A lightweight deep learning method and device for time-series InSAR turbulence layer delay correction

Abstract

This application relates to a lightweight deep learning method and device for time-series InSAR turbulence layer delay correction, and relates to the field of image data processing technology. The method includes: acquiring InSAR data of the target area; constructing different types of surface deformation signals based on geophysical modeling technology; simulating anisotropic atmospheric turbulence noise based on InSAR data and geostatistical models; superimposing and fusing the atmospheric turbulence noise characteristics and surface deformation signals to construct a simulated time-series InSAR training data set; then, designing an improved deep convolutional neural network autoencoder structure, using the L1 loss function for training, and training the preset model with the training data, which can effectively distinguish and eliminate the turbulence layer atmospheric noise in the InSAR interferogram while retaining the surface deposition signal that persists in the time series.

Description

Technical Field

[0001] The present application relates to the field of image data technology, and in particular to a lightweight deep learning method and device for time-series InSAR turbulence layer delay correction. Background Art

[0002] Interferometric synthetic aperture radar (InSAR) is an important remote sensing technology for monitoring surface deformations such as land subsidence, earthquakes, and landslides. Currently, there are three main approaches to mitigate the effects of atmospheric phase shielding (APS) on InSAR. The first relies on external auxiliary observations. For example, these methods use the zenith total delay (ZTD) measured by GNSS stations to spatially interpolate regional-scale tropospheric delay corrections, or utilize ground-based water vapor radiometers and satellite remote sensing sensors (such as MODIS and MERIS) to obtain regional water vapor distribution maps. These methods offer high accuracy, but their practical application is limited by insufficient ground station coverage and the susceptibility of optical remote sensing to cloud cover. The second approach relies on atmospheric numerical models and reanalysis datasets, such as ERA-I, WRF, and ECMWF, to perform APS correction using atmospheric delay data interpolated from the models. While these methods can mitigate large-scale APS, their limited spatial resolution makes it difficult to capture local-scale atmospheric turbulence characteristics. Furthermore, even advanced GACOS services, which integrate ECMWF and GNSS data, cannot fully describe local atmospheric heterogeneity. The third approach uses spatiotemporal filtering directly on the interferogram itself, statistically estimating and removing the atmospheric phase. This approach is sensitive to local atmospheric variations but struggles to distinguish between 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 time-series InSAR turbulence layer delay correction, which at least solves the problem of inability to effectively correct time-series InSAR turbulence layer delay. The technical solution of the present invention is as follows:

[0004] According to a first aspect of an embodiment of the present invention, a lightweight deep learning method for time-series InSAR turbulence layer delay correction is provided, the method comprising: acquiring InSAR data of a target area; the InSAR data comprising original SAR images, interferograms, and coherence maps; constructing surface deformation signals of the InSAR data in different types based on geophysical modeling technology to obtain simulated InSAR observed ground subsidence data, wherein the simulated InSAR observed ground subsidence data comprises surface deformation signals arranged in a time series; constructing anisotropic atmospheric turbulence noise features in the InSAR data based on a geostatistical model; superimposing and fusing the atmospheric turbulence noise features and the surface deformation signals to obtain a simulated time-series InSAR training dataset; and training the preset model using the simulated time-series InSAR training dataset using the mean absolute value error between output data output by the preset model during the training process and the simulated InSAR observed ground subsidence data as a loss function to obtain a target model, wherein the target model is used to remove turbulence layer noise features in different directions.

[0005] The loss function is an L1 loss function. The atmospheric turbulence noise characteristics can be represented by the atmospheric turbulence noise delay signal.

[0006] In one implementation, based on geophysical modeling technology, InSAR data signals for different types of surface deformation are constructed to obtain simulated InSAR observation ground subsidence data. This includes: using an elastic half-space spherical point source model and a finite fault displacement model to simulate the surface deformation process of InSAR data; collecting surface deformation data generated by surface deformation in each preset time sequence during the surface deformation process according to a preset time sequence; and combining the surface deformation data generated by surface deformation in each preset time sequence according to the time sequence to obtain simulated InSAR observation ground subsidence data.

[0007] In another implementation, the geostatistical model includes an anisotropic turbulent layer atmospheric delay model, wherein the anisotropic turbulent layer atmospheric delay model includes a semivariogram expressed as follows: ; The semivariogram includes the first-kind Bessel function and the second-kind modified Bessel function;

[0008]

[0009] in, Represents distance or spatial lag, that is, in the semivariogram, it measures the degree of separation in space (or time) between two sampling points in the InSAR data. Indicates the upper limit of variance, that is, when 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; Indicates smoothness, The larger the value, the smoother the semivariogram; represents the correlation length or correlation distance, which controls the range of spatial correlation of random fields; when Exceed When , the correlation between fields is significantly reduced; The second kind of modified Bessel function, different The value changes the shape of the semivariogram to characterize spatial smoothness and local variability; is the gamma function; is the first kind of Bessel function, which is used to describe random fields that may have periodic structures. The larger the function value, the greater the period or volatility of the random field.

[0010] In another implementation, based on a geostatistical model, the anisotropic atmospheric turbulence noise characteristics in the InSAR data are constructed, including: using different preset attenuation parameter groups to 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, so as to respectively correspond to the atmospheric turbulence noise characteristics in different directions represented by the assigned semi-variogram; wherein the preset attenuation parameter group includes a preset smoothness, a first preset coefficient, and a second preset coefficient.

[0011] In another implementation method, the mean absolute error between the output data of the preset model during the training process and the simulated InSAR observation ground subsidence data is used as the loss function, and the preset model is trained by simulating the time series InSAR training data set to obtain the target model, including: optimizing the model parameters of the preset model using the Adam optimizer according to the changing trend of the mean absolute error obtained in each training; and determining the model parameters corresponding to the smallest mean absolute error during the training process as the model parameters of the target model.

[0012] In another implementation method, the network structure of the preset model includes an encoder, a decoder and a feature fusion module; the encoder extracts and compresses the spatiotemporal features of the simulated time series InSAR training data set through multiple layers of convolution and pooling layers to obtain a first feature map, and the spatiotemporal features include 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 the feature dimension of the spatial features of the first feature map; the network structure includes 12 convolution layers and 2 pooling layers; the network structure is composed of a Squeeze-and-Excitation attention mechanism module, a depth-separable convolution module, a feature pyramid network module and a residual connection.

[0013] In another implementation, obtaining InSAR data of a target area includes: obtaining geographic image data of the target area and meteorological data based on a SAR imaging device; the resolution of the SAR imaging device is higher than a preset resolution; and processing the geographic image data and meteorological data based on a digital elevation model to obtain InSAR data.

[0014] Geographic image data and meteorological data constitute SAR data.

[0015] According to a second aspect of an embodiment of the present invention, a lightweight deep learning device for time-series InSAR turbulence layer delay correction is provided, the device comprising: an acquisition unit for acquiring InSAR data of a target area; the InSAR data comprising original SAR images, interferograms, and coherence maps; a construction unit for constructing, based on geophysical modeling technology, surface deformation signals of the InSAR data in different types to obtain simulated InSAR observed ground subsidence data, the simulated InSAR observed ground subsidence data comprising surface deformation signals arranged in a time series; the construction unit is further configured to construct, based on a geostatistical model, atmospheric turbulence noise features adapted to different directions in the InSAR data; a fusion unit is configured to superimpose and fuse the atmospheric turbulence noise features and the surface deformation signals to obtain a simulated time-series InSAR training data set; and a training unit is configured to train the preset model using the simulated time-series InSAR training data set using the mean absolute value error between output data output by the preset model during the training process and the simulated InSAR observed ground subsidence data as a loss function to obtain a target model, the target model being configured to remove turbulence layer noise features in different directions.

[0016] In one implementation, the construction unit is specifically used to: use an elastic half-space spherical point source model and a finite fault displacement model to simulate the surface deformation process of InSAR data; and collect surface deformation data generated by surface deformation in 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 in each preset time sequence according to the time sequence order to obtain simulated InSAR observation ground subsidence data.

[0017] In another implementation, based on a geostatistical model, the anisotropic atmospheric turbulence noise characteristics in the InSAR data are constructed, including a semivariogram expressed as follows: ; The semivariogram includes the first-kind Bessel function and the second-kind modified Bessel function;

[0018]

[0019] in, Represents distance or spatial lag, that is, in the semivariogram, it measures the degree of separation in space (or time) between two sampling points in the InSAR data. Indicates the upper limit of variance, that is, when 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; Indicates smoothness, The larger the value, the smoother the semivariogram; represents the correlation length or correlation distance, which controls the range of spatial correlation of random fields; when Exceed When , the correlation between fields is significantly reduced; The second kind of modified Bessel function, different The value changes the shape of the semivariogram to characterize spatial smoothness and local variability; is the gamma function; is the first kind of Bessel function, which is used to describe random fields that may have periodic structures. The larger the function value, the greater the period or volatility of the random field.

[0020] In another implementation, the construction unit is specifically used to: use different preset attenuation parameter groups to adjust the smoothness , the first coefficient of the first kind of Bessel function and the second coefficient of the second kind of modified Bessel function are assigned to respectively correspond to the atmospheric turbulence noise characteristics in different directions represented by the assigned semi-variogram function; wherein the preset attenuation parameter group includes a preset smoothness, a first preset coefficient and a second preset coefficient.

[0021] In another implementation, the training unit is used to: optimize the model parameters of the preset model using the Adam optimizer according to the changing trend of the mean absolute error obtained in each training; and determine the model parameters corresponding to the smallest mean absolute error during the training process as the model parameters of the target model.

[0022] 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 spatiotemporal features of the simulated time series InSAR training data set through multiple layers of convolution and pooling layers to obtain a first feature map, and the spatiotemporal features include 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 the feature dimension of the spatial features of the first feature map; the network structure includes 12 convolution 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 depth-separable convolution module, a feature pyramid network module and a residual connection.

[0023] In another implementation, the acquisition unit is specifically used to: acquire geographic image data of the target area and meteorological data based on SAR imaging equipment; the resolution of the SAR imaging equipment is higher than the preset resolution; based on the digital elevation model, the geographic image data and meteorological data are processed to obtain InSAR data.

[0024] According to a third aspect of an embodiment of the present invention, a lightweight deep learning system for time-series InSAR turbulence layer delay correction is provided, which is configured to execute a lightweight deep learning method for time-series InSAR turbulence layer delay correction as described in the first aspect and any possible implementation thereof.

[0025] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which instructions are stored. When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform a lightweight deep learning method for time-series InSAR turbulence layer delay correction as described in the first aspect and any possible implementation thereof.

[0026] According to a fifth aspect of an embodiment of the present disclosure, a computer program product is provided, which includes computer instructions. When the computer instructions are executed on an electronic device, the electronic device executes the lightweight deep learning method for time-series InSAR turbulence layer delay correction of the above-mentioned first aspect and any possible implementation thereof.

[0027] The technical solution provided by the embodiments of the present application brings at least the following beneficial effects: the present application introduces an anisotropic variogram model to obtain spatial attenuation parameters related to different directions to simulate the changing characteristics of the atmospheric phase delay in different directions, so that the simulated time-series 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 time-series InSAR training data set more accurately characterize the atmospheric delay characteristics of the target area, and thus the target model obtained based on the simulated time-series InSAR training data set can more effectively remove the noise generated by the anisotropic effect of the atmospheric turbulence layer in the InSAR data of the target area, so as to achieve effective correction of the time-series InSAR turbulence layer delay, improve the InSAR measurement accuracy, and thus greatly improve the measurement accuracy of the ground deformation signal.

[0028] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute an improper limitation of the present disclosure.

[0030] Figure 1 This is a flowchart of a lightweight deep learning method for time-series InSAR turbulence layer delay correction according to an exemplary embodiment;

[0031] Figure 2 is a schematic diagram of a network structure according to an exemplary embodiment;

[0032] Figure 3 is a schematic diagram showing a method of constructing a simulation data set according to an exemplary embodiment;

[0033] Figure 4 is a schematic diagram of a model training process according to an exemplary embodiment;

[0034] Figure 5 A plot showing a time series simulation of InSAR data before and after denoising and the area of ​​a ground subsidence area according to an exemplary embodiment;

[0035] Figure 6 1 is a time-series InSAR image before denoising according to an exemplary embodiment;

[0036] Figure 7 1 is a denoised time-series InSAR image according to an exemplary embodiment;

[0037] Figure 8 3. This is a comparison diagram showing a first oil field and a second oil field before and after the removal of ground subsidence atmospheric noise according to an exemplary embodiment;

[0038] Figure 9 is a denoising effect evaluation diagram shown according to an exemplary embodiment;

[0039] Figure 10 is a comparison diagram of time-series InSAR data and leveling data before and after denoising according to an exemplary embodiment;

[0040] Figure 11 is a time series comparison diagram of InSAR and leveling data before and after denoising according to an exemplary embodiment;

[0041] Figure 12 This is a block diagram of a lightweight deep learning device for time-series InSAR turbulence layer delay correction according to an exemplary embodiment. DETAILED DESCRIPTION

[0042] In order to enable ordinary people 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.

[0043] It should be noted that the terms "first," "second," and the like in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the numbers used in this manner are interchangeable where appropriate 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. Instead, they are merely examples of apparatus and methods consistent with certain aspects of the present disclosure as detailed in the appended claims.

[0044] Before introducing in detail the lightweight deep learning method for time-series InSAR turbulence 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 is first given.

[0045] InSAR (Interferometric Synthetic Aperture Radar) is an important remote sensing technology for monitoring surface deformations such as ground subsidence, earthquakes, and landslides. However, InSAR measurement accuracy is often affected by atmospheric disturbances, particularly tropospheric turbulence and the atmospheric phase screen (APS). This APS, caused by the uneven distribution of water vapor in the atmosphere, shifts the radar signal's phase, thereby affecting the precise measurement of ground deformation signals.

[0046] Currently, there are three main approaches to mitigate the effects of atmospheric phase shielding (APS) on InSAR. The first relies on external auxiliary observation data. While these methods offer high accuracy, they are limited in practical application due to insufficient ground station coverage and the susceptibility of optical remote sensing to cloud cover. The second approach relies on atmospheric numerical models and reanalysis datasets. While these methods can mitigate large-scale APS, their spatial resolution is limited, making it difficult to capture local-scale atmospheric turbulence characteristics. The third approach directly applies spatiotemporal filtering to the interferogram itself, statistically estimating and removing the atmospheric phase. Research has found that InSAR processing methods that apply spatiotemporal filtering based on the interferogram's inherent characteristics exploit the interferogram's statistical properties to identify and remove the effects of the atmospheric phase. While this approach can better capture local atmospheric variations and overcome the limitations of external data methods, it still fails to account for phase shifts caused by vertical stratification delays and seasonal variations. Particularly when dealing with nonlinear or sudden precipitation events, these filtering methods struggle to effectively distinguish atmospheric turbulence noise from true deformation signals. Furthermore, traditional models describing the turbulent layer are mostly based on the isotropic assumption. A major drawback of these techniques is their reliance on simplified atmospheric turbulence models. These models fail to adequately address the spatial heterogeneity and varying structure of atmospheric delays, resulting in reduced performance in realistic atmospheric environments. This is particularly true in coastal and mountainous areas, where atmospheric turbulence effects are more complex. Traditional isotropic models are unable to accurately describe atmospheric variations in these regions, limiting their application in these complex environments.

[0047] To address the above problems, this application proposes a lightweight deep learning method for time-series InSAR turbulence layer delay correction. By introducing an anisotropic variogram model, spatial attenuation parameters related to different directions are acquired to simulate the changing characteristics of atmospheric phase delay in different directions, so that the simulated time-series InSAR training dataset constructed based on the anisotropic variogram model is closer to the actual observation situation, so that the atmospheric turbulence noise characteristics in the simulated time-series InSAR training dataset can more accurately characterize the atmospheric delay characteristics of the target area, so that the target model obtained based on the simulated time-series InSAR training dataset can more effectively remove the noise generated by the anisotropic effect of the atmospheric turbulence layer in the InSAR data of the target area, so as to achieve effective correction of the time-series InSAR turbulence layer delay, improve the InSAR measurement accuracy, and thus greatly improve the measurement accuracy of the ground deformation signal.

[0048] For ease of understanding, the lightweight deep learning method for time-series InSAR turbulence layer delay correction provided by this application is specifically introduced below with reference to the accompanying drawings. This lightweight deep learning method for time-series InSAR turbulence layer delay correction is applied to the lightweight deep learning system for time-series InSAR turbulence layer delay correction described above.

[0049] Figure 1 FIG. 1 is a flow chart of a lightweight deep learning method for time-series InSAR turbulence layer delay correction according to an exemplary embodiment. Figure 1 As shown in FIG, the lightweight deep learning method for time-series InSAR turbulence layer delay correction includes the following steps.

[0050] S11, acquiring InSAR data of the target area.

[0051] InSAR data include raw SAR images, interferograms, and coherence maps;

[0052] Raw SAR image: a complex radar image (including amplitude and phase information) acquired at different times over the same area.

[0053] Interferogram: A phase difference diagram between two SAR images, which records small surface deformations or elevation changes.

[0054] Coherence diagram: reflects the correlation between two radar echo signals (value range 0~1), used to evaluate data reliability.

[0055] The above-mentioned InSAR data (interferometric synthetic aperture radar data) can be microwave remote sensing data obtained by synthetic aperture radar (SAR) satellites or aerial platforms, combined with interferometric measurement technology to generate high-precision geospatial data for detecting surface deformation, elevation changes or terrain features.

[0056] S12, based on geophysical modeling technology, constructs InSAR data on different types of surface deformation signals to obtain simulated InSAR surface deformation data.

[0057] The simulated InSAR observation ground subsidence data includes surface deformation signals arranged in a time series. The simulated InSAR observation ground subsidence data may also be referred to as InSAR observation data.

[0058] In this step, based on geophysical modeling technology, the InSAR data are simulated for different types of surface deformation signals.

[0059] S13, based on geostatistical models, construct atmospheric turbulence noise characteristics adapted to different directions in InSAR data.

[0060] In this step, the atmospheric turbulence noise characteristics adapted to different directions in the InSAR data are simulated based on the geostatistical model.

[0061] S14, superimpose and fuse the atmospheric turbulence noise characteristics and the surface deformation signal to obtain a simulated time series InSAR training data set.

[0062] S15, using the mean absolute error between the output data of the preset model output during the training process and the simulated InSAR observation ground subsidence data as the loss function, the preset model is trained by simulating the time series InSAR training data set to obtain a target model, which is used to remove the turbulent layer noise characteristics in different directions.

[0063] Through the above implementation, by introducing the anisotropic variogram model, the spatial attenuation parameters related to different directions are obtained to simulate the changing characteristics of the atmospheric phase delay in different directions, so that the simulated time-series 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 time-series InSAR training data set more accurately characterize the atmospheric delay characteristics of the target area, and thus the target model obtained based on the simulated time-series InSAR training data set can more effectively remove the noise generated by the anisotropic effect of the atmospheric turbulence layer in the InSAR data of the target area, so as to achieve effective correction of the time-series InSAR turbulence layer delay, improve the InSAR measurement accuracy, and thus greatly improve the measurement accuracy of the ground deformation signal.

[0064] As a refinement and expansion 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 explained through the following implementation steps.

[0065] In one embodiment, based on geophysical modeling technology, InSAR data is constructed in different types of surface deformation signals to obtain simulated time series InSAR observation ground subsidence data, 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 surface deformation data generated by surface deformation in each preset time series during the surface deformation process according to a preset time series; and combining the surface deformation data generated by surface deformation in each preset time series according to the time series order to obtain simulated InSAR observation ground subsidence data.

[0066] In another embodiment, the geostatistical model includes a semivariogram expressed as follows ; The semivariogram includes the first-type Bessel function and the second-type modified Bessel function.

[0067] (1).

[0068] in, Represents distance or spatial lag, that is, in the semivariogram, it measures the degree of separation in space (or time) between two sampling points in the InSAR data. Indicates the upper limit of variance, that is, when 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; Indicates smoothness, The larger the value, the smoother the semivariogram; represents the correlation length or correlation distance, which controls the range of spatial correlation of random fields; when Exceed When , the correlation between fields is significantly reduced; The second kind of modified Bessel function, different The value changes the shape of the semivariogram to characterize spatial smoothness and local variability; is the gamma function; is the first kind of Bessel function, which is used to describe random fields that may have periodic structures. The larger the function value, the greater the period or volatility of the random field.

[0069] The above semivariogram function is also called anisotropic semivariogram function.

[0070] Anisotropic variograms simulate the directional variations of atmospheric phase delay using directionally dependent spatial attenuation parameters. By specifying the relative scales and spatial variability in different directions, these models can more accurately depict the spatial distribution of atmospheric delay, making data simulations closer to actual observations and more accurately characterizing the atmospheric delay characteristics in InSAR data.

[0071] In another embodiment, based on the geostatistical model, the atmospheric turbulence noise characteristics adapted to different directions in the InSAR data are constructed, including: using different preset attenuation parameter groups, respectively adjusting the smoothness , the first coefficient of the first kind of Bessel function and the second coefficient of the second kind of modified Bessel function are assigned to respectively correspond to the atmospheric turbulence noise characteristics in different directions represented by the assigned semi-variogram function; wherein the preset attenuation parameter group includes a preset smoothness, a first preset coefficient and a second preset coefficient.

[0072] In another embodiment, the mean absolute error between the output data of the preset model output during the training process and the simulated InSAR observed ground subsidence data is used as the loss function, and the preset model is trained by simulating a time-series InSAR training data set to obtain a target model, including: optimizing the model parameters of the preset model using the Adam optimizer according to the changing trend of the mean absolute error obtained in each training; and determining the model parameters corresponding to the smallest mean absolute error during the training process as the model parameters of the target model.

[0073] In another embodiment, the network structure of the preset model includes an encoder, a decoder and a feature fusion module; the encoder extracts and compresses the spatiotemporal features of the simulated time series InSAR training data set through multiple layers of convolution and pooling layers to obtain a first feature map, and the spatiotemporal features include 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 the feature dimension of the spatial features of the first feature map; the network structure includes 12 convolution 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, depth-separable convolution, feature pyramid network and residual connection.

[0074] In another embodiment, obtaining InSAR data of a target area includes: obtaining geographic image data of the target area and meteorological data based on a SAR imaging device; the resolution of the SAR imaging device is higher than a preset resolution; and processing the geographic image data and meteorological data based on a digital elevation model to obtain InSAR data.

[0075] As an implementation, the aforementioned implementation is specifically implemented with a lightweight convolutional neural network (CNN) architecture to remove turbulence layer delay and reconstruct ground subsidence signals from InSAR time series data. By using an anisotropic synthetic dataset, the spatial heterogeneity and temporal coupling of atmospheric delays, which are common in related technical models, are addressed. The proposed network combines a Squeeze-and-Excitation (SE) attention mechanism, depthwise separable convolutions, a Feature Pyramid Network (FPN), and residual connections to effectively and automatically detect and extract ground subsidence signals from noisy InSAR data.

[0076] Therefore, by introducing the innovative SE attention mechanism and FPN, the network's feature extraction capabilities are enhanced, and the model's ability to identify weak subsidence signals is improved. This implementation is particularly suitable for resource-constrained scenarios. It can achieve efficient and accurate ground subsidence monitoring even with small-scale training data and limited computing resources, overcoming the application bottleneck of traditional methods in large-scale monitoring tasks.

[0077] As a specific implementation method, the above implementation method is specifically described in the following manner.

[0078] First, data acquisition and preprocessing. Using high-resolution SAR imagery and related auxiliary data (such as DEM and meteorological data), time-series InSAR data of the target area is collected based on DEM and meteorological data derived from the SAR imagery and digital elevation models. Preprocessing operations such as cropping and interferometry are performed on the raw data to generate time-series InSAR land subsidence data. Second, to address the limited datasets in existing technologies, synthetic time-series InSAR training samples with spatiotemporal diversity and anisotropic atmospheric turbulence noise are constructed by simulating time-series land subsidence processes under different geological conditions.

[0079] Furthermore, to improve the generalization capabilities of the model, the synthetic data consists of two main components: synthetic surface deformation trends and simulated atmospheric noise. Atmospheric noise typically manifests as regionalized variation, with both randomness and structure. The noise characteristics in InSAR imagery are estimated using the sample semivariogram and sample covariogram. The covariance function describes the attenuation and periodicity of noise with distance.

[0080] Collect data on the target area's geology, topography, and atmospheric conditions. Based on this data, utilize geophysical modeling techniques and geostatistical theory to generate simulated InSAR time series data that reflect surface deformation under varying atmospheric conditions. Atmospheric noise exhibits regionalized variable properties, including randomness, structure, spatial localization, and spatial continuity.

[0081] Turbulent layer signals are usually assumed to be isotropic when modeling. This simplification ignores the spatial correlation of atmospheric phase delay in specific directions in interferometric synthetic aperture radar (InSAR) images, that is, the anisotropic characteristics of phase changes.

[0082] Therefore, this application uses formula (1) to simulate anisotropic atmospheric delay. In InSAR data processing, the atmospheric turbulent layer delay is usually simplified to an isotropic model, that is, it is assumed that the atmospheric phase delay is uniformly distributed in all directions. However, actual observation results show that the atmospheric delay is not isotropic, but is affected by multiple factors such as water vapor transport, wind field dynamics, terrain effects and large-scale weather systems, and exhibits complex anisotropic characteristics in space. This anisotropic characteristic is particularly evident in coastal areas, mountainous areas and high-humidity climate areas. For example, the thermal difference in the sea-land boundary area causes the direction of water vapor transport to have a significant trend, making the rate of change of the atmospheric phase delay in a specific direction significantly higher than in other directions. In addition, the complex terrain conditions in mountainous areas lead to uneven distribution of turbulent characteristics and water vapor content of the wind field in different directions, so that the atmospheric delay exhibits significant directional dependence on the spatial scale. This phenomenon is particularly obvious in areas with drastic changes in weather systems or large terrain undulations, which directly affects the phase stability of InSAR time series data.

[0083] The isotropy assumption in correlation methods ignores these directional effects, potentially reducing the accuracy of InSAR deformation monitoring. This is especially true when performing multi-time series InSAR (TS-InSAR) analysis or cross-track data fusion, where the accumulation of errors can be exacerbated. The anisotropic effects of the atmospheric turbulence layer can significantly vary the phase correlations in different directions, interfering with the extraction of surface deformation signals and, in turn, affecting the reliability of InSAR measurements. Therefore, accurately characterizing the anisotropic characteristics of the atmospheric turbulence layer delay is crucial for improving the processing accuracy of InSAR data.

[0084] To achieve this goal, this implementation introduces anisotropic variograms, which simulate the directional variations of atmospheric phase delay using directionally dependent spatial attenuation parameters. These models can more accurately characterize the spatial distribution of atmospheric delay by setting the relevant scales and spatial variability in different directions, making the data simulation closer to actual observations. This allows for a more accurate characterization of the atmospheric delay characteristics in InSAR data, which can be described as Equation (1).

[0085] Third, a lightweight convolutional neural network architecture is proposed, combining the SE attention mechanism, residual networks, depthwise separable convolutions, and feature pyramid networks (FPNs) to effectively capture the spatial-temporal features in InSAR time series data. With limited training data and computing resources, it can automatically identify and extract ground subsidence signals and remove atmospheric turbulence noise.

[0086] like Figure 2As shown, the networks of the preset and target models are based on a deep convolutional neural network (CNN) architecture, incorporating advanced deep learning techniques such as the SE channel attention mechanism, residual structure, depthwise separable convolution, and feature pyramid networks (FPNs). The network structure comprises an encoder, a decoder, and a feature fusion module. The encoder efficiently extracts and compresses spatial-temporal features from time-series InSAR data through multiple layers of convolution and pooling operations. The decoder, through progressive decoding, restores the original spatial dimensions of the feature maps, enabling accurate reconstruction of the surface subsidence signal. The entire network consists of 12 convolutional layers, including two pooling layers. The pooling operation effectively eliminates the temporal dimension in the time-series imagery, allowing for more focused processing of spatial features.

[0087] Fourth, the designed deep learning network was trained using the constructed InSAR time series dataset. A supervised learning method based on the synthetic dataset was used, and the L1 loss function was used to iteratively optimize the network's performance.

[0088] The above dataset is a synthetic time series InSAR land subsidence dataset.

[0089] Data on the target area's geology, topography, and atmospheric conditions are collected to provide appropriate background information for the simulation. Based on this data, combined with geophysical modeling techniques and geostatistical theory, a simulated InSAR time series dataset is constructed to reflect surface deformation and turbulent layer noise under different atmospheric conditions. Atmospheric noise exhibits regionalized variable properties, including randomness, structure, spatial localization, and spatial continuity. Geostatistical models are introduced to simulate the anisotropic characteristics of the atmospheric turbulent layer.

[0090] By setting the spatial attenuation parameters in different directions, the changing characteristics of the atmospheric turbulent layer noise in different directions can be simulated, avoiding the simplified assumptions of the traditional isotropic model and accurately capturing the complex directional dependence of the atmospheric delay in space. Using the modified Bessel function and the first-kind Bessel function, atmospheric turbulence noise characteristics that adapt to different directions are constructed to better simulate the anisotropy of atmospheric delay and improve the accuracy of the synthetic sample data set. At the same time, considering that the real surface sedimentation signal is usually not directly accessible, the elastic half-space spherical point source model and the finite fault displacement model are used 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, the simulated turbulent noise is superimposed on the surface deformation signal to construct a synthetic InSAR training sample with temporal and spatial diversity that can reflect the anisotropic atmospheric turbulence noise characteristics for subsequent model training and optimization. The dataset production process is as follows. Figure 3 The atmospheric noise diagram shown in (a) Figure 3The time series of sedimentation diagrams shown in (b) and Figure 3 (c) shows a simulated time-series of ground subsidence images with noise. Fifth, the trained model was validated using both simulated and real-world data. The model's performance in removing atmospheric turbulence noise and reconstructing ground subsidence signals was evaluated by comparison with traditional methods. Statistical evaluation results showed significant improvements in metrics such as standard deviation and variogram analysis.

[0091] Sixth, the deformation information of the real InSAR data before and after denoising is compared and verified with the leveling data.

[0092] Seventh, in order to verify the generalization ability and practical applicability of the above implementation method, real InSAR time series data from the Yellow River Delta region were selected as test cases.

[0093] Optionally, implement the training design of the preset model under the PyTorch framework.

[0094] The specific training design process includes: Based on a pre-prepared dataset, model training was performed on a computing platform equipped with an RTX4090 graphics card with 24GB of GPU memory, an Intel i9-13900K CPU, and 128GB of RAM. The network input data consisted of nine consecutive time-series InSAR images, with the images in the training dataset uniformly sized at 192×192 pixels. To improve training stability and convergence efficiency, all input data was normalized before entering the network to eliminate scale differences between images. The network used a 3D convolution kernel (2×3×3) to simultaneously capture spatial and temporal variations, effectively learning spatiotemporal information.

[0095] The network is trained using the L1 (MAE) loss function (MAE loss MAE (Mean Absolute Error) is often also called L1-Loss, which measures the difference between the predicted value and the true value by taking the average of the absolute difference between them), aiming to accurately capture the absolute difference between the predicted result and the actual surface subsidence 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 the training process. The initial learning rate is set to 0.0001 and is dynamically adjusted through an exponential decay strategy to further improve training efficiency. During the training phase, both the training loss and the test loss of the model showed a rapid downward trend (such as Figure 4The network training set and validation set loss curves shown in (a) in the figure are stable within 50 epochs, indicating that the model can effectively generalize to the test data. In terms of evaluation indicators, the structural similarity index (SSIM) of the model increased from the initial value of 0.68 to 0.97, and the peak signal-to-noise ratio (PSNR) increased from 45 dB to 60 dB (as shown in Figure 2). Figure 4 The average SSIM and average PSNR of the training set and validation set shown in (b) in the figure verify the superior performance of the model in InSAR image denoising and surface deformation reconstruction.

[0096] In order to verify the effectiveness of the trained model on the simulation data, the simulated time series InSAR data is used for verification. Figure 5 (a) shows the target image and (b) shows the output image of the model. The simulation results of three time steps are shown, including the noisy InSAR image (input), the noise-free image of the target (target), the model output image, and the difference map 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 outline of the subsidence area is clear, and the center of the subsidence area is almost completely consistent with the target position. The profile analysis shows that the deformation of the model output and the target image in the subsidence area is almost completely consistent. The maximum deformation of the subsidence center is successfully restored and the smooth transition of the deformation is maintained. 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 between 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 time series of the subsidence value changes at the center of the subsidence area show that the denoised results of the model are highly consistent with the true values ​​of the target image, verifying the superior performance of the model in removing atmospheric noise and accurately reconstructing the subsidence signal.

[0097] In order to verify the effectiveness of the deep learning denoising method proposed in this application in practical applications, the real time series InSAR data of the Liaohe Plain were used for verification. A complex geological and climatic area was selected as the target area for the implementation of the method. The target area may be an area containing rich underground resources, such as some important oil producing areas. Specifically, the target area may include the first oil field and the second oil field at the same time. The long-term oil and gas production activities in this area have led to the extraction of underground fluids and the compaction of the strata, which in turn caused significant surface subsidence. In the time series InSAR deformation images after denoising using the deep learning method, the study 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 7In the denoised image, the background area becomes significantly more uniform, and the deformation characteristics of the subsidence area are more clear.

[0098] Figure 8 (a) and Figure 8 (b) in the figure shows the distribution of land subsidence in Liaohe Plain before and after denoising. To further verify the effectiveness of the method, Figure 8 (c) in the figure shows the comparison of profile distribution before and after denoising. The 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, the noise is almost completely eliminated, the true subsidence signal is accurately extracted, the boundary of the subsidence area is clearly visible, and the subsidence value is more accurate. When profile A-A' passes through the subsidence area of ​​Shuguang Oilfield, the maximum subsidence value is about -160 mm, and the subsidence center position and boundary shape are clearly presented; the subsidence amplitude in the Huanxiling Oilfield area is smaller, about -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 embodiment of the present application in removing noise and accurately extracting surface subsidence signals.

[0099] Based on this, the denoising effect is evaluated.

[0100] like Figure 9 Panel (a) shows a comparison of the standard deviations of the last 10 InSAR images before and after denoising. The standard deviation before denoising is significantly higher than after denoising, indicating significant noise and pixel value fluctuations in the original image. After denoising, the standard deviation is significantly reduced, demonstrating that denoising effectively reduces the impact of noise and improves image smoothness and consistency.

[0101] Figure 9 Panel (b) shows the average sliding window standard deviation at different window sizes, used to analyze the variation in standard deviation of the image at different spatial scales before and after denoising. The results show that the standard deviation after denoising is significantly lower than the standard deviation before denoising at all window sizes, further demonstrating the effective noise suppression effect of denoising at multiple scales, ensuring smoothness and consistency of the image at both local and global scales.

[0102] Figure 9(c) in Figure 1 shows the semivariogram, comparing the spatial variability of the real image before and after denoising. The denoised image maintains a low semivariogram across the entire distance range, indicating that the model effectively reduces spatial variability and removes most of the noise. Comparing the two curves, the model performs well in denoising, successfully reducing spatial variability in the image while preserving the image's key structural features. The reduced variability in the denoised image further demonstrates that the image is smoother and more consistent, significantly reducing the impact of noise, and thus improving data quality.

[0103] In order to verify the effectiveness of the denoising method proposed in this embodiment in practical applications, this embodiment compares the total shape variable distribution of InSAR data and leveling data before and after denoising from September 2022 to September 2023. Figure 10 As shown in the figure, a comparison reveals significant deviations between the undenoised InSAR data and the levelling data at some stations, particularly at stations with large subsidence amplitudes (e.g., P05 and P08). In background areas with smaller deformation (e.g., P10 to P12), the undenoised InSAR data exhibit irregular fluctuations of 20-42 mm due to atmospheric noise.

[0104] Further, if Figure 11 As shown in the figure, the deformation of the time series InSAR data and leveling data before and after denoising is compared and analyzed over time. Box plots show the deformation errors between the denoised and leveling data. The results show that the undenoised InSAR data exhibit significant deviations at most stations, especially during periods or stations with significant subsidence fluctuations, where atmospheric noise interference is strong. However, the denoised InSAR data significantly improves the agreement with the leveling data, with the overall trend and deformation values ​​closer to the actual deformation.

[0105] Finally, to validate the generalization capabilities of our method, this implementation was tested using real InSAR time series data from the target region. The trained model was directly applied to the real InSAR data from the target region without additional fine-tuning or retraining, demonstrating the robustness of this implementation.

[0106] In order to achieve the above functions, the lightweight deep learning device for time-series InSAR turbulence layer delay correction includes hardware structures and / or software modules that perform the corresponding functions. Those skilled in the art should easily realize that, in combination with the algorithm steps of each example described in the embodiments disclosed herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0107] The present disclosure also provides a Figure 12 The lightweight deep learning device for time-series InSAR turbulence layer delay correction shown includes: an acquisition unit 151, a construction unit 152, a fusion unit 153 and a training unit 154.

[0108] The acquisition unit 151 is used to acquire InSAR data of the target area; the InSAR data includes original SAR images, interference patterns and coherence patterns.

[0109] The construction unit 152 is used to construct InSAR data on different types of surface deformation signals based on geophysical modeling technology to obtain simulated time series simulated InSAR observed ground subsidence data, where the simulated time series simulated InSAR observed ground subsidence data includes surface deformation signals arranged in a time series.

[0110] The construction unit 152 is further configured to construct atmospheric turbulence noise features adapted to different directions in the InSAR data based on a geostatistical model.

[0111] The fusion unit 153 is used to superimpose and fuse the atmospheric turbulence noise characteristics and the surface deformation signal to obtain a simulated time series InSAR training data set.

[0112] The training unit 154 is used to train the preset model by simulating the time series InSAR training data set, using the mean absolute error between the output data of the preset model output during the training process and the simulated InSAR observed ground subsidence data as a loss function, to obtain a target model. The target model is used to remove the turbulent layer noise characteristics in different directions.

[0113] In one embodiment, the construction unit is specifically used to: use an elastic half-space spherical point source model and a finite fault displacement model to simulate the surface deformation process of InSAR data; and collect surface deformation data generated by surface deformation in 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 in each preset time sequence according to the time sequence to obtain simulated InSAR observed ground subsidence data.

[0114] In another embodiment, the geostatistical model includes a semivariogram represented by the formula ; The semivariogram includes the first-kind Bessel function and the second-kind modified Bessel function;

[0115]

[0116] in, Represents distance or spatial lag, that is, in the semivariogram, it measures the degree of separation in space (or time) between two sampling points in the InSAR data. Indicates the upper limit of variance, that is, when 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; Indicates smoothness, The larger the value, the smoother the semivariogram; represents the correlation length or correlation distance, which controls the range of spatial correlation of random fields; when Exceed When , the correlation between fields is significantly reduced; The second kind of modified Bessel function, different The value changes the shape of the semivariogram to characterize spatial smoothness and local variability; is the gamma function; is the first kind of Bessel function, which is used to describe random fields that may have periodic structures. The larger the function value, the greater the period or volatility of the random field.

[0117] In another implementation, the construction unit 152 is specifically configured to: use different preset attenuation parameter groups to adjust the smoothness , the first coefficient of the first kind of Bessel function and the second coefficient of the second kind of modified Bessel function are assigned to respectively correspond to the atmospheric turbulence noise characteristics in different directions represented by the assigned semi-variogram function; wherein the preset attenuation parameter group includes a preset smoothness, a first preset coefficient and a second preset coefficient.

[0118] In another implementation, the training unit 154 is used to: optimize the model parameters of the preset model using the Adam optimizer according to the changing trend of the mean absolute error obtained in each training; and determine the model parameters corresponding to the smallest mean absolute error during the training process as the model parameters of the target model.

[0119] 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 spatiotemporal features of the simulated time series InSAR training data set through multiple layers of convolution and pooling layers to obtain a first feature map, and the spatiotemporal features include 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 the feature dimension of the spatial features of the first feature map; the network structure includes 12 convolution 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, depth-separable convolution, feature pyramid network and residual connection.

[0120] In another implementation, the acquisition unit 151 is specifically used to: acquire geographic image data of the target area and meteorological data based on a SAR imaging device; the resolution of the SAR imaging device is higher than a preset resolution; and process the geographic image data and meteorological data based on a digital elevation model to obtain InSAR data.

[0121] Regarding the device in the above embodiment, the specific manner in which each unit module performs the operation has been described in detail in the embodiment of the method, and will not be elaborated here.

[0122] The present application also 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 apparatus for time-series InSAR turbulence layer delay correction or a lightweight deep learning device for time-series InSAR turbulence layer delay correction, the lightweight deep learning apparatus for time-series InSAR turbulence layer delay correction or the lightweight deep learning device for time-series InSAR turbulence layer delay correction can execute the lightweight deep learning method for time-series InSAR turbulence layer delay correction according to any of the possible implementations described above. The same technical effects can be achieved, and to avoid repetition, they are not further described here.

[0123] The present application also provides a computer program product, including a computer program or instructions, which is executed by a processor to implement a lightweight deep learning method for time-series InSAR turbulence layer delay correction, such as any of the possible implementations described above. The method achieves the same technical effects and is not further described here to avoid repetition.

[0124] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, and the true scope and spirit of the present application are indicated by the following claims.

[0125] It should be understood that the present application is not limited to the exact structures described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A lightweight deep learning method for time-series InSAR turbulence layer delay correction, characterized by: The method comprises: Acquiring InSAR data of the target area; the InSAR data includes raw SAR images, interferograms, and coherence maps; Based on geophysical modeling technology, different types of surface deformation signals in the InSAR data are constructed to obtain simulated InSAR observation ground subsidence data, wherein the simulated InSAR observation ground subsidence data includes a time series of surface deformation signals; constructing anisotropic atmospheric turbulence noise characteristics in the InSAR data based on a geostatistical model; The atmospheric turbulence noise characteristics and the surface deformation signal are superimposed and fused to construct a simulated time series InSAR training data set; The target model is obtained by training the preset model using the simulated InSAR training data set, using the mean absolute error between the output data of the preset model during the training process and the simulated InSAR observed ground subsidence data as a loss function. The target model is used to remove the turbulent layer noise characteristics; The method of constructing the InSAR data based on geophysical modeling technology in different types of surface deformation signals to obtain simulated InSAR observation ground subsidence 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 in each of the preset time sequences during the surface deformation process according to a preset time sequence; Combining the surface deformation data generated by the surface deformation in each of the preset time series in a time sequence to obtain the simulated InSAR observed ground subsidence data; The geostatistical model includes an anisotropic turbulent layer atmospheric delay model, and the anisotropic turbulent layer atmospheric delay model includes a semivariogram expressed by the following formula: ; The semivariogram includes the first kind of Bessel function and the second kind of modified Bessel function; in, Represents distance or spatial lag, that is, in the semivariogram, it measures the degree of separation in space or time between two sampling points in InSAR data. Indicates the upper limit of variance, that is, when 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; Indicates smoothness, The larger the value, the smoother the semivariogram; represents the correlation length or correlation distance, which controls the range of spatial correlation of random fields; when Exceed When , the correlation between fields is significantly reduced; The second kind of modified Bessel function, different The value changes the shape of the semivariogram to characterize spatial smoothness and local variability; is the gamma function; is the first kind of Bessel function, which is used to describe random fields that may have periodic structures. The larger the function value, the greater the period or volatility of the random field. The method of constructing anisotropic atmospheric turbulence noise characteristics in the InSAR data based on a geostatistical model includes: Using different preset attenuation parameter groups, respectively assigning values ​​to the smoothness, the first coefficient of the first-kind Bessel function, and the second coefficient of the second-kind modified Bessel function, so as to respectively correspond to the atmospheric turbulence noise characteristics in different directions represented by the semivariogram after the assignment; The preset attenuation parameter group includes a preset smoothness, a first preset coefficient and a second preset coefficient.

2. The method according to claim 1, characterized in that The method uses the mean absolute error between the output data output by the preset model during the training process and the simulated InSAR observed ground subsidence data as a loss function, trains the preset model using the simulated time-series InSAR training data set to obtain a target model, including: According to the change trend of the mean absolute error obtained in each training, the Adam optimizer is used to optimize the model parameters of the preset model; The model parameters corresponding to the smallest mean absolute error during the training process are determined as the model parameters of the target model.

3. The method according to claim 1, characterized in that The network structure of the preset model includes an encoder, a decoder and a feature fusion module; the encoder extracts and compresses the spatiotemporal features of the simulated time series InSAR training data set through multiple layers of convolution and pooling layers to obtain a first feature map, and the spatiotemporal features include 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 convolution layers and 2 pooling layers; the network structure is composed of a Squeeze-and-Excitation attention mechanism module, a depth-separable convolution module, a feature pyramid network module and a residual connection.

4. The method according to claim 1, wherein The obtaining of InSAR data of the target area includes: Acquiring SAR data of the target area based on a SAR imaging device, wherein the resolution of the SAR imaging device is higher than a preset resolution; The SAR data is processed based on a digital elevation model to obtain the InSAR data.

5. A lightweight deep learning device for time-series InSAR turbulence layer delay correction, characterized in that: The device comprises: An acquisition unit, configured to acquire InSAR data of a target area; the InSAR data includes original SAR images, interferograms, and coherence maps; A construction unit is configured to construct surface deformation signals of different types of the InSAR data based on geophysical modeling technology to obtain simulated InSAR observed ground subsidence data, wherein the simulated InSAR observed ground subsidence data includes the surface deformation signals arranged in a time series; The construction unit is further used to construct anisotropic atmospheric turbulence noise characteristics in the InSAR data based on a geostatistical model; The fusion unit is used to superimpose and fuse the atmospheric turbulence noise characteristics and the surface deformation signal to construct a simulated time series InSAR training data set; The training unit is used to train the preset model using the simulated time-series InSAR training data set, using the mean absolute error between the output data output by the preset model during the training process and the simulated InSAR observed ground subsidence data as a loss function, to obtain a target model, wherein the target model is used to remove turbulent layer noise characteristics in different directions; The construction unit is specifically configured to simulate the surface deformation process of the InSAR data using an elastic half-space spherical point source model and a finite fault displacement model; and collect surface deformation data generated by surface deformation in each of the preset time sequences during the surface deformation process according to a preset time sequence; Combining the surface deformation data generated by the surface deformation in each of the preset time series in a time sequence to obtain the simulated InSAR observed ground subsidence data; The geostatistical model includes an anisotropic turbulent layer atmospheric delay model, and the anisotropic turbulent layer atmospheric delay model includes a semivariogram expressed by the following formula: ; The semivariogram includes the first kind of Bessel function and the second kind of modified Bessel function; in: Represents distance or spatial lag, that is, in the semivariogram, it measures the degree of separation in space or time between two sampling points in InSAR data. Indicates the upper limit of variance, that is, when 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; Indicates smoothness, The larger the value, the smoother the semivariogram; represents the correlation length or correlation distance, which controls the range of spatial correlation of random fields; when Exceed When , the correlation between fields is significantly reduced; The second kind of modified Bessel function, different The value changes the shape of the semivariogram to characterize spatial smoothness and local variability; is the gamma function; is the first kind of Bessel function, which is used to describe random fields that may have periodic structures. The larger the function value, the greater the period or volatility of the random field. The building block is specifically used for: Using different preset attenuation parameter groups, respectively assigning values ​​to the smoothness, the first coefficient of the first-kind Bessel function, and the second coefficient of the second-kind modified Bessel function, so as to respectively correspond to the atmospheric turbulence noise characteristics in different directions represented by the semivariogram after the assignment; The preset attenuation parameter group includes a preset smoothness, a first preset coefficient and a second preset coefficient.

Images