An intelligent deformation prediction method based on InSAR data and multiple inducing factors
Through the long-term and short-term memory deep learning model based on InSAR data and multiple eliciting factors, the problems of insufficient utilization of multi-feature data and long-term dependency learning difficulties in the existing technology are solved, and high-precision and efficient surface deformation prediction are achieved, which is suitable for multiple type of deformation area.
Patent Information
- Application Number
- CN202411140055.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-20
- Publication Date
- 2024-12-10
- Estimated Expiration
- 2044-08-20
AI Technical Summary
The existing technology lacks the ability to utilize multi-feature data in surface deformation prediction, resulting in a lack of interpretability and low accuracy in the prediction results. Traditional recurrent neural networks have gradient disappearance or explosion in learning long-term dependence, and cannot effectively capture complex environmental factors. coupling effect.
Using a long and short-term memory deep learning model based on InSAR data and multiple evoked factors, a prediction model including multiple LSTM layers and Dropout layers is constructed by acquiring and preprocessing the deformation timing data and evoking factor data of monitoring points, using Adam optimizer and The R2 loss function is iteratively trained and optimized to improve the model's learning ability and prediction accuracy.
High-precision prediction of complex nonlinear deformation and factor data is achieved, which improves the calculation efficiency by 20%, and the prediction accuracy reaches more than 90%. It is suitable for prediction of multiple deformation areas, making up for the shortcomings of traditional models in time series research.
Smart Images
Figure CN118656684B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of surface deformation monitoring, and specifically relates to an intelligent deformation prediction method based on InSAR data and multiple inducing factors, which is suitable for predicting surface deformation using a long-short memory deep learning model. Background Art
[0002] Interferometry Synthetic Aperture Radar (InSAR) has the characteristics of all-day, all-weather and high resolution. It has been widely used in deformation monitoring such as urban ground subsidence, earthquakes, volcanoes, and permafrost. It can obtain deformation information at the centimeter to millimeter level. In addition, geological conditions, hydrological conditions, climate and other inducing factors are important factors leading to deformation. It is of great significance to carry out short-term or long-term predictions of deformation areas in the future.
[0003] In the field of artificial intelligence, machine learning is a data-driven method, and deep learning is a technology for implementing machine learning. Their prediction models are good at analyzing the nonlinear relationship between data, learning the characteristic laws of massive data from samples (InSAR deformation time series), and then using them for actual inference and decision-making. Typical machine learning models include support vector machine (SVM), random forest (RF), etc. These models have simple structures and can perform regression prediction by capturing the linear relationship between data, so they can solve simple linear regression problems. There are many types of deep learning models, and the algorithms are complex and diverse. The early feedback neural network (BP) belongs to a network model with a multi-layer perceptron (MLP) structure. The advantage is that it has a strong learning ability for the nonlinear characteristics of the sequence, but the model has a slow learning speed and lacks corresponding theoretical guidance, which makes it impossible to promote. With the advancement of deep learning technology, typical networks such as convolutional neural networks (CNN) and traditional recurrent neural networks (RNN) have been introduced one after another. Convolutional neural networks have a strong ability to analyze the nonlinear features of images due to their built-in convolution kernels, so they are very good at image processing problems. Traditional recurrent neural networks are a type of recurrent neural network that takes sequence data as input, recursively in the direction of sequence evolution, and all nodes are connected in a chain-like manner. They have a strong ability to fit nonlinear features. The above two types of models can effectively capture the characteristic laws between massive data. The convolutional structure and recurrent structure in the model support the use and update of data, and are widely used in the fields of image processing and deformation prediction.
[0004] According to current relevant deformation prediction research, most of them predict the deformation trend of the research target by learning a single data feature through the model, and rarely try to capture the data laws of multiple features, which ultimately makes the results lack of interpretability; in addition, a certain type of model cannot make full use of data resources and capture data laws, which will also affect the final results. Therefore, based on the type of study area and the measurement of data period, choosing a suitable model for prediction is the basic premise of the experiment. Multi-feature data refers to the consideration of multiple inducing factors that cause regional deformation, such as rainfall, slope, etc. This condition will have an impact on the deformation prediction effect. Model selection refers to combining the data type and its period to consider whether the selected model can achieve better prediction results. Because the terrain and environmental factors are different for different study areas, the applicability of the model may be limited, resulting in the model failing to capture effective data information;
[0005] To sum up the above problems, on the one hand, due to the complex environmental factors in different study areas and the coupling effect between multiple inducing factors, it is easy to cause regional deformation to varying degrees. There are few studies on adding multi-feature data types to the model. We should consider using multiple data features as model input to make the final prediction results more interpretable. On the other hand, traditional recurrent neural networks produce gradient vanishing or explosion phenomena when training data, and cannot learn the long-term dependence of data. The long short-term memory neural network (LSTM) developed based on the traditional RNN can control information by adding gating devices (such as forget gates, input gates, and output gates) in the hidden layer, learn the characteristic laws of longer-term data, and effectively overcome the shortcomings of traditional RNN. Summary of the invention
[0006] The purpose of the present invention is to provide a deformation intelligent prediction method based on InSAR data and multiple inducing factors in view of the above problems existing in the prior art.
[0007] The above-mentioned purpose of the present invention is achieved by the following technical means:
[0008] A deformation intelligent prediction method based on InSAR data and multiple inducing factors comprises the following steps:
[0009] Step 1, obtaining the InSAR deformation time series data and the corresponding inducing factor data of n monitoring points in the study area, and preprocessing the InSAR deformation time series data and the corresponding inducing factor data of the n monitoring points to obtain n preprocessed samples;
[0010] Step 2: Make the n preprocessed samples into training samples respectively, obtain n training samples in total, normalize the n training samples, divide the normalized n training samples into a training data set and a prediction data set, and divide the training data set into a training set and a validation set;
[0011] Step 3: Build a prediction model;
[0012] Step 4: Input the training set into the prediction model of step 3 to iteratively train the prediction model, then input the validation set into the prediction model after iterative training to obtain the prediction result of the validation set, and finally determine whether the prediction model needs to be optimized according to the prediction result of the validation set;
[0013] Step 5: Input the prediction data set into the prediction model of step 4 to perform prediction and obtain the prediction result of the prediction data set.
[0014] As mentioned above, step 1 specifically includes the following steps:
[0015] Step 1.1, obtain the InSAR deformation time series data and corresponding inducing factor data of n monitoring points in the study area, and take the InSAR deformation time series data and corresponding inducing factor data of each monitoring point as a sample to be processed, and obtain a total of n samples to be processed;
[0016] Step 1.2, load a sample to be processed into the geographic information system, read the InSAR deformation time series data of the monitoring point, and use the missing value filling method to process the InSAR deformation time series data of the monitoring point into data with equal time intervals to obtain deformation equal time interval data;
[0017] Step 1.3, sampling the induced factor data within the same time range as the InSAR deformation time series data according to the time series, obtaining the induced factor time series data with the same time series as the InSAR deformation time series data, and then using the missing value filling method to process the induced factor time series data into the induced factor equal time interval data with the same time series as the deformation equal time interval data, and finally taking the deformation equal time interval data and the corresponding induced factor equal time interval data as a preprocessed sample;
[0018] Step 1.4: Perform steps 1.2 and 1.3 on n samples to be processed in sequence to obtain n preprocessed samples.
[0019] As mentioned above, step 2 specifically includes the following steps:
[0020] Step 2.1, make n preprocessed samples into training samples respectively, and obtain n training samples in total, where the training samples include input samples X and true labels Y;
[0021] Step 2.2: Normalize the n training samples by the maximum-minimum normalization method;
[0022] Step 2.3: Divide the n normalized training samples into a training data set and a prediction data set;
[0023] Step 2.4: Slice the training data set by random sampling and divide it into a training set and a validation set.
[0024] As described above, Step 2.1 specifically includes the following steps:
[0025] Step 2.1.1: First, define the length of the input sample X as L, the length of the true label Y as p. The input sample X includes a known interval with a length of L - p and a deformation prediction interval with a length of p, and 2 ≤ L ≤ m, 1 ≤ p < L, where m is the time series length in the preprocessed samples;
[0026] Step 2.1.2: Extract the deformation amounts, rainfall amounts, and slopes corresponding to L consecutive time points in a preprocessed sample. Take the deformation amounts, rainfall amounts, and slopes corresponding to the first L - p time points among the L consecutive time points as the known interval, and take the rainfall amounts and slopes corresponding to the last p time points among the L consecutive time points as the deformation prediction interval to obtain the input sample X; finally, take the deformation amounts corresponding to the last p time points among the L consecutive time points as the true label Y; the input sample X and the true label Y together form a training sample;
[0027] Step 2.1.3: Sequentially execute Step 2.1.2 on the n preprocessed samples to obtain n training samples in total.
[0028] As above, the prediction model in Step 3 includes N LSTM layers, N - 1 Dropout layers, and one Dense layer. The Adam optimizer is used, and the loss functions are the R2 loss function and the RMSE loss function; the input sample X in the training set is input into the first LSTM layer, the output of the (i - 1)th Dropout layer is input into the ith LSTM layer, where i takes values from 2 to N; the output of the ith LSTM layer is input into the ith Dropout layer, where i takes values from 1 to N - 1; the output of the Nth LSTM layer is input into the Dense layer, and the output of the Dense layer is the output of the prediction model. At the same time, the output of the Dense layer is also input into the Adam optimizer.
[0029] As described above, Step 4 specifically includes the following steps:
[0030] Step 4.1, input the input sample X in the training set into the prediction model for iterative training, and then input the validation set into the prediction model after iterative training to obtain the validation set prediction result, and use the R2 loss function and RMSE loss function to calculate the R2 loss value and RMSE loss value between the validation set prediction result and the corresponding true label Y in the validation set;
[0031] Step 4.2: If the R2 loss value in step 4.1 is greater than or equal to the set R2 loss value, and the RMSE loss value is less than or equal to the set RMSE loss value, the prediction model does not need to be optimized and the model parameters are saved;
[0032] If the R2 loss value is less than the set R2 loss value, or the RMSE loss value is greater than the set RMSE loss value, the prediction model is optimized through artificial experience parameter adjustment and grid search method until the R2 loss value is greater than or equal to the set R2 loss value, and the RMSE loss value is less than or equal to the set RMSE loss value. The prediction model optimization is completed, the model parameters are saved, and the optimized prediction model is obtained.
[0033] Compared with the prior art, the present invention has the following beneficial effects:
[0034] (1) Universality: The prediction model of the present invention can analyze various complex nonlinear deformation and factor data and can be widely and flexibly used in the prediction of various deformation areas.
[0035] (2) High precision. Compared with traditional prediction models such as BP and RNN, the computational efficiency of the present invention is improved by 20%, and the prediction accuracy reaches more than 90%, with better prediction effect.
[0036] (3) Practicality: The method of the present invention can make up for the shortcomings of InSAR technology in time series research. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 is a flow chart of the method of the present invention;
[0038] Figure 2 is the training sample (X1~X n , X D1 ~X Dn , X S1 ~X Sn , X r1 ~X r n are input samples, deformation variables, rainfall, and slope respectively). DETAILED DESCRIPTION
[0039] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below in conjunction with embodiments. The embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.
[0040] Example:
[0041] The deformation intelligent prediction method based on InSAR data and multiple inducing factors includes the following steps:
[0042] Step 1: Obtain the InSAR deformation time series data and corresponding inducing factor data (inducing factors include rainfall and slope) of n monitoring points in the study area, and preprocess the InSAR deformation time series data and corresponding inducing factor data of n monitoring points to obtain n preprocessed samples, which specifically includes the following steps:
[0043] Step 1.1, obtain the InSAR deformation time series data and the corresponding inducing factor data of n monitoring points in the study area. The InSAR deformation time series data and the corresponding inducing factor data of each monitoring point are taken as a sample to be processed, and a total of n samples to be processed are obtained.
[0044] Step 1.2, load a sample to be processed into the geographic information system (ArcGIS), read the InSAR deformation time series data of the monitoring point and use the missing value filling method (such as cubic spline interpolation, interpolation method) to process the InSAR deformation time series data of the monitoring point into data with equal time intervals, and obtain deformation equal time interval data (that is, the interval between any two adjacent time points is the same).
[0045] Step 1.3: Sample the inducing factor data within the same time range as the InSAR deformation time series data according to the time series to obtain the inducing factor time series data with the same time series as the InSAR deformation time series data. Then use the missing value filling method to process the inducing factor time series data into inducing factor equal time interval data with the same time series as the deformation equal time interval data. Finally, use the deformation equal time interval data and the corresponding inducing factor equal time interval data as a preprocessed sample.
[0046] Step 1.4: Perform steps 1.2 and 1.3 on n samples to be processed in sequence to obtain n preprocessed samples.
[0047] Step 2: Make the n preprocessed samples into training samples respectively, obtain n training samples in total, normalize the n training samples, divide the normalized n training samples into a training data set and a prediction data set, and divide the training data set into a training set and a validation set, specifically including the following steps:
[0048] Step 2.1: Make n preprocessed samples into training samples respectively, obtaining a total of n training samples. The training samples include input sample X and true label Y, which specifically include the following steps:
[0049] Step 2.1.1: First, define the length of input sample X as L and the length of true label Y as p. The input sample X includes a known interval with length L - p and a deformation prediction interval with length p, and it satisfies 2 ≤ L ≤ m, 1 ≤ p < L, where m is the time series length in the preprocessed sample (i.e., m time points);
[0050] Step 2.1.2: Extract the deformation amount, rainfall, and slope corresponding to L consecutive time points in one preprocessed sample. Take the deformation amount, rainfall, and slope corresponding to the first L - p time points among the L consecutive time points as the known interval, and take the rainfall and slope corresponding to the last p time points among the L consecutive time points as the deformation prediction interval to obtain the input sample X; finally, take the deformation amount corresponding to the last p time points among the L consecutive time points as the true label Y. The input sample X and the true label Y together form a training sample.
[0051] The input sample X is used to be input into the prediction model. The prediction model predicts the deformation amount corresponding to the last p time points among the L consecutive time points based on the known interval and the deformation prediction interval in the input sample X to obtain a prediction result, and then compares the prediction result corresponding to the input sample X with the true label Y.
[0052] Step 2.1.3: Sequentially execute Step 2.1.2 for the n preprocessed samples, obtaining a total of n training samples.
[0053] Step 2.2: Normalize the n training samples by the maximum - minimum normalization method to improve the training efficiency and stability.
[0054] Step 2.3: Divide the n normalized training samples into a training dataset and a prediction dataset according to a ratio of 4:1.
[0055] Step 2.4: Slice the training dataset by the method of random sampling and divide it into a training set and a validation set according to a ratio of 4:1. The training set is used for model training, and the validation set is used for model validation.
[0056] Step 3: Construct a prediction model.
[0057] The prediction model includes N LSTM layers, N-1 Dropout layers, and a Dense layer. The optimizer uses the Adam optimizer, and the loss function uses the R2 loss function and the RMSE loss function. The input sample X in the training set is input into the first LSTM layer, the output of the i-1th Dropout layer is input into the ith (i ranges from 2 to N)th LSTM layer, the output of the ith (i ranges from 1 to N-1)th LSTM layer is input into the ith Dropout layer, the output of the Nth LSTM layer is input into the Dense layer, and the output of the Dense layer is the output of the prediction model. At the same time, the output of the Dense layer is also input into the Adam optimizer to optimize the prediction model.
[0058] As an implementable embodiment, the prediction model includes 3 LSTM layers, 2 Dropout layers, and 1 Dense layer. The optimizer adopts Adam optimizer, and the loss function adopts R2 loss function and RMSE loss function. The samples in the training set are input into the first LSTM layer, the output of the first LSTM layer is input into the first Dropout layer, the output of the first Dropout layer is input into the second LSTM layer, the output of the second LSTM layer is input into the second Dropout layer, the output of the second Dropout layer is input into the third LSTM layer, the output of the third LSTM layer is input into the Dense layer, and the output of the third LSTM layer is also input into the Adam optimizer to optimize the prediction model.
[0059] Step 4: Input the input sample X in the training set of step 2.4 into the prediction model constructed in step 3 to iteratively train the prediction model, then input the validation set into the prediction model after iterative training to obtain the validation set prediction result, and finally determine whether the prediction model needs to be optimized according to the validation set prediction result, which specifically includes the following steps:
[0060] Step 4.1: Input the input sample X in the training set into the prediction model for iterative training, and then input the validation set into the prediction model after iterative training to obtain the validation set prediction result. The R2 loss function and RMSE loss function are used to calculate the R2 loss value and RMSE loss value between the validation set prediction result and the corresponding true label Y in the validation set.
[0061] The R2 loss function and RMSE loss function are used to verify and evaluate the prediction accuracy. The value of the R2 loss function reflects the degree of fit between the prediction results of the validation set and the corresponding true label Y. The larger the value of the R2 loss function, the better the fitting effect; the value of the RMSE loss function reflects the difference between the prediction results of the validation set and the corresponding true label Y. The smaller the value of the RMSE loss function, the higher the prediction accuracy of the prediction model.
[0062] Step 4.2: If the R2 loss value in step 4.1 is greater than or equal to the set R2 loss value (such as 0.9), and the RMSE loss value is less than or equal to the set RMSE loss value (such as 0.001), the prediction model does not need to be optimized and the model parameters are saved;
[0063] If the R2 loss value is less than the set R2 loss value, or the RMSE loss value is greater than the set RMSE loss value, the prediction model is optimized through artificial experience parameter adjustment and grid search method until the R2 loss value is greater than or equal to the set R2 loss value, and the RMSE loss value is less than or equal to the set RMSE loss value. The prediction model optimization is completed, the model parameters are saved, and the optimized prediction model is obtained.
[0064] Step 5: Input the prediction data set into the prediction model of step 4 to perform prediction, obtain the prediction result of the prediction data set, and analyze the prediction result of the prediction data set.
[0065] In this embodiment, the prediction data set is input into the prediction model with an R2 loss value of 0.92 and an RMSE loss value of 0.0006. The final measured deformation prediction accuracy reaches more than 90%, and the calculation efficiency is improved by 20% compared with the traditional prediction model, with higher prediction accuracy and better prediction effect.
[0066] In one embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.
[0067] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.
[0068] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.
[0069] It should be noted that the embodiments described in the present invention are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the described embodiments or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the attached claims.
Claims
1. An intelligent deformation prediction method based on InSAR data and multiple inducing factors, It is characterized in that The following steps are involved: Step 1, obtaining the InSAR deformation time series data and the corresponding inducing factor data of n monitoring points in the study area, and preprocessing the InSAR deformation time series data and the corresponding inducing factor data of the n monitoring points to obtain n preprocessed samples; Step 2: Make the n preprocessed samples into training samples respectively, obtain n training samples in total, normalize the n training samples, divide the normalized n training samples into a training data set and a prediction data set, and divide the training data set into a training set and a validation set; Step 3: Build a prediction model; Step 4: Input the training set into the prediction model of step 3 to iteratively train the prediction model, then input the validation set into the prediction model after iterative training to obtain the prediction result of the validation set, and finally determine whether the prediction model needs to be optimized according to the prediction result of the validation set; Step 5: Input the prediction data set into the prediction model of step 4 to perform prediction, and obtain the prediction result of the prediction data set; The step 1 specifically comprises the following steps: Step 1.1, obtain the InSAR deformation time series data and corresponding inducing factor data of n monitoring points in the study area, the inducing factors include rainfall and slope, and the InSAR deformation time series data and corresponding inducing factor data of each monitoring point are taken as a sample to be processed, and a total of n samples to be processed are obtained; Step 1.2, load a sample to be processed into the geographic information system, read the InSAR deformation time series data of the monitoring point, and use the missing value filling method to process the InSAR deformation time series data of the monitoring point into data with equal time intervals to obtain deformation equal time interval data; Step 1.3, sampling the induced factor data within the same time range as the InSAR deformation time series data according to the time series, obtaining the induced factor time series data with the same time series as the InSAR deformation time series data, and then using the missing value filling method to process the induced factor time series data into the induced factor equal time interval data with the same time series as the deformation equal time interval data, and finally taking the deformation equal time interval data and the corresponding induced factor equal time interval data as a preprocessed sample; Step 1.4: perform steps 1.2 and 1.3 on n samples to be processed in sequence to obtain n pre-processed samples; The step 2 specifically includes the following steps: Step 2.1, make n pre-processed samples into training samples respectively, and obtain n training samples in total, where the training samples include input samples X and true labels Y; Step 2.2, normalize the n training samples by using the maximum-minimum normalization method; Step 2.3, divide the normalized n training samples into a training data set and a prediction data set; Step 2.4: Slice the training data set by random sampling and divide it into training set and validation set; The step 2.1 specifically includes the following steps: Step 2.1.1: First, define the length of the input sample X as L and the length of the true label Y as p. The input sample X includes a known interval of length L - p and a deformation prediction interval of length p, and 2 ≤ L ≤ m, 1 ≤ p < L, where m is the time series length in the preprocessed sample. Step 2.1.2: Extract the deformation amounts, rainfall amounts, and slopes corresponding to L consecutive time points in a preprocessed sample. Take the deformation amounts, rainfall amounts, and slopes corresponding to the first L - p time points among the L consecutive time points as the known interval, and take the rainfall amounts and slopes corresponding to the last p time points among the L consecutive time points as the deformation prediction interval to obtain the input sample X. Finally, take the deformation amounts corresponding to the last p time points among the L consecutive time points as the true label Y. The input sample X and the true label Y together form a training sample. Step 2.1.3: Sequentially execute Step 2.1.2 on n preprocessed samples to obtain n training samples in total. The prediction model in Step 3 includes N LSTM layers, N - 1 Dropout layers, and one Dense layer. The Adam optimizer is used, and the loss functions are the R2 loss function and the RMSE loss function. The input sample X in the training set is input into the first LSTM layer. The output of the (i - 1)th Dropout layer is input into the ith LSTM layer, where i ranges from 2 to N. The output of the ith LSTM layer is input into the ith Dropout layer, where i ranges from 1 to N - 1. The output of the Nth LSTM layer is input into the Dense layer. The output of the Dense layer is the output of the prediction model, and the output of the Dense layer is also input into the Adam optimizer. Step 4 specifically includes the following steps: Step 4.1: Input the input sample X in the training set into the prediction model for iterative training, then input the validation set into the prediction model after iterative training to obtain the prediction results of the validation set. Calculate the R2 loss value and the RMSE loss value between the prediction results of the validation set and the corresponding true label Y in the validation set using the R2 loss function and the RMSE loss function. Step 4.2: If the R2 loss value in Step 4.1 is greater than or equal to the set R2 loss value and the RMSE loss value is less than or equal to the set RMSE loss value, then the prediction model does not need to be optimized, and save the model parameters. If the R2 loss value is less than the set R2 loss value or the RMSE loss value is greater than the set RMSE loss value, then optimize the prediction model through manual experience parameter tuning and the grid search method until the R2 loss value is greater than or equal to the set R2 loss value and the RMSE loss value is less than or equal to the set RMSE loss value. The optimization of the prediction model is completed, save the model parameters, and obtain the optimized prediction model.