Complex scene deformation monitoring and classification method based on InSAR and deep learning self-attention model

Abstract

The invention relates to the technical field of deformation monitoring, in particular to a complex scene deformation monitoring and classification method based on an InSAR and a deep learning self-attention model, and the method comprises the following steps: generating a deformation interferogram by using an improved PS point and DS point combined InSAR extraction technology and depending on the assistance of an SRTM DEM image; establishing a Delaunay triangulation network and constructing an annular belt, obtaining time sequence deformation data of all PS and DS points through adaptive arc densification and omnibearing point expansion, and extracting settlement points on the bridge; the SAR-self-attention model is used for decomposing the time sequence, InSAR time sequence data is decomposed into a trend component and a seasonal component, and accurate interpretation and analysis of cross-sea bridge deformation are achieved; time sequence dynamic and seasonal modes are described by comparing a curve fitting method with a seasonal and trend decomposition method using LOESS. According to the invention, technical support is provided for building intelligent and modern bridge deformation analysis and early warning through a big data technology.

Description

Technical Field

[0001] This invention relates to the field of deformation monitoring technology, and in particular to a method for deformation monitoring and classification of complex scenes based on InSAR and a deep learning self-attention model. Background Technology

[0002] Monitoring deformation in complex environments has always been a top priority. Cross-sea bridges are a prime example of such environments. As vital transportation links connecting land and islands, cross-sea bridges are becoming increasingly important in modern society, playing a key role in enhancing transportation networks, promoting trade growth, and driving regional integration. However, these bridges are susceptible to deformation in complex coastal environments, including the effects of extreme weather events, tidal loads, and natural damage. Abnormal movement and potential failure of cross-sea bridges can have serious consequences for public safety and property in coastal cities.

[0003] In recent years, temporal interferometric synthetic aperture radar (TSInSAR) technology has significantly improved the spatiotemporal resolution and data processing accuracy of various bridge stability monitoring. However, with the emergence of ultra-long cross-sea bridges with extensive structural components and complex geological, meteorological and marine environments, such as cross-sea bridges, traditional time-series deformation fitting usually relies on linear (or linear and seasonal) models, which are not applicable in complex scenarios. Therefore, it is urgent to find an efficient way to analyze the ground deformation of ultra-long cross-sea bridges. Summary of the Invention

[0004] To achieve the above objectives, this invention provides a method for detecting and classifying deformations in complex scenes based on InSAR and a deep learning self-attention model.

[0005] A method for deformation detection and classification of complex scenes based on InSAR and a deep learning self-attention model includes the following steps:

[0006] S1: Using the improved PS and DS point combination InSAR extraction technology, and with the assistance of SRTM DEM images, joint registration is performed between Sentinel-1SAR images and Cosmo-SkyMed images to generate interferograms.

[0007] S2: Based on the generated interferogram, a Delaunay triangulation network is created and a ring is constructed. Then, through adaptive arc densification and omnidirectional point expansion, the time series deformation data of all PS and DS points are obtained, and the settlement points on the bridge are extracted.

[0008] S3: Synthesize InSAR time series samples using the aforementioned time series deformation data, and apply the SAR-self-attention model to decompose the time series, thereby decomposing the InSAR time series data into trend components and seasonal components, so as to achieve accurate interpretation and analysis of the deformation of the cross-sea bridge.

[0009] S4: Analyze the deformation time series generated above, and compare the curve fitting method and the seasonal and trend decomposition method using LOESS to describe the temporal dynamics and seasonal patterns after the time series decomposition.

[0010] Furthermore, S1 specifically includes:

[0011] With the aid of SRTMDEM images, an enhanced spectral diversity method was used to register Sentinel-1 satellites with multiple baselines.

[0012] For Cosmo-SkyMed images, co-registration is performed based on the coherence coefficient method. The monitored terrain phase is removed from the original interferogram to generate differential interferograms, and the coherence weighted phase connection method is used to improve the phase quality of each interferogram.

[0013] Furthermore, S2 specifically includes:

[0014] S21: Construct a network based on bridge geometry.

[0015] Based on the amplitude dispersion index and spatial consistency, PS candidate points were determined, and a Delaunay triangulation was constructed to connect these PS candidate points. By differentiating the interference phase by connecting adjacent points, the atmospheric phase screen was eliminated, and then the M-estimator was used to estimate the difference parameters of the Delaunay triangulation.

[0016] S22: Enhanced connectivity of bridge beams

[0017] Based on the thermal expansion characteristics of bridge beams, a ring-shaped bridge geometric network was constructed to ensure the continuity of measurement points and increase the connectivity of the entire network.

[0018] S23: Network Encryption and Dot Extension Strategy

[0019] An arc densification method based on beam geometry is adopted and compared with a fully dense network to improve network quality and computational efficiency. The radius of the two circles is set to perform adaptive arc densification to achieve full connectivity of the PS network.

[0020] S24: Time series deformation data acquisition

[0021] In the first layer network, PS points with stable phase information are identified as reference points for the second layer network. For other PS and DS points, an all-round point expansion strategy is used for detection. Each candidate point is connected to two adjacent reference PS points to ensure accurate parameter estimation of the expanded points and obtain time series deformation data of all PS and DS points.

[0022] Furthermore, the synthetic InSAR time series samples include trend components, seasonal components, and noise components, capturing typical deformations related to the physical behavior of the cross-sea bridge.

[0023] Furthermore, an additive white noise component is added to the synthetic InSAR time series samples. This noise component represents temporally uncorrelated random fluctuations. By combining the trend, seasonality, and noise component, a dataset containing several synthetic InSAR time series samples is generated.

[0024] Furthermore, the trend component and seasonal component in S3 are composed of an encoder module and a decoder module, and the activation functions used in their respective decoders are different.

[0025] Furthermore, the encoder modules in the trend component and seasonal component process the input InSAR time series data sequentially, and also include timestamp encoding technology to solve the problem of irregular time intervals in high-resolution InSAR datasets and the problem of missing time series data in medium and low-resolution InSAR datasets.

[0026] Furthermore, the decoder in the trend component and seasonal component generates trend and seasonal components based on the representation in the encoder, including a linear layer that weights and combines the encoded features, uses an activation function to capture the intricate patterns in the data, and then uses another linear layer to improve the representation and generate the predicted trend and seasonal components. The trend branch uses a tuned linear unit activation function to introduce nonlinearity and capture the positive trend deformation in the InSAR time series data.

[0027] The seasonal component was simulated using the hyperbolic tangent activation function, which simulates periodic patterns, including annual variations caused by meteorological and oceanic activities.

[0028] During training, the SAR-self-attention network uses synthetic training samples for supervised learning. The network is optimized by minimizing the mean squared error loss between the predicted trend and seasonal components and the true values. The total loss function is calculated as the sum of the losses of the three individual components, including the trend loss. trend Seasonal losses seasonal and reconstruction loss reconstruction ;

[0029] By simultaneously minimizing the losses of the three individuals, the SAR-self-attention model network decomposes InSAR time series data into trend and seasonal components, enabling accurate interpretation and analysis of the deformation of the cross-sea bridge.

[0030] Furthermore, the curve fitting method includes fitting sine functions and quadratic functions.

[0031] Among them, the fitting sine function is used to capture periodic patterns, the quadratic function is used to capture the overall trend, and the curve fitting method also includes the residual part to take into account the remaining changes;

[0032] The seasonal and trend decomposition method uses the LOESS technique to decompose the time series into trend, seasonal and residual components, and performs smooth curve fitting on local subsets of the data to capture long-term changes and cyclical patterns.

[0033] Three indices are introduced: velocity V_t, acceleration A_t, and thermal amplitude A_s, to describe the trend and seasonal components. Velocity reflects the rate of change of the trend component after decomposition, indicating the speed at which the trend changes over time, and is expressed as:

[0034]

[0035] in, Represents the i-th timestamp t i The decomposition trend component at the location, where N represents the total number of timestamps and n represents the scaling factor of 365 or 366 for leap years, is used to convert acceleration to millimeters per year;

[0036] Acceleration is another indicator that quantifies the curvature or acceleration of a trend component over time. It represents the rate of change of the trend's rate of change and is calculated by taking the second derivative of the trend component.

[0037]

[0038] It also includes the measurement of thermal amplitude, which measures the change of the decomposed seasonal components relative to the temperature difference. The thermal amplitude is calculated by dividing the amplitude of the seasonal components by the temperature change.

[0039]

[0040] in, This represents the separated seasonal component at the i-th timestamp. This represents the average value of the separated seasonal components. and These represent the average temperatures in summer and winter, respectively, and N is the total number of timestamps.

[0041] The beneficial effects of this invention are:

[0042] This invention proposes a SAR-self-attention model method that can automatically decompose InSAR time series into different deformation components. The model was trained on synthetic InSAR samples and then applied to monitor the overall deformation of a cross-sea bridge, with a focus on the movement of key substructures. These results provide technical support for building intelligent and modern bridge deformation analysis and early warning systems using big data technology. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention;

[0045] Figure 2 An improved InSAR extraction technique for combining PS and DS points, as described in this embodiment of the invention.

[0046] Schematic diagram;

[0047] Figure 3 This is a network structure diagram of a deep learning self-attention model according to an embodiment of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0049] like Figure 1-3 As shown, a method for detecting and classifying deformations in complex scenes based on InSAR and a deep learning self-attention model includes the following steps:

[0050] Step 1: Extract settlement points on the bridge using an improved combination of PS and DS point extraction techniques and InSAR.

[0051] With the aid of SRTMDEM images, Sentinel-1SAR images and Cosmo-SkyMed (CSK) images were jointly registered to generate interferograms. To increase the connectivity of the network, a ring was created after constructing the Delaunay triangulation network (DTN). The first layer of the network was adaptively densified in an arc shape using BGN, and the second layer of the network was used for omnidirectional point expansion, resulting in time-series deformation data of all PS and DS points.

[0052] Step 2: Decompose the time series using the SAR-self-attention model.

[0053] Deep learning networks rely on appropriate and sufficient training data to achieve their optimal performance. Unfortunately, real InSAR time series data lacks ground-based information on decomposition components. To address this challenge, we use the time series deformation data generated in the first step to synthesize InSAR time series samples, and then put this dataset into the SAR-self-attention model to decompose the time series, decomposing the InSAR time series data into trend and seasonal components, thereby achieving an accurate interpretation and analysis of the deformation of the cross-sea bridge.

[0054] Step 3: Deformation time series analysis

[0055] To evaluate the performance of the SAR-self-attention model, curve fitting and the seasonal and trend decomposition method (STL) using LOESS were chosen to describe the temporal dynamics and seasonal patterns that exist after time series decomposition.

[0056] Synthetic InSAR time series data:

[0057] Deep learning networks rely on appropriate and sufficient training data to achieve their optimal performance. Unfortunately, real InSAR time series data lack ground-based information on decomposed components. To address this challenge, synthetic InSAR time series samples were generated, which include trend, seasonal, and noise components to capture typical deformations associated with the physical behavior of the super-bridge.

[0058] The trend component represents the gradual change of surface deformation over time, characterizing long-term geodynamic processes. Four different trend component models were considered, which can be represented as follows:

[0059]

[0060] Where i represents a specific timestamp, A l The deformation rate, A, represents a linear trend. d and A d The deformation rate and acceleration, which control the deceleration and acceleration trends respectively, are i d The timestamp indicating the start of the observation representing the deceleration trend, and i a The deformation failure timestamp representing the accelerating trend, S d and S a It is a constant, ensuring that the trend is zero when i = 0;

[0061] The seasonal component is simulated using a sine function, representing the regularity and periodicity of deformation changes. Its calculation formula is as follows:

[0062] X s (i)=A s ·sin(2πfi+θ)#

[0063] Where As controls the seasonal amplitude, f represents the frequency (the frequency for tropical years is set to 1 / 365), and θ controls the phase shift.

[0064] To simulate atmospheric noise common in real InSAR data, additive white noise was added to the synthetic time series. This noise component represents temporally uncorrelated random fluctuations. By combining trend, seasonality, and noise components, a dataset containing 120,000 synthetic InSAR time series samples was generated, with 40,000 samples for each trend type (linear stability, deceleration, acceleration). To ensure a comprehensive evaluation of the deep learning model, the dataset was randomly divided into 80% training set and 20% test set.

[0065] Specific steps:

[0066] 1. Using an improved combination of PS and DS point extraction techniques with InSAR to extract settlement points on bridges.

[0067] First, a two-layer network integrating PS and DS interferometric techniques was used to extract settlement points on the bridge and plot the deformation map of the cross-sea bridge. Since the dynamic characteristics of the various components of the cross-sea bridge are different, if not handled properly, serious decorrelation problems may occur. In order to enhance the connectivity of PS and DS candidates, a bridge geometry-based network (BGN) with a ring structure was proposed.

[0068] With the aid of SRTMDEM images, the enhanced spectral diversity (ESD) method was used to register the multi-baseline Sentinel-1 (S1) images. For Cosmo-SkyMed (CSK) images, the classical coherence coefficient method was used for co-registration. Differential interferograms were generated by removing the monitored terrain phase from the original interferograms, and the phase quality of each interferogram was improved by using the coherent weighted phase connection method.

[0069] Next, the first layer of the network was constructed. Using an amplitude dispersion index of 0.3 and a spatial consistency of 0.6, candidate PS points (PSs) were determined. A Delaunay triangulation network (DTN) was constructed to connect the PS candidate points. By connecting adjacent points, the interference phase between two points on an arc can be differentiated, thereby eliminating the atmospheric phase screen (APS). The differential parameters of each arc of the DTN were estimated using the M-estimator.

[0070] To increase network connectivity, a ring-shaped BGN network was constructed after the DTN was built.

[0071] This method helps connect PSCs with similar characteristics. By comparing with a fully dense network, the arc densification method based on beam geometry improves network quality and computational efficiency while reducing the number of arcs. Through this network (BGN), the radii of the two circles can be set and adaptive arc densification can be performed, ultimately achieving the effect of a fully connected PS network. In this study, considering the length of the bridge element, the radii of the small circle and the large circle of the densification ring are set to 300 meters and 900 meters, respectively.

[0072] In the first layer network, PSs with stable phase information that have a temporary coherence greater than 0.72 are identified as reference points for the second layer network. The remaining PSs and DSs are detected using an all-around point expansion strategy. DS candidate points are selected based on having a statistically uniform number of pixels greater than 25. Each candidate point is connected to two adjacent reference PSs to ensure accurate parameter estimation of the expanded points. Through the second layer network, time-series deformation data of all PS and DS points are obtained, which are all settlement points on the bridge.

[0073] 2. Decompose time series data using the SAR-self-attention model.

[0074] A SAR-self-attention method is proposed, which employs a self-attention model with two branches to decompose InSAR time series data into trend and seasonal components. Figure 2 This method effectively captures the temporal dependence and variation characteristics of the input data. The trend and seasonal branches consist of encoder and decoder modules, the only difference between them being the activation function used in their respective decoders.

[0075] In each branch, the encoder module processes the input InSAR time series data sequentially. To address the issue of irregular time intervals in high-resolution InSAR datasets (such as CSK data), a timestamp encoding scheme is proposed. By encoding the relative timestamps of the observation data, the network can identify time intervals and effectively capture time dependencies. Furthermore, timestamp encoding can also address the problem of missing data caused by factors such as sensor failure and atmospheric conditions, which is common in low-resolution InSAR datasets such as S1 data. The encoded timestamps help identify whether observation data exists at specific time steps, thereby enhancing the robustness and flexibility of the SAR transformer. The timestamp encoding is added to the input embedding, and its calculation method is as follows:

[0076]

[0077] Where ts represents the timestamp, i represents the dimension of the timestamp encoding, and the parameter d model Indicates the dimension of the input embedding.

[0078] After timestamp encoding, the encoder applies a multi-head attention mechanism. This mechanism allows the network to focus on different parts of the sequence and capture long-range dependencies in the time-series data. Mathematically, the multi-head attention mechanism can be represented as:

[0079]

[0080] Where Q, K, and V are the query matrix, key matrix, and value matrix, respectively, and d K Representing the dimension of the key matrix, after the attention mechanism, the "add & normalize" operation combines the attention output with the input features, and then normalizes the resulting representation. The feedforward layer performs a non-linear transformation on the encoded representation to extract higher-level features. Finally, another addition and normalization operation is applied to merge the output of the feedforward layer with the intermediate representation of the attention mechanism.

[0081] The decoder in each branch generates trend and seasonal components based on the representations learned from the encoder. Specifically, it includes a linear layer that weights and combines the encoded features, then uses an activation function to capture intricate patterns in the data, followed by another linear layer to further refine the representation and generate predicted trend and seasonal components. The trend branch uses the Tuned Linear Unit (ReLU) activation function to introduce nonlinearity and capture positive trend deformations in InSAR time series data. On the other hand, the seasonal branch uses the hyperbolic tangent (tanh) activation function, which is suitable for simulating periodic patterns, such as annual variations caused by meteorological and oceanic activities.

[0082] During training, the SAR-self-attention network uses synthetic training samples for supervised learning. It optimizes the network by minimizing the mean squared error (MSE) loss between the predicted trend and seasonal components and the true values. The total loss function is calculated as the sum of the losses of the three individual components: trend loss and seasonal loss. trend Seasonal losses seasonal and reconstruction loss reconstruction ,

[0083] Loss total =Loss trend +Loss seasonal +Loss reconstruction #

[0084] Loss trend Measure and predict trend components Loss trend With ground-based trend component X t The difference between them

[0085]

[0086] Similarly, seasonal loss is calculated based on the predicted seasonal component. Compared with the actual seasonal components on the ground X s The difference between them

[0087]

[0088] To accurately reconstruct the original InSAR time series, the reconstruction loss is also calculated, defined as the reconstruction time series. MSE between the original InSAR time series X and the original InSAR time series X

[0089]

[0090] By minimizing these three losses simultaneously, the SAR-self-attention model network effectively decomposes InSAR time series data into trend and seasonal components, thereby enabling accurate interpretation and analysis of the deformation of the cross-sea bridge.

[0091] 3. Deformation Time Series Analysis

[0092] To evaluate the performance of the proposed SAR transformer, two other widely used benchmark methods were selected for comparison: curve fitting and Seasonal and Trend Decomposition (STL) using LOESS. The curve fitting method involves fitting a sine function to capture periodic patterns and a quadratic function to capture the overall trend. It also includes a residual component to account for the remaining variation. The STL method uses the LOESS technique to decompose the time series into trend, seasonal, and residual components. This technique performs smooth curve fitting on local subsets of the data, thereby effectively capturing long-term variations and periodic patterns.

[0093] To better describe the trend and seasonal components, three indicators are introduced: velocity V_t, acceleration A_t, and thermal amplitude A_s. Velocity reflects the rate of change of the decomposed trend components, indicating how quickly the trend changes over time, and can be expressed mathematically as follows:

[0094]

[0095] in, Represents the i-th timestamp t i The decomposition trend component at the point, where N represents the total number of timestamps and n represents the scaling factor of 365 or 366 for leap years, is used to convert the acceleration to millimeters per year (mm / year).

[0096] Acceleration is another indicator that quantifies the curvature or acceleration of a trend component over time. It represents the rate of change of the trend's rate of change and is calculated by taking the second derivative of the trend component.

[0097]

[0098] The meaning of the signal is the same as the trend velocity formula mentioned earlier, and the unit of trend acceleration is millimeters per year squared (mm / year). 2 ).

[0099] On the other hand, thermal amplitude measures the changes in the decomposed seasonal components relative to the temperature difference, which helps to understand the relationship between surface deformation patterns and temperature changes. The method for calculating thermal amplitude is to divide the amplitude of the seasonal components by the temperature change.

[0100]

[0101] in, This represents the separated seasonal component at the i-th timestamp. This represents the average value of the separated seasonal components. and These three indicators—representing the average summer and winter temperatures, respectively, and N being the total number of timestamps—provide valuable quantitative metrics that help understand the temporal dynamics and seasonal patterns present in InSAR time series data.

[0102] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.

[0103] This invention is intended to cover all such substitutions, modifications, and variations falling within the broad scope of the claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for detecting and classifying deformations in complex scenes based on InSAR and a deep learning self-attention model, characterized in that, Includes the following steps: S1: Using the improved InSAR extraction technology of PS and DS points, and with the assistance of SRTM DEM images, joint registration is performed between Sentinel-1SAR images and Cosmo-SkyMed images to generate interferograms. S2: Based on the generated interferogram, a Delaunay triangulation network is created and a ring is constructed. Then, through adaptive arc densification and omnidirectional point expansion, the time series deformation data of all PS points and DS points are obtained, and the settlement points on the bridge are extracted. S3: Synthesize InSAR time series samples using the aforementioned time series deformation data, and apply the SAR-self-attention model to decompose the time series, thereby decomposing the InSAR time series data into trend components and seasonal components, so as to achieve accurate interpretation and analysis of the deformation of the cross-sea bridge. S4: Analyze the deformation time series generated above, and compare the curve fitting method and the seasonal and trend decomposition method using LOESS to describe the temporal dynamics and seasonal patterns after the time series decomposition.

2. The method for detecting and classifying deformations in complex scenes based on InSAR and a deep learning self-attention model according to claim 1, characterized in that, S1 specifically includes: With the aid of SRTMDEM images, an enhanced spectral diversity method was used to register Sentinel-1 satellites with multiple baselines. For Cosmo-SkyMed images, co-registration is performed based on the coherence coefficient method. The monitored terrain phase is removed from the original interferogram to generate differential interferograms, and the coherence weighted phase connection method is used to improve the phase quality of each interferogram.

3. The method for complex scene deformation monitoring and classification based on InSAR and deep learning self-attention model according to claim 2, characterized in that, S2 specifically includes: S21: Construct a network based on bridge geometry. Based on the amplitude dispersion index and spatial consistency, PS candidate points were determined, and a Delaunay triangulation was constructed to connect these PS candidate points. By differentiating the interference phase by connecting adjacent points, the atmospheric phase screen was eliminated, and then the M-estimator was used to estimate the difference parameters of the Delaunay triangulation. S22: Enhanced connectivity of bridge beams Based on the thermal expansion characteristics of bridge beams, a ring-shaped bridge geometric network was constructed to ensure the continuity of measurement points and increase the connectivity of the entire network. S23: Network Encryption and Dot Extension Strategy An arc densification method based on beam geometry is adopted and compared with a fully dense network to improve network quality and computational efficiency. The radii of the two circles are set to perform adaptive arc densification to achieve full connectivity of the PS network. S24: Time series deformation data acquisition In the first layer network, PS points with stable phase information are identified as reference points for the second layer network. For other PS and DS points, an all-round point expansion strategy is used for detection. Each candidate point is connected to two adjacent reference PS points to ensure accurate parameter estimation of the expanded points and obtain time series deformation data of all PS and DS points.

4. The method for complex scene deformation monitoring and classification based on InSAR and deep learning self-attention model according to claim 3, characterized in that, The synthetic InSAR time series samples include trend components, seasonal components, and noise components, capturing typical deformations related to the physical behavior of the cross-sea bridge.

5. A method for detecting and classifying deformations in complex scenes based on InSAR and a deep learning self-attention model according to claim 4, characterized in that, The synthetic InSAR time series samples are supplemented with an additive white noise component, which represents temporally uncorrelated random fluctuations. By combining the trend, seasonality and noise component, a dataset containing several synthetic InSAR time series samples is generated.

6. The method for detecting and classifying deformations in complex scenes based on InSAR and a deep learning self-attention model according to claim 5, characterized in that, The trend component and seasonal component in S3 are composed of an encoder module and a decoder module, and the activation functions used in their respective decoders are different.

7. A method for detecting and classifying deformations in complex scenes based on InSAR and a deep learning self-attention model according to claim 6, characterized in that, The encoder modules in the trend component and seasonal component process the input InSAR time series data sequentially, and also include timestamp encoding technology to solve the problem of irregular time intervals in high-resolution InSAR datasets and the problem of missing time series data in medium and low-resolution InSAR datasets.

8. A method for detecting and classifying deformations in complex scenes based on InSAR and a deep learning self-attention model according to claim 7, characterized in that, The decoders in the trend and seasonal components generate trend and seasonal components based on the representation in the encoder, including a linear layer that weights and combines the encoded features, uses an activation function to capture the intricate patterns in the data, and uses another linear layer to refine the representation and generate the predicted trend and seasonal components. The trend branch uses a tuned linear unit activation function to introduce nonlinearity and capture the positive trend deformation in the InSAR time series data. The seasonal component was simulated using the hyperbolic tangent activation function, which simulates periodic patterns, including annual variations caused by meteorological and oceanic activities. During training, the SAR-self-attention network uses synthetic training samples for supervised learning. The network is optimized by minimizing the mean squared error loss between the predicted trend and seasonal components and the true values. The total loss function is calculated as the sum of the losses of the three individual components, including the trend loss. trend Seasonal losses seasonal and reconstruction loss reconstruction ; By simultaneously minimizing the losses of the three individuals, the SAR-self-attention model network decomposes InSAR time series data into trend and seasonal components, enabling accurate interpretation and analysis of the deformation of the cross-sea bridge.

9. A method for detecting and classifying deformations in complex scenes based on InSAR and a deep learning self-attention model according to claim 8, characterized in that, The curve fitting method includes fitting sine functions and quadratic functions. Among them, the fitting sine function is used to capture periodic patterns, the quadratic function is used to capture the overall trend, and the curve fitting method also includes the residual part to take into account the remaining changes; The seasonal and trend decomposition method uses the LOESS technique to decompose the time series into trend, seasonal and residual components, and performs smooth curve fitting on local subsets of the data to capture long-term changes and cyclical patterns. Introducing velocity V t Acceleration A t and thermal amplitude A s Three indicators are used to describe the trend and seasonal components. Velocity reflects the rate of change of the trend component after decomposition, indicating how quickly the trend changes over time, and is expressed as: in, Represents the i-th timestamp t i The decomposition trend component at the location, where N represents the total number of timestamps and n represents the scaling factor of 365 or 366 for leap years, is used to convert acceleration to millimeters per year; Acceleration is another indicator that quantifies the curvature or acceleration of a trend component over time. It represents the rate of change of the trend's rate of change and is calculated by taking the second derivative of the trend component. It also includes the measurement of thermal amplitude, which measures the change of the decomposed seasonal components relative to the temperature difference. The thermal amplitude is calculated by dividing the amplitude of the seasonal components by the temperature change. in, This represents the separated seasonal component at the i-th timestamp. This represents the average value of the separated seasonal components. and These represent the average temperatures in summer and winter, respectively, and N is the total number of timestamps.

Images