MT-InSAR deformation monitoring method based on spatiotemporal characteristics of reservoir bank landslide deformation

Abstract

The present invention discloses an MT-InSAR deformation monitoring method combined with the spatiotemporal characteristics of reservoir bank landslide deformation, comprising the following steps: step S1, selecting SAR data; step S2, constructing a small baseline set of InSAR data, including SAR image registration, interferogram combination, differential interferometry processing, phase filtering, coherence estimation and interferogram optimization combination; step S3, performing surface classification based on dual-polarization SAR intensity images in combination with random forest classification, slope and aspect constraints and deep neural networks; then performing adaptive homogeneity filtering on the interference phase based on the secondary classification results; step S4, obtaining an adaptive deformation model of the reservoir bank landslide by combining multiple Sigmoid functions and linear creep and periodic functions; step S5, first performing nonlinear search of deformation parameters by using a genetic algorithm to determine the optimal parameter set; and finally recovering the deformation time series and average deformation rate by separating atmospheric delay and decoherence noise through low-pass filtering in the spatial domain and high-pass filtering in the time domain on the basis of low-frequency deformation.

Description

Technical Field

[0001] The present invention relates to the technical field of reservoir bank landslide monitoring, and in particular to an MT-InSAR deformation monitoring method combining the temporal and spatial characteristics of reservoir bank landslide deformation. Background Art

[0002] The strain softening of reservoir bank rock and soil caused by reservoir impoundment can easily lead to the resurgence of bank landslides. Furthermore, the hydrodynamic pressure generated by rapid rises and falls in reservoir water levels can significantly reduce the stability of landslides, posing a potential threat to reservoir safety.

[0003] Traditional monitoring methods, such as GNSS measurements, leveling, and crack meter monitoring, can only obtain sparse measurements at a few locations along reservoir banks. This poses a challenge to understanding the spatiotemporal evolution of landslide movement. Synthetic aperture radar interferometry (InSAR) is a ground observation technology developed in recent decades. Its high imaging resolution, non-contact nature, and ability to trace historical deformation have attracted significant attention in monitoring reservoir bank landslide deformation. However, in addition to the deformation component, the interferometric phase also contains topographic residuals, atmospheric delay, and decoherence noise. These components interfere with the deformation phase, making it difficult to extract an accurate deformation signal. To this end, many researchers have combined SAR imagery with time-series InSAR (MT-InSAR) methods to suppress the decoherence noise and atmospheric delay in the interferogram, thereby obtaining a time series of deformation over the study area. This approach is of great significance for understanding the evolution and deformation mechanisms of reservoir landslides. Commonly used MT-InSAR methods include permanent scatterer InSAR (PS-InSAR), small baseline set InSAR (SBAS-InSAR) technology and temporary coherent point InSAR (TCP-InSAR) method.

[0004] However, the use of InSAR technology for reservoir bank landslide monitoring still faces several bottlenecks that have yet to be fully addressed. First, reservoir bank landslides are generally located in areas with few structures, resulting in a limited number of permanent scatterers in these areas. Using PS-InSAR for monitoring makes it difficult to obtain a sufficiently dense number of monitoring points, thus negating the advantages of InSAR technology. To overcome this problem, some researchers have proposed adding distributed scatterers (DS) as monitoring targets in addition to permanent scatterers to increase the density of landslide monitoring points. This method first utilizes the statistical properties of SAR intensity information over time to identify homogeneous points within the monitoring targets. Phase filtering of these homogeneous points is then performed to improve the phase quality of the DS targets. Early studies primarily used neighborhood pixels of reference pixels for homogeneity filtering, such as region growing methods. Recently, various statistical hypothesis testing methods have been proposed to select homogeneous pixels, including nonparametric and parametric tests. Nonparametric methods identify homogeneous pixels by comparing the differences between the cumulative distribution functions of time samples. For example, Ferretti et al. used the Kolmogorov-Smirnov test to select statistically homogenous pixels. Building on this, other researchers have subsequently proposed methods such as the Baumgartner-Wei-Schindler test, the Kullback-Leibler test, and the Anderson-Darling test. However, these methods require the calculation and comparison of the cumulative distribution functions of all pixels, resulting in low computational efficiency. Parametric tests directly compare the differences in the statistical properties of two samples. However, in the presence of deformation and atmospheric phase, these methods struggle to make reliable assumptions about the overall distribution. Furthermore, they ignore the temporal properties of the samples, reducing the reliability of DS selection. SAR image classification, on the other hand, can automatically classify pixels into a series of meaningful homogeneous types based on backscatter characteristics, offering a promising approach for selecting homogeneous pixels in urban areas. However, unlike urban areas, the scattering intensities of reservoir bank landslides tend to be uniform, making it easy for the entire landslide area to be classified as a single homogeneous region. Clearly, the homogeneous points selected by this method are irrational, which can easily lead to poor phase optimization results.

[0005] On the other hand, while MT-InSAR methods can effectively suppress decoherence noise, interferograms contain components such as topographic residuals and atmospheric delay in addition to the deformation phase. Spatiotemporal filtering can suppress the influence of atmospheric delay on the interferogram phase. However, the topographic residual is difficult to accurately separate from the deformation signal. Its impact on the interferogram phase is proportional to the spatial baseline. Typically, during MT-InSAR data processing, it is assumed that the deformation time series satisfies a priori deformation model. Based on this assumption, a functional relationship is established between the interferogram phase, deformation model parameters, and the topographic residual. Among deformation modeling strategies, linear models are the most commonly used priori deformation models. They can be used to describe deformation characteristics that move slowly over time (such as tectonic movement, ground subsidence, and landslide creep). For specific deformations with significant cyclical characteristics, such as permafrost, simple linear models can easily miss seasonal deformation signals. Therefore, a periodic function can be added to the linear model to invert the deformation time series. Furthermore, polynomial models can be used to fit deformation of active volcanoes and sudden changes in mining areas. It is important to note that the accuracy of both the topographic residual and the deformation time series depends on the reliability of the prior deformation model. Typically, the deformation of reservoir landslides is driven primarily by rapid changes in hydraulic conditions (e.g., water level drops, rises, and heavy rainfall), exhibiting significant step-like deformation characteristics in time series. This step-like deformation undergoes a cyclical evolutionary process of stabilization, acceleration, and stabilization. Clearly, commonly used mathematical models struggle to accurately describe the deformation patterns of reservoir landslides. Using these models to invert reservoir landslide deformation can lead to the deformation signal being misinterpreted as a topographic residual signal due to the poor agreement between the true deformation and the empirical model, thus reducing the accuracy of the deformation field. Summary of the Invention

[0006] In response to the above technical problems, the present invention provides an MT-InSAR deformation monitoring method that combines the temporal and spatial characteristics of reservoir bank landslide deformation.

[0007] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0008] A MT-InSAR deformation monitoring method combining the temporal and spatial characteristics of reservoir bank landslide deformation includes the following steps:

[0009] Step S1: According to the requirements of reservoir bank landslide deformation monitoring, multi-track high-resolution SAR data covering the monitoring area can be selected;

[0010] Step S2, constructing a small baseline set of InSAR data based on multi-period SAR images, including SAR image registration, interferogram combination, differential interferometry processing, phase filtering, coherence estimation and interferogram optimization combination;

[0011] Step S3: Based on the dual-polarization SAR intensity image, a combination of random forest classification, slope and aspect constraints, and a deep neural network is used to achieve fine surface classification of the target area; then, based on the secondary classification results, an adaptive homogeneity filtering is performed on the interferometric phase;

[0012] Step S4, combining multiple Sigmoid functions, linear creep and periodic functions to obtain an adaptive deformation model of the reservoir bank landslide;

[0013] In step S5, a genetic algorithm is first used to perform a nonlinear search for deformation parameters to determine the optimal parameter set; based on the low-frequency deformation, atmospheric delay and incoherence noise are separated by low-pass filtering in the spatial domain and high-pass filtering in the temporal domain, and finally the deformation time series and average deformation rate are restored.

[0014] As a further improvement of the above technical solution:

[0015] Preferably, the step S2 includes the following steps:

[0016] S2-1, SAR image registration: First, it is necessary to select a main image from the SAR time-series images, perform single main image registration, and crop the registered SAR image to obtain SAR data of the monitoring area;

[0017] S2-2, Interferogram combination: Estimate the spatiotemporal baseline between any two images based on the orbital data of the SAR image, select the spatiotemporal baseline threshold to combine the interferograms; construct short baseline subsets using multiple main images, and combine the interferograms using a single subset;

[0018] S2-3, differential interferometry processing: perform interferometry processing on the SAR image with a short temporal and spatial baseline to obtain an interferometric phase map;

[0019] S2-4, phase filtering: adaptive filtering to suppress the influence of noise;

[0020] S2-5, coherence estimation and interferogram optimization combination: The interference quality of the interferogram is evaluated by coherence analysis; the coherence coefficient of the interferometric phase describes the similarity between the regions with the same name in the primary and secondary images.

[0021] Preferably, the step S3 includes the following steps:

[0022] S3-1, secondary image classification: First, a random forest classification is performed using a dual-polarization intensity texture map. Then, the random forest classification results are optimized based on the slope aspect map of the reservoir bank landslide. The optimized classification results are used as training labels, and a deep neural network is used to perform secondary classification.

[0023] S3-2, Connected Domain Construction: Using SAE Convolutional Network to learn the features of ground objects and their spatial distribution;

[0024] S3-3, homogeneous pixel selection: After feature extraction, a classifier is added and supervised training is performed using a set of labeled samples. A SAE model with classification function is obtained to classify all SAR image pixels. Based on the classification results, homogeneous pixels are identified for each pixel within a predefined neighborhood.

[0025] S3-4, Interferometric Phase Optimization: In a set of homogeneous pixels, the optimal interferometric phase value for each pixel is reconstructed from the covariance matrix by maximizing the joint probability density function.

[0026] Preferably, in the construction of the connected domain, SAE constructs a deep structure by stacking multiple autoencoders and trains each network using a greedy hierarchical learning method; the feature representation of the previous network will be used as the input of the latter; suppose a normalized n-dimensional input data set x, the hidden layer Expressed as:

[0027] (1)

[0028] in, is the autoencoder function, represents the weight matrix, Indicates deviation, for The parameter set of m-dimensional hidden layer and decoder can get the output layer :

[0029] (2)

[0030] in, is the decoding function. In the formula, the learning parameter set Obtain feature representation;

[0031] By minimizing the loss function Determine the optimal training parameters:

[0032] (3)

[0033] Preferably, in the interference phase optimization, the optimal interference phase value is:

[0034] (4)

[0035] Where, is the complex covariance matrix of the target pixel, , represents the optimized interferometric phase, is the optimized interference phase.

[0036] Preferably, the step S4 includes the following steps:

[0037] S4-1, Sigmoid function parameter adjustment: adjust the shape parameters of each Sigmoid function in the function group;

[0038] S4-2, Adaptive deformation model observation equation: Based on the components of the interference phase and the acquisition time of the primary and secondary interferogram images, an observation equation is established; the equation includes the deformation phase, terrain residual phase, atmospheric delay phase, and incoherence noise;

[0039] S4-3, hypothesis testing: Test the significance of the adaptive deformation model based on the phase observation value of the small baseline set, and then test the significance of each parameter. The parameters retained after the hypothesis test are the parameters to be determined for the adaptive deformation model.

[0040] Preferably, the Sigmoid function parameter adjustment method is:

[0041] Set monitoring area at time A total of Scene SAR image; the image is composed according to the temporal and spatial baseline threshold Interference pairs ( According to the D-InSAR principle, the phase in the interferogram generated by any two SAR images is The composition is expressed as:

[0042] (5)

[0043] in, is the deformation phase of the radar line of sight (LOS) at any time; is the atmospheric delay phase; is random noise; The phase associated with the terrain residual is:

[0044] (6)

[0045] in, For the interferometer vertical baseline, is the topographic residual of the pixel, and are the radar incident angle and wavelength respectively, is the slant distance from the SAR sensor to the ground; in addition, the MT-InSAR method represents the deformation phase in the interferogram as:

[0046] (7)

[0047] in, is the time increment between two SAR images, represents the LOS deformation rate, 、 are the parameters of the periodic sine and cosine functions respectively;

[0048] The Sigmoid function is expressed as:

[0049] (8)

[0050] Among them, the shape parameter and They represent the center position and slope of the curve respectively; by adjusting the size of the parameters, the inflection point moment and time span of the step characteristic accelerated deformation are described.

[0051] Preferably, the observation equation of the adaptive deformation model is:

[0052] The deformation model is expressed as:

[0053] (9)

[0054] in, for The cumulative deformation of time, is the parameter for adjusting the amplitude of the Sigmoid function; each Sigmoid function represents a deformation acceleration signal, which has a year as the period according to the deformation law of reservoir landslide;

[0055] for and Any high coherence pixel in the differential interferogram generated by the SAR image at any moment can be used to construct the observation equation:

[0056] (10)

[0057] in, For the interferometer vertical baseline, is the topographic residual of the pixel, is the slant range from the SAR sensor to the ground, for deformation, for deformation, is the residual phase.

[0058] Preferably, the step S5 specifically includes the following steps:

[0059] S5-1, GA algorithm parameter estimation: GA algorithm is used to encode the problem to be solved into a string through a genetic coding method such as binary coding or tree coding, and an initial population is randomly generated in the solution space. The genes in the population include linear rate, terrain residual, periodic function parameters and step function coefficients;

[0060] S5-2, High-frequency deformation extraction: Calculate the low-frequency signal of each interferometer pair using the solved deformation parameters and terrain residuals. Subtract the modeled deformation phase and terrain residual phase from the interferometer phase observation equation to obtain the residual phase; the residual phase contains atmospheric delay, decoherence noise, and high-frequency deformation phase. Perform spatial low-pass filtering and temporal high-pass filtering on the residual phase to recover the high-frequency deformation phase.

[0061] S5-3, deformation time series estimation: add the low-frequency deformation phase and the high-frequency deformation phase to obtain the complete deformation phase in the interference phase; solve the deformation time series based on the conversion relationship between deformation and phase;

[0062] S5-4, Average deformation rate: The deformation time series is used to reveal the temporal evolution characteristics of the reservoir bank landslide; the spatial pattern of deformation is described by the average deformation rate.

[0063] Preferably, in step S5, the low-frequency deformation phase and the high-frequency deformation phase are added to obtain a complete deformation phase in the interference phase; Unwrapped Interference Phase and Phase time series of SAR images The relationship between them is:

[0064] (16)

[0065] in, Represents the design matrix between the interferogram and the SAR image; the least squares method is used to solve the time series of the deformation phase:

[0066] (17)

[0067] The temporal deformation of the reservoir bank landslide is restored through the conversion relationship between phase and deformation.

[0068] The MT-InSAR deformation monitoring method provided by the present invention, which combines the temporal and spatial characteristics of reservoir bank landslide deformation, has the following advantages over existing technologies:

[0069] (1) The present invention combines the spatiotemporal characteristics of reservoir bank landslide deformation with the MT-InSAR deformation monitoring method. Based on dual-polarization SAR intensity imagery, it combines random forest classification, slope and aspect constraints, and deep neural networks to achieve fine surface classification of the target area. Adaptive homogeneity filtering is then performed on the interferometric phase, which improves phase quality while ensuring the accuracy of the InSAR deformation monitoring results.

[0070] (2) The present invention combines the spatiotemporal characteristics of reservoir bank landslide deformation with the MT-InSAR deformation monitoring method. A combined random forest and deep neural network interferometric phase optimization method is proposed to address the problems of poor interferometric phase quality and difficulty in balancing the accuracy and density of monitoring point targets in the reservoir bank landslide study area. By adjusting the amplitude, boundary, and center position of multiple Sigmoid functions, this method has been widely used in the field of reservoir landslide displacement prediction.

[0071] (3) The MT-InSAR deformation monitoring method of the present invention, which combines the spatiotemporal characteristics of reservoir bank landslide deformation, can perform refined InSAR deformation monitoring for complex reservoir bank landslide scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 This is the overall flow chart of the MT-InSAR monitoring method for reservoir bank landslides of the present invention.

[0073] Figure 2 It is a schematic diagram of adjusting the shape parameters of the Sigmoid function to describe step deformation according to the present invention.

[0074] Figure 3 This is the InSAR average deformation rate map of a typical reservoir bank landslide in the existing technology.

[0075] Figure 4 This is a GNSS accuracy verification diagram of the InSAR time series deformation of a typical reservoir bank landslide in existing technology. DETAILED DESCRIPTION

[0076] The following is a detailed description of the specific embodiments of the present invention. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0077] like Figure 1 As shown, the present invention proposes an MT-InSAR deformation monitoring method combining the spatiotemporal characteristics of reservoir bank landslide deformation, which specifically includes the following steps:

[0078] Step S1, SAR data acquisition

[0079] According to the needs of reservoir bank landslide deformation monitoring, multi-track high-resolution SAR data covering the monitoring area can be selected.

[0080] Step S2: Small baseline set InSAR interferometry processing

[0081] The data products provided by the SAR satellite platform are single-look complex (SLC) images, which contain amplitude and phase information. The phase is used to measure the distance between the observed target and the sensor. Deformation is the change in distance during the time interval between the acquisition of two images. Interferometric processing of two SAR images with repeated orbits to recover deformation phase information is the basis of all time-series InSAR processing technologies. Generally, in addition to the deformation phase, a single interferogram also includes other components such as terrain phase, atmospheric delay and incoherence noise. The present invention constructs a small baseline set of InSAR data based on multi-time SAR images, including SAR image registration, interferogram combination, differential interferogram processing, phase filtering, coherence estimation and interferogram optimization combination.

[0082] The interferometric phase processing flow before MT-InSAR parameter estimation and temporal deformation inversion in the present invention includes the following steps:

[0083] S2-1, SAR image registration

[0084] SAR images from repeated orbits exhibit spatial and temporal variations, and these geometric deviations can affect the interferometric phase. First, a master image must be selected from the SAR time series and then registered. Image registration accuracy reaches sub-pixel levels, typically reaching 0.0001 pixels. The registered SAR image is then cropped to obtain SAR data for the monitored area.

[0085] S2-2, Interference pattern combination

[0086] The spatiotemporal baseline between any two images is estimated based on the orbital data of the SAR images. To effectively suppress the effects of decoherence and topographic residuals on the interferometric phase, a spatiotemporal baseline threshold is selected to combine the interferograms. In the reservoir bank landslide scenario, a short baseline subset is constructed using multiple master images. Furthermore, to avoid rank deficiency during temporal deformation inversion, the interferograms are combined using a single subset.

[0087] S2-3, differential interferometry processing

[0088] Interferometric processing of SAR images with short temporal and spatial baselines produces interferometric phase maps. 30m resolution SRTM DEM data can be used for geocoding SAR images and removing terrain phases. Polynomial fitting can be used to remove orbital errors.

[0089] S2-4, phase filtering

[0090] The phase of incoherent noise in interferograms is random, and filtering can effectively suppress its effects. Adaptive (ADF) filtering, based on the minimum mean square error (MSE) criterion, automatically adjusts the autocorrelation function between noise and phase to optimize the filtering effect. Furthermore, non-local filtering methods can maintain good edge characteristics of deformations.

[0091] S2-5, Coherence estimation and interference pattern optimization combination

[0092] The interferogram quality is evaluated through coherence analysis. The coherence coefficient of the interferometric phase describes the degree of similarity between the primary and secondary images in the same region, and ranges from [0, 1]. A value of 0 indicates complete incoherence, while a value of 1 indicates identical scattering characteristics between the two images. Interferograms with low coherence in deformed regions should be removed to avoid inaccuracies in subsequent parameter estimation and temporal deformation.

[0093] Step S3, neural network interference phase optimization

[0094] In the reservoir bank landslide area, the presence of decorrelation factors such as noisy speckle noise, large surface motion and spatial phase gradient, and strong atmospheric effects reduces the interferometric phase quality, making it difficult to balance the accuracy and density of the extracted monitoring point targets.

[0095] To address the challenges of poor interferometric phase quality in reservoir bank landslide areas and the difficulty in balancing target accuracy and density of monitoring points, this paper employs a neural network-based phase optimization method for InSAR interferometric phase analysis to increase the number of effective monitoring points. First, based on dual-polarization SAR intensity imagery, a fine surface classification of the target area is achieved using a combination of random forest classification, slope and aspect constraints, and a deep neural network. Then, based on the secondary classification results, the interferometric phase is adaptively homogeneously filtered, improving phase quality while ensuring the accuracy of InSAR deformation monitoring results.

[0096] The specific implementation plan is:

[0097] S3-1, secondary image classification

[0098] First, random forest classification is performed using the dual-polarization intensity texture map. Then, the random forest classification results are optimized based on the slope aspect map constraints of the reservoir bank landslide. The optimized classification results are used as training labels, and a deep neural network is used to run a secondary classification.

[0099] Random forest classification is a typical supervised learning algorithm, an ensemble algorithm based on decision trees. Its "randomness" is reflected in two aspects: first, backsampling from the original sample set; second, reverse sampling from the original features, from which the optimal features are selected. A subset of samples and features forms a decision tree, and multiple decision trees form a "forest." Within this "forest," all trees are integrated and vote on the classification result. The class with the highest number of votes is ultimately output as the classification result.

[0100] Deep neural networks have many layers and a wide range of capabilities, allowing them to be mapped to arbitrary functions, resulting in better classification results. In practical applications, 30% of labeled samples are randomly selected for model training, with the remainder used as a validation set. During training, the monitored area is divided into multiple categories based on the selection of homogeneous areas and the backscatter intensity characteristics of typical land cover.

[0101] S3-2, Connected Domain Construction

[0102] This paper uses SAE convolutional networks to learn the characteristics of ground object categories and their spatial distribution. The model uses an unsupervised approach to learn features from unlabeled raw data, which can better describe the original image features and improve the accuracy of image classification results.

[0103] Specifically, SAE builds a deep structure by stacking multiple autoencoders and trains each network using a greedy hierarchical learning method. The feature representation of the previous network will serve as the input of the latter.

[0104] Given a normalized n-dimensional input dataset x, the hidden layer It can be expressed as:

[0105] (1)

[0106] in, is the autoencoder function, represents the weight matrix, Indicates deviation, for The output layer can be obtained from the m-dimensional hidden layer and decoder :

[0107] (2)

[0108] in, is the decoding function. In the formula, the learning parameter set Get feature representation.

[0109] By minimizing the loss function Determine the optimal training parameters:

[0110] (3)

[0111] S3-3, homogeneous pixel selection

[0112] After unsupervised training, the SAE model is only a feature extractor and lacks classification capabilities. A classifier must be added after feature extraction and supervised training is performed using a set of labeled samples. The classifier can be an SVM, random forest, or softmax model.

[0113] This training process fine-tunes the network parameters using error backpropagation, ultimately yielding a SAE model with classification capabilities that can be used to classify all SAR image pixels. Based on the SAE network's classification results, homogeneous pixels are identified for each pixel within a predefined neighborhood. For efficiency, a single rectangular window is used for all pixels. Within this window, neighbors within the target pixel's eight-connected domain are labeled as homogeneous.

[0114] S3-4, Interferometric Phase Optimization

[0115] In a set of homogeneous pixels, the optimal interferometric phase value for each pixel can be reconstructed from the covariance matrix by maximizing the joint probability density function:

[0116] (4)

[0117] Where, is the complex covariance matrix of the target pixel, , represents the optimized interferometric phase, is the optimized interference phase.

[0118] The optimized interferometric phase effectively suppresses the incoherence noise of low-coherence monitoring point targets, thereby improving the phase reliability and point density.

[0119] Step S4: InSAR adaptive deformation model of reservoir bank landslide

[0120] In the MT-InSAR method, the rationality of the spatiotemporal evolution characteristics of deformation depends on the reliability of the prior deformation model. The MT-InSAR method suppresses the influence of spatiotemporal decorrelation and atmospheric delay on the quality of the interference phase by combining interference patterns. In reservoir landslide scenes with few artificial buildings and severe decoherence, this method is usually based on multiple main images. The present invention adds a Sigmoid function that can better describe the significant step characteristics of reservoir bank landslides affected by reservoir water level and heavy rainfall on the basis of linear and periodic functions to construct an adaptive function group that describes complex deformation characteristics.

[0121] The present invention combines multiple Sigmoid functions and traditional linear creep and periodic functions to obtain an adaptive deformation model of reservoir bank landslide.

[0122] In the present invention, constructing an InSAR adaptive deformation model includes the following steps:

[0123] S4-1, Sigmoid function parameter adjustment

[0124] The Sigmoid function is a monotonically increasing, bounded, and symmetrical S-shaped curve function between 0 and 1. It contains two nonlinear shape parameters, such as the inflection point of accelerated deformation and the time span in equation (6). In order to make the function group reasonably describe the stability, accelerated deformation, and stable periodic motion characteristics of the reservoir bank landslide, the shape parameters of each Sigmoid function in the function group are adjusted in this embodiment. The specific parameter adjustment effect is as follows: Figure 2 .

[0125] Specifically, the monitoring area is set at time A total of These images are combined into a single image based on the temporal and spatial baseline threshold. Interference pairs ( ). According to the D-InSAR principle, the phase in the interferogram generated by any two SAR images is The composition is expressed as:

[0126] (5)

[0127] in, is the deformation phase of the radar line of sight (LOS) at any time; is the atmospheric delay phase; is random noise. The phase associated with the terrain residual is:

[0128] (6)

[0129] in, For the interferometer vertical baseline, is the topographic residual of the pixel, and are the radar incident angle and wavelength respectively, is the slant distance from the SAR sensor to the ground. In addition, in the interferogram, traditional MT-InSAR methods usually use linear deformation or periodic models to represent the deformation phase:

[0130] (7)

[0131] in, is the time increment between two SAR images, represents the LOS deformation rate, 、 are the parameters of the periodic sine and cosine functions respectively. Considering Scene SAR images can be combined into Based on the high coherence points, the observation equations of each pixel or differential arc segment can be established. The least squares method can be used to solve the terrain residual and deformation parameters. However, the deformation model is the key to parameter estimation and time series deformation inversion. If the deformation model is unreasonable, it will affect the estimation accuracy of deformation and DEM error. The Sigmoid function is a monotonically increasing, bounded and symmetric S-shaped curve function between 0 and 1, which can be expressed as:

[0132] (8)

[0133] Among them, the shape parameter and Representing the center position and slope of the curve, respectively; by adjusting the size of these two parameters, they can be used to describe the inflection point and time span of the step-like accelerated deformation. Therefore, this function can reasonably reflect the movement characteristics of reservoir landslides in time series.

[0134] S4-2, Observation equations of adaptive deformation model

[0135] An observation equation is established based on the components of the interferometric phase and the acquisition time of the primary and secondary interferogram images. This equation includes the deformation phase, terrain residual phase, atmospheric delay phase, and incoherence noise. In the deformation phase, considering that the sigmoid function ranges from 0 to 1, an amplitude parameter is set before each sigmoid function to adjust the magnitude of the step deformation. Due to different MT-InSAR frameworks, the observation equations vary. The SBAS-InSAR framework, based on unwrapped phase, establishes a pixel-by-pixel phase equation. However, based on PS-InSAR and TCP-InSAR, the phase difference equation between arc segments is established, which includes the differences in phase components between the pixels that make up the arc segment.

[0136] Specifically, based on the traditional MT-InSAR deformation model, the Sigmoid function is embedded as a function group to perform parameter estimation and time series deformation inversion. The new deformation model is expressed as:

[0137] (9)

[0138] in, for The cumulative deformation of time, is the parameter for adjusting the amplitude of the Sigmoid function. In the formula, each Sigmoid function represents a deformation acceleration signal, which has a yearly period according to the deformation law of reservoir landslide.

[0139] Combining the basic principles of D-InSAR, and Any high coherence pixel in the differential interferogram generated by the SAR image at any moment can be used to construct the observation equation:

[0140] (10)

[0141] in, For the interferometer vertical baseline, is the topographic residual of the pixel, is the slant range from the SAR sensor to the ground, for deformation, for deformation, is the residual phase.

[0142] S4-3, Hypothesis Testing

[0143] Because real deformations are complex and varied, to accurately describe deformation characteristics over long time series, this paper first tests the significance of the adaptive deformation model using the F distribution at a significance level of 0.01 based on the phase observations of the small baseline set. It then uses the Tu distribution at a significance level of 0.01 to test the significance of each parameter within the segment, thereby eliminating insignificant parameters. The parameters that remain after the hypothesis test are the parameters to be determined for the adaptive deformation model.

[0144] Step S5: InSAR deformation estimation of reservoir bank landslide

[0145] InSAR deformation includes average deformation rate and time series, which are used to reveal the spatial and temporal evolution characteristics of reservoir bank landslides, respectively. Estimating deformation parameters, including deformation parameters and topographic residuals, is a key step in MT-InSAR. Given the nonlinearity of the sigmoid function, the shape parameters in the observation equation cannot be directly solved using least squares methods.

[0146] To address the problem of nonlinear parameters in deformation models, the present invention first uses a genetic algorithm to perform a nonlinear search for deformation parameters to determine the optimal parameter set. Based on low-frequency deformation, atmospheric delay and incoherence noise are separated through low-pass filtering in the spatial domain and high-pass filtering in the temporal domain, ultimately recovering the deformation time series and average deformation rate. The method includes the following steps:

[0147] S5-1, GA algorithm parameter estimation

[0148] The present invention uses the GA algorithm to encode the problem to be solved into the form of a string through a genetic coding method of binary coding or tree coding, and randomly generates an initial population in the solution space. The genes in the population include linear rate, terrain residual, periodic function parameters and step function coefficients.

[0149] In this method, the least squares norm of the residual phase is used as a fitness function to evaluate the population's convergence criteria. First, the size of each initial individual gene is set, i.e., the search range of the solution space for each parameter to be determined. Second, the fitness function value corresponding to each gene is calculated and a determination is made as to whether the iterative termination convergence criteria are met. If not, the population is subjected to an optimization operation involving selection, crossover, and mutation to generate a new population, and the fitness function value is repeatedly calculated. The optimization operation is terminated until the iterative termination criteria of the fitness function are met, and the population gene generated by this process is used as the global optimal solution for the deformation model's estimated parameters.

[0150] Specifically, the derivative of the Sigmoid function with respect to a variable can be expressed by itself:

[0151] (11)

[0152] The traditional method uses Taylor series expansion to linearize the observation equation and obtain the error equation of the interference phase:

[0153] (12)

[0154] in, are the parameters to be determined for the deformation model and DEM error; The phase of the deformation model corresponding to the approximate value of the nonlinear parameter; is the coefficient matrix of the deformation parameters in the observation equation, expressed as:

[0155] (13)

[0156] Combining the basic principles of the Gauss-Markov model, the traditional method uses the least squares method to obtain the deformation parameters ( ):

[0157] (14)

[0158] Obtain the deformation parameters and terrain residuals in the adaptive deformation model.

[0159] Although the series expansion method can obtain the parameters to be solved by the least squares method, the actual deformation in the reservoir bank landslide is complex and changeable, making it difficult to obtain the approximate value of the nonlinear parameters for each pixel. The model error is easily introduced during the parameter inversion process. The GA algorithm first encodes the problem to be solved into a string using a genetic coding method such as binary coding or tree coding, and randomly generates an initial population in the solution space. The genes in the population contain linear rate , terrain residual , periodic function parameters , and the step function coefficients In the present invention, the least square norm of the residual phase is set as the fitness function to evaluate the convergence condition of the population, as shown in the following formula:

[0160] (15)

[0161] First, the size of each initial individual gene is set, that is, the search range of the solution space for each parameter to be determined. Second, the fitness function value corresponding to each gene is calculated and the iterative termination condition of Equation (15) is determined to be satisfied. If not, the population is optimized through selection, crossover, and mutation to generate a new population, and the fitness function value is repeatedly calculated. Once the iterative termination condition of the fitness function is satisfied, the optimization operation is terminated, and the population gene generated by this process is used as the global optimal solution for the deformation model estimation parameters.

[0162] S5-2, high-frequency deformation extraction

[0163] In this embodiment, the low-frequency signal of each interferometer pair is calculated using the solved deformation parameters and terrain residual. The modeled deformation phase and terrain residual phase are subtracted from the interferometer phase observation equation to obtain the residual phase. The residual phase includes atmospheric delay, incoherence noise, and high-frequency deformation phase. High-frequency deformation (non-modeled deformation) signals, such as atmospheric delay and incoherence noise, have different spatiotemporal characteristics.

[0164] In this embodiment, the residual phase is first decomposed onto the SAR time series using the least squares method. High-frequency deformation exhibits low-frequency characteristics in both the spatial and temporal domains. By combining the low-frequency spatial and high-frequency temporal characteristics of atmospheric delay and the high-frequency characteristics of decoherence noise in both space and time, the residual phase in the SAR image time series is subjected to spatial low-pass filtering and temporal high-pass filtering to recover the high-frequency deformation phase.

[0165] S5-3, Deformation Time Series Estimation

[0166] By adding the low-frequency and high-frequency deformation phases, the complete deformation phase in the interferometric phase can be obtained. Based on the interferogram combination, a functional relationship is established between the time-series SAR imagery and the interferogram. Since all interferograms belong to the same baseline subset, the least squares method can be used to directly solve the time series of the deformation phase. Based on the conversion relationship between deformation and phase, the deformation time series can be solved.

[0167] The low-frequency deformation phase and the high-frequency deformation phase are added together to obtain the complete deformation phase in the interference phase. Unwrapped Interference Phase and Phase time series of SAR images The relationship between can be modeled as:

[0168] (16)

[0169] in, Represents the design matrix between the interferogram and the SAR image. All interferograms belong to the same baseline subset, and the least squares method can directly solve the time series of the deformation phase:

[0170] (17)

[0171] Through the conversion relationship between phase and deformation, the temporal deformation of the reservoir bank landslide can be restored.

[0172] S5-4, average deformation rate

[0173] Deformation time series can be used to reveal the temporal evolution characteristics of reservoir bank landslides. However, the spatial pattern of deformation is usually described by the average deformation rate.

[0174] In this embodiment, the cumulative deformation of the time series SAR images is solved by least squares element by element, and the average deformation rate in the time series is obtained to reveal the spatial evolution of the landslide.

[0175] Finally, the spatiotemporal evolution characteristics of the reservoir bank landslide were obtained.

[0176] The above examples are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, they are not intended to limit the present invention. Therefore, any simple modifications, equivalent variations, and modifications to the above examples that do not depart from the technical solution of the present invention and are based on the technical essence of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A MT-InSAR deformation monitoring method combining the spatiotemporal characteristics of reservoir bank landslide deformation, characterized by: The specific steps include: Step S1: According to the requirements of reservoir bank landslide deformation monitoring, multi-track high-resolution SAR data covering the monitoring area can be selected; Step S2, constructing a small baseline set of InSAR data based on multi-period SAR images, including SAR image registration, interferogram combination, differential interferometry processing, phase filtering, coherence estimation and interferogram optimization combination; Step S3: Based on the dual-polarization SAR intensity image, a combination of random forest classification, slope and aspect constraints, and a deep neural network is used to achieve fine surface classification of the target area; then, based on the secondary classification results, an adaptive homogeneity filtering is performed on the interferometric phase; Step S4, combining multiple Sigmoid functions, linear creep and periodic functions to obtain an adaptive deformation model of the reservoir bank landslide; In step S5, a genetic algorithm is first used to perform a nonlinear search for deformation parameters to determine the optimal parameter set; based on the low-frequency deformation, atmospheric delay and incoherence noise are separated by low-pass filtering in the spatial domain and high-pass filtering in the temporal domain, and finally the deformation time series and average deformation rate are restored; The step S3 includes the following steps: S3-1, secondary image classification: First, a random forest classification is performed using a dual-polarization intensity texture map. Then, the random forest classification results are optimized based on the slope aspect map of the reservoir bank landslide. The optimized classification results are used as training labels, and a deep neural network is used to perform secondary classification. S3-2, Connected Domain Construction: Using SAE Convolutional Network to learn the features of ground objects and their spatial distribution; S3-3, homogeneous pixel selection: After feature extraction, a classifier is added and supervised training is performed using a set of labeled samples. A SAE model with classification function is obtained to classify all SAR image pixels. Based on the classification results, homogeneous pixels are identified for each pixel within a predefined neighborhood. S3-4, Interferometric Phase Optimization: In a set of homogeneous pixels, the optimal interferometric phase value for each pixel is reconstructed from the covariance matrix by maximizing the joint probability density function; The step S4 includes the following steps: S4-1, Sigmoid function parameter adjustment: adjust the shape parameters of each Sigmoid function in the function group; S4-2, Adaptive deformation model observation equation: Based on the components of the interference phase and the acquisition time of the primary and secondary interferogram images, an observation equation is established; the equation includes the deformation phase, terrain residual phase, atmospheric delay phase, and incoherence noise; S4-3, hypothesis testing: test the significance of the adaptive deformation model based on the phase observation value of the small baseline set, and then test the significance of each parameter. The parameters retained after the hypothesis test are the parameters to be determined for the adaptive deformation model; The Sigmoid function parameter adjustment method is: The study area is set at time A total of Scene SAR image; according to the temporal and spatial baseline threshold, the image is combined into M interferometric pairs. According to the D-InSAR principle, the phase in the interferogram generated by any two SAR images is The composition is expressed as: （5） in, is the deformation phase of the radar line of sight at any time; is the atmospheric delay phase; is random noise; The phase associated with the terrain residual is: （6） in, For the interferometer vertical baseline, is the topographic residual of the pixel, and are the radar incident angle and wavelength respectively, is the slant distance from the SAR sensor to the ground; in addition, the MT-InSAR method represents the deformation phase in the interferogram as: （7） in, is the time increment between two SAR images, Represents the deformation rate in the direction of radar line of sight, 、 are the parameters of the periodic sine and cosine functions respectively; The Sigmoid function is expressed as: （8） Among them, the shape parameter and They represent the center position and slope of the curve respectively; by adjusting the size of the parameters, the inflection point moment and time span of the step characteristic accelerated deformation are described.

2. The MT-InSAR deformation monitoring method according to claim 1, characterized in that: The step S2 includes the following steps: S2-1, SAR image registration: First, it is necessary to select a main image from the SAR time-series images, perform single main image registration, and crop the registered SAR image to obtain SAR data of the monitoring area; S2-2, Interferogram combination: Estimate the spatiotemporal baseline between any two images based on the orbital data of the SAR image, select the spatiotemporal baseline threshold to combine the interferograms; construct short baseline subsets using multiple main images, and combine the interferograms using a single subset; S2-3, differential interferometry processing: perform interferometry processing on the SAR image with a short temporal and spatial baseline to obtain an interferometric phase map; S2-4, phase filtering: adaptive filtering to suppress the influence of noise; S2-5, coherence estimation and interferogram optimization combination: The interference quality of the interferogram is evaluated by coherence analysis; the coherence coefficient of the interferometric phase describes the similarity between the regions with the same name in the primary and secondary images.

3. The MT-InSAR deformation monitoring method according to claim 1, characterized in that: In the connected domain construction, SAE constructs a deep structure by stacking multiple autoencoders and trains each network using a greedy hierarchical learning method; the feature representation of the previous network will be used as the input of the latter; suppose a normalized n-dimensional input data set x, the hidden layer is ,in , Expressed as: （1） in, is the autoencoder function, represents the weight matrix, Indicates deviation, for The parameter set of m-dimensional hidden layer and decoder can get the output layer : （2） in, For the decoding function, in the formula, the learning parameter set Obtain feature representation; By minimizing the loss function Determine the optimal training parameters: （3）。 4. The MT-InSAR deformation monitoring method according to claim 3, characterized in that: In the interference phase optimization, the optimal interference phase value is: （4） Where, is the complex covariance matrix of the target pixel, , represents the optimized interferometric phase, is the optimized interference phase.

5. The MT-InSAR deformation monitoring method according to claim 1, characterized in that: The observation equation of the adaptive deformation model is: The deformation model is expressed as: （9） in, for The cumulative deformation of time, is the parameter for adjusting the amplitude of the Sigmoid function; each Sigmoid function represents a deformation acceleration signal, which has a year as the period according to the deformation law of reservoir landslide; for and Any high coherence pixel in the differential interferogram generated by the SAR image at any moment can be used to construct the observation equation: （10） in, For the interferometer vertical baseline, is the topographic residual of the pixel, is the slant range from the SAR sensor to the ground, for deformation, for deformation, is the residual phase.

6. The MT-InSAR deformation monitoring method according to claim 1, characterized in that: The step S5 specifically includes the following steps: S5-1, GA algorithm parameter estimation: GA algorithm is used to encode the problem to be solved into a string through a genetic coding method such as binary coding or tree coding, and an initial population is randomly generated in the solution space. The genes in the population include linear rate, terrain residual, periodic function parameters and step function coefficients; S5-2, High-frequency deformation extraction: Calculate the low-frequency signal of each interferometer pair using the solved deformation parameters and terrain residuals. Subtract the modeled deformation phase and terrain residual phase from the interferometer phase observation equation to obtain the residual phase; the residual phase contains atmospheric delay, decoherence noise, and high-frequency deformation phase. Perform spatial low-pass filtering and temporal high-pass filtering on the residual phase to recover the high-frequency deformation phase. S5-3, deformation time series estimation: add the low-frequency deformation phase and the high-frequency deformation phase to obtain the complete deformation phase in the interference phase; solve the deformation time series based on the conversion relationship between deformation and phase; S5-4, Average deformation rate: The deformation time series is used to reveal the temporal evolution characteristics of the reservoir bank landslide; the spatial pattern of deformation is described by the average deformation rate.

7. The MT-InSAR deformation monitoring method according to claim 1, characterized in that: In step S5, the low-frequency deformation phase and the high-frequency deformation phase are added to obtain a complete deformation phase in the interference phase; Unwrapped Interference Phase and Phase time series of SAR images The relationship between them is: （16） in, Represents the design matrix between the interferogram and the SAR image; the least squares method is used to solve the time series of the deformation phase: （17） The temporal deformation of the reservoir bank landslide is restored through the conversion relationship between phase and deformation.

Images