CN114255401B

CN114255401B - A method for predicting land subsidence based on spatiotemporal series data

Info

Publication number: CN114255401B
Application number: CN202111566594.9A
Authority: CN
Inventors: 高明亮; 冷靖; 宫辉力; 陈蓓蓓; 李小娟; 周超凡; 史珉; 关金环
Original assignee: Capital Normal University
Current assignee: Capital Normal University
Priority date: 2021-12-20
Filing date: 2021-12-20
Publication date: 2025-07-15
Anticipated expiration: 2041-12-20
Also published as: CN114255401A

Abstract

The application provides a ground subsidence prediction method based on space-time sequence data, which comprises the steps of obtaining SAR images, track data and DEM files in a preset time period of a preset geographic area, obtaining n time sequence subsidence data sets arranged in time sequence based on the SAR images, the track data and the DEM files, wherein each time sequence subsidence data set comprises historical subsidence values of m predicted points, and predicting subsidence values of future time nodes of the predicted point i based on the historical subsidence values of the predicted point i and the historical subsidence values of the predicted point having spatial correlation with the predicted point i, wherein the value of i is 1 to m. The application can improve the rationality and accuracy of prediction.

Description

Ground subsidence prediction method based on space-time sequence data

Technical Field

The application relates to the field of ground subsidence, in particular to a ground subsidence prediction method based on space-time sequence data.

Background

Ground subsidence refers to a geological phenomenon that an underground unconsolidated rock formation is consolidated and compressed due to natural factors or ergonomic activities and results in a reduction of ground elevation in a certain area, and belongs to a slowly-varying geological disaster. Among them, as an important means for disaster prevention and control, monitoring and prediction technologies of ground subsidence have been greatly developed.

The conventional time series prediction model mainly adopts time series data to predict the ground subsidence, however, the geographical things or attributes are related in space distribution, and because the time series data generally does not contain space information, an accurate ground subsidence prediction result cannot be obtained based on the conventional time series prediction model.

Disclosure of Invention

Aiming at the technical problems, the application adopts the following technical scheme:

the embodiment of the invention provides a ground subsidence prediction method based on space-time sequence data, which comprises the following steps:

s100, acquiring SAR images, track data and DEM files in a preset time period of a preset geographic area;

s200, obtaining n time sequence sedimentation data sets arranged in time sequence based on the SAR image, the track data and the DEM file, wherein each time sequence sedimentation data set comprises historical sedimentation values of m predicted points;

S300, predicting the sedimentation value of a future time node of the predicted point i based on the historical sedimentation value of the predicted point i and the historical sedimentation value of the predicted point having spatial correlation with the predicted point i, wherein the value of i is 1 to m.

Embodiments of the present invention also provide a non-transitory computer readable storage medium having stored therein at least one instruction or at least one program loaded and executed by a processor to implement the foregoing method.

An embodiment of the present invention also provides an electronic device, which includes a processor and the non-transitory computer readable storage medium described above.

Embodiments of the present invention also provide a computer program product comprising a computer program, characterized in that the computer program is executed by a processor to implement the aforementioned method.

According to the ground subsidence prediction method based on the space-time sequence data, provided by the embodiment of the application, the influence of the space correlation on the ground subsidence prediction is considered, namely, the space-time correlation of the data is considered, so that the time and space correlation of the space-time data can be captured at the same time, the ground subsidence prediction can be more accurately and effectively performed, and a more accurate prediction result can be obtained.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for the description of the embodiments will be briefly described below, and it is apparent that the drawings in the following description are only some embodiments of the present application, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flow chart of a ground subsidence prediction method based on time-space sequence data provided by an embodiment of the application;

FIG. 2 is a schematic diagram of an interferometric processing of SAR images;

FIG. 3 is a schematic diagram of performing time-series deformation estimation on SAR images after interference processing;

FIG. 4 is a computational schematic of a door structure;

fig. 5 (a) to 5 (c) are schematic diagrams of convolution kernels and processing using the convolution kernels.

Detailed Description

The following description of the embodiments of the present application will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present application, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to fall within the scope of the application.

Fig. 1 is a flowchart of a ground subsidence prediction method based on space-time sequence data according to an embodiment of the present application. As shown in fig. 1, the ground subsidence prediction method based on space-time sequence data provided by the embodiment of the application comprises the following steps:

S100, SAR images, track data and DEM files in a preset time period of a preset geographic area are acquired.

In the embodiment of the present invention, the preset geographic area may be a user-defined geographic area. The preset time period may be a user-defined time as long as enough historical sedimentation data can be obtained to ensure accurate prediction can be achieved. SAR images, track data and DEM files can be obtained in an existing mode.

S200, obtaining n time sequence sedimentation data sets arranged in time sequence based on the SAR image, the track data and the DEM file, wherein each time sequence sedimentation data set comprises historical sedimentation values of m predicted points.

In the embodiment of the invention, SAR images are integrated with software packages such as ISCE, MINTPY and the like in a Linux environment to perform SAR image interference processing and time sequence deformation estimation, so as to obtain n time sequence sedimentation data sets arranged in time sequence. The method specifically comprises the following steps:

Step one, based on track data and a DEM file, adopting ISCE integrated environment to perform interference processing of SAR images. As shown in fig. 2, the method mainly comprises the steps of baseline estimation, registration, resampling, small baseline group pairing, interferogram generation and the like, and specifically comprises the following steps:

(1) Selecting a proper main image from the SAR image in the SLC format;

(2) And registering the main image and the auxiliary image to a unified coordinate system by using the acquired track data. The auxiliary image is an image other than the main image in the SAR image. Those skilled in the art will recognize that any registration method for registering the primary and secondary images to a unified coordinate system is within the scope of the present invention.

(3) After registration is completed, combining an external DEM file to simulate a main image intensity map and an inverted terrain phase, and respectively carrying out differential interference on the auxiliary image and the main image to generate a single main image interferogram set. Those skilled in the art will recognize that any method for implementing a single primary image interferogram set is within the scope of the present invention.

(4) And screening small baseline pairs by setting time baselines, null baselines and coherence thresholds, generating small baseline pair combinations, and carrying out differential processing on interference pattern pairs corresponding to the small baselines in the generated small baseline pair combinations respectively to generate a small baseline interference pattern set.

For any pixel x (r, c) in the interferogram in the set of generated small baseline interferograms, r, c are the coordinates of the point in the azimuth-to-elevation coordinate system, respectively, and the phase composition of the point can be expressed as:

Wherein, the For the phase contribution of the surface deformation,For the phase contribution of the track error,To the phase contribution of the DEM error,To the phase contribution of the atmospheric delay error,The phase contribution of components such as background scattering, thermal noise, and phase loss interference noise.

Step two, performing time sequence processing by MintPy, as shown in fig. 3, mainly comprising the following steps:

1) And processing the S1A track lifting interferograms in the small baseline interferogram set obtained by using the minimum spanning tree, the space-time baseline threshold value and the like to obtain an interferogram networking.

2) And carrying out phase unwrapping and error removal on the obtained interference pattern networking. For example, the interferogram can be phase unwrapped using a minimum cost stream and then phase error corrected using phase closure and bridging.

3) Inversion is carried out on the interference pattern networking passing through the step 2). In one exemplary embodiment, for example, the interferogram grid may be inverted using SBAS and weighted least squares.

4) And carrying out atmospheric layering and ionosphere delay correction on the inverted interferogram networking. In one exemplary embodiment, for example, an integrated PyAPS open source software package may be utilized to perform atmospheric layering and ionospheric delay correction on the inverted interference pattern based on a global atmosphere model (e.g., ERA-5) or the like.

5) And (3) performing terrain error removal on the interference pattern networking after the step 4). In one exemplary embodiment, topography error removal may be performed on the interference pattern network after 4), for example, using pixel-by-pixel geometric correction and time-series deformation models, etc.

6) And 5) performing earth surface deformation time sequence inversion on the interference pattern networking after the step 5). In one exemplary embodiment, for example, the surface deformation time series inversion may be performed on the interferogram grid after 5) using geocoding to obtain a deformation time series, i.e., a time series sedimentation data set arranged in time series.

In the embodiment of the invention, the historical sedimentation value is the accumulated sedimentation amount.

In the embodiment of the invention, each time sequence sedimentation data set is embodied in a matrix form, and the space correlation with the predicted point i refers to the predicted point corresponding to the unit connected with the unit where the predicted point i is located. In one exemplary embodiment, the historical settlement values of the predicted point i and the historical settlement values of the predicted point having a spatial correlation with the predicted point i are obtained using a set convolution kernel. Therefore, the predicted sedimentation value of each predicted point considers the neighborhood space relation, so that the predicted result is more accurate. In the embodiment of the invention, the proper convolution kernel size can be determined by continuously training test and comparing the predicted actual effect, and convolution kernels with different sizes can be selected according to the requirement.

Further, in the embodiment of the invention, the sedimentation value of the future time node of the predicted point i can be predicted by using a set space-time sequence prediction model. Preferably, the spatio-temporal sequence prediction model may be a ConvLSTM model. Further, in the embodiment of the present invention, S300 further includes:

S310, dividing n time sequence sedimentation data sets into n1 training sets and n2 test sets, dividing the n1 training sets into k1 training sets according to a set time sequence length L, and dividing the n2 test sets into k2 test sets according to the set time sequence length L.

And converting the acquired time sequence sedimentation data set into tensors according to the date, so that convolution operation is convenient. The data set is divided into training and test sets according to a certain proportion according to time sequence, for example, n1:n2=9:1. The time sequence length L may be a custom setting, and may be an empirical value, for example, l=12, which indicates that the ground subsidence data of the first 12 days is used to predict the ground subsidence data of the 13 th day, and the historical subsidence data in the training set are arranged according to the format of [123456789101112], [2345678910111213], to obtain k1 training groups. Similarly, the historical sedimentation data in the test set was arranged in a format of [123456789101112], [2345678910111213]. To obtain k2 test groups.

S320, constructing ConvLSTM models and setting super parameters, and training by using the k1 training groups to obtain a ConvLSTM model after training.

In the training process, k1 training groups are sequentially input into a ConvLSTM model which is constructed, the model extracts data through convolution operation according to a set convolution kernel, and the obtained convolution result is input, updated in state, forgotten and output according to a gate structure meeting the following conditions:

Wherein i is an input gate, f is a forgetting gate, C is a unit state, o is an output gate, X _t represents an input at time t in a time sequence, H _t represents an implicit state output of a corresponding unit, W is a weight coefficient matrix, b is a bias term, σ is a sigmoid activation function, tanh is a hyperbolic tangent activation function, o represents a Hadamard product, X represents convolution, and X, C, H, i, f, o are all three-dimensional tensors.

The super parameters may include the number of network layers, the number of neurons per network layer, the size of the convolution kernel, the learning rate (learning_rate) of the entire network, the weight decay (weight_decay) of the entire network, etc.

As shown in fig. 4, the implementation principle of the conditions (1) - (5) is that three values are input firstly, one is input X _t at the moment, the second is the cell state value C _t-1 at the last moment, and finally the cell output h _t-1 at the last moment. The final outputs are the cell state value C _t and the cell output H _t at this time. Gating calculations are performed separately in the units. Firstly, three values of 0-1, namely the gate values of an input gate (foget gate), a forget gate (input gate) and an output gate (output gate), are calculated through weight calculation and a sigmoid activation function (sigma). And then, carrying out weight calculation by using the input value X and the output H of the last unit, obtaining the state input at the moment through a tanh activation function, integrating the calculation, forgetting to multiply the state at the moment, and inputting the state input at the moment by the gate to obtain the final state C. And the output gate is only associated with the output, and the final output is the output gate multiplied by tanh (C).

The specific process of convolution calculation may be expressed, for example, as that the convolution kernel is a p×p matrix, and the input source data with, for example, q×q size is sequentially scanned and subjected to inner product to obtain the output. When the convolution kernel moves by one position, a convolution value is obtained at the output layer, the whole convolution result is obtained after the scanning is completed, and the final obtained output data size is (Q-P+1) × (Q-P+1).

In an embodiment of the present invention, convLSTM models have multiple hidden layers. To avoid overfitting, normalization is performed when the data is input to scale the raw data so that all sedimentation values in each dataset lie between 0 and 1.

S330, evaluating the trained ConvLSTM models, and determining the model with the evaluation result meeting the first preset condition as the ConvLSTM model to be tested.

In S330, the trained ConvLSTM model may be evaluated based on the mean absolute value MAE, the mean square error MSE, and the root mean square error RMSE;

Wherein, the Wherein yf ⁽ⁱ⁾ is a predicted sedimentation value of an ith predicted point in a training set corresponding to a time node corresponding to a jth training set, the training set corresponding to the time node corresponding to the jth training set is a training set corresponding to the time node predicted by using the jth training set, for example, the jth training set is [2345678910111213], the corresponding training set is a training set with a time node of 14, and y ⁽ⁱ⁾ is an actual sedimentation value of the ith predicted point in the training set corresponding to the time node corresponding to the jth training set.

If the MAE, MSE and RMSE are all smaller than the first preset threshold, that is, the evaluation result meets the first preset condition, the prediction result is accurate, and the corresponding trained ConvLSTM model is determined to be the ConvLSTM model to be tested. The first preset threshold may be determined based on the data accuracy.

S340, inputting the k2 test groups into the ConvLSTM models to be tested, and obtaining test results.

K2 test groups are sequentially input into the ConvLSTM model to be tested, and a corresponding prediction result is obtained, namely each prediction point can obtain a corresponding prediction sedimentation value.

S350, comparing the test result with a reference result corresponding to the test set, and determining the to-be-tested ConvLSTM model as a target ConvLSTM model if the comparison result meets a second preset condition.

In S350, the reference result is the corresponding historical sedimentation value in the test set. The test result and the reference result may be evaluated based on the mean absolute value MAE, the mean square error MSE and the root mean square error RMSE in S330, and detailed description thereof is omitted for avoiding redundant description. If the MAE, MSE and RMSE are all smaller than the second preset threshold, namely the evaluation result meets the second preset condition, the prediction result is accurate, and the ConvLSTM model to be tested is determined to be the target ConvLSTM model. The second preset threshold may be determined based on data accuracy, and in one example may be the same as the first preset threshold, and in another example may be different from the first preset threshold. If the evaluation result meets the second preset condition, the determined target ConvLSTM model can be visually output, and if the evaluation result does not meet the second preset condition, the adjustment can be performed by adjusting the size of the convolution kernel and the like, and training is performed again.

In an exemplary embodiment of the present invention, as shown in fig. 5 (a), the convolution kernel of the obtained object ConvLSTM model object may be a 3*3 matrix, that is, a matrix of 3*3 convolution kernel capable of ensuring that the prediction result meets the preset accuracy, that is, 8 prediction points having spatial correlation with each prediction point. In a specific application, the input source data, for example, 9*9 size as shown in fig. 5 (b), is scanned sequentially by using 3*3 matrix and then subjected to inner product to obtain output. Each time the convolution kernel moves by one position, a convolution value is obtained at the output layer, and when scanning is completed, a result after the whole convolution is obtained, and finally, the size of the obtained output data is 7*7, as shown in fig. 5 (c).

S360, predicting the sedimentation value of the future time node of the predicted point i by using the target ConvLSTM model.

In this step, for example, if it is necessary to predict the sedimentation value at the 13 th day of the predicted point i, the time series data set of the first 12 days may be input to the target ConvLSTM model.

Embodiments of the present application also provide a non-transitory computer readable storage medium that may be disposed in an electronic device to store at least one instruction or at least one program for implementing one of the methods embodiments, the at least one instruction or the at least one program being loaded and executed by the processor to implement the methods provided by the embodiments described above.

Embodiments of the present application also provide an electronic device comprising a processor and the aforementioned non-transitory computer-readable storage medium.

Embodiments of the present application also provide a computer program product comprising program code for causing an electronic device to carry out the steps of the method according to the various exemplary embodiments of the application as described in the specification, when said program product is run on the electronic device.

While certain specific embodiments of the application have been described in detail by way of example, it will be appreciated by those skilled in the art that the above examples are for illustration only and are not intended to limit the scope of the application. Those skilled in the art will also appreciate that many modifications may be made to the embodiments without departing from the scope and spirit of the application. The scope of the application is defined by the appended claims.

Claims

1. A method for predicting land subsidence based on spatiotemporal series data, characterized in that the method comprises the following steps:

S100, obtaining SAR images, orbit data and DEM files within a preset time period of a preset geographical area;

S200, based on the SAR image, the orbit data and the DEM file, obtaining n time-series settlement data sets arranged in chronological order, each time-series settlement data set including historical settlement values of m prediction points;

S300, predicting the settlement value of the prediction point i at a future time node based on the historical settlement value of the prediction point i and the historical settlement values of the prediction points spatially associated with the prediction point i, where i ranges from 1 to m;

The ConvLSTM model is used to predict the settlement value of the future time node of the prediction point i;

S300 further includes:

S310, dividing n time series sedimentation data sets into n1 training sets and n2 test sets, and dividing the n1 training sets into k1 training groups according to a set time series length L, and dividing the n2 test sets into k2 test groups according to the set time series length L;

S320, constructing a ConvLSTM model and setting hyperparameters, and training the constructed ConvLSTM model using k1 training groups to obtain a trained ConvLSTM model;

S330, evaluating the trained ConvLSTM model, and determining a model whose evaluation result meets the first preset condition as a ConvLSTM model to be tested;

S340, inputting k2 test groups into the ConvLSTM model to be tested to obtain a test result;

S350, comparing the test result with the reference results corresponding to the k2 test groups, and if the comparison result satisfies the second preset condition, determining the ConvLSTM model to be tested as the target ConvLSTM model;

S360, using the target ConvLSTM model to predict the settlement value of the future time node of the prediction point i.

2. The method according to claim 1 is characterized in that the historical settlement values of the prediction point i and the historical settlement values of the prediction points spatially associated with the prediction point i are obtained using a set convolution kernel.

3. The method according to claim 1, characterized in that in S320, data is extracted through convolution operation according to the set convolution kernel, and the obtained convolution result is input, state updated, forgotten, and output according to a gate structure that meets the following conditions:

Among them, i is the input gate, f is the forget gate, C is the unit state, is the output gate, _Xt represents the input at the tth moment in the time series, _Ht represents the hidden state output of the corresponding unit, W is the weight coefficient matrix, b is the bias term, σ is the sigmoid activation function, tanh is the hyperbolic tangent activation function, represents Hadamard product and * represents convolution.

4. The method according to claim 1, characterized in that, in S330, the trained ConvLSTM model is evaluated based on mean absolute value MAE, mean square error MSE and root mean square error RMSE;

in, Wherein, yf ⁽ⁱ⁾ is the predicted settlement value of the i-th prediction point in the training set corresponding to the time node corresponding to the j-th training group, and y ⁽ⁱ⁾ is the actual settlement value of the i-th prediction point in the training set corresponding to the time node corresponding to the j-th training group;

If the MAE, MSE and RMSE are all less than the preset thresholds, the corresponding trained ConvLSTM model is determined as the ConvLSTM model to be tested.

5. The method according to claim 1 is characterized in that the ConvLSTM model has multiple hidden layers.

6. The method according to claim 1, characterized in that the convolution kernel of the target ConvLSTM model is a 3*3 matrix.

7. A non-transitory computer-readable storage medium, wherein at least one instruction or at least one program is stored in the storage medium, and the at least one instruction or the at least one program is loaded and executed by a processor to implement the method as claimed in any one of claims 1 to 6.

8. An electronic device, comprising a processor and the non-transitory computer-readable storage medium of claim 7.