Lightweight deep learning method and device for time sequence InSAR turbulent layer delay correction

Abstract

The invention relates to a lightweight deep learning method and device for time sequence InSAR turbulent layer delay correction, and relates to the technical field of image data processing. The method comprises the following steps: acquiring InSAR data of a target area; based on a geophysical modeling technology, constructing different types of earth surface deformation signals; simulating anisotropic atmospheric turbulence noise based on InSAR data and a geostatistical model; overlapping and fusing the atmospheric turbulence noise features and the earth surface deformation signals, and constructing a simulation time sequence InSAR training data set; then, an improved deep convolutional neural network auto-encoder structure is designed, an L1 loss function is adopted for training, a preset model is trained through training data, turbulent layer atmospheric noise in an InSAR interferogram can be effectively distinguished and eliminated, and meanwhile surface subsidence signals continuously existing in a time sequence are reserved;

Description

Technical Field

[0001] The present application relates to the technical field of image data, and particularly to a lightweight deep learning method and device for correcting the delay of the temporal InSAR turbulent layer. Background Art

[0002] InSAR (Interferometric Synthetic Aperture Radar) is an important remote sensing technology for monitoring surface deformations such as ground subsidence, earthquakes, and landslides. Currently, there are mainly three categories of methods to mitigate the impact of Atmospheric Phase Screen (APS) on InSAR. The first category relies on external auxiliary observation data. For example, the tropospheric delay correction at the regional scale is achieved by spatial interpolation using the Zenith Total Delay (ZTD) measured by GNSS stations, or the regional water vapor content distribution map is obtained by means of ground-based water vapor radiometers and satellite remote sensing sensors (such as MODIS and MERIS). These methods have high accuracy, but are limited by the insufficient coverage of ground stations and the susceptibility of optical remote sensing to cloud cover, so there are limitations in practical applications. The second category of methods is based on atmospheric numerical models and reanalysis datasets, such as ERA-I, WRF, and ECMWF, and uses the interpolated atmospheric delay data in the model for APS correction. Although this category of methods can weaken the large-scale APS, the spatial resolution is limited, and it is difficult to capture the local-scale atmospheric turbulence characteristics. In addition, even the advanced GACOS service that integrates ECMWF and GNSS data cannot fully describe the heterogeneity of the local atmosphere. The third category of methods directly performs spatio-temporal filtering on the interferogram itself to statistically estimate and eliminate the atmospheric phase. This method is sensitive to local atmospheric changes, but it is difficult to effectively distinguish the phase changes caused by turbulent noise and short-term deformation events. Summary of the Invention

[0003] The present invention provides a lightweight deep learning method and device for correcting the delay of the temporal InSAR turbulent layer, so as to at least solve the problem that the delay of the temporal InSAR turbulent layer cannot be effectively corrected. The technical solution of the present invention is as follows: According to the first aspect of the embodiments of the present invention, a lightweight deep learning method for time-series InSAR turbulence layer delay correction is provided. The method includes: obtaining InSAR data of a target area; the InSAR data includes an original SAR image, an interferogram, and a coherence map; based on geophysical modeling technology, constructing InSAR data for different types of surface deformation signals to obtain simulated InSAR observed ground settlement data, and the simulated InSAR observed ground settlement data includes surface deformation signals arranged in a time series; based on a geostatistical model, constructing anisotropic atmospheric turbulence noise characteristics in the InSAR data; superimposing and fusing the atmospheric turbulence noise characteristics and the surface deformation signals to obtain a simulated time-series InSAR training dataset; using the mean absolute error between the output data output by a preset model during training and the simulated InSAR observed ground settlement data as a loss function, and training the preset model through the simulated time-series InSAR training dataset to obtain a target model, where the target model is used to remove turbulence layer noise characteristics in different directions.

[0004] The loss function is an L1 loss function. The atmospheric turbulence noise characteristics can be characterized by an atmospheric turbulence noise delay signal.

[0005] In one implementation, based on geophysical modeling technology, constructing InSAR data for different types of surface deformation signals to obtain simulated InSAR observed ground settlement data includes: using an elastic half-space spherical point source model and a finite fault displacement model to simulate the surface deformation process of the InSAR data; and collecting surface deformation data generated by the surface deformation at each preset time sequence during the surface deformation process according to a preset time sequence; combining the surface deformation data generated by the surface deformation at each preset time sequence in the order of time sequence to obtain simulated InSAR observed ground settlement data.

[0006] In another implementation, the geostatistical model includes an anisotropic turbulence layer atmospheric delay model, and the anisotropic turbulence layer atmospheric delay model includes a semivariogram represented by the following formula ; the semivariogram includes a first-kind Bessel function and a second-kind modified Bessel function;

[0007] where represents the distance or spatial lag, that is, in the semivariogram, it measures the separation degree of two sampling points in the InSAR data in space (or time) represents the upper limit of variance, that is, after the distance between the two sampling points exceeds a preset range, the semivariogram tends to a stable value, which can also be regarded as the maximum variance of the field; represents the smoothness, The larger the value, the smoother the semivariogram; represents the correlation length or correlation distance, which controls the range over which the random field is spatially correlated; when exceeds the correlation between the fields decreases significantly; The second kind of modified Bessel function, different values will change the shape of the semivariogram to characterize spatial smoothness and local variability; is the gamma function; is the Bessel function of the first kind, which is used to describe the random field that may have a periodic structure. The larger the function value, the more likely it indicates that the random field has a period or volatility over a large range.

[0008] In another implementation, based on the geostatistical model, the anisotropic atmospheric turbulence noise characteristics in the InSAR data are constructed, including: using different preset attenuation parameter groups to sequentially assign values to the smoothness, the first coefficient of the Bessel function of the first kind, and the second coefficient of the modified Bessel function of the second kind, so as to respectively obtain the atmospheric turbulence noise characteristics in different directions characterized by the semivariogram after assignment; wherein, the preset attenuation parameter group includes a preset smoothness, a first preset coefficient, and a second preset coefficient.

[0009] In another implementation, taking the mean absolute error between the output data output by the preset model during the training process and the simulated InSAR observed ground settlement data as the loss function, the preset model is trained by the simulated time-series InSAR training dataset to obtain the target model, including: optimizing the model parameters of the preset model using the Adam optimizer according to the change trend of the mean absolute error obtained each time during training; determining the model parameters corresponding to the smallest mean absolute error during the training process as the model parameters of the target model.

[0010] In another implementation, the network structure of the preset model includes an encoder, a decoder, and a feature fusion module; the encoder extracts and compresses the spatio-temporal features of the simulated time-series InSAR training dataset through multiple convolutional and pooling layers to obtain a first feature map, and the spatio-temporal features include the change features in the spatial domain and the time domain; the decoder is used to gradually decode the spatial features of the first feature map to obtain a second feature map, and the feature dimension of the spatial features of the second feature map is higher than that of the spatial features of the first feature map; the network structure includes 12 convolutional layers and 2 pooling layers; the network structure is composed of a Squeeze-and-Excitation attention mechanism module, a depthwise separable convolution module, a feature pyramid network module, and a residual connection.

[0011] In another implementation, obtaining InSAR data of a target area includes: based on a SAR image device, obtaining geographical image data of the target area and obtaining meteorological data; the resolution of the SAR image device is higher than a preset resolution; based on a digital elevation model, processing the geographical image data and the meteorological data to obtain InSAR data.

[0012] The geographical image data and the meteorological data constitute SAR data.

[0013] According to the second aspect of the embodiments of the present invention, a lightweight deep learning device for correcting the temporal InSAR turbulence layer delay is provided. The device includes: an acquisition unit for acquiring InSAR data of a target area; the InSAR data includes an original SAR image, an interferogram, and a coherence map; a construction unit for constructing surface deformation signals of the InSAR data in different types based on geophysical modeling techniques to obtain simulated InSAR observed ground settlement data, and the simulated InSAR observed ground settlement data includes surface deformation signals arranged in a time series; the construction unit is further configured to construct anisotropic atmospheric turbulence noise characteristics in the InSAR data based on a geostatistical model; a fusion unit for superimposing and fusing the atmospheric turbulence noise characteristics and the surface deformation signals to obtain a simulated temporal InSAR training dataset; a training unit for using the mean absolute error between the output data output by a preset model during the training process and the simulated InSAR observed ground settlement data as a loss function, and training the preset model through the simulated temporal InSAR training dataset to obtain a target model, and the target model is used to remove turbulence layer noise characteristics in different directions.

[0014] In one implementation, the construction unit is specifically configured to: adopt an elastic half-space spherical point source model and a finite fault displacement model to simulate the surface deformation process of the InSAR data; and collect surface deformation data generated by surface deformation at each preset time sequence during the surface deformation process according to a preset time sequence; combine the surface deformation data generated by surface deformation at each preset time sequence in chronological order to obtain simulated InSAR observed ground settlement data.

[0015] In another implementation, constructing anisotropic atmospheric turbulence noise characteristics in the InSAR data based on a geostatistical model includes a semivariogram represented by the following formula ; the semivariogram includes a first-kind Bessel function and a second-kind modified Bessel function;

[0016] where represents the distance or spatial lag, that is, in the semivariogram, it measures the separation degree of two sampling points in the InSAR data in space (or time) Denotes the upper limit of variance, that is, after the distance between the two sampling points exceeds the preset range, the semivariogram tends to a stable value, which can also be regarded as the maximum variance of the field; Denotes smoothness, The larger the value, the smoother the semivariogram; Denotes the correlation length or correlation distance, which controls the range within which the random fields are spatially correlated; when exceeds the correlation between the fields decreases significantly; The second-kind modified Bessel function, different values will change the shape of the semivariogram to characterize spatial smoothness and local variability; is the gamma function; is the first-kind Bessel function, which is used to describe the random field that may have a periodic structure, The larger the function value, the more it indicates that the random field has a period or volatility within a larger range.

[0017] In another implementation, the building unit is specifically configured to: adopt different preset attenuation parameter groups, and sequentially assign values to the smoothness , the first coefficient of the first-kind Bessel function, and the second coefficient of the second-kind modified Bessel function, respectively, to obtain the characteristics of atmospheric turbulence noise in different directions characterized by the semivariogram after assignment; wherein, the preset attenuation parameter group includes a preset smoothness, a first preset coefficient, and a second preset coefficient.

[0018] In another implementation, the training unit is configured to: optimize the model parameters of the preset model by using the Adam optimizer according to the change trend of the mean absolute error obtained in each training; determine the model parameters corresponding to the minimum mean absolute error during the training process as the model parameters of the target model.

[0019] In another implementation, the network structure of the preset model includes an encoder, a decoder, and a feature fusion module; the encoder extracts and compresses the spatio-temporal features of the simulated temporal InSAR training dataset through multiple convolutional and pooling layers to obtain a first feature map, and the spatio-temporal features include the change features in the spatial domain and the temporal domain; the decoder is used to gradually decode the spatial features of the first feature map to obtain a second feature map, and the feature dimension of the spatial features of the second feature map is higher than the feature dimension of the spatial features of the first feature map; the network structure includes 12 convolutional layers and 2 pooling layers; the network structure is constructed by optimizing the interference network using the minimum spanning tree algorithm and is composed of a Squeeze-and-Excitation attention mechanism module, a depthwise separable convolution module, a feature pyramid network module, and a residual connection.

[0020] In another implementation, the obtaining unit is specifically configured to: obtain the geographic image data of the target area based on the SAR image device, and obtain the meteorological data; the resolution of the SAR image device is higher than the preset resolution; and process the geographic image data and the meteorological data based on the digital elevation model to obtain InSAR data.

[0021] According to the third aspect of the embodiments of the present invention, a lightweight deep learning system for temporal InSAR turbulent layer delay correction is provided. The system is configured to execute the lightweight deep learning method for temporal InSAR turbulent layer delay correction as described in the first aspect and any possible implementation manner thereof.

[0022] According to the fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided. Instructions are stored on the computer-readable storage medium. When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the lightweight deep learning method for temporal InSAR turbulent layer delay correction as described in the first aspect and any possible implementation manner thereof.

[0023] According to the fifth aspect of the embodiments of the present disclosure, a computer program product is provided. The computer program product includes computer instructions. When the computer instructions run on an electronic device, the electronic device is enabled to execute the lightweight deep learning method for temporal InSAR turbulent layer delay correction as described in the first aspect and any possible implementation manner thereof.

[0024] The technical solutions provided by the embodiments of the present application at least bring the following beneficial effects: By introducing the anisotropic variogram model, the present application obtains the spatial attenuation parameters related to different directions to simulate the variation characteristics of the atmospheric phase delay in different directions, so that the simulated temporal InSAR training data set constructed based on the anisotropic variogram model is closer to the actual observation situation, so that the atmospheric turbulence noise characteristics in the simulated temporal InSAR training data set can more accurately characterize the atmospheric delay characteristics of the target area. Therefore, the target model obtained based on the simulated temporal InSAR training data set can more effectively remove the noise generated by the anisotropic effect of the atmospheric turbulent layer in the InSAR data of the target area, so as to effectively correct the temporal InSAR turbulent layer delay, improve the InSAR measurement accuracy, and thus greatly improve the measurement accuracy of the ground deformation signal.

[0025] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings here are incorporated into the specification and form a part of this specification, showing embodiments consistent with the present disclosure, and are used together with the specification to explain the principles of the present disclosure, and do not constitute an undue limitation on the present disclosure.

[0027] Figure 1 is a flowchart of a lightweight deep learning method for temporal InSAR turbulence layer delay correction shown according to an exemplary embodiment; Figure 2 is a schematic diagram of a network structure shown according to an exemplary embodiment; Figure 3 is a schematic diagram of constructing a simulation dataset shown according to an exemplary embodiment; Figure 4 is a schematic diagram of a model training process shown according to an exemplary embodiment; Figure 5 is a comparison graph of temporal simulation InSAR data and ground subsidence area before and after denoising shown according to an exemplary embodiment; Figure 6 is a temporal InSAR image before denoising shown according to an exemplary embodiment; Figure 7 is a temporal InSAR image after denoising shown according to an exemplary embodiment; Figure 8 is a comparison graph of ground subsidence atmospheric noise before and after removal in the first oil field and the second oil field shown according to an exemplary embodiment; Figure 9 is an evaluation graph of denoising effect shown according to an exemplary embodiment; Figure 10 is a comparison graph of temporal InSAR data and leveling data before and after denoising shown according to an exemplary embodiment; Figure 11 is a comparison graph of InSAR and leveling data time series before and after denoising shown according to an exemplary embodiment; Figure 12 is a block diagram of a lightweight deep learning device for temporal InSAR turbulence layer delay correction shown according to an exemplary embodiment. Detailed implementation manners

[0028] To enable those of ordinary skill in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0029] It should be noted that the terms "first", "second", etc. in the description, claims and above-mentioned drawings of the present disclosure are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0030] Before introducing in detail the lightweight deep learning method for temporal InSAR turbulent layer delay correction provided by the embodiments of the present application, a brief introduction to the application scenarios involved in the embodiments of the present application will be given first.

[0031] InSAR (Interferometric Synthetic Aperture Radar) is an important remote sensing technology for monitoring surface deformations such as ground subsidence, earthquakes, and landslides. However, the measurement accuracy of InSAR is often affected by atmospheric disturbances, especially tropospheric turbulence effects and atmospheric phase screen (APS) effects. This atmospheric phase screen is caused by the non-uniform distribution of water vapor in the atmosphere, resulting in a phase shift of the radar signal, thereby affecting the accurate measurement of the ground deformation signal.

[0032] At present, there are mainly three categories of methods to mitigate the impact of atmospheric phase screening (APS) on InSAR. The first category relies on external auxiliary observation data. These methods have high accuracy, but are limited in practical applications due to insufficient ground station coverage and the vulnerability of optical remote sensing to cloud cover. The second category of methods is based on atmospheric numerical models and reanalysis datasets. Although these methods can weaken large-scale APS, their spatial resolution is limited, making it difficult to capture the characteristics of atmospheric turbulence at the local scale. The third category of methods directly performs spatio-temporal filtering on the interferogram itself to estimate and eliminate the atmospheric phase statistically. It has been found through research that the InSAR processing method based on the spatio-temporal filtering of the characteristics of the interferogram itself uses the statistical characteristics of the interferogram to identify and remove the influence of the atmospheric phase. Although this method can better capture local atmospheric changes and overcome the limitations of external data methods, it still fails to consider the phase changes brought about by vertical stratification delay and seasonal variations. Especially when dealing with non-linear or sudden settlement events, it is difficult for the filtering methods in related technologies to effectively distinguish between atmospheric turbulence noise and real deformation signals. In addition, most of the traditional models describing the atmosphere in the turbulent layer are based on the isotropic assumption. A major defect of related technologies is their reliance on simplified atmospheric turbulence models. These models fail to fully address the spatial heterogeneity and varying structure of atmospheric delay, resulting in a decline in performance under actual atmospheric backgrounds. Especially in regions with high humidity and large temperature gradients such as coastal and mountainous areas, the atmospheric turbulence effect is more complex, and the traditional isotropic models cannot accurately describe the atmospheric changes in these regions, limiting their application in these complex environments.

[0033] To address the above problems, this application proposes a lightweight deep learning method for temporal InSAR turbulent layer delay correction. By introducing an anisotropic variogram model, the spatial decay parameters related to different directions are obtained to simulate the variation characteristics of atmospheric phase delay in different directions, making the simulated temporal InSAR training dataset constructed based on this anisotropic variogram model closer to the actual observation situation, so that the atmospheric turbulence noise characteristics in the simulated temporal InSAR training dataset can more accurately represent the atmospheric delay characteristics of the target area. Thus, the target model obtained based on this simulated temporal InSAR training dataset can more effectively remove the noise generated by the anisotropic effect of the atmospheric turbulent layer in the InSAR data of the target area, so as to achieve effective correction of the temporal InSAR turbulent layer delay, improve the InSAR measurement accuracy, and thus greatly improve the measurement accuracy of ground deformation signals.

[0034] For the sake of easy understanding, the following specifically introduces the lightweight deep learning method for temporal InSAR turbulent layer delay correction provided by this application in combination with the accompanying drawings. This lightweight deep learning method for temporal InSAR turbulent layer delay correction is applied to the above-mentioned lightweight deep learning system for temporal InSAR turbulent layer delay correction.

[0035] Figure 1 is a flowchart of a lightweight deep learning method for temporal InSAR turbulent layer delay correction shown according to an exemplary embodiment, as Figure 1 shown, the lightweight deep learning method for temporal InSAR turbulent layer delay correction includes the following steps.

[0036] S11, Obtain InSAR data of the target area.

[0037] The InSAR data includes the original SAR image, interferogram, and coherence map; Original SAR image: Complex radar images (including amplitude and phase information) acquired at different times in the same area.

[0038] Interferogram: A phase difference map of two SAR images, recording minute surface deformations or elevation changes.

[0039] Coherence map: Reflects the correlation of two radar echo signals (value range 0 - 1), used to evaluate data reliability.

[0040] The above InSAR data (interferometric synthetic aperture radar data) can be microwave remote sensing data obtained through a synthetic aperture radar (SAR) satellite or an airborne platform, combined with interferometry technology to generate high-precision geospatial data for detecting surface deformations, elevation changes, or terrain features.

[0041] S12, Based on geophysical modeling techniques, construct surface deformation signals of InSAR data for different types to obtain simulated InSAR surface deformation data.

[0042] The simulated InSAR ground subsidence data includes surface deformation signals arranged in a time series. This simulated InSAR ground subsidence data can also be simply referred to as InSAR observation data.

[0043] In this step, based on geophysical modeling techniques, simulate the InSAR data for different types of surface deformation signals.

[0044] S13, Based on the geostatistical model, construct the characteristics of atmospheric turbulence noise in InSAR data that adapt to different directions.

[0045] In this step, based on the geostatistical model, simulate the characteristics of atmospheric turbulence noise in InSAR data that adapt to different directions.

[0046] S14, Superimpose and fuse the characteristics of atmospheric turbulence noise and surface deformation signals to obtain a simulated temporal InSAR training dataset.

[0047] S15. Using the mean absolute error between the output data of the preset model during the training process and the simulated InSAR observed ground settlement data as the loss function, the preset model is trained with the simulated time series InSAR training dataset to obtain a target model, which is used to remove the noise characteristics of the turbulent layer in different directions.

[0048] Through the above implementation, by introducing the anisotropic variogram model, the spatial attenuation parameters related to different directions are obtained to simulate the variation characteristics of the atmospheric phase delay in different directions, making the simulated time series InSAR training dataset constructed based on this anisotropic variogram model closer to the actual observation situation, so that the atmospheric turbulence noise characteristics in the simulated time series InSAR training dataset can more accurately represent the atmospheric delay characteristics of the target area. Therefore, the target model obtained based on this simulated time series InSAR training dataset can more effectively remove the noise generated by the anisotropic effect of the atmospheric turbulent layer in the InSAR data of the target area, so as to effectively correct the delay of the time series InSAR turbulent layer, improve the InSAR measurement accuracy, and thus greatly improve the measurement accuracy of the ground deformation signal.

[0049] As a refinement and extension of the specific implementation of the above embodiment, in order to fully illustrate the specific implementation process of this embodiment, the above implementation is further described through the following implementation steps.

[0050] In one implementation, based on geophysical modeling technology, simulated time series InSAR observed ground settlement data of InSAR data for different types of surface deformation signals are constructed, including: using an elastic half-space spherical point source model and a finite fault displacement model to simulate the surface deformation process of InSAR data; and collecting the surface deformation data generated by the surface deformation during each preset time series according to the preset time series; combining the surface deformation data generated by the surface deformation during each preset time series in the order of time series to obtain the simulated InSAR observed ground settlement data.

[0051] In another implementation, the geostatistical model includes a semivariogram represented by the following formula ; the semivariogram includes the first kind of Bessel function and the second kind of modified Bessel function.

[0052] (1).

[0053] Among them, represents the distance or spatial lag, that is, in the semivariogram, it measures the separation degree of two sampling points in the InSAR data in space (or time) represents the variance upper limit, that is, after the distance between the two sampling points exceeds the preset range, the semivariogram The stable value that it tends to can also be regarded as the maximum variance of the field; denotes the smoothness, The larger the value, the smoother the semivariogram; denotes the correlation length or correlation distance, which controls the range within which the random field is spatially correlated; when exceeds the correlation between fields decreases significantly; The second-kind modified Bessel function, different values will change the shape of the semivariogram to characterize the spatial smoothness and local variability; is the gamma function; is the Bessel function of the first kind, which is used to describe the random field that may have a periodic structure. The larger the function value, the more it indicates that the random field has a period or volatility within a larger range.

[0054] The above semivariogram is also called the anisotropic semivariogram function.

[0055] The anisotropic semivariogram function simulates the variation characteristics of the atmospheric phase delay in different directions through direction-dependent spatial attenuation parameters. These models can more accurately characterize the spatial distribution of atmospheric delay by setting the correlation scales and spatial variabilities in different directions, making the data simulation closer to the actual observation situation, so as to more accurately characterize the atmospheric delay characteristics in InSAR data.

[0056] In another implementation, based on the geostatistical model, the characteristics of atmospheric turbulence noise in InSAR data adapted to different directions are constructed, including: using different preset attenuation parameter groups to sequentially assign values to the smoothness , the first coefficient of the Bessel function of the first kind, and the second coefficient of the second-kind modified Bessel function, respectively, to obtain the characteristics of atmospheric turbulence noise in different directions characterized by the assigned semivariogram; among them, the preset attenuation parameter group includes a preset smoothness, a first preset coefficient, and a second preset coefficient.

[0057] In another implementation, taking the mean absolute error between the output data of the preset model and the simulated InSAR observed ground settlement data during the training process as the loss function, the preset model is trained by the simulated time-series InSAR training dataset to obtain the target model, including: optimizing the model parameters of the preset model using the Adam optimizer according to the change trend of the mean absolute error obtained each time; determining the model parameters corresponding to the minimum mean absolute error during the training process as the model parameters of the target model.

[0058] In another implementation, the network structure of the preset model includes an encoder, a decoder, and a feature fusion module; the encoder extracts and compresses the spatio-temporal features of the simulated temporal InSAR training dataset through multiple convolutional and pooling layers to obtain a first feature map, and the spatio-temporal features include the change features in the spatial domain and the temporal domain; the decoder is used to gradually decode the spatial features of the first feature map to obtain a second feature map, and the feature dimension of the spatial features of the second feature map is higher than that of the spatial features of the first feature map; the network structure includes 12 convolutional layers and 2 pooling layers; the network structure is constructed by optimizing the interference network using the minimum spanning tree algorithm and through the Squeeze-and-Excitation attention mechanism, depthwise separable convolution, feature pyramid network, and residual connection.

[0059] In another implementation, obtaining the InSAR data of the target area includes: based on the SAR image device, obtaining the geographical image data of the target area and obtaining the meteorological data; the resolution of the SAR image device is higher than the preset resolution; based on the digital elevation model, processing the geographical image data and the meteorological data to obtain the InSAR data.

[0060] As an implementation, the above implementation is specifically implemented in combination with a lightweight convolutional neural network (CNN) architecture for removing the turbulent layer delay and reconstructing the ground settlement signal in the InSAR time series data. By using an anisotropic synthetic dataset, the problems of atmospheric delay spatial heterogeneity and temporal coupling existing in the related technology models are solved. This network combines the Squeeze-and-Excitation (SE) attention mechanism, depthwise separable convolution, feature pyramid network (FPN), and residual connection, and can effectively automatically detect and extract the ground settlement signal from the noise-polluted InSAR data.

[0061] Therefore, by introducing the innovative SE attention mechanism and FPN, the feature extraction ability of the network is enhanced, and the recognition ability of the model for weak settlement signals is improved. This implementation is particularly suitable for resource-constrained scenarios and can still achieve efficient and accurate ground settlement monitoring under the conditions of small-scale training data and limited computing resources, overcoming the application bottleneck of traditional methods in large-scale monitoring tasks.

[0062] As a specific implementation, the above implementation is specifically described in the following way.

[0063] First, data collection and preprocessing. Through high-resolution SAR images and related auxiliary data (such as DEM, meteorological data, etc.), based on the DEM data and meteorological data obtained from the SAR image data and digital elevation model, collect the time-series InSAR data of the target area. Perform preprocessing operations such as cropping and interference on the original data to generate time-series InSAR ground settlement data. Second, to solve the problem of scarce data sets in the existing technology, by simulating the time-series ground settlement process under different geological conditions, construct synthetic time-series InSAR training samples with spatio-temporal diversity and considering anisotropic atmospheric turbulence noise.

[0064] Furthermore, to improve the generalization ability of the model, the synthetic data mainly consists of two parts: the synthetic surface deformation trend and the simulated atmospheric noise. The atmospheric noise usually behaves as a regionalized variable, with randomness and structure. Estimate the noise characteristics in the InSAR image through the sample semi-variogram and sample co-variogram. The covariance function is used to describe the attenuation and periodic variation of the noise with distance.

[0065] Collect the geological, topographical, and atmospheric condition data of the target area. Based on the collected data, use geophysical modeling techniques and geostatistical theory to generate simulated InSAR time-series data that can reflect the surface deformation under different atmospheric conditions. The atmospheric noise has the properties of a regionalized variable, including randomness, structure, spatial limitation, and spatial continuity, etc.

[0066] The turbulent layer signal is usually assumed to be isotropic during modeling. This simplification ignores the spatial correlation of the atmospheric phase delay in a specific direction in the interferometric synthetic aperture radar (InSAR) image, that is, the anisotropic characteristics of the phase change.

[0067] Therefore, the present application uses the method of formula (1) to simulate anisotropic atmospheric delay. In InSAR data processing, the delay of the atmospheric turbulence layer is usually simplified to an isotropic model, that is, it is assumed that the atmospheric phase delay is uniformly distributed in all directions. However, the actual observation results show that the atmospheric delay is not isotropic, but is affected by various factors such as water vapor transport, wind field dynamics, terrain effects, and large-scale weather systems, showing complex anisotropic characteristics in space. This anisotropic characteristic is particularly evident in coastal areas, mountainous areas, and high-humidity climate regions. For example, the thermal difference in the land-sea boundary region leads to a significant trend in the direction of water vapor transport, making the change rate of the atmospheric phase delay in a specific direction significantly higher than that in other directions. In addition, the complex terrain conditions in mountainous areas result in uneven distribution of the turbulence characteristics and water vapor content of the wind field in different directions, so that the atmospheric delay shows significant direction dependence on the spatial scale. This phenomenon is particularly obvious in areas with drastic changes in weather systems or large topographic undulations, directly affecting the phase stability of InSAR time series data.

[0068] The isotropic assumption in the related methods ignores these directional effects, which may lead to a decrease in the accuracy of InSAR deformation monitoring. Especially when performing multi-temporal InSAR (TS-InSAR) analysis or cross-track data fusion, the accumulation of errors may be further aggravated. The anisotropic effect of the atmospheric turbulence layer will make the phase correlation in different directions significantly different, resulting in interference in the extraction of the surface deformation signal, and thus affecting the reliability of InSAR measurement. Therefore, accurately characterizing the anisotropic characteristics of the atmospheric turbulence layer delay is crucial for improving the processing accuracy of InSAR data.

[0069] To achieve this goal, this embodiment introduces an anisotropic variogram, and simulates the change characteristics of the atmospheric phase delay in different directions through direction-related spatial attenuation parameters. These models can more accurately characterize the spatial distribution of atmospheric delay by setting the correlation scales and spatial variabilities in different directions, making the data simulation closer to the actual observation situation. To more accurately characterize the atmospheric delay characteristics in InSAR data, it can be described by formula (1).

[0070] Thirdly, a lightweight convolutional neural network structure is proposed, which combines the SE attention mechanism, residual network, depthwise separable convolution, and feature pyramid network (FPN) to effectively capture the spatio-temporal features in InSAR time series data. Under limited training data and computing resources, automatically identify and extract ground settlement signals and remove atmospheric turbulence noise.

[0071] Such as Figure 2As shown in the figure, the networks of the preset model and the target model are based on the deep convolutional neural network (CNN) architecture, and incorporate advanced deep learning techniques such as the SE channel attention mechanism, residual structure, depthwise separable convolution, feature pyramid network, and feature pyramid network (FPN) fusion. The network structure includes an encoder, a decoder, and a feature fusion module. The encoder can efficiently extract and compress the spatio-temporal features in the temporal InSAR data through multiple convolutional operations and pooling layers; the decoder restores the original spatial dimension of the feature map by gradually decoding, so as to accurately reconstruct the surface settlement signal. The entire network has 12 convolutional layers, including 2 pooling layers. The pooling operation effectively eliminates the time dimension in the temporal images, so as to more intensively process the spatial features.

[0072] Fourthly, use the constructed InSAR time series dataset to train the designed deep learning network. Adopt a supervised learning method based on a synthetic dataset and use the L1 loss function to iteratively optimize the performance of the network.

[0073] The above dataset is a synthetic time series InSAR ground settlement dataset.

[0074] Collect data on the geology, terrain, and atmospheric conditions of the target area to provide appropriate background information for simulation. Based on the collected data, combined with geophysical modeling techniques and geostatistical theory, construct a simulated InSAR time series dataset to reflect surface deformation and turbulent layer noise under different atmospheric conditions. Atmospheric noise has the properties of regionalized variables, including characteristics such as randomness, structuralness, spatial limitation, and spatial continuity. Introduce a geostatistical model to simulate the anisotropic characteristics of the atmospheric turbulent layer.

[0075] By setting the spatial attenuation parameters in different directions, the variation characteristics of the atmospheric turbulent layer noise in different directions can be simulated, avoiding the simplified assumptions of traditional isotropic models and accurately capturing the complex directional dependence presented by atmospheric delay in space. Use the modified Bessel function and the Bessel function of the first kind to construct the characteristics of atmospheric turbulent noise adapted to different directions, so as to better simulate the anisotropy of atmospheric delay and improve the accuracy of the synthetic sample dataset. At the same time, considering that the true surface settlement signal is usually not directly available, use the elastic half-space spherical point source model and the finite fault displacement model to simulate the surface deformation process. By randomly setting the deformation time, different types of surface deformation patterns are simulated to generate time-series surface deformation samples that conform to the actual observation data. Finally, superimpose the simulated turbulent noise and the surface deformation signal to construct a synthetic InSAR training sample with spatio-temporal diversity and reflecting the characteristics of anisotropic atmospheric turbulent noise for subsequent model training and optimization. The dataset production process is as shown in Figure 3 the atmospheric noise map shown in (a) in Figure 3The time-series ground settlement map shown in (b) and Figure 3 The simulated time-series ground settlement image map with noise shown in (c). Fifth, the trained model is verified using simulated data and real data. By comparing with traditional methods, the performance of the model in removing atmospheric turbulence noise and reconstructing the ground settlement signal is evaluated. Statistical evaluation results show that the model has shown significant performance improvement in indicators such as standard deviation and variogram analysis.

[0076] Sixth, the deformation information of the real InSAR data before and after denoising is compared and verified with the leveling data.

[0077] Seventh, to verify the generalization ability and practical applicability of the above embodiments, real InSAR time-series data in the Yellow River Delta region is selected as a test case.

[0078] Optionally, the training design of the preset model is implemented under the PyTorch framework.

[0079] The specific training design process includes: Based on the prepared dataset, model training is carried out on an operation platform with an RTX4090 graphics card with 24GB of video processing unit (GPU) memory, an Intel i9-13900K CPU, and 128GB of running memory; the input data of the network is 9 consecutive time-series InSAR images, and the image size in the training dataset is uniformly 192×192 pixels. To improve the training stability and convergence efficiency of the network, all input data is normalized before entering the network to eliminate the scale differences between different images. The network captures the change features in both the spatial domain and the time domain through a three-dimensional convolutional kernel (2×3×3), thereby effectively learning spatio-temporal information.

[0080] The training of the network uses the L1 (MAE) loss function (MAE loss MAE (Mean Absolute Error, average absolute error) is usually also called L1-Loss, which measures the difference between the predicted value and the real value by taking the average of the absolute differences), aiming to accurately capture the absolute difference between the prediction result and the actual ground settlement signal and reduce the negative impact of extreme errors on network training. During the training process, the Adam optimizer is used to accelerate the convergence and optimization of network parameters to ensure stable and efficient updates during training. The initial learning rate is set to 0.0001 and dynamically adjusted through an exponential decay strategy to further improve the training efficiency. During the training phase, both the training loss and the test loss of the model show a rapid downward trend (such as Figure 4The network training set and validation set loss curves shown in (a) therein tend to stabilize within 50 epochs, indicating that the model can effectively generalize to the test data. In terms of evaluation metrics, the structural similarity index (SSIM) of the model increases from the initial value of 0.68 to 0.97, and the peak signal-to-noise ratio (PSNR) increases from 45 dB to 60 dB (as shown in Figure 4 the average SSIM and average PSNR graphs of the training set and validation set shown in (b) therein), verifying the superior performance of the model in InSAR image denoising and surface deformation reconstruction.

[0081] To verify the effectiveness of the trained model on the simulated data, simulated time-series InSAR data was used for verification. As shown in Figure 5 in (a) the target image and in (b) the output image of the model in Figure 5, the simulation results at three time steps are respectively shown, including the noisy InSAR image (input), the target noise-free image (target), the model output image, and the difference image between the target and the output image. The results show that the model can effectively remove noise and reconstruct the surface subsidence signal. The output image is highly consistent with the target image, the contour of the subsidence area is clear, and the center of the subsidence area is almost exactly the same as the target position. The profile analysis shows that the deformation amount of the model output and the target image in the subsidence area is almost exactly the same, successfully recovering the maximum deformation amount at the subsidence center and maintaining the smooth transition of the deformation. Compared with the noisy image, the model greatly reduces the error caused by turbulent noise. In the statistical analysis of the subsidence area, the difference in the number of subsidence pixels output by the model and the target image is less than 2.7%, and compared with the noisy image, the model significantly reduces the non-real subsidence area. The change in the time-series subsidence value at the center point of the subsidence area shows that the result after the model denoising is highly consistent with the true value of the target image, verifying the superior performance of the model in removing atmospheric noise and accurately reconstructing the subsidence signal.

[0082] To verify the effectiveness of the deep learning denoising method proposed in this application in practical applications, real-time InSAR data of the Liaohe Plain was used for verification. A complex geological and climatic area was selected as the target area for the implementation of the method. This target area can be an area rich in underground resources. For example, this target area can be some important oil-producing areas. Specifically, this target area can include the First Oilfield and the Second Oilfield at the same time. Long-term oil and gas extraction activities in this area have led to the extraction of underground fluids and the compaction of strata, thereby triggering significant surface subsidence. In the time-series InSAR deformation images after denoising using the deep learning method, the research period is from April 2022 to January 2024. Compared with the original automated processing results ( Figure 6 ), the atmospheric noise in the denoised image is almost completely eliminated, and the deformation signal shows a highly clear and stable spatial distribution ( Figure 7). In the denoised image, the background area becomes significantly uniform, and the deformation characteristics of the subsidence area are more distinct.

[0083] Figure 8 (a) in Figure 8 and (b) in Figure 8 respectively show the distribution of land subsidence in the Liaohe Plain before and after denoising. To further verify the effectiveness of the method,

[0084] (c) in

[0085] shows the comparison of the profile distributions before and after denoising. This profile A - A' passes through the subsidence areas of Shuguang Oilfield and Huanxiling Oilfield for comparative analysis. The profile line before denoising (gray) shows strong high - frequency oscillation characteristics. In contrast, the profile line after denoising (orange) is significantly smoother, with almost all noise eliminated, the true subsidence signal being accurately extracted, the boundary of the subsidence area being clearly visible, and the subsidence values being more accurate. When the profile A - A' passes through the subsidence area of Shuguang Oilfield, the maximum subsidence value is approximately - 160 mm, and the subsidence center position and boundary shape are clearly presented; the subsidence amplitude in the area of Huanxiling Oilfield is smaller, approximately - 80 mm, but the deformation trend is also clear. The denoising process significantly improves the deformation reconstruction effect of the subsidence area, verifying the performance of the implementation mode of this application in removing noise and accurately extracting the surface subsidence signal. Figure 9 (a) in

[0086] Figure 9 shows the comparison of the standard deviations of the last 10 frames of InSAR images before and after denoising. The standard deviation before denoising is significantly higher than that after denoising, indicating that there is significant noise and pixel value fluctuation in the original image. After denoising, the standard deviation is significantly reduced, proving that the denoising process effectively reduces the influence of noise and improves the smoothness and consistency of the image.

[0087] Figure 9In (c), the semivariogram plot is shown to compare the spatial variability of the true image before denoising and the denoised image. The denoised image maintains lower semivariance values across the entire distance range, indicating that the model effectively reduces spatial variability and removes most of the noise. By comparing the two curves, the model performs well in denoising, successfully reducing the spatial variability in the image and retaining the main structural features of the image. The reduced variability of the denoised image further indicates that the image is smoother and more consistent, with the noise impact significantly weakened, thus improving the quality of the data.

[0088] To verify the effectiveness of the denoising method proposed in this embodiment in practical applications, this embodiment compares the total deformation distribution of InSAR data and leveling data before and after denoising during the period from September 2022 to September 2023. As Figure 10 shown, through comparison, it is found that there are significant deviations between the un-denoised InSAR data and the leveling data at some stations, especially at stations with large settlement amplitudes (such as P05 and P08). In the background areas with small deformation values (such as P10 to P12), the un-denoised InSAR data is affected by atmospheric noise and shows irregular fluctuations of 20 - 42 mm.

[0089] Furthermore, as Figure 11 shown, the deformation of the time-series InSAR data and the leveling data before and after denoising is analyzed through time-evolution comparison. The box plot shows the deformation error between the data before and after denoising and the leveling data. The results show that there are significant deviations in the un-denoised InSAR data at most stations, especially during time periods or at stations with drastic changes in settlement, where the interference of atmospheric noise is strong. The coincidence degree between the denoised InSAR data and the leveling data is significantly improved, and the overall trend and deformation values are closer to the true deformation.

[0090] Finally, to verify the generalization ability of the method in this paper, this embodiment selects the real InSAR time-series data of the target area for testing. Using the trained model, it is directly applied to the real InSAR data of the target area without additional fine-tuning or re-training, demonstrating the high robustness of this embodiment.

[0091] To implement the above functions, the lightweight deep learning device for time-series InSAR turbulent layer delay correction includes the corresponding hardware structures and / or software modules for performing each function. Those skilled in the art should easily realize that, in combination with the algorithm steps of each example described in the embodiments disclosed herein, this application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a certain function is executed in the way of hardware or computer software driving the hardware depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0092] The embodiments of the present disclosure also provide a lightweight deep learning device for time-series InSAR turbulent layer delay correction as shown in Figure 12 which includes: an acquisition unit 151, a construction unit 152, a fusion unit 153, and a training unit 154.

[0093] The acquisition unit 151 is used to acquire InSAR data of the target area; the InSAR data includes the original SAR image, the interferogram, and the coherence map.

[0094] The construction unit 152 is used to construct different types of surface deformation signals of the InSAR data based on geophysical modeling technology to obtain simulated time-series InSAR observed ground settlement data, and the simulated time-series InSAR observed ground settlement data includes surface deformation signals arranged in a time series.

[0095] The construction unit 152 is also used to construct the characteristics of atmospheric turbulence noise in the InSAR data that adapt to different directions based on the geostatistical model.

[0096] The fusion unit 153 is used to superimpose and fuse the characteristics of atmospheric turbulence noise and the surface deformation signals to obtain a simulated time-series InSAR training dataset.

[0097] The training unit 154 is used to use the mean absolute error between the output data output by the preset model during the training process and the simulated InSAR observed ground settlement data as the loss function, and train the preset model through the simulated time-series InSAR training dataset to obtain a target model, and the target model is used to remove the characteristics of turbulent layer noise in different directions.

[0098] In one embodiment, the construction unit is specifically configured to: simulate the surface deformation process of InSAR data by using an elastic half-space spherical point source model and a finite fault displacement model; collect surface deformation data generated by surface deformation at each preset time sequence during the surface deformation process according to a preset time sequence; and combine the surface deformation data generated by surface deformation at each preset time sequence in the order of time sequence to obtain simulated InSAR observed ground settlement data.

[0099] In another embodiment, the geostatistical model includes a semivariogram represented by the following formula ; the semivariogram includes a first kind of Bessel function and a second kind of modified Bessel function;

[0100] where represents the distance or spatial lag, that is, in the semivariogram, it measures the separation degree of two sampling points in the InSAR data in space (or time) represents the upper limit of variance, that is, after the distance between the two sampling points exceeds a preset range, the semivariogram tends to a stable value, which can also be regarded as the maximum variance of the field; represents the smoothness, the larger the value, the smoother the semivariogram; represents the correlation length or correlation distance, which controls the range of spatial correlation of the random field; when exceeds the second kind of modified Bessel function, different values will change the shape of the semivariogram to characterize the spatial smoothness and local variability; is the gamma function; is the first kind of Bessel function, which is used to describe the random field that may have a periodic structure, the larger the function value, the more the random field has a period or volatility in a larger range.

[0101] In another implementation manner, the construction unit 152 is specifically configured to: use different preset attenuation parameter groups to sequentially assign values to the smoothness , the first coefficient of the first kind of Bessel function and the second coefficient of the second kind of modified Bessel function, so as to respectively obtain the characteristics of atmospheric turbulence noise in different directions represented by the assigned semivariogram; where the preset attenuation parameter group includes a preset smoothness, a first preset coefficient and a second preset coefficient.

[0102] In another implementation manner, the training unit 154 is configured to: optimize the model parameters of a preset model by using an Adam optimizer according to the change trend of the average absolute error obtained in each training; and determine the model parameters corresponding to the minimum average absolute error during the training process as the model parameters of the target model.

[0103] In another implementation manner, the network structure of the preset model includes an encoder, a decoder, and a feature fusion module; the encoder extracts and compresses the spatio-temporal features of the simulated temporal InSAR training dataset through multiple convolutional and pooling layers to obtain a first feature map, and the spatio-temporal features include the change features in the spatial domain and the temporal domain; the decoder is configured to gradually decode the spatial features of the first feature map to obtain a second feature map, and the feature dimension of the spatial features of the second feature map is higher than that of the spatial features of the first feature map; the network structure includes 12 convolutional layers and 2 pooling layers; the network structure is constructed by optimizing the interference network by using a minimum spanning tree algorithm and through a Squeeze-and-Excitation attention mechanism, depthwise separable convolution, a feature pyramid network, and a residual connection.

[0104] In another implementation manner, the obtaining unit 151 is specifically configured to: obtain the geographical image data of a target area and obtain meteorological data based on a SAR image device; the resolution of the SAR image device is higher than a preset resolution; and process the geographical image data and the meteorological data based on a digital elevation model to obtain InSAR data.

[0105] Regarding the device in the above embodiments, the specific manners in which each unit module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0106] The embodiment of the present application further provides a computer-readable storage medium. When the instructions in the computer-readable storage medium are executed by a processor of a lightweight deep learning device for temporal InSAR turbulence layer delay correction or a lightweight deep learning device for temporal InSAR turbulence layer delay correction, the lightweight deep learning device for temporal InSAR turbulence layer delay correction or the lightweight deep learning device for temporal InSAR turbulence layer delay correction can execute the lightweight deep learning method for temporal InSAR turbulence layer delay correction in any of the above possible implementation manners. And the same technical effects can be achieved. To avoid repetition, it will not be elaborated here.

[0107] The embodiment of the present application further provides a computer program product, including a computer program or instructions, and the computer program or instructions are executed by a processor to perform the lightweight deep learning method for temporal InSAR turbulence layer delay correction in any of the above possible implementation manners. And the same technical effects can be achieved. To avoid repetition, it will not be elaborated here.

[0108] Those skilled in the art will readily conceive of other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any variations, uses, or adaptations of the present application, which follow the general principles of the present application and include the common general knowledge or conventional technical means in the technical field not disclosed in the present application. The specification and examples are only regarded as exemplary, and the true scope and spirit of the present application are pointed out by the following claims.

[0109] It should be understood that the present application is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present application is only limited by the appended claims.

Claims

1. A lightweight deep learning method for time-series InSAR turbulent layer delay correction, characterized in that, The method includes: Obtaining InSAR data of a target area; the InSAR data includes an original SAR image, an interferogram, and a coherence map; Based on geophysical modeling techniques, constructing different types of surface deformation signals in the InSAR data to obtain simulated InSAR observed ground settlement data, where the simulated InSAR observed ground settlement data includes time-series surface deformation signals; Based on a geostatistical model, constructing the anisotropic atmospheric turbulence noise characteristics in the InSAR data; Superposing and fusing the atmospheric turbulence noise characteristics and the surface deformation signals to construct a simulated time-series InSAR training dataset; Using the mean absolute error between the output data output by a preset model during training and the simulated InSAR observed ground settlement data as a loss function, training the preset model through the simulated InSAR training dataset to obtain a target model, where the target model is used to remove the turbulence layer noise characteristics; The constructing, based on geophysical modeling techniques, different types of surface deformation signals in the InSAR data to obtain simulated InSAR observed ground settlement data includes: Adopting an elastic half-space spherical point source model and a finite fault displacement model to simulate the surface deformation process of the InSAR data; and collecting surface deformation data generated by surface deformation during each of the preset time sequences according to a preset time sequence; Combining the surface deformation data generated by surface deformation during each of the preset time sequences in chronological order to obtain the simulated InSAR observed ground settlement data.

2. The method according to claim 1, wherein The geostatistical model includes an anisotropic turbulent layer atmospheric delay model, and the anisotropic turbulent layer atmospheric delay model includes a semivariogram represented by the following formula ; the semivariogram includes a first-kind Bessel function and a second-kind modified Bessel function; Among them, represents the distance or spatial lag, that is, in the semivariogram, it measures the separation degree of two sampling points in the InSAR data in space (or time). represents the upper limit of variance, that is, after the distance between the two sampling points exceeds the preset range, the semivariogram tends to a stable value, which can also be regarded as the maximum variance of the field; represents the smoothness, the larger the value, the smoother the semivariogram; represents the correlation length or correlation distance, which controls the range of spatial correlation of the random field; when exceeds the correlation between fields decreases significantly; the second kind of modified Bessel function, different values will change the shape of the semivariogram to characterize the spatial smoothness and local variability; is the gamma function; is the Bessel function of the first kind, which is used to describe the random field that may have a periodic structure. The larger the function value, the more likely it indicates that the random field has a period or volatility in a larger range.

3. The method according to claim 2, wherein The constructing, based on a geostatistical model, the anisotropic atmospheric turbulence noise characteristics in the InSAR data includes: Using different preset attenuation parameter groups to sequentially assign values to the smoothness, the first coefficient of the first kind of Bessel function, and the second coefficient of the second kind of modified Bessel function, respectively, to obtain the atmospheric turbulence noise characteristics in different directions characterized by the variogram after assignment respectively; Wherein, the preset attenuation parameter group includes a preset smoothness, a first preset coefficient, and a second preset coefficient.

4. The method according to any one of claims 1 to 3, characterized in that, The using the mean absolute error between the output data output by a preset model during training and the simulated InSAR observed ground settlement data as a loss function, training the preset model through the simulated time-series InSAR training dataset to obtain a target model includes: According to the change trend of the mean absolute error obtained by each training, using an Adam optimizer to optimize the model parameters of the preset model; Determining the model parameters corresponding to the minimum mean absolute error during training as the model parameters of the target model.

5. The method according to claim 3, wherein The network structure of the preset model includes an encoder, a decoder, and a feature fusion module; the encoder extracts and compresses the spatio-temporal features of the simulated temporal InSAR training dataset through multiple convolutional and pooling layers to obtain a first feature map, where the spatio-temporal features include the change features in the spatial domain and the temporal domain; the decoder is used to gradually decode the spatial features of the first feature map to obtain a second feature map, and the feature dimension of the spatial features of the second feature map is higher than that of the spatial features of the first feature map; the network structure includes 12 convolutional layers and 2 pooling layers; the network structure is composed of a Squeeze-and-Excitation attention mechanism module, a depthwise separable convolution module, a feature pyramid network module, and a residual connection.

6. The method according to any one of claims 1 to 3, characterized in that The obtaining of the InSAR data of the target area includes: Based on the SAR imaging device, obtaining the SAR data of the target area, where the resolution of the SAR imaging device is higher than the preset resolution; Based on the digital elevation model, processing the SAR data to obtain the InSAR data.

7. A lightweight deep learning device for time-series InSAR turbulent layer delay correction, characterized in that, The device includes: An obtaining unit, configured to obtain the InSAR data of the target area; the InSAR data includes the original SAR image, the interferogram, and the coherence map; A constructing unit, configured to construct the surface deformation signals of different types of the InSAR data based on the geophysical modeling technology to obtain the simulated InSAR observed ground settlement data, where the simulated InSAR observed ground settlement data includes the surface deformation signals arranged in a time series; The constructing unit is further configured to construct the anisotropic atmospheric turbulence noise characteristics in the InSAR data based on the geostatistical model; A fusion unit is configured to superimpose and fuse the atmospheric turbulence noise characteristics and the surface deformation signals to construct a simulated temporal InSAR training dataset; A training unit is configured to use the mean absolute error between the output data output by the preset model during the training process and the simulated InSAR observed ground settlement data as a loss function, and train the preset model through the simulated temporal InSAR training dataset to obtain a target model, where the target model is used to remove the turbulence layer noise characteristics in different directions; Specifically, the constructing unit is configured to use an elastic half-space spherical point source model and a finite fault displacement model to simulate the surface deformation process of the InSAR data; and collect the surface deformation data generated by the surface deformation during each of the preset time series according to the preset time series; Combining the surface deformation data generated by the surface deformation during each of the preset time series in the order of time series to obtain the simulated InSAR observed ground settlement data.

8. The device according to claim 7, characterized in that, Specifically, the constructing unit is configured to: Use an elastic half-space spherical point source model and a finite fault displacement model to simulate the surface deformation process of the InSAR data; and collect the surface deformation data generated by the surface deformation during each of the preset time series according to the preset time series; Combine the ground deformation data generated by the ground deformation in each of the preset time sequences in chronological order to obtain the simulated time sequence InSAR training data.

9. The device according to claim 7, characterized in that, The geostatistical model includes an anisotropic turbulent layer atmospheric delay model, and the anisotropic turbulent layer atmospheric delay model includes a semivariogram represented by the following formula ; the semivariogram includes a first-kind Bessel function and a second-kind modified Bessel function; Wherein: represents the distance or spatial lag, that is, in the semivariogram, it measures the separation degree of two sampling points in the InSAR data in space (or time); represents the upper limit of variance, that is, after the distance between the two sampling points exceeds a preset range, the semivariogram; tends to a stable value, which can also be regarded as the maximum variance of the field; represents the smoothness; the larger the value, the smoother the semivariogram; represents the correlation length or correlation distance, which controls the range of spatial correlation of the random field; when exceeds the correlation between fields is significantly reduced; the second kind of modified Bessel function, different values will change the shape of the semivariogram to characterize the spatial smoothness and local variability; is the gamma function; is the first kind of Bessel function, which is used to describe the random field that may have a periodic structure, the larger the function value, it indicates that the random field has a period or volatility in a larger range.

Images