High time-frequency time series deformation estimation method for bank landslide

Abstract

The embodiment of the present disclosure provides a reservoir bank landslide high-time-frequency time sequence deformation estimation method, which belongs to the technical field of measurement and specifically comprises the following steps: obtaining the slant range low-time-frequency time sequence deformation measurement value of the reservoir bank landslide by using the multi-time-phase InSAR technology; combining the daily rainfall, the reservoir water level height and the InSAR low-time-frequency time sequence deformation variable and their respective time information to generate a prediction data set; establishing a deep learning network according to the time different frequency characteristics of the high-time-frequency hydrological observation and the InSAR low-time-frequency time sequence deformation value, dividing the prediction data set into a training data set and a test data set, training the deep learning network embedded in the lag feature extraction module by using the training data set, and obtaining a prediction model; and evaluating the deformation prediction accuracy by using four evaluation indexes of root mean square error, mean absolute error, mean absolute error percentage and Pearson correlation coefficient and adjusting the best response parameters of the model. Through the scheme of the present disclosure, high-precision prediction under high-time-frequency is realized.

Description

Technical Field

[0001] This disclosure relates to the field of measurement technology, and in particular to a method for estimating high-frequency temporal deformation of reservoir bank landslides. Background Technology

[0002] Reservoir bank slopes undergo creep as water levels rise and fall. Landslides, characterized by their suddenness and destructiveness, can slide into the reservoir at high speed within seconds, generating massive surges and posing a significant threat to life and property. Monitoring the evolution of historical surface deformation can help predict the risk of premature landslide instability. The Global Navigation Satellite System (GNSS) is a technology for real-time monitoring of deformation in landslide-prone areas. However, GNSS monitoring stations located in landslide deformation zones are vulnerable to damage, and available GNSS observation data is sparsely distributed spatially, making it impossible to obtain global deformation information about landslides. Unlike GNSS, InSAR (Interferometric Synthetic Aperture Radar) technology can obtain high spatial resolution landslide deformation fields. However, limited by the revisit cycle of SAR satellites, this technology cannot capture the daily development of the deformation area.

[0003] In recent decades, landslide deformation prediction methods have been broadly categorized into model-driven and data-driven methods. Effective model-driven methods estimate the optimal deformation state by constructing quantitative models of the relationship between deformation and hydrological observations (such as Kalman filtering). However, some studies have pointed out that the cumulative effect of long-term rainfall can trigger landslide movement, indicating that the behavior of hydrological variables and landslide behavior are not a one-to-one mapping; there is not only a nonlinear mapping but also a time lag. Studies of the Kleine landslide show a lag time of 13.5 days between landslide displacement and maximum rainfall, while landslide movement in southwestern Oregon exhibits a lag time of 27-49 days during the rainy season. This suggests that a time lag factor must be determined before predicting each landslide study, but the time lag factor may differ for each monitoring point, thus requiring automated selection of the time lag factor. Data-driven prediction benefits from effective univariate deformation prediction methods, which can directly predict future deformation from historical deformation sequences (such as Long Short-Term Memory (LSTM) networks and Support Vector Machines). However, these prediction models can only obtain prediction results consistent with the frequency of the input historical deformation data. However, the current revisit period for SAR data is approximately 11 to 24 days, and this time-resolution deformation field loses temporal details.

[0004] It is evident that there is an urgent need for a high-precision, high-time-frequency method for estimating the high-time-frequency temporal deformation of reservoir bank landslides. Summary of the Invention

[0005] In view of this, the present disclosure provides a method for estimating high-frequency temporal deformation of reservoir bank landslides, which at least partially solves the problem of poor prediction accuracy in the prior art.

[0006] This disclosure provides a method for estimating high-frequency temporal deformation of reservoir bank landslides, including:

[0007] Step 1: Use multi-temporal InSAR technology to obtain the slope range-direction low-frequency temporal deformation measurement values ​​of the reservoir bank landslide and geocode them;

[0008] Step 2: Combine daily rainfall, reservoir water level, and InSAR low-frequency time-series shape variables and their respective temporal information to generate a prediction dataset:

[0009] Step 3: Based on the time difference characteristics between high-frequency hydrological observations and low-frequency InSAR time-series deformation values, a deep learning network is established. The prediction dataset is divided into a training dataset and a test dataset. The training dataset is used to train the deep learning network embedded with the lag feature extraction module to obtain the prediction model. The feature extraction module consists of a deep learning network LSTM and a time-series deformation / hydrological modulation factor.

[0010] Step 4: Input the test dataset into the prediction model, evaluate the deformation prediction accuracy using four evaluation metrics: root mean square error, mean absolute error, mean absolute error percentage, and Pearson correlation coefficient, and adjust the model's optimal response parameters.

[0011] According to a specific implementation of this disclosure, the step of obtaining the slant range-directed low-frequency temporal deformation measurement value of the reservoir bank landslide using multi-temporal InSAR technology includes:

[0012] The temporal deformation of the reservoir bank landslide was obtained by processing the acquired SAR images through image registration, main and secondary image interferometry, phase filtering, phase unwrapping, atmospheric phase removal, and multi-temporal InSAR deformation parameter estimation.

[0013] According to a specific implementation of an embodiment of this disclosure, step 2 specifically includes:

[0014] Multiple feature points with coherence greater than a preset value in the interferometric data are extracted from the deformation measurements to perform high-time-frequency estimation of the landslide deformation field, and a prediction dataset is constructed for each feature point.

[0015] According to a specific implementation of this disclosure, the prediction dataset includes InSAR time series data and its time information, rainfall time series data and its time information, and reservoir water level time series data and its time information.

[0016] According to a specific implementation of an embodiment of this disclosure, step 3 specifically includes:

[0017] Extract time-series features from the prediction dataset;

[0018] Based on the preset time delay coefficient range, the deformation time delay characteristics, rainfall time delay characteristics, and water level time delay characteristics of the prediction dataset are extracted;

[0019] The nonlinear relationship between deformation time lag features, rainfall time lag features and the state to be updated is estimated, and weights and biases are calculated accordingly. The influence weights of each feature are iteratively updated by calculating the loss value between the true state value and the calculated state value, so as to obtain the optimal nonlinear relationship between deformation features, rainfall time lag features and the state to be updated for each feature point, and form a prediction model.

[0020] The high-frequency temporal deformation estimation scheme for reservoir bank landslides in this embodiment includes: Step 1, obtaining the slant range to low-frequency temporal deformation measurement values ​​of the reservoir bank landslide using multi-temporal InSAR technology and geocoding them; Step 2, combining daily rainfall, reservoir water level, and InSAR low-frequency temporal deformation values ​​and their respective time information to generate a prediction dataset; Step 3, establishing a deep learning network based on the time-inconsistency characteristics of high-frequency hydrological observations and InSAR low-frequency temporal deformation values, dividing the prediction dataset into a training dataset and a test dataset, and training the deep learning network embedded in the lag feature extraction module with the training dataset to obtain a prediction model, wherein the feature extraction module consists of a deep learning network LSTM and a temporal deformation / hydrological modulation factor; Step 4, inputting the test dataset into the prediction model, evaluating the deformation prediction accuracy using four evaluation indicators: root mean square error, mean absolute error, mean absolute error percentage, and Pearson correlation coefficient, and adjusting the optimal response parameters of the model.

[0021] The beneficial effects of the embodiments of this disclosure are as follows: With the support of a simple deep learning model, the InSAR deformation and hydrological observation data are fused to estimate the high-frequency time series deformation of landslides. Moreover, the lag effect of precipitation or reservoir water level in the landslide deformation of the reservoir bank is taken into account, thus achieving high-precision prediction at high time frequency. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1A flowchart illustrating a method for estimating high-frequency temporal deformation of reservoir bank landslides provided in this embodiment of the present disclosure;

[0024] Figure 2 This is a schematic diagram illustrating the specific implementation process of a method for estimating high-frequency temporal deformation of reservoir bank landslides provided in this embodiment of the disclosure;

[0025] Figure 3 A schematic diagram of a network architecture provided in this disclosure embodiment;

[0026] Figure 4 A schematic diagram of simulated InSAR time-series deformation / daily reservoir water level / daily rainfall data provided in this embodiment of the disclosure;

[0027] Figure 5 This is a high-time-frequency estimation effect diagram of the simulated landslide deformation field under different SAR satellite periods provided in the embodiments of this disclosure;

[0028] Figure 6 A graph showing the relationship between the high time-frequency estimation accuracy of landslide deformation field and SAR satellite access period provided in the embodiments of this disclosure. Detailed Implementation

[0029] The embodiments of this disclosure will now be described in detail with reference to the accompanying drawings.

[0030] The following specific examples illustrate the implementation of this disclosure. Those skilled in the art can easily understand other advantages and effects of this disclosure from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. This disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0031] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0032] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this disclosure. The drawings only show the components related to this disclosure and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0033] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0034] This disclosure provides a method for estimating high-frequency temporal deformation of reservoir bank landslides, which can be applied to the prediction of reservoir bank landslides in natural disaster prevention scenarios.

[0035] See Figure 1 This is a flowchart illustrating a method for estimating high-frequency temporal deformation of reservoir bank landslides according to an embodiment of this disclosure. Figure 1 and Figure 2 As shown, the method mainly includes the following steps:

[0036] Step 1: Use multi-temporal InSAR technology to obtain the slope range-direction low-frequency temporal deformation measurement values ​​of the reservoir bank landslide and geocode them;

[0037] Furthermore, the step of obtaining the slant range-direction low-frequency temporal deformation measurement values ​​of the reservoir bank landslide using multi-temporal InSAR technology includes:

[0038] The temporal deformation of the reservoir bank landslide was obtained by processing the acquired SAR images through image registration, main and secondary image interferometry, phase filtering, phase unwrapping, atmospheric phase removal, and multi-temporal InSAR deformation parameter estimation.

[0039] The purpose of this invention is to solve three problems: 1. The difficulty in fusing heterogeneous, heterogeneous, and heterogeneous time-series data. Landslide surface deformation and hydrological data acquired by InSAR technology are correlated but different types of data. Heterogeneous, heterogeneous, and heterogeneous time-series data are more difficult to fuse than heterogeneous, heterogeneous time-series data of the same type (e.g., fusing InSAR observation data and GNSS observation data or multi-track InSAR observation data); 2. The complexity of the mapping relationship between multiple types of time-series data. The response of hydrological observations to landslide surface deformation is very complex, affected by differences in slope lithology, different environments of the slope, different slope shapes, reservoir water level fluctuations, and rainfall intensity and frequency. In practice, this manifests as, for example, varying degrees of lag between surface deformation fluctuations and rainfall fluctuations and reservoir water level fluctuations in different scenarios; and different parts of the landslide being affected by rainfall and water levels to varying degrees; 3. High-frequency dynamic estimation of landslides based on deep learning combined with two types of data. Traditional linear state estimation algorithms, such as KF, can effectively achieve dynamic daily forecasting. However, these algorithms can only solve linear state estimation problems and cannot achieve good results in nonlinear states. In landslide scenarios, there is a lag between deformation and hydrological observation. Therefore, parameters need to be set according to specific scenarios. Traditional algorithms have the disadvantage of not being able to automatically adapt to the scenario.

[0040] Its basic idea can be summarized as follows:

[0041] Fusion of heterogeneous, heterogeneous, and heterogeneous time-series data based on deep learning;

[0042] LSTM-based extraction of time delay features from time series data.

[0043] Nonlinear state estimation based on DNN.

[0044] In specific implementation, step 1 includes:

[0045] By selecting the MTI InSAR technology, the temporal deformation of the deformation region can be obtained after several steps, including image registration, master-slave image interferometry, phase unwrapping, error signal processing, and temporal deformation calculation.

[0046] Step 2: Combine daily rainfall, reservoir water level, and InSAR low-frequency time-series shape variables and their respective temporal information to generate a prediction dataset:

[0047] Furthermore, step 2 specifically includes:

[0048] Multiple feature points with coherence greater than a preset value in the interferometric data are extracted from the deformation measurements to perform high-time-frequency estimation of the landslide deformation field, and a prediction dataset is constructed for each feature point.

[0049] Optionally, the prediction dataset includes InSAR time series data and its time information, rainfall time series data and its time information, and reservoir water level time series data and its time information.

[0050] In practice, several feature points with a coherence greater than 0.3 in the interferometric data obtained from the deformation calculation of the experimental area can be extracted for high-frequency estimation of the landslide deformation field. A dataset is constructed for each feature point, and each dataset includes three parts: 1. InSAR time series data and its time information (year, month, day); 2. Rainfall time series data and its time information (year, month, day); 3. Reservoir water level time series data and its time information (year, month, day). Of course, the preset values ​​can be adjusted according to actual needs.

[0051] Step 3: Based on the time difference characteristics between high-frequency hydrological observations and low-frequency InSAR time-series deformation values, a deep learning network is established. The prediction dataset is divided into a training dataset and a test dataset. The training dataset is used to train the deep learning network embedded with the lag feature extraction module to obtain the prediction model. The feature extraction module consists of a deep learning network LSTM and a time-series deformation / hydrological modulation factor.

[0052] Based on the above embodiments, step 3 specifically includes:

[0053] Extract time-series features from the prediction dataset;

[0054] Based on the preset time delay coefficient range, the deformation time delay characteristics, rainfall time delay characteristics, and water level time delay characteristics of the prediction dataset are extracted;

[0055] The nonlinear correlation between deformation time lag characteristics, hydrological observation time lag characteristics and the state to be updated is estimated, and the weights and biases are calculated accordingly. The influence weights of each feature are iteratively updated by calculating the loss value between the true state value and the calculated state value, so as to obtain the optimal nonlinear relationship between deformation characteristics, hydrological observation time lag characteristics and the state to be updated for each feature point, and form a prediction model.

[0056] In specific implementation, the deep learning network architecture is as follows: Figure 3 As shown, temporal features are extracted from the data. The impact of rainfall on landslide deformation usually exhibits a lag effect, but the specific lag is influenced by differences in slope lithology, the environment in which the slope is located, the shape of the slope, and the intensity and frequency of rainfall. To overcome these difficulties, this embodiment embeds a lag feature extraction module into the network. This module consists of a deep learning network LSTM and temporal deformation or hydrological modulation factors (e.g., rainfall / reservoir water level). It utilizes the ability of LSTM to preserve effective information in temporal data to reconstruct temporal features, thus taking into account the lag difference effect of landslides.

[0057] Assuming the input data categories include InSAR historical data sequences, rainfall, and reservoir water levels, module 1 is responsible for processing InSAR time-series data to extract deformation time-lag features, module 2 is responsible for processing rainfall data to extract rainfall time-lag features, and module 3 is responsible for processing reservoir water level data to extract water level time-lag features. Extracting time-lag features requires determining time-lag coefficients. One advantage of this method is that these coefficients do not need to be highly accurate; they only need to be within a large range, which will be very convenient in practical applications.

[0058] The method estimates the nonlinear relationship between deformation features, hydrological observation time lag features, and the state to be updated. Considering that the state value to be updated is mainly affected by historical cumulative deformation, as well as by different time lag features of rainfall and reservoir water level rise and fall, the method uses a DNN network to learn the nonlinear relationship between features and state values. The network calculates weights and biases based on the relationship between features and state values, and iteratively updates the influence weights of each feature by calculating the loss value between the true state value and the calculated state value. Finally, it obtains the optimal nonlinear relationship between deformation features, hydrological observation time lag features, and the state to be updated for this point.

[0059] Step 4: Input the test dataset into the prediction model, evaluate the deformation prediction accuracy using four evaluation metrics: root mean square error, mean absolute error, mean absolute error percentage, and Pearson correlation coefficient, and adjust the model's optimal response parameters.

[0060] In practice, after obtaining the prediction model, the test dataset can be input into the prediction model. The deformation prediction accuracy is evaluated by four evaluation indicators: root mean square error, mean absolute error, mean absolute error percentage, and Pearson correlation coefficient. The optimal response parameters of the model are then adjusted. The prediction model with adjusted parameters can then be used to predict the real-time collected data to obtain more accurate prediction results.

[0061] The high-frequency time-series deformation estimation method for reservoir bank landslides provided in this embodiment estimates the high-frequency time-series deformation of landslides by fusing InSAR deformation and hydrological observation data with the support of a simple deep learning model. Moreover, it considers the lag effect of precipitation or reservoir water level in reservoir bank landslide deformation, and achieves high-precision prediction at high time frequencies.

[0062] The following will illustrate this solution through a specific embodiment. The present invention generates InSAR deformation at different time resolutions by simulating the revisit period of SAR data. These InSAR deformation data are landslide-like deformation sequences dominated by trend deformation and periodic deformation, in which the periodic contribution of displacement comes from seasonal water level fluctuations and rainfall. The simulated scenario uses 25 historical SAR data sets (assumed to be sentinel data) from January 1, 2017 to October 30, 2018. After InSAR data processing, a reservoir bank landslide hazard zone was identified. Starting from October 30, 2018, reservoir water level monitors and rainfall monitors were installed in the landslide hazard zone. SAR update data, reservoir water level data, and rainfall data from October 30, 2018 to April 23, 2021 were collected for model training. Starting from April 23, 2021, the deformation status of the day was estimated daily based on the acquired reservoir water level and rainfall data. Every 12 days, after obtaining updated SAR data, the deformation estimation accuracy was evaluated and the landslide deformation prediction results were corrected until June 22, 2022. Rainfall data was set to be slightly higher in June-August each year than in other months, with a prolonged and severe rainstorm event simulated in the summer of 2021. Reservoir water level data was set to show concentrated reservoir drainage from late April to early May each year, and water supply from late August to early September. Finally, a dataset composed of simulated rainfall data, simulated reservoir water level data, and corresponding InSAR deformation data was used to perform simulation data predictions. Figure 4 As shown.

[0063] The data registration accuracy during data processing is ensured to be above one-thousandth of a pixel, and the average coherence of the interferometric points is greater than 0.5. After processing the SAR data using TCP-InSAR technology, the time-series cumulative deformation results of the landslide deformation zone are obtained. The RMS of deformation points in a 10×10 window far from the deformation region is statistically analyzed, with an accuracy better than 15 mm / year.

[0064] Then, from the InSAR results obtained by deformation calculation in the experimental area, several feature points with interferometric data coherence greater than 0.3 can be extracted for high-time-frequency estimation of the landslide deformation field. A dataset is constructed for each feature point, and each dataset includes three parts: 1. InSAR time series data and its time information (year, month, day); 2. Rainfall time series data and its time information (year, month, day); 3. Reservoir water level time series data and its time information (year, month, day). In areas where rainfall and reservoir water level fluctuations lead to landslide instability, the surface displacement of unstructured soil and rock blocks on the landslide surface increases rapidly when the cumulative rainfall reaches its peak. After incomplete statistical analysis of the cross-correlation between rainfall and landslide surface displacement, it was found that the lag time of rainfall on landslide deformation is approximately 10-22 days. The rainfall delay coefficient mentioned in the method can be taken as a slightly larger value; this invention sets it to 30. The entire time series dataset can be divided in a 2:1 ratio to obtain a training set and a test set.

[0065] like Figure 3 As shown, a feature extraction module is set in the network to adaptively extract the time-delay features of landslide deformation, including deformation system features and time-delay features under the influence of hydrological modulation factors; this module consists of a Long Short-Term Memory (LSTM) algorithm. Since LSTM can store effective information about time data, it is also often used for univariate time prediction. At the end of this step, we obtain the deformation delay and rainfall / reservoir water level delay characteristics. Through feature extraction, the nonlinear correlation between historical deformation and rainfall / reservoir water level features and deformation is estimated. The l-layer DNN network calculates the weights and biases of neurons according to the relationship between features and state values, and uses the ReLU activation function to modify the linear units and simplify the calculation process. Then, the weights of each feature are iteratively updated by calculating the loss value between the true value and the calculated value; the training process is shown in Equation (1). X(E) is the training dataset, θ is the initial parameter, and F and G are the relationship between feature extraction and parameter estimation, respectively. These are the training parameters.

[0066]

[0067] For this task, the hyperparameter settings are as follows: a single LSTM layer is applied for feature extraction, and ten DNN layers are applied for parameter estimation with different dimensions (i.e., [input_dim, 64], [64, 128], [128, 256], [256, 512], [512, 256], [256, 128], [128, 64], [64, 32], [32, 16], [16, 1]). The mean squared error (MSE) loss function scores the prediction performance, and the Adam optimization algorithm iteratively updates the network parameters based on the training data with a learning rate of 1e-4.

[0068] To facilitate understanding of this invention, the theoretical basis of this invention is provided below:

[0069] First, the historical landslide sequence data, rainfall or other influencing factors, and their respective temporal information are reconstructed and input into BLOCK1...BLOCKn for temporal feature extraction. The functional relationship is as follows:

[0070]

[0071]

[0072]

[0073]

[0074] h e =tanh(s e )q e (6)

[0075] Among them, the forgetting gate f e (Time e) controls the weights of the current information and normalizes them to the range of 0-1 using σ(*) and tanh(*) units; external input gate unit g e Integrate and add new information; output gate unit q e The information that needs to be output for control requires that the internal state of the cell be processed by the self-updating unit s. e b, U, and W are the bias, input weight, and recurrent weight in the LSTM cell, respectively.

[0076] For BLOCK1, when e iterates through [t-1,...,tn], h e They also have different deformation time delay characteristics. For BLOCK2, when e traverses [t-1,...,tm], h e They also have different precipitation time lag characteristics, expressed as

[0077]

[0078] in The deformation time-delay eigenvector, Let X be the precipitation time-lag feature vector. F1 and F2 are the functional relationships between the input data and its features, respectively, and θ1 and θ2 are the parameters obtained after training. X is the input data. E1 = {[t-1], [t-1, t-2], ..., [t-1, t-2, ..., tm]} and E2 = {[t-1], [t-1, t-2], ..., [t-1, t-2, ..., tn]} represent the time series of information involved in the features.

[0079] After feature extraction, it is also necessary to estimate the nonlinear correlation between deformation delay characteristics, precipitation delay characteristics, and the state to be updated. The l-layer DNN network calculates the weights and biases of neurons based on the relationship between features and state values, and uses the ReLU activation function to correct linear elements and simplify the calculation process; then, the weights of each feature are iteratively updated by calculating the loss value between the true value and the calculated value, and the loss function is shown in the following formula;

[0080]

[0081] Abbreviated as:

[0082] H L =G(H0,θ3)=G{F(X(E);θ)} (9)

[0083]

[0084] In the formula, b and W are the bias and input weights of the DNN cell, respectively; H L This is the final predicted value.

[0085] Figure 5 The results of high-time-frequency estimation of landslide deformation from simulated data are displayed. The high-time-frequency estimation performance was demonstrated for TerraSAR-X, Sentinel-1, ALOS-2, COSMO-skymed, and Radarsat-2 simulation data with revisit periods of 11, 12, 14, 16, and 24 days. The simulated data (InSAR deformation / rainfall / reservoir water level) simulates real-world conditions with random errors. High-time-frequency estimation aims to maintain high stability and reliability even with data source errors. KF and KF-ARIMA calculations of the data... If it were a one-to-one mapping, the daily rainfall and reservoir water level would be immediately reflected in the deformation estimate. However, in normal circumstances, the relationship between hydrological environment and ground deformation is not one-to-one. Minor environmental changes usually do not reflect in deformation; high-density, high-intensity environmental changes are what need attention. Therefore, the model should have the ability to selectively discard information. In addition, there is a lag effect between deformation and hydrological environment to varying degrees. The results show that the high time-frequency estimation of the simulated data in these five revisit periods is less affected by random noise than KF and KF-ARIMA, and the estimation results have high stability and accuracy. The figure shows the absolute difference between the prediction results of KF, KF-ARIMA and the present invention and the actual InSAR update values. In terms of prediction accuracy, the method of the present invention has the highest accuracy, followed by the KF-ARIMA method, and the KF method has the lowest accuracy. The accuracy of the present invention is improved compared with other methods in data of different revisit periods.

[0086] Figure 6 To investigate whether the present invention has the ability to reduce the propagation of prediction errors during the InSAR deformation "window period," the prediction accuracy (root mean square error) with a revisit period of 1 to 50 days was statistically analyzed. Figure 5The prediction accuracy of this invention is significantly higher than that of KF and KF-ARIMA. KF's maximum RMSE is 12 mm, KF-ARIMA's is 6 mm, and this invention's is less than 1 mm. With increasing revisit periods (window periods from 1 to 50 days), the KF method gradually exhibits severe bias and error accumulation, indicating that the predetermined quantitative relationship between rainfall and deformation is no longer applicable in heavy rainfall. While the KF-ARIMA method improves upon error accumulation to some extent, its deformation estimation accuracy is not enhanced under the limitations of the traditional state estimation method KF. In contrast, this invention, by combining real-time updated rainfall / reservoir level fluctuation data with historical deformation sequences to generate a deformation sequence, avoids the accumulation of high-frequency time-frequency estimation errors in InSAR deformation caused by SAR revisit periods, achieving higher prediction accuracy.

[0087] The units described in the embodiments of this disclosure can be implemented in software or in hardware.

[0088] It should be understood that the various parts of this disclosure can be implemented in hardware, software, firmware, or a combination thereof.

[0089] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. A method for estimating high-frequency temporal deformation of reservoir bank landslides, characterized in that, include: Step 1: Use multi-temporal InSAR technology to obtain the slope range-direction low-frequency temporal deformation measurement values ​​of the reservoir bank landslide and geocode them; The steps for obtaining low-frequency temporal deformation measurements of the slant range of the reservoir bank landslide using multi-temporal InSAR technology include: The temporal deformation of the reservoir bank landslide was obtained by processing the acquired SAR images through image registration, main and secondary image interferometry, phase filtering, phase unwrapping, atmospheric phase removal, and multi-temporal InSAR deformation parameter estimation. Step 2: Combine daily rainfall, reservoir water level, and InSAR low-frequency time series deformation variables and their respective time information to generate a prediction dataset; Step 2 specifically includes: Multiple feature points with coherence greater than a preset value in the interferometric data are extracted from the deformation measurements to perform high-time-frequency estimation of the landslide deformation field. A prediction dataset is constructed for each feature point, wherein the prediction dataset includes InSAR time series data and its time information, rainfall time series data and its time information, and reservoir water level time series data and its time information. Step 3: Based on the time difference characteristics between high-frequency hydrological observations and low-frequency InSAR time-series deformation values, a deep learning network is established. The prediction dataset is divided into a training dataset and a test dataset. The training dataset is used to train the deep learning network embedded with the lag feature extraction module to obtain the prediction model. The feature extraction module consists of a deep learning network LSTM and a time-series deformation / hydrological modulation factor. Step 3 specifically includes: Extract time-series features from the prediction dataset; Based on the preset time delay coefficient range, the deformation time delay characteristics, rainfall time delay characteristics, and water level time delay characteristics of the prediction dataset are extracted; The nonlinear relationship between deformation time lag characteristics, hydrological observation time lag characteristics and the state to be updated is estimated, and the weights and biases are calculated accordingly. The influence weights of each feature are iteratively updated by calculating the loss value between the true state value and the calculated state value, so as to obtain the optimal nonlinear relationship between deformation characteristics, hydrological observation time lag characteristics and the state to be updated for each feature point and form a prediction model. Step 4: Input the test dataset into the prediction model, evaluate the deformation prediction accuracy using four evaluation metrics: root mean square error, mean absolute error, mean absolute error percentage, and Pearson correlation coefficient, and adjust the model's optimal response parameters.

Images