the data point satisfying the following formula:

[0018] ,

[0019] wherein, G is the gravity value of the current data point, μ is the local mean of the high-resolution shipborne gravity data within the spatial window, and σ is the standard deviation of the high-resolution shipborne gravity data within the spatial window.

[0020] Further, the second-stage quality control comprises:

[0021] obtaining the standard earth gravity model EGM2008, and rejecting the data point satisfying the following formula:

[0022] ,

[0023] wherein, G is the gravity value of the current data point, GEGM is the gravity value of the current data point of the standard earth gravity model EGM2008, and σEGM is the standard deviation of the EGM2008 model.

[0024] Further, the step S104 comprises:

[0025] Step S1041: inputting the preprocessed low-resolution satellite altimetry gravity data in the training dataset as input data into the principal regression network P, obtaining the corresponding super-resolution reconstructed gravity data GSR , and the expression is:

[0026] ,​​​​​​​​​

[0027] Super-resolution reconstructed gravity data including first super-resolution reconstructed gravity data and second super-resolution reconstructed gravity data , the first super-resolution reconstructed gravity data is:

[0028] ,

[0029] wherein, is the pre-processed low-resolution satellite altimetry gravity data in the paired sample dataset;

[0030] the second super-resolution reconstructed gravity data is:

[0031] ,

[0032] wherein, is the pre-processed low-resolution satellite altimetry gravity data in the unpaired sample dataset;

[0033] Step S1042: inputting the super-resolution reconstructed gravity data into the dual regression network D, to obtain low-resolution mapped reconstructed low-resolution gravity data , the expression being:

[0034] ,

[0035] wherein, is the pre-processed low-resolution satellite altimetry gravity data in the training dataset;

[0036] reconstructed low-resolution gravity data including first reconstructed low-resolution gravity data and second reconstructed low-resolution gravity data , the first reconstructed low-resolution gravity data is:

[0037] ,

[0038] wherein, is the first super-resolution reconstructed gravity data; is the pre-processed low-resolution satellite altimetry gravity data in the paired sample dataset;

[0039] the second reconstructed low-resolution gravity data is:

[0040] ,

[0041] wherein, is the second super-resolution reconstructed gravity data; is the pre-processed low-resolution satellite altimetry gravity data in the paired sample data set;

[0042] Step S1043: judging whether the pre-processed low-resolution satellite altimetry gravity data in the input training data set belongs to the paired sample data set, if yes, executing step S1044-step S1047; if not, executing step S1045-step S1047;

[0043] Step S1044: inputting the pre-processed high-resolution shipborne gravity data in the paired sample data set into the primary regression network P, obtaining the low-resolution mapped shipborne low-resolution gravity data , and inputting the low-resolution mapped shipborne low-resolution gravity data into the dual regression network D, obtaining the high-resolution mapped shipborne high-resolution gravity data , the expression is:

[0044] ;

[0045] Step S1045: constructing a target function, and calculating the error of all output data in the primary regression network P and the dual regression network D and the corresponding label data;

[0046] Step S1046: judging whether the target function satisfies a preset condition, if not, executing step S1047, if yes, completing the training of the dual regression deep learning model, and obtaining the trained dual regression deep learning model;

[0047] Step S1047: back-propagating the target error to the primary regression network P and the dual regression network D, updating the learnable parameters of the dual regression deep learning model, and continuing to execute step S1041 to step S1046.

[0048] Further, the target function is:

[0049] ,

[0050] wherein, is the target error, is a binary indicator function, if the pre-processed low-resolution satellite altimetry gravity data in the training data set in step S1043 belongs to the paired sample data set, then ; if the pre-processed low-resolution satellite altimetry gravity data in step S1043 does not belong to the paired sample data set, but belongs to the unpaired sample data set, then ; ; ​​for the cyclic consistency error, used to describe the super-resolution reconstructed gravity data by the dual regression network D on the pre-processed low-resolution satellite altimetry gravity data in the training dataset for the degraded reconstruction error, used to describe the reconstructed low-resolution gravity data obtained by the degraded reconstruction on the first super-resolution reconstructed gravity data for the degraded reconstruction error, used to describe the reconstructed low-resolution gravity data obtained by the degraded reconstruction on the first super-resolution reconstructed gravity data between the pixel and structural consistency; for the supervised reconstruction error, used to describe the pre-processed low-resolution satellite altimetry gravity data in the paired sample dataset by the primary regression network P for the degraded reconstruction error, used to describe the reconstructed low-resolution gravity data obtained by the degraded reconstruction on the first super-resolution reconstructed gravity data for the degraded reconstruction error, used to describe the reconstructed low-resolution gravity data obtained by the degraded reconstruction on the first super-resolution reconstructed gravity data between the data and structural consistency error; for the dual regression error, used to describe the high-resolution shipborne gravity data in the paired sample dataset by the dual regression network D for the degraded reconstruction error, used to describe the reconstructed low-resolution gravity data obtained by the degraded reconstruction on the first super-resolution reconstructed gravity data for the degraded reconstruction error, used to describe the reconstructed low-resolution gravity data obtained by the degraded reconstruction on the first super-resolution reconstructed gravity data between the pixel and structural consistency; for the dual consistency error, used to describe the first reconstructed low-resolution gravity data obtained by the degraded reconstruction on the first super-resolution reconstructed gravity data by the dual regression network D for the degraded reconstruction error, used to describe the reconstructed low-resolution gravity data obtained by the degraded reconstruction on the first super-resolution reconstructed gravity data for the degraded reconstruction error, used to describe the reconstructed low-resolution gravity data obtained by the degraded reconstruction on the first super-resolution reconstructed gravity data for the degraded reconstruction error, used to describe the reconstructed low-resolution gravity data obtained by the degraded reconstruction on the first super-resolution reconstructed gravity data between the pixel and structural consistency;

[0051] further, the cyclic consistency error is:

[0052] ,

[0053] wherein, is the reconstructed low-resolution gravity data; is the pre-processed low-resolution satellite altimetry gravity data in the training dataset; represents the pixel L1 loss, SSIM represents the structural similarity index measure, is a learnable parameter;

[0054] the supervised reconstruction error is:

[0055] ,

[0056] wherein, is the first super-resolution reconstructed gravity data, is the pre-processed high-resolution shipborne gravity data in the paired sample dataset; represents the pixel L1 loss, and SSIM represents a structural similarity index measure, is a learnable parameter;

[0057] the dual regression error is:

[0058] ,

[0059] wherein, is the shipborne low-resolution gravity data; is the pre-processed low-resolution satellite altimetry gravity data in the paired sample dataset; represents the pixel L1 loss, and SSIM represents a structural similarity index measure, is a learnable parameter;

[0060] the dual consistency error is:

[0061] ,

[0062] wherein, is the first reconstructed low-resolution gravity data; is the shipborne low-resolution gravity data; represents the pixel L1 loss, and SSIM represents a structural similarity index measure, is a learnable parameter.

[0063] Based on the above invention content, the present application has the following technical effects relative to the prior art:

[0064] The present application proposes a multi-source, multi-physical-field gravity data super-resolution reconstruction method, which is based on a semi-supervised dual regression deep learning model, constructs a closed-loop deep learning architecture composed of a primary regression network P (realizing the mapping of a low-resolution gravity field to a high-resolution gravity field) and a dual regression network D (realizing the mapping of a high-resolution gravity field to a low-resolution gravity field), and jointly uses a paired sample dataset (containing paired pre-processed low-resolution satellite altimetry gravity data and corresponding pre-processed high-resolution shipborne gravity data) and an unpaired sample dataset (containing only pre-processed low-resolution satellite altimetry gravity data); the closed-loop architecture calculates a first super-resolution reconstructed gravity data and the pre-processed high-resolution shipborne gravity data in the paired sample dataset the supervised reconstruction error between the low resolution mapped shipborne low resolution gravity data and the preprocessed low resolution satellite altimetry data in the paired sample dataset the dual regression error between the first reconstructed low resolution gravity data and the low resolution mapped shipborne low resolution gravity data and the low resolution mapped shipborne low resolution gravity data the dual consistency error constrains the paired sample data, and the reconstructed low resolution gravity data is calculated by low resolution mapping and the preprocessed low resolution satellite altimetry gravity data in the training dataset the cyclic consistency error constrains the paired sample data and the unpaired sample data, respectively; the above semi-supervised dual regression learning mechanism makes full use of the low resolution satellite data covering a large area of sea.

[0065] The application simulates the degradation process from high resolution gravity data to low resolution gravity data through a dual regression network D, and adaptively models key factors (such as noise and filtering effect) in the degradation process; this mechanism significantly enhances the ability of the model to capture regional degradation characteristics, improves the reliability of knowledge transfer between labeled and unlabeled areas, and greatly improves the generalization ability of the reconstruction result to different regional data.

[0066] Based on the semi-supervised learning mechanism and adaptive degradation modeling, the low resolution satellite gravity data of the target area is input into the trained main regression network P, and high resolution reconstructed gravity data can be obtained. This method can fully exploit the potential information of a large amount of unlabeled satellite gravity data under the condition of limited high resolution shipborne data, breaking through the trade-off bottleneck between accuracy and efficiency of traditional methods, and providing important support for improving the modeling ability of marine gravity field. And the method of the application effectively improves the reconstruction accuracy and robustness under the premise of significantly reducing the dependence on shipborne gravity data, providing high-precision, high-resolution gravity field data support for marine resource exploration and seabed topographic mapping. BRIEF DESCRIPTION OF DRAWINGS

[0067] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0068] Figure 1 is a flowchart of a multi-source, multi-physical field gravity data super-resolution reconstruction method provided by the embodiments of the application;

[0069] Figure 2 FIG. 1 is a schematic diagram of a framework for gravity data super-resolution reconstruction based on a dual regression deep learning model according to an embodiment of the present application. DETAILED DESCRIPTION

[0070] In order to better understand the technical solutions of the present application, the embodiments of the present application will be described in detail below with reference to the drawings. It should be clear that the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0071] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0072] It should be understood that the term "and / or" used herein is only to describe the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents that the front and rear associated objects are a "or" relationship.

[0073] Referring to Figure 1 FIG. 2 is a schematic diagram of a flow of a multi-source, multi-physical field gravity data super-resolution reconstruction method according to an embodiment of the present application. As shown in the figure, it mainly includes the following steps. Figure 1

[0074] Step S101: Obtain gravity data, which includes low-resolution satellite altimetry gravity data and high-resolution shipborne gravity data.

[0075] Step S102: Preprocess the low-resolution satellite altimetry gravity data and the high-resolution shipborne gravity data, obtain preprocessed low-resolution satellite altimetry gravity data and preprocessed high-resolution shipborne gravity data, and construct paired sample data set and unpaired sample data set according to the preprocessed low-resolution satellite altimetry gravity data and the preprocessed high-resolution shipborne gravity data.

[0076] The paired sample data set is composed of preprocessed low-resolution satellite altimetry gravity data and preprocessed high-resolution shipborne gravity data The unpaired sample data set is composed of preprocessed low-resolution satellite altimetry gravity data . ​

[0077] The paired sample dataset and the unpaired sample dataset constitute a training dataset. The low resolution satellite altimetry gravity data in the training dataset includes the preprocessed low resolution satellite altimetry gravity data in the paired sample dataset and the preprocessed low resolution satellite altimetry gravity data in the unpaired sample dataset .

[0078] The preprocessing includes data cleaning, quality control, outlier detection and rejection.

[0079] The quality control includes first stage quality control and second stage quality control on the high resolution shipborne gravity data, the first stage quality control includes:

[0080] calculating the local mean μ and the standard deviation σ of the high resolution shipborne gravity data within a spatial window, and rejecting the data points satisfying the following formula:

[0081] ,

[0082] wherein, G is the gravity value of the current data point, μ is the local mean of the high resolution shipborne gravity data within the spatial window, and σ is the standard deviation of the high resolution shipborne gravity data within the spatial window.

[0083] The second stage quality control includes:

[0084] obtaining the standard earth gravity model EGM2008, and rejecting the data points satisfying the following formula:

[0085] ,

[0086] wherein, G is the gravity value of the current data point, GEGM2008 is the gravity value of the current data point of the standard earth gravity model EGM2008, and σ is the standard deviation of the EGM2008 model.

[0087] resampling the preprocessed shipborne gravity data to a uniform grid by bilinear interpolation to realize spatial alignment with the preprocessed low resolution satellite altimetry gravity data, and generating the paired sample dataset and the unpaired sample dataset by using the spatially aligned preprocessed low resolution satellite altimetry gravity data and the preprocessed high resolution shipborne gravity data.

[0088] ​​​​​​​​​Step S103: constructing a dual regression deep learning model, the dual regression deep learning model comprising: a primary regression network P and a dual regression network D; the primary regression network P and the dual regression network D can be any deep neural network type, including but not limited to the following classic types of networks: convolutional neural network, recurrent neural network, fully connected neural network, Transformer, and Mamba.

[0089] The primary regression network P is responsible for super-resolution gravity data reconstruction, realizing the mapping between the preprocessed low-resolution satellite altimetry gravity data and the super-resolution reconstructed gravity data, denoted as a first mapping, and the expression is:

[0090] ,

[0091] wherein, is the preprocessed low-resolution satellite altimetry gravity data in the training data set, is the super-resolution reconstructed gravity data.

[0092] The super-resolution reconstructed gravity data includes the first super-resolution reconstructed gravity data and the second super-resolution reconstructed gravity data , the first super-resolution reconstructed gravity data is:

[0093] ,

[0094] wherein, is the preprocessed low-resolution satellite altimetry gravity data in the paired sample data set;

[0095] The second super-resolution reconstructed gravity data is:

[0096] ,

[0097] wherein, is the preprocessed low-resolution satellite altimetry gravity data in the unpaired sample data set.

[0098] The dual regression network D is responsible for degradation process simulation, realizing the mapping between the super-resolution reconstructed gravity data and the reconstructed low-resolution gravity data, and the mapping between the preprocessed high-resolution shipborne gravity data and the shipborne low-resolution gravity data. The mapping between the super-resolution reconstructed gravity data and the reconstructed low-resolution gravity data is denoted as a second mapping, and the mapping between the preprocessed high-resolution shipborne gravity data and the shipborne low-resolution gravity data is denoted as a third mapping.

[0099] The expression of the second mapping is:

[0100] ,

[0101] wherein, is the reconstructed low resolution gravity data, is the super-resolution reconstructed gravity data.

[0102] reconstructing low resolution gravity data comprising a first reconstructed low resolution gravity data and a second reconstructed low resolution gravity data , the first reconstructed low resolution gravity data is:

[0103] ,

[0104] wherein, is the first super-resolution reconstructed gravity data;

[0105] the second reconstructed low resolution gravity data is:

[0106] ,

[0107] wherein, is the second super-resolution reconstructed gravity data.

[0108] the third mapping is expressed as:

[0109] ,

[0110] wherein, is the shipborne low resolution gravity data, is the pre-processed high resolution shipborne gravity data.

[0111] The first mapping is a mapping of low resolution gravity data LRQ to high resolution gravity data HRQ, and the second mapping and the third mapping are mappings of high resolution gravity data HRQ to low resolution gravity data LRQ. The low resolution gravity data LRQ comprises pre-processed low resolution satellite altimetry gravity data, reconstructed low resolution gravity data, and shipborne low resolution gravity data; and the high resolution gravity data HRQ comprises super-resolution reconstructed gravity data and pre-processed high resolution shipborne gravity data.

[0112] If the training data is a paired sample data set, the dual regression deep learning model can realize training and learning of the dual regression deep learning model by synchronously constraining the three mapping relationships of the first mapping, the second mapping, and the third mapping.

[0113] If the training data is a non-paired sample data set, the dual regression deep learning model realizes training and learning of the dual regression deep learning model by constraining both mapping relationships of the first mapping and the second mapping.

[0114] The primary regression network P and the dual regression network D constitute a cyclic consistency mapping relationship (LRQ→HRQ→LRQ) from low-resolution gravity data LRQ to high-resolution gravity data HRQ and then to low-resolution gravity data LRQ. The closed-loop structure enables the dual regression deep learning model to fully utilize paired sample data sets and a large number of non-paired sample data sets, significantly reducing the dependence on high-cost high-resolution shipborne gravity data. By introducing adaptive degradation modeling of the dual regression network as additional supervision information, the ability of the model to capture regional bias is further improved, promoting knowledge transfer between labeled regions (paired sample regions) and unlabeled regions (non-paired sample regions), significantly reducing the dependence on large amounts of shipborne gravity data, thereby enhancing the accuracy, robustness, and generalization ability of the reconstruction results.

[0115] Step S104: training the dual regression deep learning model using the training data set to obtain a trained dual regression deep learning model.

[0116] In one embodiment, the preprocessed low-resolution satellite altimetry gravity data in the training data set is input into the primary regression network P, and the preprocessed high-resolution shipborne gravity data in the training data and the super-resolution reconstruction gravity data generated by the primary regression network P are input into the dual regression network D. The primary regression network P and the dual regression network D are trained simultaneously, specifically including:

[0117] Step S1041: inputting the preprocessed low-resolution satellite altimetry gravity data in the training data set into the primary regression network P to obtain corresponding super-resolution reconstruction gravity data Step S1042: inputting the preprocessed high-resolution shipborne gravity data in the training data and the super-resolution reconstruction gravity data generated by the primary regression network P into the dual regression network D to obtain corresponding super-resolution reconstruction gravity data , the expression is:

[0118] ,

[0119] The super-resolution reconstruction gravity data includes first super-resolution reconstruction gravity data and second super-resolution reconstruction gravity data , the first super-resolution reconstruction gravity data is:

[0120] ,

[0121] wherein, the preprocessed low-resolution satellite altimetry gravity data in the paired sample dataset;

[0122] the second super-resolution reconstructed gravity data is:

[0123] ,

[0124] wherein, the preprocessed low-resolution satellite altimetry gravity data in the unpaired sample dataset.

[0125] Step S1042: inputting the super-resolution reconstructed gravity data into the dual regression network D to obtain the low-resolution mapped reconstructed low-resolution gravity data , and the expression is:

[0126] ,

[0127] wherein, the preprocessed low-resolution satellite altimetry gravity data in the training dataset;

[0128] the reconstructed low-resolution gravity data includes the first reconstructed low-resolution gravity data and the second reconstructed low-resolution gravity data , the first reconstructed low-resolution gravity data is:

[0129] ,

[0130] wherein, the first super-resolution reconstructed gravity data; the preprocessed low-resolution satellite altimetry gravity data in the paired sample dataset;

[0131] the second reconstructed low-resolution gravity data is:

[0132] ,

[0133] wherein, the second super-resolution reconstructed gravity data; the preprocessed low-resolution satellite altimetry gravity data in the unpaired sample dataset.

[0134] Step S1043: judging whether the preprocessed low-resolution satellite altimetry gravity data whether it belongs to a paired sample dataset, if it belongs, then step S1044-step S1047 is executed; if it does not belong, then step S1045-step S1047 is executed;

[0135] Step S1044: the preprocessed high-resolution shipborne gravity data Inputting the dual regression network D, obtaining the low-resolution mapped shipborne low-resolution gravity data , the expression is:

[0136] ,

[0137] Step S1045: constructing a target function, calculating the error of all output data in the primary regression network P and the dual regression network D and its corresponding label data;

[0138] Considering the error of the reconstructed gravity data in terms of pixel and structural consistency, an error function that combines local detail fidelity and global perceptual similarity is defined, the expression is as follows:

[0139] ,

[0140] wherein, represents the pixel L1 loss, SSIM represents the structural similarity index measure, is a learnable parameter that dynamically balances the contribution of pixel L1 loss and structural similarity index measure.

[0141] It should be understood that the above error function is only one possibility provided by the embodiments of the present application, and any error function or function combination under the dual regression learning framework proposed in the present application is within the protection scope of the present application.

[0142] The error function is used to calculate the cyclic consistency error , the supervised reconstruction error , the dual regression error and the dual consistency error .

[0143] The target function for training the dual regression learning model is:

[0144] ,

[0145] wherein, is the target error, is a binary indicator function, if the preprocessed low-resolution satellite altimetric gravity data in the training data set in step S1043 belongs to a paired sample dataset, then ; if the pre-processed low-resolution satellite altimetry gravity data in step S1043 does not belong to the paired sample dataset but to the unpaired sample dataset, then ; is a cyclic consistency error, which is used to describe the deviation error between the reconstructed low-resolution gravity data obtained by performing degradation reconstruction on the first super-resolution reconstructed gravity data and the pre-processed low-resolution satellite altimetry gravity data in the training dataset; ; is a supervised reconstruction error, which is used to describe the data and structure consistency error between the first super-resolution reconstructed gravity data obtained by performing super-resolution reconstruction on the pre-processed low-resolution satellite altimetry gravity data in the paired sample dataset and the pre-processed high-resolution shipborne gravity data in the paired sample dataset; ; is a dual regression error, which is used to describe the error degree between the shipborne low-resolution gravity data obtained by performing degradation reconstruction on the high-resolution shipborne gravity data in the paired sample dataset and the pre-processed low-resolution satellite altimetry data in the paired sample dataset; ; is a dual consistency error, which is used to describe the deviation error between the first reconstructed low-resolution gravity data obtained by performing degradation reconstruction on the first super-resolution reconstructed gravity data and the shipborne low-resolution gravity data obtained by performing degradation reconstruction on the high-resolution shipborne gravity data in the paired sample dataset by the dual regression network D; ; The scalars λ, β and γ are learnable parameters for controlling the relative contributions of the cyclic consistency error, the supervised reconstruction error, the dual regression error and the dual consistency error.

[0146] The cyclic consistency error is:

[0147] ,

[0148] wherein, is the reconstructed low-resolution gravity data; is the pre-processed low-resolution satellite altimetry gravity data in the training dataset; represents the pixel L1 loss, and SSIM represents the structural similarity index measure,​​​​ is a learnable parameter.

[0149] the supervised reconstruction error is:

[0150] ,

[0151] wherein, is the first super-resolution reconstructed gravity data, is the pre-processed high-resolution shipborne gravity data in the paired sample dataset; represents the pixel L1 loss, and SSIM represents the structural similarity index measure, is a learnable parameter.

[0152] the dual regression error is:

[0153] ,

[0154] wherein, is the shipborne low-resolution gravity data; is the pre-processed low-resolution satellite altimetry data in the paired sample dataset; represents the pixel L1 loss, and SSIM represents the structural similarity index measure, is a learnable parameter.

[0155] the dual consistency error is:

[0156] ,

[0157] wherein, is the first reconstructed low-resolution gravity data; is the shipborne low-resolution gravity data; represents the pixel L1 loss, and SSIM represents the structural similarity index measure, is a learnable parameter.

[0158] Step S1046: determining whether the objective function meets a preset condition, if not, performing step S1047, and if yes, completing the training of the dual regression deep learning model to obtain the trained dual regression deep learning model;

[0159] The preset condition includes one of the following standards: the number of iterations reaches a specified upper limit of the number of iterations; the target error is lower than a specified error limit value; the descending rate of the target error is lower than a set descending rate threshold; and the running time reaches a predetermined upper limit of time.

[0160] Step S1047: setting the target error The back propagation to the primary regression network P and the dual regression network D updates the learnable parameters of the dual regression deep learning model, and the steps S1041 to S1046 are continuously executed.

[0161] Figure 2 is a schematic diagram of a framework for gravity data super-resolution reconstruction based on a dual regression deep learning model provided by the embodiments of the present application. The primary regression network P is responsible for realizing the mapping of the low-resolution gravity field to the high-resolution gravity field (LRQ→HRQ), that is, the super-resolution reconstruction of gravity data; the dual regression network D is responsible for realizing the mapping of the high-resolution gravity field to the low-resolution gravity field (HRQ→LRQ), that is, the simulation of the degradation process. Through the synergistic constraints of the cycle consistency loss, the reconstruction loss, the dual regression, and the dual consistency, the synergistic optimization of the primary regression network P and the dual regression network D is realized; at the same time, the adaptive degradation process simulation realized by the dual regression network D can better capture the regional bias and resolution loss mode introduced in the satellite data processing process, ensuring that the model can still effectively learn the rules of data fusion in the case of only a small amount of labeled data.

[0162] Step S105: input the preprocessed low-resolution satellite gravity data of the target area to be reconstructed into the primary regression network P in the trained dual regression deep learning model, and obtain the super-resolution reconstruction gravity data of the target area.

[0163] The present application provides a multi-source, multi-physical-field gravity data super-resolution reconstruction method, which is based on a semi-supervised dual regression deep learning model, constructs a closed-loop deep learning architecture composed of a primary regression network P (realizing the mapping of the low-resolution gravity field to the high-resolution gravity field) and a dual regression network D (realizing the mapping of the high-resolution gravity field to the low-resolution gravity field), and jointly uses paired sample data sets (containing paired preprocessed low-resolution satellite gravity data and corresponding preprocessed high-resolution shipborne gravity data) and unpaired sample data sets (containing only preprocessed low-resolution satellite gravity data). The closed-loop architecture constrains the paired sample data by calculating the supervised reconstruction error between the first super-resolution reconstruction gravity data and the preprocessed high-resolution shipborne gravity data in the paired sample data set the dual regression error between the low-resolution mapped shipborne low-resolution gravity data and the preprocessed low-resolution satellite gravity data in the paired sample data set the dual consistency error between the low-resolution mapped first reconstruction low-resolution gravity data and the low-resolution mapped shipborne low-resolution gravity data The dual consistency error between the low-resolution mapped first reconstruction low-resolution gravity data and the low-resolution mapped shipborne low-resolution gravity data is calculated. and the pre-processed low-resolution satellite altimetry gravity data in the training data set respectively constrain learning of paired sample data and non-paired sample data; the above semi-supervised dual regression learning mechanism fully utilizes low-resolution satellite data covering a wide range of sea areas.

[0164] The application example ensures that the model can effectively learn the data fusion rule even if only a small amount of labeled data is available based on the common constraint of cycle consistency and dual consistency error through the proposed dual regression network learning mechanism, organically fuses wide-area and massive low-resolution satellite altimetry data and local and sparse high-resolution shipborne gravity observation data, significantly reduces the dependence on a large amount of labeled data set, improves the efficiency and accuracy of data fusion, and realizes reconstruction of a high-resolution marine gravity field.

[0165] The application simulates the degradation process from high-resolution gravity data to low-resolution gravity data through the dual regression network D, and adaptively models key factors (such as noise and filtering effect) in the degradation process; this mechanism significantly enhances the ability of the model to capture regional degradation characteristics, improves the reliability of knowledge transfer between labeled areas and unlabeled areas, and thus greatly improves the generalization ability of the reconstruction result to different regional data.

[0166] Based on the semi-supervised learning mechanism and adaptive degradation modeling, the low-resolution satellite gravity data of the target area is input into the trained primary regression network P, and high-resolution reconstructed gravity data can be obtained. This method can fully exploit the potential information of a large amount of unlabeled satellite gravity data under the condition of limited high-resolution shipborne data, breaks through the trade-off bottleneck between accuracy and efficiency of traditional methods, and also provides important support for improving the modeling capability of marine gravity field. Moreover, the method of the application effectively improves the reconstruction accuracy and robustness under the premise of significantly reducing the dependence on shipborne gravity data, and provides high-resolution gravity field data support for marine geophysical exploration and seabed topographic mapping.

[0167] In the embodiments of the application, "at least one" means one or more, and "multiple" means two or more. The "and / or" describes the association relationship of the associated objects, which means that there can be three kinds of relationships, for example, A and / or B can represent the cases of A alone, A and B together, and B alone. Wherein A and B can be singular or plural. The character " / " generally represents an "or" relationship between the front and rear associated objects. "At least one of the following" and the like means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b and c can represent: a, b, c, a-b, a-c, b-c or a-b-c, wherein a, b and c can be single or multiple.

[0168] The above description is only specific embodiments of the present application, and any skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered by the protection scope of the present application.

Claims

1. A multi-source, multi-physical-field gravity data super-resolution reconstruction method, characterized in that, The method comprises the following steps: Step S101: acquiring gravity data, the gravity data comprising low-resolution satellite altimetry gravity data and high-resolution shipborne gravity data; Step S102: pre-processing the low-resolution satellite altimetry gravity data and the high-resolution shipborne gravity data to obtain pre-processed low-resolution satellite altimetry gravity data and pre-processed high-resolution shipborne gravity data, and constructing a paired sample data set and an unpaired sample data set according to the pre-processed low-resolution satellite altimetry gravity data and the pre-processed high-resolution shipborne gravity data; The paired sample data set and the unpaired sample data set constitute a training data set; The paired sample dataset consists of pre-processed low resolution satellite altimetry gravity data and pre-processed high resolution shipborne gravity data The unpaired sample dataset consists of pre-processed low resolution satellite altimetry gravity data ​ Step S103: constructing a dual regression deep learning model, the dual regression deep learning model comprising a primary regression network P and a dual regression network D; Step S104: training the dual regression deep learning model using the training data set to obtain a trained dual regression deep learning model, specifically comprising: Step S1041: pre-processed low-resolution satellite altimetry gravity data in the training data set is input into the main regression network P As input data, the main regression network P is input, and the corresponding super-resolution reconstructed gravity data is obtained The expression is: ， Super-resolution reconstructed gravity data including first super-resolution reconstructed gravity data and second super-resolution reconstructed gravity data , the first super-resolution reconstructed gravity data is: ， wherein, is the pre-processed low resolution satellite altimetry gravity data in the paired sample dataset; Second super-resolution reconstructed gravity data is: ， wherein, is the pre-processed low resolution satellite altimetry gravity data in the non-paired sample dataset; Step S1042: super-resolution reconstruction of gravity data , input the dual regression network D, obtain the low-resolution mapped reconstructed low-resolution gravity data , the expression is: ； Reconstructing low resolution gravity data including a first reconstructed low resolution gravity data and a second reconstructed low resolution gravity data , the first reconstructed low resolution gravity data is: ； Second reconstruction of low resolution gravity data is: ； Step S1043: judging whether the preprocessed low-resolution satellite altimetry gravity data in the input training data set belongs to paired sample data set, if yes, executing step S1044-step S1047; if not, executing step S1045-step S1047; whether belongs to paired sample data set, if yes, executing step S1044-step S1047; if not, executing step S1045-step S1047; Step S1044: pre-processing the high-resolution shipborne gravity data Input the dual regression network D, and obtain the low-resolution mapped shipborne low-resolution gravity data The expression is: ； Step S1045: constructing an objective function, and calculating errors of all output data in the primary regression network P and the dual regression network D and corresponding label data; Step S1046: determining whether the objective function meets a preset condition, if not, performing step S1047, and if yes, completing training of the dual regression deep learning model to obtain a trained dual regression deep learning model; Step S1047: Calculate the target error The back propagation is performed to the primary regression network P and the dual regression network D, the learnable parameters of the dual regression deep learning model are updated, and steps S1041 to S1046 are continuously executed. Step S105: inputting pre-processed low-resolution satellite gravity data of a target area to be reconstructed into the primary regression network P in the trained dual regression deep learning model to obtain super-resolution reconstructed gravity data of the target area.

2. The method of claim 1, wherein, Low resolution satellite altimetry gravity data in the training data set Pre-processed low resolution satellite altimetry gravity data in the paired sample data set Pre-processed low resolution satellite altimetry gravity data in the unpaired sample data set .

3. The method of claim 1, wherein, The pre-processing comprises data cleaning, quality control, abnormal value detection and elimination.

4. The method of claim 3, wherein, The quality control comprises first-stage quality control and second-stage quality control on the high-resolution shipborne gravity data.

5. The method of claim 4, wherein, The first-stage quality control comprises: The computation of high resolution shipboard gravity data in The local mean μ and standard deviation σ within a spatial window will satisfy the following formula Reject: ， in, Here, μ represents the gravity value at the current data point, and μ is the high-resolution shipboard gravity data. The local mean within the spatial window, σ represents the high-resolution shipborne gravity data in [the context of the space window]. Standard deviation within the spatial window.

6. The method of claim 4, wherein, The second-stage quality control comprises: The standard Earth gravity model EGM2008 is acquired, which satisfies the following formula Cull: ， wherein, is the gravity value of the current data point, is the gravity value of the current data point of the standard Earth gravity model EGM2008, is the standard deviation of the EGM2008 model.

7. The method of claim 1, wherein, The objective function is: ， in, For the target error, It is a binary indicator function, which is the preprocessed low-resolution satellite altimetry and gravity data in the training dataset described in step S1043. If it belongs to the paired sample dataset, then If the low-resolution satellite altimetry gravity data preprocessed in step S1043 If it does not belong to the paired sample dataset, but to the unpaired sample dataset, then ; Cyclic consistency error is used to describe the error in super-resolution reconstructed gravity data obtained through the dual regression network D. Reconstructed low-resolution gravity data obtained through degradation reconstruction Compared with the preprocessed low-resolution satellite altimetry and gravity data in the training dataset The degree of deviation between them in terms of pixel and structural consistency; To supervise reconstruction errors, the main regression network P is used to describe the preprocessed low-resolution satellite altimetry and gravity data in the paired sample dataset. The first super-resolution reconstructed gravity data obtained from super-resolution reconstruction Preprocessed high-resolution shipborne gravity data from the paired sample dataset Errors in data and structural consistency between them; The dual regression error is used to describe the high-resolution shipborne gravity data in the paired sample dataset processed by the dual regression network D. Shipborne low-resolution gravity data obtained from degradation reconstruction Preprocessed low-resolution satellite altimetry data in the paired sample dataset The degree of error between them; The dual consistency error is used to describe the error in reconstructing gravity data from the first super-resolution network D. The first reconstructed low-resolution gravity data obtained through degradation reconstruction High-resolution shipboard gravity data in the paired sample dataset were compared with those obtained through a dual regression network D. Shipborne low-resolution gravity data obtained from degradation reconstruction The deviation error between; scalars λ, β, and γ are learnable parameters used to control the relative contributions of cyclic consistency error, supervised reconstruction error, dual regression error, and dual consistency error.

8. The method of claim 7, wherein, The cycle-consistency error Is: ， wherein, to reconstruct the low resolution gravity data; to train the pre-processed low resolution satellite altimetry gravity data in the data set; represents the pixel L1 loss, SSIM represents the structural similarity index measure, is a learnable parameter; the supervised reconstruction error is: ， wherein, is the first super-resolution reconstructed gravity data, is the pre-processed high-resolution shipborne gravity data in the paired sample dataset; represents the pixel L1 loss, SSIM represents the structural similarity index measure, is a learnable parameter; The dual regression error Is: ， wherein, is the shipboard low-resolution gravity data; is the pre-processed low-resolution satellite altimetry data in the paired sample dataset; represents the pixel L1 loss, SSIM represents the structural similarity index measure, is a learnable parameter; The dual consistency error Is: ， wherein, is the first reconstructed low resolution gravity data; is the shipborne low resolution gravity data; represents the pixel LI loss, SSIM represents the structural similarity index measure, is a learnable parameter.

Images