CN116148855A - Time-series InSAR Atmospheric Phase Removal and Deformation Calculation Method and System - Google Patents

Abstract

The invention discloses a method and a system for removing and resolving deformation of a time sequence InSAR atmospheric phase, wherein the method firstly acquires a time sequence SAR image and DEM data of a monitoring area, and carries out preprocessing, differential interference, filtering and phase unwrapping; secondly, constructing a sample library containing differential interferograms of atmospheric phases; generating an antagonistic neural network CGAN based on the conditions, amplifying the sample library, and constructing a complete version of sample library; then constructing an atmosphere phase removal transune network model based on a transune network, and training and testing to remove the atmosphere phase in the differential interferogram; and finally, performing time sequence InSAR deformation calculation based on the differential interference pattern with the atmospheric phase removed so as to acquire the earth surface deformation information of the monitoring area. The invention can break through the technical bottleneck that the atmospheric phase error cannot be completely eliminated in the existing InSAR technology, and simultaneously improves the accuracy of time sequence InSAR deformation calculation.

Description

Method and system for atmospheric phase removal and deformation resolution of time-series InSAR

Technical Field

The invention relates to the technical field of InSAR deformation monitoring, and in particular to a method and system for time-series InSAR atmospheric phase removal and deformation resolution.

Background Art

Synthetic aperture radar interferometry (InSAR) is widely used for deformation monitoring on the ground. However, the presence of atmospheric phase errors in InSAR measurements makes InSAR technology still challenging in estimating accurate surface deformation. When radar signals propagate in a layered turbulent atmosphere and ionosphere, the measurement accuracy of a single interferogram is susceptible to atmospheric influences. Ionospheric influences are caused by changes in electron density in the ionosphere, which can cause phase distortion of radar signals and have a significant impact on data with larger wavelengths such as P-band and L-band SAR data. Tropospheric delay is caused by changes in temperature, pressure, and humidity, and can be divided into turbulent component delay and stratification delay. Turbulent component delay is caused by changes in water vapor distribution in the troposphere at short time resolutions and spatial scales of several kilometers, and this delay is usually difficult to model and remove. The atmospheric stratification delay is related to the terrain and shows spatial correlation on length scales of tens of kilometers. In this context, a 20% spatial or temporal change in relative humidity will bring tens of centimeters of atmospheric errors to the InSAR interferogram, resulting in incorrect interpretation of the interference phase in the interferogram and inaccurate extraction of deformation.

At present, many methods have been developed in the field of InSAR technology to estimate and eliminate the impact of atmospheric errors on InSAR interferograms, which can be divided into three categories: the first category is based on empirical model methods, that is, using elevation linear models or power law models to remove atmospheric stratification delays related to terrain, the second category is based on external data methods to generate atmospheric delay maps, where external data includes meteorological data (ERA5/WRF/EWMC), global satellite navigation system (GNSS) data, spectral data (MODIS/MERIS/Sentinel-3), and the third category is based on phase methods, such as linear elevation correction and time series InSAR methods. The invention patent with publication number CN114624708A provides a ground-based radar atmospheric correction method and system in a complex environment. The method obtains atmospheric phase estimation results by extracting permanent scatterer points and performing regional division, triangulation and atmospheric phase interpolation, and removes them in the accumulated phase to achieve atmospheric correction. However, this method is a phase-based atmospheric correction method in the field of ground-based radar technology, and it ignores time variable signals and underestimates the seasonal characteristics of atmospheric delay. In recent years, deep learning has shown its application potential in InSAR technology in phase unwrapping, atmospheric phase estimation, DEM super-resolution reconstruction, coherence estimation and phase filtering, deformation detection and prediction, etc. The invention patent with publication number CN114280608A provides a method and system for removing atmospheric effects related to DInSAR elevation. The method uses an MLP neural network model to simulate the atmospheric phase related to elevation; the unwrapped phase is subtracted from the simulated atmospheric phase related to elevation, thereby completing the atmospheric effect removal. However, this method can only remove the atmospheric stratification delay related to elevation, and cannot effectively remove the turbulent atmospheric delay.

By summarizing the above methods, we found that the following problems still exist: (1) The empirical model can remove the atmospheric stratification delay related to terrain, but cannot remove the atmospheric turbulence component; (2) The method based on external data depends on the temporal and spatial resolution of the external data. For example, the method based on GNSS data depends on the density of the GNSS network, which makes it difficult to capture the local atmospheric turbulence changes in the frozen soil area. The MODIS/MERIS/Sentinel-3 images have low temporal and spatial resolution and cannot be applied to cloud conditions. In addition, the data acquisition time is inconsistent with the SAR data acquisition time. The combined data (GACOS) method ignores the local turbulence and inhomogeneous atmospheric components; (3) The time domain filtering algorithm depends on the parameter setting and is susceptible to seasonal changes in the InSAR time series deformation, and cannot remove the atmospheric turbulence component. In summary, the current atmospheric phase error removal methods in time series InSAR technology mostly rely on external data and algorithm model performance, and cannot effectively remove some turbulent atmospheric errors.

In order to more effectively suppress and remove the influence of atmospheric errors in time-series InSAR technology, we developed a method and system for atmospheric phase removal and deformation solution of time-series InSAR, designed a method for InSAR atmospheric phase sample augmentation based on the conditional generative adversarial neural network CGAN with self-attention mechanism, and introduced the TransUNet network to construct a network framework for atmospheric phase removal based on time-series InSAR technology.

Summary of the invention

The purpose of the present invention is to provide a method and system for removing atmospheric phase and calculating deformation of time-series InSAR in view of the shortcomings of the prior art. The present invention can effectively separate atmospheric delay error and deformation of time-series InSAR and improve the accuracy of InSAR deformation calculation.

The objective of the present invention is achieved through the following technical solutions: In a first aspect, an embodiment of the present invention provides a method for atmospheric phase removal and deformation resolution of a time series InSAR, comprising the following steps:

(1) Obtain time series SAR image data and digital elevation model data of the monitoring area, and perform preprocessing, differential interferometry, filtering, and phase unwrapping on the data;

(2) constructing a first sample library of differential interferograms containing atmospheric phase according to the thresholds of the time baseline and the space baseline;

(3) Based on the conditional generative adversarial neural network (CGAN), the first sample library containing the atmospheric phase is augmented and constructed to obtain a complete sample library of differential interferograms;

(4) Based on the TransUNet network, the atmospheric phase removal TransUNet network model is constructed, and the model is trained and tested to remove the atmospheric phase from all differential interferograms;

(5) Based on the differential interferogram after removing the atmospheric phase, the deformation of the time-series InSAR is solved to obtain the surface deformation information of the monitoring area.

Optionally, step (1) includes the following sub-steps:

(1.1) Obtain time series SAR image data of the monitoring area and extract digital elevation model data of the monitoring area according to the latitude and longitude range of the monitoring area;

(1.2) Data preprocessing: First, the original time series SAR image data file is imported and converted into a single-view complex data set. Then, the orbit parameters are updated in combination with the downloaded refined track data file. Finally, the digital elevation model data and refined track data file are used for geometric positioning registration to obtain the registered time series SAR images.

(1.3) Differential interferometry: First, the time series SAR images after registration are interferometrically processed, and the phase difference between the main image and the auxiliary image is performed to obtain the interferogram. Then, the terrain phase and the flat ground phase of each interferogram are calculated using the digital elevation model data according to the inverse distance weighted interpolation algorithm, and the two phases are subtracted to obtain the differential interferogram.

(1.4) In order to suppress the influence of noise in the differential interferogram, nonlinear adaptive Goldstein spatial domain filtering is selected to filter the differential interferogram to obtain a filtered differential interferogram;

(1.5) Phase unwrapping: The filtered differential interferogram is subjected to phase unwrapping using the minimum cost flow method to obtain a phase unwrapped differential interferogram.

Optionally, step (2) includes the following sub-steps:

，并从差分干涉图集

中挑选含有大气相位的差分干涉图集

；(2.1) Select the differential interferogram from the phase unwrapped differential interferogram according to the threshold of the time baseline and the spatial baseline to generate a differential interferogram set

, and from the differential interferogram

Select the differential interferogram containing the atmospheric phase

;

进行图像样本和标签制作，对每个差分干涉图进行图像切分处理生成图像样本，并对生成的图像样本进行归一化处理，以构建含有大气相位的差分干涉图的第一样本库并记为

。(2.2) Differential interference atlas

Image samples and labels are prepared, each differential interferogram is segmented to generate image samples, and the generated image samples are normalized to construct the first sample library of differential interferograms containing the atmospheric phase and recorded as

.

Optionally, the step (3) includes the following sub-steps:

对条件生成对抗神经网络CGAN进行训练和测试，以生成额外的含有大气相位的第二样本库

；(3.1) The first sample library based on differential interferograms containing atmospheric phase

The conditional generative adversarial neural network (CGAN) is trained and tested to generate an additional second sample library containing atmospheric phases.

;

与第一样本库

合并生成含有大气相位的差分干涉图的第三样本库I；根据相位解缠后的差分干涉图选择未受到大气相位影响的差分干涉图，并进行图像切分和数据集的制作，以生成未含有大气相位的差分干涉图的第四样本库P；根据第三样本库I和第四样本库P构建差分干涉图的完整版样本库并记为

。(3.2) The second sample library

With the first sample library

The third sample library I of the differential interferogram containing the atmospheric phase is generated by merging; the differential interferogram not affected by the atmospheric phase is selected according to the differential interferogram after phase unwrapping, and image segmentation and data set preparation are performed to generate the fourth sample library P of the differential interferogram not containing the atmospheric phase; the complete version of the sample library of the differential interferogram is constructed according to the third sample library I and the fourth sample library P and recorded as

.

Optionally, the conditional generative adversarial neural network CGAN includes a generator and a discriminator.

Optionally, the complete version of the sample library includes an InSAR differential interferogram sample data set not containing atmospheric phase corresponding to the fourth sample library, a network-generated InSAR differential interferogram sample data set containing atmospheric phase corresponding to the second sample library, and an InSAR differential interferogram sample data set containing real atmospheric phase corresponding to the first sample library.

Optionally, step (4) includes the following sub-steps:

(4.1) Based on the TransUNet network, the atmospheric phase removal TransUNet network model is constructed to reconstruct the differential interferogram after the output atmospheric phase is removed;

作为大气相位去除TransUNet网络模型的输入数据集，按用户输入的比例将其划分为训练数据集和测试数据集，将训练数据集加载到大气相位去除TransUNet网络模型中进行训练，以获取训练好的权重参数信息，使用测试数据集加载训练好的权重参数信息，以获取去除大气相位的差分干涉图样本；(4.2) Complete sample library

As the input data set of the atmospheric phase removal TransUNet network model, it is divided into a training data set and a test data set according to the ratio input by the user, the training data set is loaded into the atmospheric phase removal TransUNet network model for training to obtain the trained weight parameter information, and the trained weight parameter information is loaded using the test data set to obtain the differential interferogram sample after removing the atmospheric phase;

(4.3) The differential interferogram samples after removing the atmospheric phase are sliced and merged to generate a differential interferogram after removing the atmospheric phase.

Optionally, the TransUNet network combines the two network structures of UNet and Tranformers, and is composed of an encoder and a decoder to form a U-shaped structure.

Optionally, the step (5) is specifically as follows: performing time-series InSAR deformation solution based on the differential interferogram after removing the atmospheric phase, using the small baseline subset method to calculate the time-series deformation variable and cumulative deformation variable of the monitoring area, and using weighted least squares estimation to achieve a robust solution of the time-series deformation to obtain the surface deformation information of the monitoring area.

A second aspect of an embodiment of the present invention provides a system for time-series InSAR atmospheric phase removal and deformation resolution, which is used to implement the above-mentioned time-series InSAR atmospheric phase removal and deformation resolution method, including:

Data preprocessing module, used for data import and image registration of acquired time series SAR images;

The data differential interferometry, filtering and phase unwrapping processing module is used to perform differential interferometry, nonlinear adaptive Goldstein spatial filtering and minimum cost flow phase unwrapping according to the registered time series SAR images;

The construction and augmentation module of the differential interferogram sample library is used to generate samples of differential interferograms containing atmospheric phase, augment the atmospheric phase samples based on the conditional generative adversarial neural network CGAN, and construct the differential interferogram sample library;

An atmospheric phase removal module, used to remove the atmospheric phase from the constructed differential interferogram sample library based on the TransUNet network; and

The time-series InSAR deformation calculation module is used to perform time-series InSAR deformation calculation on the differential interferogram after removing the atmospheric phase, so as to obtain the deformation information of the surface.

The beneficial effect of the present invention is that the present invention can break through the technical bottleneck of the existing InSAR technology that cannot completely eliminate the atmospheric phase error. The proposed TransUNet network can effectively reduce the influence of the atmosphere in the InSAR interferogram, showing the great potential of the atmospheric phase elimination method based on deep learning, and improving the accuracy of the time-series InSAR deformation solution. The invention can also be used in the application field of high-precision deformation extraction based on InSAR technology, suitable for deformation extraction applications of natural surfaces, and serve the fields of geographical national conditions monitoring of ground subsidence and geological disaster census in my country.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a method for removing atmospheric phase and calculating deformation of a time-series InSAR according to the present invention;

FIG2 is a schematic diagram of a conditional generative adversarial neural network (CGAN) network structure according to an embodiment of the present invention;

FIG3 is a schematic diagram of a sample containing an atmospheric phase according to an embodiment of the present invention; wherein (a) in FIG3 is a real sample image containing an atmospheric phase; and (b) in FIG3 is a sample image containing an atmospheric phase generated by CGAN;

FIG4 is a schematic diagram of a TransUNet network atmospheric phase removal structure according to an embodiment of the present invention;

FIG5 is a comparison diagram of differential interferograms of an embodiment of the present invention; wherein (a) in FIG5 is an original differential interferogram containing an atmospheric phase; (b) in FIG5 is a differential interferogram after removing the atmospheric phase based on ERA5 meteorological data; and (c) in FIG5 is a differential interferogram after removing the atmospheric phase based on the TransUNet network.

FIG6 is a comparison diagram of linear deformation rates of time-series InSAR deformation solution according to an embodiment of the present invention; wherein (a) in FIG6 is the linear deformation rate of the SBAS method based on time-space domain filtering to remove the atmospheric phase; (b) in FIG6 is the linear deformation rate of the SBAS method based on ERA5 meteorological data to remove the atmospheric phase; (c) in FIG6 is the linear deformation rate of the SBAS method based on the TransUNet network to remove the atmospheric phase;

FIG. 7 is a schematic diagram of the structure of the system for time-series InSAR atmospheric phase removal and deformation resolution of the present invention.

DETAILED DESCRIPTION

The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

The method for removing atmospheric phase and calculating deformation of time-series InSAR of the present invention, as shown in FIG1 , specifically comprises the following steps:

(1) Obtain time series SAR (Synthetic Aperture Radar) image data and DEM (Digital Elevation Model) data of the monitoring area, and perform preprocessing, differential interferometry, filtering, and phase unwrapping on the data.

(1.1) Obtain the time series SAR image data of the monitoring area and extract the DEM data of the monitoring area according to the latitude and longitude range of the monitoring area.

Specifically, the monitoring area in this embodiment is located near Namtso Lake in the central part of the Tibet Autonomous Region. First, a time series SAR image data set of the monitoring area is obtained. The time series SAR image data of this embodiment is the Sentinel-1A data of the TOPS mode open source of the European Space Agency. A total of 83 scenes of IW mode VV polarization images with an orbit number of 150 (descending orbit) and a frame number of 490 are selected for time series InSAR processing, and the time coverage range is from March 16, 2017 to March 24, 2020. The specific parameters of the data are: 120km width, descending orbit mode, and the resolution in the range and azimuth directions is about 2.3m and 13.9m. At the same time, according to the latitude and longitude range of the SAR image 89.807°E~92.785°E, 29.654°N~31.776°N, a SRTM DEM data with a resolution of 30 meters is selected, and all DEM data are spliced using ENVI5.3 software to form the DEM data of the monitoring area.

(1.2) Data preprocessing: First, import the original time series SAR image data file and convert the format to generate a single-view complex data set. Then, update the orbit parameters in combination with the downloaded refined orbit data file. Finally, use the DEM data and refined orbit data file for geometric positioning and registration to obtain the registered time series SAR images.

Specifically, the acquired time series SAR image data set of the monitoring area is preprocessed, and the tiff file of the original time series SAR data is imported for format conversion to generate single-view complex data. At the same time, the orbit parameters are updated in combination with the downloaded precise track data file for subsequent DEM registration. Then, the DEM data and precise track data file are used to perform geometric positioning registration on the SAR image data. In order to achieve the registration accuracy of Sentinel-1 data in azimuth direction of one thousandth, enhanced spectral diversity (ESD) registration is also required to resample all auxiliary images to the framework of the main image, so as to obtain the registered time series SAR image.

(1.3) Differential interferometry: First, the aligned time series SAR images are interferometrically processed, and the phase difference between the main image and the auxiliary image is performed to obtain the interferogram. Then, the terrain phase and the flat ground phase of each interferogram are calculated using the DEM data according to the inverse distance weighted interpolation algorithm, and these two phases are subtracted to obtain the differential interferogram.

It should be understood that, in this embodiment, the open source software GMTSAR is used to perform differential interferometry processing on the registered time series SAR images.

(1.4) In order to suppress the influence of noise in the differential interferogram, nonlinear adaptive Goldstein spatial domain filtering is selected to filter the differential interferogram to obtain the filtered differential interferogram.

(1.5) Phase unwrapping: The filtered differential interferogram is subjected to phase unwrapping using the minimum cost flow method to obtain a phase unwrapped differential interferogram.

Specifically, the phase unwrapping process can be performed on the filtered differential interferogram using the minimum cost flow algorithm. The path difference of the microwaves in the two imagings can be obtained based on the phase value, so as to calculate the topography, landforms and slight changes on the surface of the target area, which can be used for the establishment of digital elevation models, crustal deformation detection, etc., that is, to obtain terrain elevation data from the interference fringes.

It should be understood that after preprocessing, differential interference, filtering and phase unwrapping of the time series SAR images, a series of time series differential interference patterns are formed, so a differential interference pattern data set after phase unwrapping can be finally obtained.

(2) Constructing a first sample library of differential interferograms containing atmospheric phase according to the thresholds of the time baseline and the space baseline.

，并从差分干涉图集

中挑选含有大气相位的差分干涉图集

。(2.1) Select the differential interferogram from the phase unwrapped differential interferogram according to the threshold of the time baseline and the spatial baseline to generate a differential interferogram set

, and from the differential interferogram

Select the differential interferogram containing the atmospheric phase

.

，接着从差分干涉图集

采用人工目视解译方法选择含有大气相位的差分干涉图集

。Specifically, the differential interferograms are selected based on the multiple differential interferograms finally generated in step (1.5). In this embodiment, the small baseline threshold method is used for selection, that is, the time baseline is 50 days and the space baseline is 100 meters to select the differential interferogram, so as to preliminarily determine the differential interferogram set.

, then from the differential interference atlas

Selection of differential interferograms containing atmospheric phase using manual visual interpretation

.

进行图像样本和标签制作，由于原始的差分干涉图幅宽较大，需要进行图像分块处理形成神经网络的训练样本，即对每个差分干涉图进行图像切分处理，例如本实施例中的训练样本的大小为64×64，然后对生成的图像样本进行归一化处理，最后构建含有大气相位的差分干涉图的第一样本库并记为

，作为后续条件生成对抗神经网络CGAN的输入数据集。(2.2) Construct a sample library of differential interferograms containing atmospheric phase:

Image samples and labels are prepared. Since the original differential interferogram has a large width, it is necessary to perform image segmentation to form a training sample for the neural network. That is, each differential interferogram is segmented. For example, the size of the training sample in this embodiment is 64×64. Then, the generated image samples are normalized. Finally, the first sample library containing the differential interferogram of the atmospheric phase is constructed and recorded as

, as the input data set for the subsequent conditional generation of the adversarial neural network CGAN.

(3) Based on the conditional generative adversarial neural network (CGAN), the first sample library containing the atmospheric phase is augmented and constructed to obtain a complete sample library of differential interferograms.

对条件生成对抗神经网络CGAN进行训练和测试，以生成额外的含有大气相位的第二样本库

。(3.1) The first sample library based on differential interferograms containing atmospheric phase

The conditional generative adversarial neural network (CGAN) is trained and tested to generate an additional second sample library containing atmospheric phases.

.

In order to generate more similar samples containing atmospheric phase to train subsequent neural networks, a conditional generative adversarial neural network (CGAN) is used to extract atmospheric phase features and generate samples of differential interferograms containing atmospheric phase.

输入到条件生成对抗神经网络CGAN中，先到生成器，生成器的输入为服从正态分布的随机噪声向量z与含有大气相位的特征向量

，输出为生成的伪大气相位样本

。再到判别器，判别器的输入为来自真实数据集的大气相位样本x与生成的伪大气相位样本

，判别器的作用要判别大气相位样本是否真实，输出为判别器判断大气相位样本是否真实的概率。自注意力机制可以获得图像的全局几何特征，且减小对外部信息的依赖，能够更好地捕获大气相位特征的内部相关性。同时针对GAN网络存在的训练不稳定和梯度容易消失/爆炸问题，引入Wasserstein距离代替传统JS散度，将GAN网络的损失函数写成：

In this embodiment, the network structure of the conditional generative adversarial neural network CGAN is shown in FIG2.

The input to the conditional generative adversarial neural network CGAN is first to the generator, and the input of the generator is a random noise vector z that follows a normal distribution and a feature vector containing the atmospheric phase.

, the output is the generated pseudo atmospheric phase sample

Then to the discriminator, the input of the discriminator is the atmospheric phase sample x from the real data set and the generated pseudo atmospheric phase sample

, the role of the discriminator is to determine whether the atmospheric phase sample is real, and the output is the probability of the discriminator judging whether the atmospheric phase sample is real. The self-attention mechanism can obtain the global geometric features of the image and reduce the dependence on external information, which can better capture the internal correlation of the atmospheric phase features. At the same time, in order to solve the problems of unstable training and easy gradient disappearance/explosion in the GAN network, the Wasserstein distance is introduced to replace the traditional JS divergence, and the loss function of the GAN network is written as:

是真实大气相位数据块的分布，

是由生成器模型生成伪大气相位样本（

）大气相位数据块的分布，

为

和

两个点之间沿直线的均匀样本分布，

表示真实大气相位样本x服从

分布的期望，

表示判别器判断生成器生成的大气相位样本是否真实的概率，

表示生成器模型生成大气相位样本

服从

分布的期望，

表示判别器判断真实的大气相位样本是否真实的概率，

表示生成器模型生成大气相位样本

服从

分布的期望，

表示

梯度算子，

表示2-范数。在训练过程中判别器网络和生成器网络相互竞争，以便生成器网络能够尽可能接近真实含有大气误差相位数据的分布。in,

is the distribution of the real atmospheric phase data block,

The pseudo atmospheric phase samples are generated by the generator model (

) Distribution of atmospheric phase data blocks,

for

and

A uniform distribution of samples along a straight line between two points,

Indicates that the real atmospheric phase sample x obeys

The expectation of the distribution,

represents the probability that the discriminator judges whether the atmospheric phase sample generated by the generator is real,

Represents the generator model generating atmospheric phase samples

obey

The expectation of the distribution,

represents the probability that the discriminator judges whether the real atmospheric phase sample is real,

Represents the generator model generating atmospheric phase samples

obey

The expectation of the distribution,

express

Gradient operator,

Denotes the 2-norm. During the training process, the discriminator network and the generator network compete with each other so that the generator network can be as close as possible to the distribution of the real phase data containing atmospheric errors.

，该样本库

用于条件生成对抗神经网络CGAN的训练和测试，在训练时将第一样本库

以8:2比例来划分训练集和测试集，利用浪潮服务器（NF5468M6）的PyTorch框架对条件生成对抗神经网络CGAN进行训练和测试，该服务器运行内存128GB，拥有8张Nvidia A40 40GB显卡。在训练过程中使用Adam优化器迭代地更新神经网络权重，具体参数为平滑常数

，平滑常数

，学习率

。最终生成一组额外的含有大气相位的第二样本库

。示例性地，利用条件生成对抗神经网络CGAN生成的含有大气相位样本如图3中的（b）所示，与图3中的（a）所示的真实的含有大气相位的样本图进行对比，其中，横坐标表示距离向的像素，纵坐标表示方位向的像素，均显示了像素的个数，两幅图进行对比可以表明条件生成对抗神经网络CGAN在建模含有大气相位样本的能力，且能较好地模拟出差分干涉图上大气相位的特征。The first sample library of differential interferograms containing atmospheric phase generated based on step (2.2)

, the sample library

Used for training and testing of conditional generative adversarial neural network CGAN, the first sample library is used during training

The training set and test set are divided into 8:2 ratios, and the conditional generative adversarial neural network CGAN is trained and tested using the PyTorch framework of the Inspur server (NF5468M6). The server has 128GB of running memory and 8 Nvidia A40 40GB graphics cards. During the training process, the Adam optimizer is used to iteratively update the neural network weights. The specific parameters are smoothing constants.

, smoothing constant

, learning rate

Finally, an additional second sample library containing atmospheric phases is generated.

. Exemplarily, the sample containing the atmospheric phase generated by the conditional generative adversarial neural network CGAN is shown in (b) of Figure 3, which is compared with the real sample image containing the atmospheric phase shown in (a) of Figure 3, where the horizontal axis represents the pixels in the distance direction and the vertical axis represents the pixels in the azimuth direction, both of which show the number of pixels. Comparison of the two images shows the ability of the conditional generative adversarial neural network CGAN in modeling samples containing the atmospheric phase, and can better simulate the characteristics of the atmospheric phase on the differential interferogram.

与第一样本库

合并生成含有大气相位的差分干涉图的第三样本库I，根据相位解缠后的差分干涉图选择未受到大气相位影响的差分干涉图，并进行图像切分和数据集的制作，以生成未含有大气相位的差分干涉图的第四样本库P，根据第三样本库I和第四样本库P构建差分干涉图的完整版样本库并记为

。(3.2) The second sample library

With the first sample library

The third sample library I of the differential interferogram containing the atmospheric phase is generated by merging, and the differential interferogram not affected by the atmospheric phase is selected according to the differential interferogram after phase unwrapping, and image segmentation and data set preparation are performed to generate the fourth sample library P of the differential interferogram not containing the atmospheric phase. The complete sample library of the differential interferogram is constructed according to the third sample library I and the fourth sample library P and recorded as

.

与步骤（2.2）生成的第一样本库

合并构成含有大气相位的差分干涉图的第三样本库I，以增加后续大气相位去除网络的训练数据的多样性。然后根据步骤（1.5）生成的相位解缠后的差分干涉图选择未受到大气相位影响的差分干涉图，并进行图像切分和数据集的制作，生成未含有大气相位的差分干涉图的第四样本库P，最终根据第三样本库I和第四样本库P构建差分干涉图的完整版样本库，并将其记为

，将其作为后续TransUNet网络进行大气相位去除的输入数据集。Specifically, the second sample library generated in step (3.1)

The first sample library generated in step (2.2)

The third sample library I of the differential interferogram containing the atmospheric phase is combined to increase the diversity of the training data of the subsequent atmospheric phase removal network. Then, according to the differential interferogram after phase unwrapping generated in step (1.5), the differential interferogram not affected by the atmospheric phase is selected, and image segmentation and data set preparation are performed to generate the fourth sample library P of the differential interferogram not containing the atmospheric phase. Finally, a complete sample library of the differential interferogram is constructed based on the third sample library I and the fourth sample library P, and it is recorded as

, and use it as the input dataset for the subsequent TransUNet network to remove the atmospheric phase.

包括第四样本库对应的未含有大气相位的InSAR差分干涉图样本数据集、第二样本库对应的网络生成含有大气相位的InSAR差分干涉图样本数据集和第一样本库对应的含有真实大气相位的InSAR差分干涉图样本数据集。It should be understood that the complete sample library

It includes an InSAR differential interferogram sample data set without atmospheric phase corresponding to the fourth sample library, a network-generated InSAR differential interferogram sample data set containing atmospheric phase corresponding to the second sample library, and an InSAR differential interferogram sample data set containing real atmospheric phase corresponding to the first sample library.

(4) Based on the TransUNet network, an atmospheric phase removal TransUNet network model is constructed, and the model is trained and tested to remove the atmospheric phase from all differential interferograms.

(4.1) Based on the TransUNet network, an atmospheric phase removal TransUNet network model is constructed to reconstruct the differential interferogram after the output atmospheric phase is removed.

，该网络结构的主体是TransUNet网络，它充分结合UNet和Tranformers这两种网络结构，由encoder和decoder组成U型结构。且在编码器encoder结构上运用了Transformers的encoder结构，因此可从差分干涉图数据集中学习大气相位的多尺度特征。最后添加单个残差单元和卷积层用于重构输出大气相位去除后的图像，残差单元即即输入的差分干涉图与去除大气相位后的差分干涉图之间的差值。首先输入一个含有大气相位误差的图像，如果图像为单通道，则通过repeat函数复制两次，将通道扩充为三通道。将三通道图像通过ResNetV2网络结构进行下采样，将图像编码为高级特征表示。然后创建一个feature列表将每个下采样后的特征图保存下来。最后经过ResNetV2网络结构输出。feature列表包含三个尺寸大小的特征图，分别是[B,C,112,112]，[B,4C,56,56]，[B,8C,28,28]。具体ResNetV2网络结构：先使用卷积核（7x7，s=2，p=3）的卷积操作进行root下采样为[B,C,112,112]，然后通过maxpooling层下采样为[B,C,56,56]。最后通过三个block下采样输出特征图。然后对下采样后的特征图[B,16C,14,14]进行embedding，利用卷积核将图像切成一个个patch，并加入position embedding得到输出特征图。接着利用Transformer的Encoder，Conv2d和decoder输出为[B,32,224,224]大小的图像。其中decoder为通过双线性上采样使特征图尺寸扩大一倍，然后与之前卷积后shortcut的特征图进行concat，最后通过两个Conv2d将特征图映射到低维空间。将网络输出的特征图进行分割，在TransUNet网络最后一层添加一个残差单元，即输入的干涉图与没有大气相位的干涉图的差值，最后输出为去除大气相位的样本。In this embodiment, the atmospheric phase removal structure based on the TransUNet network is shown in FIG4 . The input of the network is the complete sample library.

The main body of the network structure is the TransUNet network, which fully combines the two network structures of UNet and Tranformers, and is composed of an encoder and a decoder to form a U-shaped structure. The encoder structure of Transformers is used on the encoder encoder structure, so the multi-scale features of the atmospheric phase can be learned from the differential interferogram dataset. Finally, a single residual unit and a convolutional layer are added to reconstruct the output image after the atmospheric phase is removed. The residual unit is the difference between the input differential interferogram and the differential interferogram after the atmospheric phase is removed. First, input an image with an atmospheric phase error. If the image is a single channel, it is copied twice by the repeat function to expand the channel to three channels. The three-channel image is downsampled through the ResNetV2 network structure, and the image is encoded as a high-level feature representation. Then create a feature list to save each downsampled feature map. Finally, it is output through the ResNetV2 network structure. The feature list contains feature maps of three sizes, namely [B, C, 112, 112], [B, 4C, 56, 56], and [B, 8C, 28, 28]. The specific ResNetV2 network structure: First, use the convolution operation of the convolution kernel (7x7, s=2, p=3) to perform root downsampling to [B, C, 112, 112], and then downsample to [B, C, 56, 56] through the maxpooling layer. Finally, downsample the output feature map through three blocks. Then embed the downsampled feature map [B, 16C, 14, 14], use the convolution kernel to cut the image into patches, and add position embedding to get the output feature map. Then use the Transformer's Encoder, Conv2d and decoder to output an image of size [B, 32, 224, 224]. The decoder doubles the size of the feature map by bilinear upsampling, then concatenates it with the feature map of the shortcut after the previous convolution, and finally maps the feature map to a low-dimensional space through two Conv2d. The feature map output by the network is segmented, and a residual unit is added to the last layer of the TransUNet network, which is the difference between the input interferogram and the interferogram without the atmospheric phase, and the final output is the sample without the atmospheric phase.

作为大气相位去除TransUNet网络模型的输入数据集，按用户输入的比例将其划分为训练数据集和测试数据集，将训练数据集加载到大气相位去除TransUNet网络模型中进行训练，以获取训练好的权重参数信息，使用测试数据集加载训练好的权重参数信息，以获取去除大气相位的差分干涉图样本。(4.2) Complete sample library

As the input data set of the atmospheric phase removal TransUNet network model, it is divided into a training data set and a test data set according to the ratio input by the user. The training data set is loaded into the atmospheric phase removal TransUNet network model for training to obtain the trained weight parameter information. The trained weight parameter information is loaded using the test data set to obtain the differential interferogram samples after removing the atmospheric phase.

作为大气相位去除TransUNet网络模型的输入数据集，按用户输入的比例划分为训练数据集和测试数据集，例如可以按照8:2的比例划分为训练数据集和测试数据集，将训练数据集加载到大气相位去除TransUNet网络模型中进行训练，可以利用浪潮服务器（NF5468M6）的PyTorch框架对TransUNet网络络进行训练和测试，该服务器运行内存128GB，拥有8张Nvidia A40 40GB显卡。在训练过程中使用Adam优化器迭代地更新神经网络权重，具体参数为平滑常数

，平滑常数

，学习率

。训练以后可以得到训练好的权重参数信息，之后使用测试数据集加载预训练好的权重参数信息，从而得到去除大气相位的差分干涉图样本。Specifically, the complete sample library constructed in step (3.2)

As the input data set of the atmospheric phase removal TransUNet network model, it is divided into a training data set and a test data set according to the ratio input by the user. For example, it can be divided into a training data set and a test data set according to the ratio of 8:2. The training data set is loaded into the atmospheric phase removal TransUNet network model for training. The PyTorch framework of the Inspur server (NF5468M6) can be used to train and test the TransUNet network. The server has 128GB of running memory and 8 Nvidia A40 40GB graphics cards. During the training process, the Adam optimizer is used to iteratively update the neural network weights. The specific parameters are smoothing constants.

, smoothing constant

, learning rate

After training, the trained weight parameter information can be obtained, and then the pre-trained weight parameter information is loaded using the test data set to obtain the differential interferogram sample with the atmospheric phase removed.

(4.3) The differential interferogram samples after removing the atmospheric phase are sliced and merged to generate a differential interferogram after removing the atmospheric phase.

Specifically, the differential interferogram samples after removing the atmospheric phase generated in step (4.2) were sliced and merged to generate the differential interferogram after removing the atmospheric phase. In order to further compare the ability of removing the atmospheric phase in the differential interferogram based on the TransUNet network, the ERA5-Interim reanalysis data of the ECMWF comprehensive forecast system model was downloaded to obtain the meteorological data corresponding to the monitoring area, and the ERA5 external meteorological data was used to remove the atmospheric phase from the differential interferogram during the period of 20170917-201011. Exemplarily, a differential interferogram comparison diagram is shown in FIG5 , wherein the horizontal axis represents pixels in the distance direction, and the vertical axis represents pixels in the azimuth direction, both of which show the number of pixels. FIG5 (a) is the original differential interferogram, which contains a serious atmospheric turbulence delay error, which will have a serious impact on the subsequent time series InSAR deformation solution; FIG5 (b) is the differential interferogram after removing the atmospheric phase based on the ERA5 meteorological data. It can be seen that part of the atmospheric turbulence error phase can be removed based on the ERA5 meteorological data, but it cannot be completely removed, and the differential interferogram still has a serious atmospheric error (see the lower right corner of FIG5 (b)); FIG5 (c) is the differential interferogram after removing the atmospheric phase based on the TransUNet network. It can be seen that the atmospheric phase removal TransUNet network model constructed based on the TransUNet network can effectively remove the atmospheric phase error of the differential interferogram for subsequent time series InSAR deformation solution. This method also improves the accuracy of subsequent time series InSAR deformation solution.

(5) Based on the differential interferogram after removing the atmospheric phase, the deformation of the time-series InSAR is solved to obtain the surface deformation information of the monitoring area.

In this embodiment, the time-series InSAR deformation solution is performed based on the differential interferogram after removing the atmospheric phase, the SBAS (Small Baseline Subset) method is used to calculate the time-series deformation variable and cumulative deformation variable of the monitoring area, and the weighted least squares estimation is used to achieve a robust solution of the time-series deformation to obtain the surface deformation information of the monitoring area.

Specifically, the time-series InSAR deformation is solved according to the differential interferogram after removing the atmospheric phase in step (4.3), and the SBAS method is used to calculate the time-series deformation and cumulative deformation of the monitoring area. The solution algorithm uses weighted least squares estimation to achieve robust solution of time-series deformation, and finally obtains the surface deformation information of the monitoring area. In order to evaluate the effect of deformation solution of the SBAS method based on the TransUNet network to remove the atmospheric phase, a comparative analysis is conducted using the SBAS method based on the spatial and temporal domain filtering to remove the atmospheric phase and the SBAS method based on the ERA5 meteorological data to remove the atmospheric phase. As shown in FIG6 , a linear deformation rate diagram of the time-series InSAR deformation solution in this embodiment is shown, wherein the horizontal axis represents pixels in the range direction, and the vertical axis represents pixels in the azimuth direction, and both show the number of pixels. From the figure, it can be found that the deformation result obtained by the SBAS method based on the TransUNet network to remove the atmospheric phase is relatively smooth, as shown in (c) in FIG6 , and the deformation result does not include the local atmospheric disturbance delay affecting a specific area. The linear deformation rate diagram obtained by the SBAS method based on time-space domain filtering to remove the atmospheric phase is shown in (a) in FIG6 . No improvement is seen in the deformation result obtained by using the SBAS method based on the ERA5 meteorological data to remove the atmospheric phase, as shown in (b) in FIG6 , which is related to the low spatial resolution of the ERA5 meteorological data.

It is worth mentioning that the embodiment of the present invention further provides a system for time-series InSAR atmospheric phase removal and deformation resolution, which is used to implement the above-mentioned time-series InSAR atmospheric phase removal and deformation resolution method.

In this embodiment, the system includes a data preprocessing module, a data differential interference and filtering and phase unwrapping processing module, a differential interference pattern sample library construction and augmentation module, an atmospheric phase removal module and a time-series InSAR deformation solution module, as shown in FIG7 .

In this embodiment, the data preprocessing module is used to perform data import and image registration on the acquired time series SAR images.

In this embodiment, the data differential interference and filtering and phase unwrapping processing module is used to perform differential interference, nonlinear adaptive Goldstein spatial filtering, and minimum cost flow phase unwrapping according to the aligned time series SAR images. It should be noted that the user can automatically configure different process algorithm parameters according to needs.

In this embodiment, the differential interferogram sample library construction and augmentation module is used to generate samples of differential interferograms containing atmospheric phase, augment the atmospheric phase samples based on the conditional generative adversarial neural network CGAN, and construct the differential interferogram sample library. It should be noted that the user can select the division ratio of the training data set and the test data set according to needs, and modify the network training parameters.

In this embodiment, the atmospheric phase removal module is used to remove the atmospheric phase from the constructed differential interferogram sample library based on the TransUNet network. It should be noted that the user can select the division ratio of the training data set and the test data set according to the needs, and modify the network training parameters.

In this embodiment, the time-series InSAR deformation calculation module is used to perform time-series InSAR deformation calculation on the generated differential interferogram after removing the atmospheric phase, so as to obtain deformation information of the ground surface.

The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. A method for time-series InSAR atmospheric phase removal and deformation resolution, characterized in that it comprises the following steps: (1) Obtain time series SAR image data and digital elevation model data of the monitoring area, and perform preprocessing, differential interferometry, filtering, and phase unwrapping on the data; (2) constructing a first sample library of differential interferograms containing atmospheric phase according to the thresholds of the time baseline and the space baseline; (3) Based on the conditional generative adversarial neural network (CGAN), the first sample library containing the atmospheric phase is augmented and constructed to obtain a complete sample library of differential interferograms; (4) Based on the TransUNet network, the atmospheric phase removal TransUNet network model is constructed, and the model is trained and tested to remove the atmospheric phase from all differential interferograms; (5) Based on the differential interferogram after removing the atmospheric phase, the deformation of the time-series InSAR is solved to obtain the surface deformation information of the monitoring area. 2. The method for time-series InSAR atmospheric phase removal and deformation resolution according to claim 1, characterized in that step (1) comprises the following sub-steps: (1.1) Obtain time series SAR image data of the monitoring area and extract digital elevation model data of the monitoring area according to the latitude and longitude range of the monitoring area; (1.2) Data preprocessing: First, the original time series SAR image data file is imported and converted into a single-view complex data set. Then, the orbit parameters are updated in combination with the downloaded refined track data file. Finally, the digital elevation model data and refined track data file are used for geometric positioning registration to obtain the registered time series SAR images. (1.3) Differential interferometry: First, the time series SAR images after registration are interferometrically processed, and the phase difference between the main image and the auxiliary image is performed to obtain the interferogram. Then, the terrain phase and the flat ground phase of each interferogram are calculated using the digital elevation model data according to the inverse distance weighted interpolation algorithm, and the two phases are subtracted to obtain the differential interferogram. (1.4) In order to suppress the influence of noise in the differential interferogram, nonlinear adaptive Goldstein spatial domain filtering is selected to filter the differential interferogram to obtain a filtered differential interferogram; (1.5) Phase unwrapping: The filtered differential interferogram is subjected to phase unwrapping using the minimum cost flow method to obtain a phase unwrapped differential interferogram. 3. The method for time-series InSAR atmospheric phase removal and deformation resolution according to claim 1, wherein step (2) comprises the following sub-steps: (2.1) Select the differential interferogram from the phase unwrapped differential interferogram according to the threshold of the time baseline and the spatial baseline to generate a differential interferogram set

, and from the differential interferogram

Select the differential interferogram containing the atmospheric phase

; (2.2) Differential interference atlas

Image samples and labels are prepared, each differential interferogram is segmented to generate image samples, and the generated image samples are normalized to construct the first sample library of differential interferograms containing the atmospheric phase and recorded as

. 4. The method for time-series InSAR atmospheric phase removal and deformation resolution according to claim 1, wherein step (3) comprises the following sub-steps: (3.1) The first sample library based on differential interferograms containing atmospheric phase

The conditional generative adversarial neural network (CGAN) is trained and tested to generate an additional second sample library containing atmospheric phases.

; (3.2) The second sample library

With the first sample library

The third sample library I of the differential interferogram containing the atmospheric phase is generated by merging; the differential interferogram not affected by the atmospheric phase is selected according to the differential interferogram after phase unwrapping, and image segmentation and data set preparation are performed to generate the fourth sample library P of the differential interferogram not containing the atmospheric phase; the complete version of the sample library of the differential interferogram is constructed according to the third sample library I and the fourth sample library P and recorded as

. 5. The method for time-series InSAR atmospheric phase removal and deformation solution according to claim 4 is characterized in that the conditional generative adversarial neural network (CGAN) includes a generator and a discriminator. 6. The method for removing atmospheric phase and resolving deformation of time-series InSAR according to claim 4, characterized in that the complete version of the sample library comprises an InSAR differential interferogram sample data set not containing atmospheric phase corresponding to the fourth sample library, an InSAR differential interferogram sample data set containing atmospheric phase generated by the network corresponding to the second sample library, and an InSAR differential interferogram sample data set containing real atmospheric phase corresponding to the first sample library. 7. The method for time-series InSAR atmospheric phase removal and deformation resolution according to claim 1, wherein step (4) comprises the following sub-steps: (4.1) Based on the TransUNet network, the atmospheric phase removal TransUNet network model is constructed to reconstruct the differential interferogram after the output atmospheric phase is removed; (4.2) Complete sample library

As the input data set of the atmospheric phase removal TransUNet network model, it is divided into a training data set and a test data set according to the ratio input by the user, the training data set is loaded into the atmospheric phase removal TransUNet network model for training to obtain the trained weight parameter information, and the trained weight parameter information is loaded using the test data set to obtain the differential interferogram sample after removing the atmospheric phase; (4.3) The differential interferogram samples after removing the atmospheric phase are sliced and merged to generate a differential interferogram after removing the atmospheric phase. 8. The method for atmospheric phase removal and deformation solution of time-series InSAR according to claim 7 is characterized in that the TransUNet network combines the two network structures of UNet and Tranformers, and is composed of an encoder and a decoder to form a U-shaped structure. 9. The method for time-series InSAR atmospheric phase removal and deformation resolution according to claim 1 is characterized in that step (5) specifically comprises: performing time-series InSAR deformation resolution based on the differential interferogram after removing the atmospheric phase, using a small baseline subset method to calculate the time-series deformation variable and cumulative deformation variable of the monitoring area, and using weighted least squares estimation to achieve a robust resolution of the time-series deformation to obtain the surface deformation information of the monitoring area. 10. A system for time-series InSAR atmospheric phase removal and deformation resolution, used to implement the method for time-series InSAR atmospheric phase removal and deformation resolution according to any one of claims 1 to 9, characterized in that it comprises: Data preprocessing module, used for data import and image registration of acquired time series SAR images; The data differential interferometry, filtering and phase unwrapping processing module is used to perform differential interferometry, nonlinear adaptive Goldstein spatial filtering and minimum cost flow phase unwrapping according to the registered time series SAR images; The construction and augmentation module of the differential interferogram sample library is used to generate samples of differential interferograms containing atmospheric phase, augment the atmospheric phase samples based on the conditional generative adversarial neural network CGAN, and construct the differential interferogram sample library; An atmospheric phase removal module, used to remove the atmospheric phase from the constructed differential interferogram sample library based on the TransUNet network; and The time-series InSAR deformation calculation module is used to perform time-series InSAR deformation calculation on the differential interferogram after removing the atmospheric phase, so as to obtain the deformation information of the surface.

Images