US-12625255-B2 - Complex scene deformation monitoring and classifying method based on InSAR and deep learning self-attention model

US12625255B2US 12625255 B2US12625255 B2US 12625255B2US-12625255-B2

Abstract

Disclosed is a complex scene deformation monitoring and classifying method based on interferometric synthetic aperture radar (InSAR) and a deep learning self-attention model, including: generating a deformation interferogram by using an InSAR extraction technology combining persistent scatterer (PS) points and distributed scatter (DS) points and relying on a shuttle radar topography mission digital elevation model (SRTM DEM) image; constructing a Delaunay triangulation network and an annular zone, obtaining a time series deformation data of PS points and DS points through adaptive arc densification and omni-directional point expansion, and extracting settlement points on a bridge; decomposing a time series by an SAR-self-attention model, decomposing InSAR time series data into a trend component and a seasonal component, and interpreting and analyzing deformation of the sea-crossing bridge; and describing time series dynamics and a seasonal pattern by comparing a curve fitting method with a seasonal and trend decomposition method.

Inventors

Peifeng MA
Zherong WU
Yi Zheng

Assignees

Peifeng MA

Dates

Publication Date: 20260512
Application Date: 20240124
Priority Date: 20231121

Claims (9)

1 . A complex scene deformation monitoring and classifying method based on interferometric synthetic aperture radar (InSAR) and a deep learning self-attention model, comprising: S 1 : performing a co-registration between a first SAR image from a first satellite constellation and a second SAR image from a second satellite constellation by using an improved InSAR extraction technology of PS points and DS points and relying on an assistance of a shuttle radar topography mission digital elevation model (SRTM DEM) image, and generating an interferogram; S 2 : constructing a Delaunay triangulation network and an annular zone on a basis of the generated interferogram, obtaining time series deformation data of all PS points and DS points through adaptive arc densification and omni-directional point expansion, and extracting settlement points on a bridge; S 3 : synthesizing an InSAR time series sample through the time series deformation data, decomposing a time series by using an SAR-self-attention model, so as to decompose InSAR time series data into a trend component and a seasonal component, and accurately interpreting and analyzing deformation of a sea-crossing bridge; and S 4 : analyzing the generated deformation time series, describing time dynamics and a seasonal pattern after the time series is decomposed by comparing a curve fitting method with a seasonal and trend decomposition method using locally weighted scatterplot smoothing (LOESS).
2 . The complex scene deformation monitoring and classifying method based on InSAR and the deep learning self-attention model according to claim 1 , wherein the S1 specifically comprises: performing a registration on the first SAR image with multiple baselines with the assistance of the SRTM DEM image through an enhanced spectral diversity method, as for the second SAR image, performing a co-registration based on a coherence coefficient method, removing a monitored topographical phase from an original interferogram to generate differential interferograms, and improving a phase quality of each differential interferogram through a coherence weighted phase linking method.
3 . The complex scene deformation monitoring and classifying method based on InSAR and the deep learning self-attention model according to claim 2 , wherein the S2 specifically comprises: S21: constructing a bridge geometry-based network, specifically, determining a PS candidate point based on an amplitude dispersion index and a spatial consistency, connecting the PS candidate point by constructing the Delaunay triangulation network, differentiating an interference phase by connecting adjacent points, to eliminate an atmospheric phase screen, and then estimating a difference parameter of the Delaunay triangulation network by M-estimator; S22: enhancing a bridge beam connectivity, specifically, constructing a ring bridge geometric network based on a thermal expansion feature of a bridge beam, to guarantee a continuity of measurement points and increase a connectivity of an entire network; S23: performing network densifying and a point expansion strategy, specifically, comparing an arc densifying method based on beam geometry with a complete dense network, improving a network quality and a computing efficiency, setting radius of two circles, performing adaptive arc densifying, and implementing full connection of a PS network; and S24: obtaining the time series deformation data, specifically, in a first-layer network, identifying PS points with stable phase information as reference points of a second-layer network, detecting other PS points and DS points by using an omni-directional point expansion strategy, connecting each candidate point to two adjacent reference PS points to guarantee accurate parameter estimation of expansion points, and obtaining the time series deformation data of all the PS and DS points.
4 . The complex scene deformation monitoring and classifying method based on InSAR and the deep learning self-attention model according to claim 3 , wherein the synthesized InSAR time series sample comprise a trend component, a seasonal component and a noise element, and typical deformation related to physical behaviors of the sea-crossing bridge are captured.
5 . The complex scene deformation monitoring and classifying method based on InSAR and the deep learning self-attention model according to claim 4 , wherein an additive white noise component is added to the synthesized InSAR time series sample, and the noise component represents random fluctuations uncorrelated in time, and a data set comprising several synthesized InSAR time series samples is generated by combining the trend component, the seasonal component and the noise component.
6 . The complex scene deformation monitoring and classifying method based on InSAR and the deep learning self-attention model according to claim 5 , wherein the trend component and the seasonal component in the S3 comprise an encoder module and a decoder module, and activation functions used in each decoder are different.
7 . The complex scene deformation monitoring and classifying method based on InSAR and the deep learning self-attention model according to claim 6 , wherein encoder modules in the trend component and the seasonal component process input InSAR time series data in sequence, and further comprise a timestamp encoding technology for solving irregular time intervals in a high-resolution InSAR dataset and missing time series data in a medium-low resolution InSAR dataset.
8 . The complex scene deformation monitoring and classifying method based on InSAR and the deep learning self-attention model according to claim 7 , wherein the decoders in the trend component and the seasonal component generate a trend element and a seasonal element according to representation in the encoders, comprise a linear layer, perform a weighted combination on encoded features, use the activation functions to capture complex patterns in the data, use another linear layer to perfect the representation, and generate a predicted trend element and seasonal element, and a trend branch uses a rectified linear unit activation function and introduces nonlinearity to capture positive trend deformation in the InSAR time series data; the seasonal component uses a hyperbolic tangent activation function, and the hyperbolic tangent activation function simulates a periodic pattern, comprising annual variations due to meteorological and oceanographic activities; during training, the SAR-self-attention network uses a synthesized training sample for supervised learning, the network is optimized by minimizing a mean square error loss between a predicted trend component and seasonal component and a true value, and a total loss function is calculated as a sum of the three individual losses, comprising a trend loss Loss trend , a seasonal loss Loss seasonal and a reconstruction loss Loss reconstruction ; and by minimizing the three individual losses simultaneously, the SAR-self-attention model network decomposes the InSAR time series data into the trend component and the seasonal component, and accurately interprets and analyzes the deformation of the sea-crossing bridge.
9 . The complex scene deformation monitoring and classifying method based on InSAR and the deep learning self-attention model according to claim 8 , wherein the curve fitting method comprises fitting a sinusoidal function and a quadratic function, wherein fitting the sinusoidal function is to capture the periodic pattern, fitting the quadratic function is to capture a general trend, and the curve fitting method further comprises a residual part, to account for residual variations; the seasonal and trend decomposition method decomposes the time series into a trend part, a seasonal part and a residual part using the LOESS technology, and performs smooth curve fitting on a local subset of the data, so as to capture long-term variations and a periodic pattern; three indexes of velocity V t , acceleration A t and thermal amplitude A s are introduced to describe the trend element and the seasonal element, and the velocity reflects a variation rate of the trend element after decomposition, representing a velocity of a trend varying over time, and is expressed as: V t = n N - 1 ⁢ ∑ i = 1 N - 1 ( X i + 1 t - X i t t i + 1 - t i ) ; wherein X i t represents a decomposed trend component at an i-th timestamp at t i , N represents a total number of timestamps, and n represents a scaling factor of 365 or 366 for a leap year, and is used to convert the acceleration to millimeter/year; the acceleration is another indicator quantifying a curvature or an acceleration of the trend element varying over time, representing a variation velocity of a trend variation rate, and is calculated by performing second derivative on the trend element: A t = - n 2 N - 2 ⁢ ∑ i = 1 N - 2 [ X i + 2 t - 2 ⁢ X i + 1 t + X i t ( t i + 2 - t i + 1 ) ⁢ ( t i + 1 - t i ) ] ; and the curve fitting method further comprises measuring a thermal amplitude, wherein the measuring a thermal amplitude is to measure a variation condition of the decomposed seasonal component relative to a temperature difference, and the thermal amplitude is calculated by dividing an amplitude of the seasonal element by a temperature variation: A s = 2 · max ⁡ ( ❘ "\[LeftBracketingBar]" X i s - X s _ ❘ "\[RightBracketingBar]" ) T _ s - T w _ , i = 1 ⁢ to ⁢ N ; wherein X i s represents a separated seasonal element of the i-th timestamp, X s represents a mean value of separated seasonal elements, X s and T w represent mean temperatures of summer and winter respectively, and N is the total number of timestamps.

Description

TECHNICAL FIELD The present disclosure relates to the technical field of deformation monitoring, and particularly relates to a complex scene deformation monitoring and classifying method based on interferometric synthetic aperture radar (InSAR) and a deep learning self-attention model. BACKGROUND Various types of deformation monitoring in complex scenes have always been a top priority, and sea-crossing bridges are a typical complex scene. As an important transportation link connecting land and islands, the sea-crossing bridges have become increasingly important in modern societies, and they have a critical function in enhancing transportation networks, promoting trade growth, and facilitating regional integration. However, these bridges are susceptible to deformation in complex coastal environments, including extreme weather events, tidal loading, and natural damage. The unusual movement and potential failure of the sea-crossing bridges have serious impacts on public safety and property in coastal cities. In recent years, the time-series interferometric synthetic aperture radar (TSInSAR) technology has significantly improved the temporal and spatial resolution and data processing accuracy in monitoring the stability of various types of bridges. However, with the emergence of ultra-long sea-crossing bridges with a wide range of structural compositions in complex geological, meteorological, and oceanic environments, conventional time-series deformation fitting typically relies on linear (or linear and seasonal) models, which are not applicable in complex situations, and thus there is an urgent need to efficiently analyze ground deformations of ultra-large sea-crossing bridges. SUMMARY Based on the above objectives, the present disclosure provides a complex scene deformation monitoring and classifying method based on interferometric synthetic aperture radar (InSAR) and a deep learning self-attention model. The complex scene deformation monitoring and classifying method based on InSAR and the deep learning self-attention includes: S1: performing a co-registration between a Sentinel-1 SAR image and a Cosmo-SkyMed image by using an improved InSAR extraction technology combining PS points and DS points and relying on an assistance of a shuttle radar topography mission digital elevation model (SRTM DEM) image, and generating an interferogram;S2: constructing a Delaunay triangulation network and an annular zone on a basis of the generated interferogram, obtaining time series deformation data of all PS points and DS points through adaptive arc densification and omni-directional point expansion, and extracting settlement points on a bridge;S3: synthesizing an InSAR time series sample through the time series deformation data, decomposing a time series by using an SAR-self-attention model, so as to decompose InSAR time series data into a trend component and a seasonal component, and accurately interpreting and analyzing deformation of a sea-crossing bridge; andS4: analyzing the generated deformation time series, describing time dynamics and a seasonal pattern after the time series is decomposed by comparing a curve fitting method with a seasonal and trend decomposition method using locally weighted scatterplot smoothing (LOESS). Further, S1 specifically includes: performing a registration on Sentinel 1 with multiple baselines with the aid of the SRTM DEM image through an enhanced spectral diversity method,as for the Cosmo-SkyMed image, performing a co-registration based on a coherence coefficient method, removing a monitored topographical phase from an original interferogram, to generate differential interferograms, and improving a phase quality of each interferogram through a coherence weighted phase linking method. Further, the S2 specifically includes: S21: constructing a bridge geometry-based network, specifically,determining a PS candidate point based on an amplitude dispersion index and a spatial consistency, connecting the PS candidate point by constructing the Delaunay triangulation network, differentiating an interference phase by connecting adjacent points, to eliminate an atmospheric phase screen, and then estimating a difference parameter of the Delaunay triangulation network by M-estimator;S22: enhancing a bridge beam connectivity, specifically,constructing a ring bridge geometric network based on a thermal expansion feature of a bridge beam, to guarantee a continuity of measurement points and increase a connectivity of the entire network;S23: performing network densifying and a point expansion strategy, specifically,comparing an arc densifying method based on beam geometry with a complete dense network, improving a network quality and a computing efficiency, setting radius of two circles, performing adaptive arc densifying, and implementing full connection of a PS network; andS24: obtaining the time series deformation data, specifically,in a first-layer network, identifying PS points with stable phase info