Classification and prediction method based on InSAR (Interferometric Synthetic Aperture Radar) time sequence large-range deformation results

Abstract

The invention provides a classification and prediction method based on InSAR (Interferometric Synthetic Aperture Radar) time sequence large-range deformation results, and relates to the field of land subsidence prediction. The prediction method comprises the following steps of: acquiring and verifying long-time sequence earth surface deformation data of a predetermined area by using a time sequence InSAR (Interferometric Synthetic Aperture Radar) technology; respectively completing trend clustering and spatial clustering for the long-time sequence earth surface deformation data, and extracting a settlement hot spot area; establishing a settlement hot spot area deformation prediction model based on the settlement hot spot area; and predicting the land subsidence trend of the subsidence hot spot region in the future predetermined time length by using the prediction model, and outputting a subsidence prediction map. Compared with a conventional method, the multi-point land subsidence prediction model can be established, the subsidence hot spot area is accurately extracted through unsupervised classification, and finally the future subsidence prediction result of the hot spot area is provided. Experimental results show that the method has remarkable advantages in the aspect of improving the operation efficiency and precision of the prediction model.

Description

Technical Field

[0001] The present invention relates to the field of ground subsidence prediction, and in particular to a classification and prediction method based on InSAR time series large-scale deformation results. Background Art

[0002] Spaceborne synthetic radar aperture interferometry (InSAR) technology is an effective remote sensing tool for mapping ground deformation caused by various natural causes or human activities. InSAR technology has evolved from theory to differential interferometry and then to time-series interferometry. The theory of related surface deformation measurement has become very mature and its application in urban monitoring has become more and more extensive.

[0003] In order to predict the ground subsidence in advance and take measures to prevent accidents and disasters, it is an effective method to establish an effective ground subsidence prediction model. Existing machine learning methods face two major challenges: first, high-dimensional data sets will not only slow down the learning process, but may also be mixed with misleading noise. Therefore, how to accurately extract key features from these data is crucial; second, because data such as groundwater level changes are often difficult to match the data granularity of subsidence monitoring points, rigid interpolation processing will damage the quality of training data. These two problems lead to difficulties in feature extraction and sample matching of machine learning models, affecting the accuracy of prediction and the generalization performance of the model, especially when predicting subsidence in areas with variable geological conditions. The deep learning-based method uses computers to explore deformation characteristics, which can reduce the need for complex parameters such as geology and hydrology. Among them, the long short-term memory network LSTM model improves the gradient explosion problem of recurrent neural networks, can better retain long-term memory, and capture long-term dependencies in nonlinear sequences, which corresponds to the surface time series InSAR deformation. Therefore, the LSTM model has obvious advantages in time series InSAR subsidence prediction.

[0004] However, most of the current LSTM time series prediction models only predict subsidence at a single observation point, without considering the mutual influence between points, and the model is easily affected by errors. If a large-scale subsidence data set is directly input into the model for learning, the efficiency and accuracy of the model will be seriously affected. Therefore, how to establish an efficient and accurate multi-point ground subsidence prediction model is an urgent problem to be solved. Summary of the invention

[0005] Purpose of the invention: To propose a classification and prediction method based on large-scale deformation results of InSAR time series, so as to solve the problem of large-scale InSAR time series deformation prediction.

[0006] The present invention proposes a classification and prediction method based on InSAR time series large-scale deformation results, the steps are as follows:

[0007] Use time-series InSAR technology to obtain and verify long-term surface deformation data of the predetermined area;

[0008] Perform trend clustering and spatial clustering for the long-term series surface deformation data, and extract settlement hotspot areas based on the clustering results;

[0009] Based on the settlement hotspot area, a deformation prediction model for the settlement hotspot area is established; the deformation prediction model for the settlement hotspot area is used to predict the ground settlement trend of the settlement hotspot area within a predetermined time period in the future, and a settlement prediction map is output.

[0010] In a further embodiment, the process of acquiring and verifying long-time series surface deformation data of a predetermined area using time-series InSAR technology specifically includes:

[0011] Using the same type of SAR sensor to acquire a time-series SAR image of a predetermined area;

[0012] Using time-series InSAR technology to fuse the SAR images into large-scale time-series InSAR data;

[0013] The accuracy of the large-scale time-series InSAR data is verified by using leveling and GNSS data, the optimal time-series InSAR data results are selected from the large-scale time-series InSAR data and spatially interpolated to obtain long-term surface deformation data in the predetermined area.

[0014] In a further embodiment, the process of selecting the optimal time series InSAR data result from the large-scale time series InSAR data and performing spatial interpolation on it specifically includes:

[0015] Detect missing values ​​in the large-scale time-series InSAR data and use the kernel function to calculate the covariance matrix K containing all data points:

[0016]

[0017] In the formula, Indicates the midpoint of a large range of time series InSAR data and Point The covariance between them; K is an m×m matrix, where m is the number of observations in the data set;

[0018] Construct a posterior distribution and calculate the posterior mean :

[0019]

[0020] In the formula, is the covariance matrix between missing data points and observed data points; is the covariance matrix between the observed data points; Represents observed data.

[0021] In a further embodiment, the process of completing trend clustering and space clustering for the long-term series surface deformation data includes:

[0022] Calculate the distance metric between large-scale time-series InSAR data and cluster the data points in a trend direction:

[0023]

[0024]

[0025] In the formula, is a smoothing hyperparameter; e is an exponential operator; is the activation function, Indicates the calculated distance; Represents data samples and data samples The Euclidean distance between represents an index pair, where represents the weight of the distance between the kth point in the sequence x and all the points in the sequence y; represents the soft dynamic time warping function, x and y represent two sample time series;

[0026] For each data point p in the large-scale time series InSAR data, if -If there are less than the predetermined minimum number of points in the neighborhood, point p is marked as a noise point; if point p is not a noise point, point p is considered to be a density-connected point, and point p and all its density-reachable points are recursively included in the same cluster to complete the spatial clustering of all data points.

[0027] In a further embodiment, the process of extracting the settlement hotspot area based on the clustering result specifically includes:

[0028] Calculate standard distance To observe the behavior of each cluster in space, the formula is as follows:

[0029]

[0030] In the formula, Represents the point pairs in the cluster; represents the cluster mean center; n represents the total number of clusters;

[0031] Based on the standard distances in the X and Y directions, a two-dimensional standard deviation ellipse is calculated and drawn for each cluster;

[0032] A distance threshold is set to decide whether to retain a cluster. The retained clusters are considered as settlement hotspots, while the removed clusters are considered as noise.

[0033] In a further embodiment, based on the settlement hotspot area, the process of establishing a settlement hotspot area deformation prediction model specifically includes:

[0034] Build a Bi-LSTM network model and automatically find its optimal parameters, including the number of hidden layers, loss function, number of neurons in each layer, activation function, number of iterations, and learning rate;

[0035] In the time dimension, a 60% training set, a 20% validation set, and a 20% test set are divided from the sample set of the clustering result;

[0036] Training the Bi-LSTM network model, during which the predicted output of each result node is compared with the actual InSAR observation value, and the prediction error is back-propagated back to the network by an error back-propagation algorithm;

[0037] After multiple iterations, the weights and internal parameters of the Bi-LSTM network model were adjusted, and finally the deformation prediction model of the settlement hotspot area was obtained.

[0038] In a further embodiment, the process of using the deformation prediction model for the settlement hotspot area to predict the ground settlement trend of the settlement hotspot area within a predetermined time period in the future and outputting a settlement prediction map specifically includes:

[0039] The trained deformation prediction model for settlement hotspot areas is used to predict the test set, and the root mean square error (RMSE) and mean absolute error (MAPE) are calculated to evaluate the performance of the deformation prediction model for settlement hotspot areas on the test set:

[0040] If the error is within the allowable range, the deformation prediction model for the settlement hotspot area will use the previous prediction result as part of the next input, and finally realize the prediction of the ground settlement trend of the settlement hotspot area within a predetermined period of time in the future, and output the settlement prediction map.

[0041] In addition, the present invention also proposes an urban deformation classification and trend prediction system, which includes four components: a data acquisition unit, a data processing unit, a prediction model building unit, and a result output unit.

[0042] The data acquisition unit uses the time-series InSAR technology to acquire and verify the long-term surface deformation data of the predetermined area.

[0043] The data processing unit performs trend clustering and space clustering for the long-term series surface deformation data, and extracts settlement hotspot areas based on the clustering results.

[0044] The prediction model building unit establishes a deformation prediction model for the settlement hotspot area based on the settlement hotspot area.

[0045] The result output unit uses the deformation prediction model of the settlement hotspot area to predict the ground settlement trend of the settlement hotspot area within a predetermined time period in the future, and outputs a settlement prediction map.

[0046] When the urban deformation classification and trend prediction system is running, it can automatically execute the above-disclosed classification and prediction method based on the InSAR time series large-scale deformation results.

[0047] In addition, the present invention also proposes an electronic device, which includes a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the above-disclosed classification and prediction method based on InSAR time series large-scale deformation results.

[0048] In addition, the present invention also proposes a computer-readable storage medium, which stores at least one executable instruction. When the executable instruction is executed on an electronic device, the electronic device executes the above-disclosed classification and prediction method based on InSAR time series large-scale deformation results.

[0049] Beneficial effects: The present invention discloses a classification and prediction method based on InSAR time series large-scale deformation results, which can not only provide reliable urban deformation classification and InSAR time series deformation prediction, but also does not require prior parameters of the study area. In addition, the settlement hotspot areas are extracted through cluster analysis, which significantly improves the accuracy and efficiency of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a flow chart of the classification and prediction method based on the InSAR time series large-scale deformation results in the embodiment.

[0051] Figure 2 It is the SBAS spatiotemporal baseline diagram in the embodiment.

[0052] Figure 3 It is a trend classification space result diagram in the embodiment.

[0053] Figure 4 It is a trend classification time series result diagram in the embodiment.

[0054] Figure 5 This is a diagram of the spatial classification results in the embodiment.

[0055] Figure 6It is a bar chart of the standard distances of each cluster in the embodiment.

[0056] Figure 7 It is a two-dimensional standard deviation ellipse diagram of each cluster in the embodiment.

[0057] Figure 8 It is a map of the sedimentation hotspot areas retained in the embodiment.

[0058] Fig. 9 It is a scatter plot of the correlation between the predicted value of the settlement hotspot area No. 13 and the InSAR value in the embodiment.

[0059] Fig.10 This is a settlement prediction map for the settlement hotspot area in the embodiment after 120 days.

[0060] Fig.11 This is a diagram showing the timing results of prediction of some settlement hotspot areas in the embodiment. DETAILED DESCRIPTION

[0061] In the following description, a large number of specific details are provided to provide a more thorough understanding of the present invention. However, it is apparent to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, in order to avoid confusion with the present invention, some technical features known in the art are not described.

[0062] In view of the objective problems existing in the background technology, the present invention proposes a classification and prediction method based on the InSAR time series large-scale deformation results. First, the time series InSAR technology is used to obtain the large-scale time series deformation in urban areas, and the K-means algorithm with Soft-DTW is used to extract different settlement trend categories, and then the DBSCAN algorithm is used to segment the categories with the same settlement trend in terms of spatial distance. Then, the settlement hotspot areas are screened out by calculating the standard distance for the many subsets after segmentation. Then, a Bi-LSTM prediction model is established for each hotspot area, and after evaluating the accuracy, a series of continuous future prediction results are finally generated.

[0063] The examples show that the present invention can provide reliable InSAR time series deformation prediction without the need for prior parameters of the study area. In addition, the settlement hotspot area is extracted through cluster analysis, which significantly improves the accuracy and efficiency of the prediction model.

[0064] The technical solution of the present invention is described in detail below through specific embodiments.

[0065] Embodiment 1:

[0066] This embodiment discloses a classification and prediction method based on InSAR time series large-scale deformation results, and its flow chart is shown in Figure 1, this method is implemented by the following steps:

[0067] S1: Acquiring and verifying urban surface deformation data using time-series InSAR technology

[0068] In this step, the same type of SAR sensor is first used to obtain time-series SAR images of the urban area, and then the time-series InSAR technology is used to obtain the long-term surface deformation information of the study area, and the accuracy of the time-series InSAR results is verified using leveling, GNSS and other data.

[0069] At present, the main methods used for time series InSAR processing include permanent scatterer synthetic aperture radar interferometry (PS-InSAR) and synthetic aperture radar interferometry based on small baseline sets (SBAS-InSAR). The core concept of PS-InSAR technology is to select one image as the main image by using multiple single-view SAR images covering the same study area, and then align other SAR images with the main image. In the time series, based on the stability of amplitude and phase information, permanent scatterers (PS) are identified and selected as targets. Through interference and de-terraining processing, the differential interferometric phase based on the PS target is obtained, and the differential interferometric phase of adjacent PS targets is secondarily differentiated. According to the characteristics of different phase components in the interferometric phase after two differences, a deformation phase model is constructed to estimate the deformation and residual terrain information.

[0070] SBAS-InSAR technology is based on the interference pairs of multiple main images, and uses high coherence points to restore the time series deformation information of the study area. The workflow of this technology is as follows: First, the time-space baseline of multiple SAR images covering different time points in a specific area is calculated, and the interference pairs are selected according to the appropriate time-space baseline threshold. Then, the selected interference pairs are subjected to differential interference processing and phase unwrapping. Finally, according to the formation of subsets of freely combined interference patterns in the interference pattern, the least squares method or singular value decomposition (SVD) method is used to estimate the deformation parameters of the phase equation composed of all interference patterns. The advantage of SBAS-InSAR technology is that it overcomes the problem of poor coherence of some interference patterns that may be caused by the selection of a single reference image in the PS-InSAR method, reduces the demand for SAR data, and improves computational efficiency.

[0071] Finally, the best time-series InSAR results from the two methods were selected and spatially interpolated using the Kriging interpolation method to obtain the fitted settlement results over a large area in urban areas.

[0072] S2: Classify and process large-scale time-series InSAR data from two dimensions: trend and space, and extract subsidence hotspots

[0073] (1) Missing value filling:

[0074] Due to system reasons, some images of the Sentinel-1 satellite may be missing. In order to keep the time interval of the data completely consistent so that the model can learn the sedimentation trend more accurately, it is necessary to interpolate the missing parts.

[0075] First, detect the missing time in the data set, interpolate the continuous function based on the discrete data, so that the continuous curve passes through all discrete points, and estimate the approximate value of the function at the missing point to complete the data filling. The missing value filling method uses Gaussian process regression:

[0076] Use the kernel function to calculate the covariance matrix K containing all data points. For each point in the data set X and , the kernel function gives the covariance between them:

[0077]

[0078] Here, K is an m×m matrix and m is the number of observations in the dataset.

[0079] In the case of missing data, the goal is to fill in the missing values ​​in the data matrix Y. To do this, a posterior distribution can be constructed that takes into account both the observed data and the missing data. Represents the observed data, using represents missing data, then the posterior mean It can be calculated by the following formula:

[0080]

[0081] in, is the covariance matrix between missing and observed data points, is the covariance matrix between the observed data points.

[0082] (2) Trend clustering:

[0083] This step uses statistical criteria to classify time series sedimentation into different cluster labels. The K-means algorithm initializes m clusters at random locations, and the selection of m is based on the elbow method. For each new data point, the method calculates the distance between the point and each cluster, and the point is assigned to the cluster with the smallest distance. After each data point is clustered into one of the clusters, the cluster center moves, and the process ends when the convergence threshold of the cluster center movement is reached. The main limitation of using K-means is that it does not consider the temporality of the data set and only regards the input time series as a set of features with the number of time steps. In most cases, although two time series have very similar shapes as a whole, these shapes are not aligned on the x-axis. The unique advantage of using DTW (Dynamic Time Warping) is that it efficiently handles the nonlinearity and speed changes of time series data by extending and shortening the time series, providing flexible sequence matching without the need for consistent sequence length.

[0084]

[0085] In the formula, x and y represent two sample time series, d is the Euclidean distance, Represents an index pair.

[0086] Similarly, using DTW as a minimization function for calculating distance also has its limitations, that is, it cannot be differentiable everywhere. Another variant of DTW, Soft-DTW, replaces the min function in DTW with a softmin function. By introducing a smoothing mechanism, it reduces noise sensitivity and provides more robust time series matching, which is especially suitable for scenarios such as InSAR where data fluctuations exist.

[0087]

[0088]

[0089] Where γ is a smoothing hyperparameter, e is an exponential operator, Indicates the calculated distance.

[0090] Therefore, this step uses the Soft-DTW algorithm to configure the K-means clustering algorithm as a distance metric to calculate the distance between InSAR time series data and classify the time series data sets based on trends.

[0091] (3) Spatial clustering:

[0092] Two points far apart may have similar subsidence trends due to different reasons, thus affecting the researchers' judgment. In order to eliminate such influences, the output of the K-means algorithm is passed to the density-based spatial clustering of applications with noise (DBSCAN) algorithm. The DBSCAN algorithm can divide areas with sufficient density of points into the same cluster. The algorithm includes two input parameters: distance threshold And the minimum number of points min. The algorithm steps of DBSCAN can be summarized as follows: for each point p, if -If there are less than min points in the neighborhood, point p is marked as a noise point; if p is not a noise point, p is considered to be a density-connected point, and p and all its density-reachable points are recursively included in the same cluster. This process will continue until no more points can be included in the current cluster. The algorithm is completed when all points have been visited and each point has been assigned to a cluster or marked as noise.

[0093] Compared with directly applying DBSCAN to the original InSAR data for classification, the input of this algorithm is based on the results of trend classification. The purpose is to make a finer division of the unsupervised clustering results to facilitate the subsequent extraction of subsidence hotspots.

[0094] (4) Extraction of settlement hotspot areas:

[0095] After obtaining many clustering results from semi-supervised learning, the next goal is to find important signals from the generated clustering graph and remove unimportant signals. This step is done by calculating the standard distance To observe the behavior of each cluster in space, the formula is as follows:

[0096]

[0097] in( , ) represents the point pairs in the cluster, ( , ) represents the cluster mean center, and n represents the total number of clusters. and , i.e., the standard distance in the X and Y directions, calculates and plots the two-dimensional standard deviation ellipse of each cluster, which can visually examine the spatial variability of a single cluster. The distance threshold is set to decide whether to retain a cluster. The retained clusters will be regarded as subsidence hotspots, indicating obvious surface deformation activities in the city, while the removed clusters represent unimportant signals and are regarded as noise.

[0098] S3: Establish a deformation prediction model for settlement hotspots and use it to predict future ground settlement trends

[0099] (1) Bi-LSTM network construction:

[0100] The core of deep learning LSTM model prediction lies in the three gate structures in the hidden layer. The gate is a way to selectively let information pass.

[0101] The first step is for the "Forget Gate" to decide what information to forget from the cell state. By looking at the output of the previous moment and the current input , using the Sigmoid function to decide whether to save or delete information. Its expression is as follows:

[0102]

[0103] The second step is to use the "input gate" to decide what information to store in the cell state. Decide which values ​​to update, The layer is responsible for creating a vector of candidate values , which is added to the cell state The expression is as follows:

[0104]

[0105]

[0106] Previous cell state pass Forget the old memories and go through Updated to after adding new memory . Its expression is as follows:

[0107]

[0108] The third step is for the "output gate" to decide what information needs to be output. The principle is the same as above. Then change the cell state Normalize to [-1, 1] using the tanh function and multiply it by Output to get the part that the model needs to output . Its expression is as follows:

[0109]

[0110]

[0111] in, , , , They represent input gate, forget gate, cell state and output gate respectively; is the weight matrix under the output gate, is the deviation matrix under the output gate; t represents the current moment; t-1 represents the previous moment; Represents the output at the previous moment; Represents the input at the current moment; Indicates the cell state at the current moment; Indicates the cell state at the previous moment; Represents the output at the current moment; and They are sigmoid and hyperbolic tangent functions respectively. In summary, through the interaction between the forget gate, input gate, output gate and activation function, a complete LSTM neuron can realize the screening of effective information in time series data.

[0112] The bidirectional long short-term memory neural network (Bi-directional-LSTM) is composed of two layers of LSTM, forward and reverse. It makes predictions through bidirectional learning data, which improves the accuracy of the prediction results. The traditional LSTM is a one-way transmission and can only extract past sedimentation information for training and prediction. It is only affected by past sedimentation data. Bi-LSTM increases the ability to extract information in the reverse direction and can extract sequence information in both the forward and backward directions at the same time, improve the correlation between the forward and backward data learning, and jointly output the sedimentation data.

[0113] (2) Bi-LSTM model training:

[0114] The data input to the input layer should be a normalized small cluster sample set. Data normalization ensures that all features are on the same scale, which helps the model better understand the relative importance of each feature, thereby improving the accuracy of prediction; normalized data helps speed up the convergence of the neural network training process. This example selects the MinMax algorithm to limit the data to the range of 0 to 1 while maintaining the relative size relationship of the sample features.

[0115]

[0116] In the formula, Represents the value of a single data. is the minimum value of the column where the data is located. It is the maximum value of the column where the data is located.

[0117] In the time dimension, the sample set is divided into the first 60% training set for model training, and the first 60%-80% as the validation set to facilitate the adjustment of model parameters during the training process. The last 20% is selected as the test set to finally evaluate the generalization ability of the model.

[0118] Before model training, the grid algorithm is used to optimize hyperparameters to find the best hyperparameters, including the learning step size of the LSTM network, the number of hidden layers, the number of neurons in each layer, the learning rate, etc. The remaining parameters such as the loss function, activation function, and the number of iterations of the early stopping threshold are set based on experience.

[0119] During the training process, the predicted output of each Bi-LSTM result node is compared with the actual InSAR observation value. The prediction error is back-propagated back to the network through the error back-propagation algorithm. Through multiple iterations of the Adam optimization algorithm, the Bi-LSTM model continuously adjusts the weights and internal parameters in the network, gradually optimizes the model to minimize the prediction error, and finally generates a deep learning Bi-LSTM prediction model with stable weight parameters.

[0120] (3) Model accuracy assessment:

[0121] This step uses two indicators, root mean square error and mean absolute error percentage, to test the learning or prediction results of the training set, validation set, and test set.

[0122] The root mean square error (RMSE) quantifies the absolute degree to which the predicted value deviates from the actual value:

[0123]

[0124] In the formula, represents the predicted value, Verify the sample value.

[0125] The mean absolute error (MAPE) provides the average percentage deviation between the model's predictions and the actual values:

[0126]

[0127] In the formula, represents the predicted value, Verify the sample value.

[0128] (4) Bi-LSTM model prediction:

[0129] After the accuracy is achieved, starting from the last few time steps of the test set, for each new prediction, the previous prediction result will be used as part of the next input, thereby recursively using the model's predictions to generate 10 consecutive future predictions and storing the results for further analysis.

[0130] Embodiment 2:

[0131] Taking the process disclosed in Example 1 as an application example, the process of the present invention is further explained:

[0132] Step 1: Use time-series InSAR technology to obtain large-scale time-series deformation of the city.

[0133] The 78-scene ascending orbit Sentine-1A data were used, with VV polarization and IW imaging modes, covering the period from October 2, 2020 to October 23, 2023. The small baseline subsat (SBAS) technique was used to solve the 1000-meter cumulative deformation sequence along a subway line in Nanjing, and a total of 54,252 points were obtained. The reliability of the InSAR observation results was confirmed by leveling data. Figure 2 It is a time-space baseline map.

[0134] Step 2: Conduct trend and spatial cluster analysis to identify and extract deformation hotspots.

[0135] After testing, the InSAR time series data were found to be missing on some dates, including ten missing one-period images, two missing two-period images, and one missing three-period image. After interpolation processing, a total of 94 periods of data were obtained.

[0136] The Soft-DTW algorithm is used to deploy the Time Series K-means (TSKM) algorithm to perform time series clustering. The classification results are shown in the following figure: Figure 3 As shown, in terms of timing Figure 4 shown.

[0137] The parameters ε of the DBSCAN algorithm are set to 0.5 and min is set to 5, and the clusters with similar settlement obtained by the TSKM algorithm are spatially divided into small clusters. The classification results are shown in Figure 2. Figure 5 shown.

[0138] Analyze the clustering results and calculate the standard distance in each direction for all clusters. Figure 6 The standard distance bar graph for each cluster. The two-dimensional standard deviation ellipse for each cluster is shown in Figure 7 In this example, a 0.5 km distance threshold is used to decide whether to keep or delete a cluster. The final retained settlement hotspot area is shown in Figure 8 shown.

[0139] Step 3: Train the deep learning model and use the Bi-LSTM model to predict the deformation data of the settlement hotspot area.

[0140] In this example, the TensorFlow framework of Python language is used to build a Bi-LSTM neural network model. The number of hidden layers is set to 2, the loss function is the mean square error MSE, the number of neurons is 64, the number of iterations is 50, the early stopping threshold is 15, and the dropout parameter is 0.1. In the first 64 training set data, 12 input data are used to predict the next data through a sliding window. The 15-period validation set data is used to evaluate the model accuracy of each model trained in the training set, and the best model is saved. After the training is completed, the determination coefficient R is introduced 2 The two evaluation indicators of RMSE and RMSE are used to judge whether the model is overfitting. The results of evaluating the quality of the regression model are as follows: Fig. 9 As shown in the figure, since there is no significant difference between the two indicators of the training set and the test set, it can be considered that the model has good generalization ability and prediction performance.

[0141] The accuracy of the prediction results is as follows:

[0142] Table 1 Prediction accuracy of each subsidence hotspot area

[0143] Cluster number Training set RMSE / mm Test set RMSE / mm Training set MAPE Test set MAPE 10 0.57 0.93 41% 66% 13 0.48 0.73 36% 52% 15 0.44 0.72 33% 55% 17 0.38 0.54 30% 41% 19 0.48 0.91 37% 71% 21 0.43 0.87 30% 67% 34 0.73 1.87 54% 145% 35 0.38 0.66 29% 51% 38 0.33 0.62 24% 48% 41 0.33 0.72 25% 53% 43 0.34 0.75 25% 58% 45 0.40 0.66 29% 51% 49 0.49 0.59 36% 45%

[0144] The maximum RMSE of the test set of 13 clusters in the study area is 1.87 mm, the minimum is 0.54 mm, and the average is 0.79 mm. The maximum MAPE is 145%, the minimum is 41%, and the average is 57%. From the above results, it can be seen that the algorithm can provide accurate deformation prediction for urban areas.

[0145] After the accuracy reaches the target, starting from the last few time steps of the test set, for each new prediction, the previous prediction result will be used as part of the next input, thus generating 10 consecutive future predictions. The precipitation prediction results for the precipitation hotspot area after the last image are as follows: Fig.10 As shown in Figure 2, the prediction time series results of some subsidence hotspot areas are as follows: Fig.11 shown.

[0146] Embodiment 3:

[0147] The present embodiment discloses a system for urban deformation classification and trend prediction, which includes four components: a data acquisition unit, a data processing unit, a prediction model construction unit, and a result output unit. The data acquisition unit uses time-series InSAR technology to obtain and verify the long-term series surface deformation data of a predetermined area. The data processing unit performs trend clustering and space clustering for the long-term series surface deformation data, and extracts the settlement hotspot area based on the clustering results. The prediction model construction unit establishes a deformation prediction model for the settlement hotspot area based on the settlement hotspot area. The result output unit uses the deformation prediction model for the settlement hotspot area to predict the ground settlement trend of the settlement hotspot area within a predetermined period of time in the future, and outputs a settlement prediction map.

[0148] The urban deformation classification and trend prediction system is implemented by an electronic device that automatically executes the classification and prediction method based on the InSAR time series large-scale deformation results disclosed in the above embodiment. The electronic device includes a processor, a memory, a communication interface and a communication bus. The processor, the memory and the communication interface communicate with each other through the communication bus. The memory is used to store at least one executable instruction, which enables the processor to execute Figure 1 The classification and prediction method based on the InSAR time series large-scale deformation results disclosed in Examples 1 to 2, the specific process is not repeated here.

[0149] In addition, in the actual application process of the classification and prediction method based on the InSAR time series large-scale deformation results disclosed in the above embodiment, its operation logic can be written into a set of executable instructions, and the executable instructions can be written into a storage medium and run on an electronic device.

[0150] More specific examples of the computer-readable storage medium mentioned in this embodiment may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. The computer-readable storage medium may include a data signal propagated in a baseband or as part of a carrier wave, wherein a readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The readable signal medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, device, or device.

[0151] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes may be made to it in form and detail without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A classification and prediction method based on InSAR time series large-scale deformation results, characterized in that: The steps include: Use time-series InSAR technology to obtain and verify long-term surface deformation data of the predetermined area; Perform trend clustering and spatial clustering for the long-term series surface deformation data, and extract settlement hotspot areas based on the clustering results; Based on the settlement hotspot area, a deformation prediction model for the settlement hotspot area is established; the deformation prediction model for the settlement hotspot area is used to predict the ground settlement trend of the settlement hotspot area within a predetermined time period in the future, and a settlement prediction map is output.

2. The classification and prediction method based on InSAR time series large-scale deformation results according to claim 1 is characterized in that: The method of obtaining and verifying long-time series surface deformation data of a predetermined area by using time-series InSAR technology specifically includes: Using the same type of SAR sensor to acquire a time-series SAR image of a predetermined area; Using time-series InSAR technology to fuse the SAR images into large-scale time-series InSAR data; The accuracy of the large-scale time-series InSAR data is verified by using leveling and GNSS data, the optimal time-series InSAR data results are selected from the large-scale time-series InSAR data and spatially interpolated to obtain long-term surface deformation data in the predetermined area.

3. The classification and prediction method based on InSAR time series large-scale deformation results according to claim 2 is characterized in that: Selecting the best time series InSAR data result from the large-scale time series InSAR data and performing spatial interpolation on it specifically includes: Detect missing values ​​in the large-scale time-series InSAR data and use the kernel function to calculate the covariance matrix K containing all data points: ； In the formula, Indicates the midpoint of a large range of time series InSAR data and Point The covariance between them; K is an m×m matrix, where m is the number of observations in the data set; Construct a posterior distribution and calculate the posterior mean : ； In the formula, is the covariance matrix between missing data points and observed data points; is the covariance matrix between the observed data points; Represents observed data.

4. The classification and prediction method based on InSAR time series large-scale deformation results according to claim 2 is characterized in that: The trend clustering and space clustering are completed for the long-term series surface deformation data, including: Calculate the distance metric between large-scale time-series InSAR data and cluster the data points in a trend direction: ； ； In the formula, is a smoothing hyperparameter; e is an exponential operator; is the activation function, Indicates the calculated distance; Represents data samples and data samples The Euclidean distance between represents a path, which is a mapping from sequence x to sequence y, where represents the weight of the distance between the 𝑘th point in the sequence x and all the points in the sequence y; represents the soft dynamic time warping function, x and y represent two sample time series; For each data point p in the large-scale time series InSAR data, if -If there are less than the predetermined minimum number of points in the neighborhood, point p is marked as a noise point; if point p is not a noise point, point p is considered to be a density-connected point, and point p and all its density-reachable points are recursively included in the same cluster to complete the spatial clustering of all data points.

5. The classification and prediction method based on InSAR time series large-scale deformation results according to claim 4 is characterized in that: Based on the clustering results, the settlement hotspots are extracted, including: Calculate standard distance To observe the behavior of each cluster in space, the formula is as follows: ； In the formula, Represents the point pairs in the cluster; represents the cluster mean center; n represents the total number of clusters; Based on the standard distances in the X and Y directions, a two-dimensional standard deviation ellipse is calculated and drawn for each cluster; A distance threshold is set to decide whether to retain a cluster. The retained clusters are considered as settlement hotspots, while the removed clusters are considered as noise.

6. The classification and prediction method based on InSAR time series large-scale deformation results according to claim 5 is characterized in that: Based on the settlement hotspot area, a deformation prediction model for the settlement hotspot area is established, which specifically includes: Build a Bi-LSTM network model and automatically find its optimal parameters, including the number of hidden layers, loss function, number of neurons in each layer, activation function, number of iterations, and learning rate; In the time dimension, a 60% training set, a 20% validation set, and a 20% test set are divided from the sample set of the clustering result; Training the Bi-LSTM network model, during which the predicted output of each result node is compared with the actual InSAR observation value, and the prediction error is back-propagated back to the network by an error back-propagation algorithm; After multiple iterations, the weights and internal parameters of the Bi-LSTM network model were adjusted, and finally the deformation prediction model of the settlement hotspot area was obtained.

7. The classification and prediction method based on InSAR time series large-scale deformation results according to claim 6 is characterized in that: The deformation prediction model for the settlement hotspot area is used to predict the ground settlement trend of the settlement hotspot area within a predetermined time period in the future, and a settlement prediction map is output, specifically including: The trained deformation prediction model for settlement hotspot areas is used to predict the test set, and the root mean square error (RMSE) and mean absolute error (MAPE) are calculated to evaluate the performance of the deformation prediction model for settlement hotspot areas on the test set: If the error is within the allowable range, the deformation prediction model for the settlement hotspot area will use the previous prediction result as part of the next input, and finally realize the prediction of the ground settlement trend of the settlement hotspot area within a predetermined period of time in the future, and output the settlement prediction map.

8. An electronic device, characterized in that: The device comprises: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the classification and prediction method based on InSAR time series large-scale deformation results as described in any one of claims 1 to 7 is implemented.

9. A computer-readable storage medium, characterized in that: The storage medium stores at least one executable instruction. When the executable instruction is executed on the electronic device, the electronic device executes the classification and prediction method based on InSAR time series large-scale deformation results as described in any one of claims 1 to 7.

Images