A land subsidence prediction method based on InSAR time series data and a CNN-LSTM model

Abstract

A ground subsidence prediction method based on InSAR time series data and CNN-LSTM model. The method first obtains the InSAR ground surface deformation time series data of the monitoring points in the research area, after preprocessing, the sliding window technology is used to construct the training sample set containing the historical observation characteristics and the future prediction label; then the training sample is input into the CL-TSF model constructed. The CL-TSF model adopts the fusion architecture of series connection of CNN and LSTM, specifically containing multiple one-dimensional convolution layers, batch normalization layers and maximum pooling layers for deep mining of local features of data and noise resistance processing, and coupling introduction of multiple LSTM units of Dropout regularization mechanism to capture the long distance time series dependence relationship of deformation data, finally through the full connection layer to realize the multi-step subsidence prediction. The method of the application significantly improves the calculation efficiency, prediction accuracy and generalization ability of the model of the ground subsidence prediction.

Description

Technical Field

[0001] This invention belongs to the field of geological disaster monitoring and computer deep learning technology, specifically relating to a method for predicting ground deformation time series based on synthetic aperture radar interferometry (InSAR) time series data and utilizing convolutional neural network (CNN) and long short-term memory network (LSTM) models. Background Technology

[0002] With the acceleration of urbanization and large-scale mining of mineral resources, land subsidence has become a common, slowly evolving geological hazard. Land subsidence not only leads to building cracks and damage to underground pipelines, but in severe cases, it can even trigger collapses due to uneven settlement, posing a significant threat to people's lives and property, as well as major engineering projects in mining areas and cities. Therefore, conducting high-precision land subsidence monitoring and evolution trend prediction is of paramount importance for disaster prevention and mitigation, as well as the maintenance of engineering projects throughout their entire lifecycle.

[0003] Currently, land subsidence monitoring mainly relies on traditional leveling surveys, GPS monitoring, and the rapidly developing Synthetic Aperture Radar Interferometry (InSAR) technology. Among these, time-series InSAR technologies (such as PS-InSAR and SBAS-InSAR) have become the primary means of acquiring long-term data on large-scale land subsidence due to their advantages of being available 24 / 7, covering a wide area, and having high spatial resolution. However, acquiring massive amounts of subsidence monitoring data is only the first step. How to extract evolutionary patterns from these historical time-series data, which contain noise and complex fluctuations, and accurately predict future subsidence trends remains a challenging and hot research topic.

[0004] Existing land subsidence prediction methods are mainly divided into two categories: physical-driven models and data-driven models. Physical-driven models (such as fluid-structure interaction models) have clear mechanisms, but they often rely on detailed hydrogeological parameters (such as permeability coefficient and compression modulus), which are extremely difficult to obtain accurately in practical engineering, severely limiting their widespread application. Traditional data-driven methods (such as regression analysis, grey model GM(1,1), and ARIMA models) are computationally simple, but most are based on linear assumptions, making it difficult to capture the nonlinear, dynamic fluctuations, and complex time-dependent characteristics commonly present in land subsidence. This results in low prediction accuracy and poor generalization ability when facing complex conditions such as mining disturbances and seasonal rainfall.

[0005] In recent years, deep learning technology has demonstrated excellent performance in handling time series prediction problems, but single models still have limitations. While a single convolutional neural network (CNN) can effectively extract local features from InSAR data and filter out some high-frequency noise through convolutional kernels, it lacks a time gating mechanism and is weak in handling the time dependencies of long sequences, making it difficult to capture the long-term cumulative trend of sedimentation. On the other hand, while a single long short-term memory network (LSTM) is good at capturing the correlation between long-term sequences, its feature extraction efficiency is low when directly dealing with raw InSAR data containing atmospheric delay or phase noise. It is also susceptible to gradient problems or overfitting due to noise interference, making it difficult to balance local feature mining and global trend prediction.

[0006] In view of this, existing land subsidence prediction methods suffer from insufficient feature mining, difficulty in capturing long-term dependencies, and insufficient prediction accuracy when processing nonlinear and noisy InSAR time series data. Therefore, there is an urgent need for a comprehensive prediction model that can simultaneously integrate the local feature extraction capabilities of CNN and the time series memory capabilities of LSTM, so as to improve the robustness and accuracy of land subsidence prediction and provide a scientific basis for geological disaster early warning. Summary of the Invention

[0007] The purpose of this invention is to address the aforementioned problems in existing technologies, namely, the insufficient feature extraction, poor noise resistance, and difficulty in capturing long-term dependencies, leading to low prediction accuracy, when a single deep learning model processes InSAR time-series data. This invention provides a land subsidence prediction method based on InSAR time-series data and a CNN-LSTM model. The method of this invention fully combines the efficient feature extraction capability of CNNs with the temporal memory advantage of LSTMs, compensating for the shortcomings of InSAR technology in deformation trend prediction. Compared to single convolutional neural networks (CNNs) and long short-term memory networks (LSTMs), the method of this invention effectively overcomes the limitations of CNN models in capturing long-term dependencies and LSTM models in extracting local features. By fusing the feature extraction advantages of CNNs with the temporal memory capabilities of LSTMs, it significantly improves the computational efficiency, prediction accuracy, and generalization ability of the land subsidence prediction model.

[0008] The above-mentioned objectives of the present invention are achieved by the following technical means: A method for predicting land subsidence based on InSAR time series data and a CNN-LSTM model includes the following steps: Step S1, Data Acquisition and Solving: The small baseline set (SBAS-InSAR) technique is used to perform interferometric processing on the P-scene SAR image of the study area to obtain the cumulative deformation time series data of a large number of monitoring points in the study area at T observation times; Step S2, Data Preprocessing and Sample Construction: Outlier removal and missing value imputation are performed on the cumulative deformation time series data of the monitoring points. The data is mapped to the standard interval using the normalization algorithm. The time series supervision samples are constructed based on the sliding window algorithm. The data of the first L time steps are used as input features, and the data of the last S time steps are used as prediction labels. The sample set is divided into training set, validation set and test set. Step S3: Construct the CL-TSF spatiotemporal prediction model: Build a serially coupled model that includes a feature extraction module and a temporal memory module. The CNN module is used to extract local high-dimensional features of the time series data and filter out noise, while the LSTM module is used to capture the long-distance temporal dependencies of the sedimentation data. Step S4, Model Training: Iteratively train the CL-TSF spatiotemporal prediction model using the training set constructed in step S2, and update the model parameters using the Adam optimizer and mean squared error (MSE) loss function until the model converges; Step S5, Multi-step prediction and inverse normalization: Input the cumulative deformation time series to be predicted into the trained CL-TSF model, output the predicted values ​​for the next S time points, and restore them to the true cumulative deformation values ​​through inverse normalization. Step S6, Result Visualization: Combine the predicted deformation value obtained in step S5 with the geographic coordinate information of the monitoring points, and use MATLAB to generate a spatiotemporal distribution map of ground subsidence in the study area at S future times, so as to realize the visualization of the subsidence trend.

[0009] Furthermore, the specific process of obtaining the cumulative deformation time series data in step S1 is as follows: Step S1.1: Select SAR images covering the study area of ​​P scene, filter interferometric image pairs according to the spatiotemporal baseline threshold, remove the terrain phase using the external digital elevation model, and generate differential interferogram; Step S1.2: The differential interferogram is enhanced by filtering using the Goldstein filtering algorithm, and the phase is unwrapped using the minimum cost flow (MCF) algorithm; Step S1.3: By selecting stable control points, orbit refinement and re-leveling are performed. Atmospheric delay phase is estimated and removed. The Singular Value Decomposition (SVD) method is used to solve for the cumulative deformation time series D={d1, d2, d3, ..., dt} of each monitoring point at T observation times. T}

[0010] Furthermore, the specific implementation process of step S2 is as follows: Step S2.1, Data Normalization: The original deformation time series data of each monitoring point is standardized using the MinMaxScaler method, mapping the data to the [0,1] interval. The calculation formula is as follows: (1.1) in, This represents the original deformation value at the current moment. For the normalized data, This is the entire time series data set for this monitoring point. and These are the minimum and maximum values ​​of the sequence, respectively. Step S2.2: Constructing Time-Series Supervised Samples: The univariate time series is converted into supervised learning sample pairs using the sliding window algorithm; the input sequence length is set to L, and the prediction step size is S; for the normalized cumulative variable time series D={d1, d2, d3, ... d... T}, the input features of the t-th sample and corresponding tags Represented as: (1.2) (1.3) The window slides forward with a step size of 1 until it covers the entire sequence, generating a total of TL-S+1 samples.

[0011] Step S2.3, Dataset Partitioning: Adjust the constructed sample set to a three-dimensional tensor format (Samples, TimeSteps, Features), and finally divide the sample set into training set, validation set and test set according to time sequence.

[0012] Furthermore, the CL-TSF spatiotemporal prediction model in step S3 adopts a serially coupled architecture, with the following specific structural parameters: The CL-TSF spatiotemporal prediction model includes, in order of data flow, a feature extraction module, a temporal memory module, and a fully connected output module. The feature extraction module comprises two sets of stacked convolutional units: the first set includes a filter, a one-dimensional convolutional layer (Conv1D), and corresponding batch normalization and max pooling layers; the second set includes a filter, a one-dimensional convolutional layer, and corresponding batch normalization and max pooling layers; the one-dimensional convolution operation formula is as follows: (1.4) Here, convolution is the operation. For the first lLayer output features, As weight, For bias, For convolution operations, It is the ReLU activation function. This refers to the position index within the convolution kernel. for l-1 The input feature value of the layer at the corresponding position, and j is the position index of the output feature sequence; The temporal memory module contains two LSTM layers: the first LSTM layer contains neurons, is set to return a sequence to output the complete sequence, and is followed by a random deactivation layer (Dropout); the second LSTM layer contains neurons, is set to return a complete sequence to output the final state, and is followed by a random deactivation layer (Dropout). The fully connected output module contains a fully connected layer with the number of neurons equal to the prediction step size S and the activation function being a linear function.

[0013] Furthermore, the one-dimensional convolutional layer output feature sequence obtained by formula (1.4) is used as the input of the LSTM layer, and this feature sequence is denoted as X = { x 1, x 2, ... x t ,……},in x t express t The LSTM layer takes features extracted from a one-dimensional convolutional layer at each time step; the LSTM layer captures long-range dependencies through a gating mechanism, and its internal unit update formula is as follows: (1.5) (1.6) (1.7) (1.8) (1.9) (1.10) in, This is the current input time. This is the hidden state from the previous moment. In cellular state, , , These are the forget gate, input gate, and output gate, respectively. For the Sigmoid function, Represents the Hadama product; through the forgetting gate This determines how much historical information to retain, thereby addressing the gradient vanishing problem in long sequences. , , , These represent the weights corresponding to the forget gate, input gate, candidate cell state, and output gate, respectively. , , , These represent the biases corresponding to the forget gate, input gate, candidate cell state, and output gate, respectively. The candidate cell states for the LSTM layer.

[0014] Furthermore, the model training process in step S4 is as follows: Step S4.1, Parameter Configuration: Select the Adam optimizer and set the learning rate; select the mean squared error (MSE) as the loss function, as shown in the following formula: (1.11) in This represents the batch sample size. For true normalized values, This is a predicted value; Step S4.2, Iterative Optimization: Input the training set into the model, set the batch size and the maximum number of iterations (Epochs); use the backpropagation algorithm to calculate the gradient and update the weights; Step S4.3, Early Stopping Strategy: During training, use the test set for validation and monitor the validation loss. If the loss value does not decrease within several consecutive rounds, trigger the early stopping mechanism and save the current optimal model parameters.

[0015] Furthermore, step S6 uses root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) as evaluation indicators, and the calculation formulas are as follows: (1.12) (1.13) (1.14) in, To predict the total number of samples, The true deformation value after inverse normalization (unit: mm). This is the predicted deformation value after inverse normalization.

[0016] Compared with the prior art, the present invention has the following advantages: (1) High data utilization: This invention makes full use of the high-density long-term time series data obtained by SBAS-InSAR, and ensures that the model has enough historical information for learning by using the strategy of predicting the next 5 periods from the first 65 periods.

[0017] (2) High model accuracy: The CL-TSF model constructed in this invention integrates the feature extraction capability of CNN and the temporal memory capability of LSTM. CNN effectively filters out residual atmospheric and phase noise in InSAR data, while LSTM accurately captures the nonlinear cumulative trend of settlement. Experiments show that when predicting the settlement values ​​of the last 5 periods, the RMSE and MAPE indices of this model are significantly better than those of a single LSTM or CNN model.

[0018] (3) Intuitive results: This invention not only outputs single-point prediction values, but also generates a regional settlement spatiotemporal distribution map through step S6, which can intuitively identify potential settlement funnel areas and provide direct decision-making basis for disaster prevention and mitigation. Attached Figure Description

[0019] Figure 1 A flowchart illustrating a ground subsidence prediction method based on InSAR time-series data and a CNN-LSTM model, provided as an embodiment of the present invention. Figure 2 An architecture based on InSAR time series data and a CNN-LSTM model is provided for embodiments of the present invention. Figure 3 An internal structure of a hybrid model unit based on InSAR time series data and CNN-LSTM model at time step t, provided in an embodiment of the present invention. Figure 4 A detailed network architecture based on InSAR time series data and a CNN-LSTM model is provided for embodiments of the present invention. Figure 5 This invention provides a spatial deformation map based on InSAR time series data and a CNN-LSTM model at a representative time step, as provided in an embodiment of the invention.

[0020] Figure 6 The error distribution of the model of this invention in all test samples in different models is shown below: (a-1) is the MAE evaluation result of CNN-LSTM, (a-2) is the MAPE evaluation result of CNN-LSTM, (a-3) is the RMSE evaluation result of CNN-LSTM; (b-1) is the MAE evaluation result of CNN, (b-2) is the MAPE evaluation result of CNN, (b-3) is the RMSE evaluation result of CNN; (c-1) is the MAE evaluation result of LSTM, (c-2) is the MAPE evaluation result of LSTM, (c-3) is the RMSE evaluation result of LSTM.

[0021] Figure 7 Comparison of multi-step prediction results for monitoring point 11 in step 5 of this embodiment of the invention Figure 8 For the comparison of multi-step prediction results of monitoring point 421 in step 5 of the embodiment of the present invention Figure 9 For the comparison of multi-step prediction results of monitoring point 691 in step 5 of the embodiment of the present invention Figure 10 This is a comparison of the 70th consecutive multi-step prediction and observation of the subsidence field in step 5 of the present invention. Figure 11 The spatial distribution of the prediction error (difference between observed settlement and predicted settlement) in step 5 of this embodiment of the invention, period 70. Detailed Implementation

[0022] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The terms such as "upper," "lower," "left," "right," and "middle" used in this specification are merely for clarity of description and are not intended to limit the scope of the invention. Any changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention.

[0024] like Figure 1 The diagram shows a flowchart of a ground subsidence prediction method based on InSAR time-series data and a CNN-LSTM model, provided by an embodiment of the present invention. The method mainly includes steps such as data acquisition and processing, preprocessing and sample construction, model construction, model training, multi-step prediction, and result visualization. The specific implementation process is as follows: S1. Data Acquisition and Solving This embodiment first utilizes SBAS-InSAR technology to process 70 SAR images of the study area. Through processes such as radiometric correction, interferogram generation, flattening, Goldstein filtering, and minimum cost flow (MCF) phase unwrapping, and combined with external DEM to remove topographic phase, the cumulative deformation time series of each monitoring point within the study area at 70 observation times is calculated using the SVD method. This high-precision time series data constitutes the foundation dataset for subsequent deep learning models.

[0025] S2, Data Preprocessing and Sample Construction To adapt to the input requirements of deep learning models, the monitoring point data was first cleaned, and missing values ​​were filled in using cubic spline interpolation. Then, the MinMaxScaler method was used to normalize the data to the [0, 1] interval to accelerate model convergence.

[0026] Next, a sliding window technique is used to construct supervised learning samples. In this embodiment, the input window length L=65 and the prediction step size S=5. That is, data from period t to t+64 are used as input features to predict data from period t+65 to t+69. Finally, the sample set is divided into a training set (80%), a validation set (10%), and a test set (10%) according to time sequence.

[0027] S3. Constructing the CL-TSF spatiotemporal prediction model This embodiment constructs a CL-TSF spatiotemporal prediction model based on the cascaded coupling of CNN and LSTM, aiming to simultaneously capture local features and long-term temporal dependencies of InSAR time series data. The specific design details of the model are provided by [the relevant authority / organization]. Figures 2 to 4 Detailed display: such as Figure 2 As shown, this is an architecture based on InSAR time series data and a CNN-LSTM model provided by an embodiment of the present invention. This architecture demonstrates the complete process of data flow from the input layer, first passing through a CNN feature extraction module to extract high-dimensional noise-resistant features, then being fed into an LSTM time series memory module for long short-term memory modeling, and finally outputting the predicted values ​​for the next 5 time steps through a Prediction Module. Figure 3 The figure shows the internal structure of a hybrid model unit based on InSAR time series data and a CNN-LSTM model at time step t, provided in an embodiment of the present invention. The figure details the internal data flow: the data first undergoes convolution operations through convolution kernels (…). After activation, the signal is passed to the LSTM unit; the LSTM unit passes the signal through the forget gate ( ), Input gate ( ) and output gate ( Working together to update cell state ( ) and hidden state ( )like Figure 4The diagram illustrates a detailed network architecture based on InSAR time-series data and a CNN-LSTM model, as provided in this embodiment of the invention. The specific configuration of each layer's parameters, based on the code implementation of this embodiment, is as follows: CNN Feature Extraction Module: Contains two stacked sets of one-dimensional convolutional units. The first set includes a Conv1D layer with 32 filters and a kernel size of 3, followed by a BatchNormalization layer and a MaxPooling1D layer (pooling window 2). The second set includes a Conv1D layer with 16 filters and a kernel size of 3, followed by a BatchNormalization layer and a MaxPooling1D layer. This module is used to extract local fluctuation features from the original InSAR sequence and compress noise.

[0028] The LSTM temporal memory module consists of two LSTM layers. The first LSTM layer has 32 neurons, with `return_sequences=True` to pass the complete time series, followed by a dropout layer with a dropout rate of 0.2. The second LSTM layer has 16 neurons, with `return_sequences=False` to output the final hidden state, also followed by a dropout layer with a dropout rate of 0.2. This module is used to capture long-range temporal dependencies and prevent overfitting. The prediction output module contains a fully connected (Dense) layer with 5 neurons and a Linear activation function, used to directly output normalized predictions for the next 5 periods.

[0029] S4, Model Training The training environment was built using a Python deep learning framework. The Adam adaptive moment estimator was selected, with a learning rate set to 0.001; mean squared error (MSE) was used as the loss function. The batch size was set to 64, and the maximum number of epochs was set to 100. During training, the training set was input into the model for backpropagation to update the weights, and the validation set was used to monitor the validation loss. If the validation loss did not decrease within several consecutive epochs, an early stopping strategy was triggered, halting training and saving the current optimal model weight parameters.

[0030] S5. Model Prediction and Accuracy Evaluation The optimal model saved in step S4 is used to predict the test set, and the output prediction results are inversely normalized using the scaler parameters saved in step S2 to obtain the true predicted ground settlement value.

[0031] like Figure 6 The figure shows the error distribution of the model of this invention across all test samples in different models. Figure 6 (a-1) to (a-3) in the model correspond to the CL-TSF model of this invention. Figure 6 (b-1) to (b-3) in the model correspond to a single CNN model. Figure 6 (c-1) to (c-3) in the model correspond to a single LSTM model. A comparison clearly shows that: Figure 6 The scatter points in (a-1) to (a-3) are most densely distributed, and the errors of most samples are concentrated near the value of 0. The RMSE is only 5.260, which is significantly better than the single CNN model (RMSE=11.366) and the single LSTM model (RMSE=6.634).

[0032] To verify the model's generalization ability under different geological deformation modes, this embodiment selects three representative monitoring points for detailed analysis: like Figure 7 As shown in the figure (monitoring point 11), this point exhibits a relatively stable linear settlement trend, and the model of this invention almost completely coincides with the actual observation value.

[0033] like Figure 8 As shown (monitoring point 421), the settlement curve at this point exhibits obvious nonlinear fluctuation characteristics, and the model successfully predicted the local fluctuation changes.

[0034] like Figure 9 As shown (monitoring point 691), this point is a significant settlement point with a large cumulative deformation, yet the model still maintains extremely high tracking accuracy.

[0035] S6. Results Visualization and Spatiotemporal Analysis To evaluate the model's predictive ability in the spatial dimension, this embodiment generates a spatiotemporal distribution map of regional subsidence. For example... Figure 5 As shown, this is a spatial deformation map based on InSAR time series data and a CNN-LSTM model at a representative time step, provided by an embodiment of the present invention. The first six sub-maps ("Input") show a portion of the input sequence, depicting the temporal evolution of the surface deformation field. The color gradually changes from dark red (early stage ts0) to cyan (late stage ts65), fully presenting the dynamic information encoded in the 65-step input sequence. The last two sub-maps ("Target") show the actual ground deformation field corresponding to times ts66 and ts67. These data constitute the predicted target. Furthermore, a spatial field comparative analysis is conducted on the prediction results of the next five time steps (i.e., the 66th to the 70th period). Here, only the deformation evolution of the 70th period (i.e., the final prediction step) is shown for illustration.

[0036] like Figure 10 The figure shows a comparison between the predicted and observed subsidence fields in step 5 (70th step) of this embodiment of the invention. The left column represents the actual InSAR observed subsidence field, and the right column represents the subsidence field predicted by the CL-TSF model. Visual comparison reveals that the predicted layer is highly consistent with the actual observed layer in terms of the location of the subsidence center, the boundary of the subsidence range, and the transition of the deformation gradient. This indicates that the model is not only accurate in predicting over time but also highly reliable in extrapolating the spatial evolution.

[0037] To quantify the spatial prediction error, the prediction residual (observed value minus predicted value) for each pixel in the last period was calculated and its spatial distribution was plotted: like Figure 11 The figure shows the spatial distribution of prediction error (the difference between observed and predicted settlement) in step 5 of this embodiment over 70 periods. The vast majority of the study area is represented by colors indicating low error, with only a few small error patches appearing in the edge areas with poor coherence. This further confirms that the method of this invention can achieve high-precision regional spatiotemporal prediction of land settlement, providing accurate data support for urban planning and disaster prevention and mitigation.

[0038] In summary, this embodiment, by constructing a CL-TSF coupled model, fully utilizes the feature extraction and noise resistance capabilities of CNNs and the temporal memory capabilities of LSTMs, effectively solving the prediction challenges caused by high noise and strong nonlinearity in InSAR data. It not only demonstrates excellent performance in single-point prediction but also exhibits extremely high consistency in extrapolating the evolution of regional subsidence fields.

[0039] Obviously, the present invention has been described in detail above with general descriptions and specific embodiments. However, some modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention are within the scope of protection claimed by the present invention.

Claims

1. A method for predicting land subsidence based on InSAR time-series data and a CNN-LSTM model, characterized in that, Includes the following steps: Step S1, Data Acquisition and Solving: The P-scene SAR image of the study area is interferometrically processed using the small baseline set technique to obtain the cumulative deformation time series data of a large number of monitoring points in the study area at T observation times; Step S2, Data Preprocessing and Sample Construction: Outlier removal and missing value imputation are performed on the cumulative deformation time series data of the monitoring points. The data is mapped to the standard interval using the normalization algorithm. The time series supervision samples are constructed based on the sliding window algorithm. The data of the first L time steps are used as input features, and the data of the last S time steps are used as prediction labels. The sample set is divided into training set, validation set and test set. Step S3: Construct the CL-TSF spatiotemporal prediction model: Build a serially coupled model that includes a feature extraction module and a temporal memory module. The CNN module is used to extract local high-dimensional features of the time series data and filter out noise, while the LSTM module is used to capture the long-distance temporal dependencies of the sedimentation data. Step S4, Model Training: Iteratively train the CL-TSF spatiotemporal prediction model using the training set constructed in step S2, and update the model parameters using the Adam optimizer and mean squared error loss function until the model converges. Step S5, Multi-step prediction and inverse normalization: Input the cumulative deformation time series to be predicted into the trained CL-TSF model, output the predicted values ​​for the next S time points, and restore them to the true cumulative deformation values ​​through inverse normalization. Step S6, Result Visualization: Combine the predicted deformation value obtained in step S5 with the geographic coordinate information of the monitoring points, and use MATLAB to generate a spatiotemporal distribution map of ground subsidence in the study area at S future times, so as to realize the visualization of the subsidence trend.

2. The ground subsidence prediction method based on InSAR time series data and CNN-LSTM model according to claim 1, characterized in that, The specific process for obtaining the cumulative deformation time series data in step S1 is as follows: Step S1.1: Select SAR images covering the study area of ​​P scene, filter interferometric image pairs according to the spatiotemporal baseline threshold, remove the terrain phase using an external digital elevation model, and generate differential interferogram; Step S1.2: The differential interferogram is enhanced by filtering using the Goldstein filtering algorithm, and the phase unwrapping is performed using the minimum cost flow algorithm; Step S1.3: By selecting stable control points, perform orbit refinement and re-leveling, estimate and remove the atmospheric delay phase, and use the singular value decomposition method to calculate the cumulative deformation time series D={d1, d2, d3, ..., dt} for each monitoring point at T observation times. T } 3. The ground subsidence prediction method based on InSAR time series data and CNN-LSTM model according to claim 1, characterized in that, The specific implementation process of step S2 is as follows: Step S2.1, Data Normalization: The original deformation time series data of each monitoring point is standardized using the max-min normalization method, mapping the data to the [0,1] interval. The calculation formula is as follows: (1.1)； in, This represents the original deformation value at the current moment. For the normalized data, This is the entire time series data set for this monitoring point. and These are the minimum and maximum values ​​of the sequence, respectively. Step S2.2: Constructing Time-Series Supervised Samples: The univariate time series is converted into supervised learning sample pairs using the sliding window algorithm; the input sequence length is set to L, and the prediction step size is S; for the normalized cumulative variable time series D={d1, d2, d3, ... d... T }, the input features of the t-th sample and corresponding tags Represented as: (1.2)； (1.3)； The window slides forward with a step size of 1 until it covers the entire sequence, generating a total of TL-S+1 samples. Step S2.3, Dataset Partitioning: Adjust the constructed sample set to a three-dimensional tensor format, and finally divide the sample set into training set, validation set and test set according to time sequence.

4. The ground subsidence prediction method based on InSAR time series data and CNN-LSTM model according to claim 1, characterized in that, The CL-TSF spatiotemporal prediction model in step S3 adopts a serially coupled architecture, and the specific structural parameters are as follows: The CL-TSF spatiotemporal prediction model includes, in order of data flow, a feature extraction module, a temporal memory module, and a fully connected output module. The feature extraction module comprises two stacked convolutional units: the first group includes a filter, a one-dimensional convolutional layer, and corresponding batch normalization and max pooling layers; the second group includes a filter, a one-dimensional convolutional layer, and corresponding batch normalization and max pooling layers; the one-dimensional convolution operation formula is as follows: (1.4)； Here, convolution is the operation. For the first l Layer output features, As weight, For bias, For convolution operations, It is the ReLU activation function. This refers to the position index within the convolution kernel. for l-1 The input feature value of the layer at the corresponding position, and j is the position index of the output feature sequence; The temporal memory module contains two LSTM layers: the first LSTM layer contains neurons, is set to return a sequence to output the complete sequence, and is followed by a random deactivation layer (Dropout); the second LSTM layer contains neurons, is set to return a complete sequence to output the final state, and is followed by a random deactivation layer (Dropout). The fully connected output module contains a fully connected layer with the number of neurons equal to the prediction step size S and the activation function being a linear function.

5. The ground subsidence prediction method based on InSAR time series data and CNN-LSTM model according to claim 4, characterized in that, The one-dimensional convolutional layer output feature sequence obtained by formula (1.4) is used as the input of the LSTM layer, and this feature sequence is denoted as X = { x 1, x 2, ... x t ,……},in x t express t The feature input is extracted by the one-dimensional convolutional layer at each time step; The LSTM layer captures long-range dependencies through a gating mechanism, and its internal unit update formula is as follows: (1.5)； (1.6)； (1.7)； (1.8)； (1.9)； (1.10)； in, This is the current input time. This is the hidden state from the previous moment. In cellular state, , , These are the forget gate, input gate, and output gate, respectively. For the Sigmoid function, Represents the Hadama product; through the forgetting gate This determines how much historical information to retain, thereby addressing the gradient vanishing problem in long sequences. , , , These represent the weights corresponding to the forget gate, input gate, candidate cell state, and output gate, respectively. , , , These represent the biases corresponding to the forget gate, input gate, candidate cell state, and output gate, respectively. The candidate cell states for the LSTM layer.

6. The ground subsidence prediction method based on InSAR time series data and CNN-LSTM model according to claim 1, characterized in that, The model training process in step S4 is as follows: Step S4.1, Parameter Configuration: Select the Adam optimizer and set the learning rate; select mean squared error as the loss function, as shown in the following formula: (1.11)； in This represents the batch sample size. For true normalized values, This is a predicted value; Step S4.2, Iterative Optimization: Input the training set into the model, set the batch size and the maximum number of iterations; calculate the gradient and update the weights using the backpropagation algorithm; Step S4.3, Early Stopping Strategy: During training, the test set is used for validation, and the loss of the validation set is monitored. If the loss value does not decrease within several consecutive rounds, the early stopping mechanism is triggered and the current optimal model parameters are saved.

7. The ground subsidence prediction method based on InSAR time series data and CNN-LSTM model according to claim 1, characterized in that, Step S6 uses Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) as evaluation indicators, and the calculation formulas are as follows: (1.12)； (1.13)； (1.14)； in, To predict the total number of samples, The true deformation value after denormalization. This is the predicted deformation value after inverse normalization.

Images