SBAS-InSAR ground subsidence prediction method considering ground load

Abstract

The application provides a SBAS-InSAR ground subsidence prediction method considering ground load, first, the overall framework of the ground subsidence model considering the ground load is determined based on an LSTM long short-term memory neural network, then the parameters of the ground subsidence model are configured, the training sample and the verification sample are selected, the ground subsidence model is trained, and the accuracy of the ground subsidence model is evaluated; the trained ground subsidence model is used for predicting and analyzing the subsequent subsidence of the feature points, the application can better evaluate the influence of the building load on the ground subsidence, more accurately predict the ground subsidence trend, and provide a scientific basis for city planning, land development and infrastructure maintenance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of geological disaster monitoring and prevention, and mainly relates to a SBAS-InSAR ground subsidence prediction method considering ground load. BACKGROUND

[0002] Under the background of urbanization and population growth, the sustainable development of urban infrastructure has become an important task. However, due to natural factors, human activities or underground engineering, etc., the damage and structural instability of buildings caused by ground subsidence have become a serious problem. In order to ensure the safety and stability of the city, it is crucial to accurately assess and predict the ground subsidence situation and trend.

[0003] At present, common ground subsidence monitoring methods mainly include GPS measurement, satellite altimetry, precise leveling, etc. However, these methods have certain limitations in terms of spatial coverage, temporal resolution and economic cost, etc., and it is difficult to meet the demand of large-scale regional ground subsidence monitoring. In recent years, the synthetic aperture radar differential interferometric measurement technology (D-InSAR) has emerged, which can obtain high-precision, low-cost ground deformation information in all-weather and all-day. The small baseline set technology (SBAS) is a long-time sequence analysis method based on this, which effectively suppresses the influence of time decorrelation, spatial decorrelation and atmospheric delay. The application of this method in ground subsidence has achieved a series of results.

[0004] At present, various machine learning algorithms are used in InSAR ground subsidence result prediction. Compared with traditional neural network models, the long short-term memory network model LSTM can effectively capture long-term dependencies and process time-related information in sequences, and is particularly suitable for processing sequence data. Therefore, it has also been gradually used in time series ground deformation prediction. In the region with rapid urbanization process, building load is one of the main reasons for causing ground subsidence. However, in the current research, when the obtained ground subsidence is predicted in the time dimension, the influence of the factors causing ground subsidence on the accuracy of the prediction result is not considered.

[0005] In order to solve the above problems, the present application combines the long short-term memory network model LSTM with the SBAS-InSAR ground subsidence monitoring technology, and proposes a new method to predict the ground subsidence. SUMMARY

[0006] To solve the above technical problems, the present application proposes a SBAS-InSAR ground subsidence prediction method considering ground load, based on the ground subsidence results obtained by SBAS-InSAR technology, analyzes the correlation between the results and the ground building area load, and based on the LSTM long short-term memory neural network, constructs a subsidence prediction model considering the ground building load.

[0007] To achieve the above object, the technical scheme adopted by the present application is:

[0008] A SBAS-InSAR land subsidence prediction method considering ground load, characterized in that it comprises the following steps:

[0009] Step 1: download all Sentinel-1A SLC format SAR data and all Landsat8 OLI and Sentinel-2A / B optical data covering the prediction area, then pre-process the Sentinel-1A to obtain time series SAR data covering the study area, pre-process the Landsat8 OLI and Sentinel-2A / B to obtain accurate optical data, then select the same name control points on the SAR data and optical data respectively for spatial registration, and ensure the consistency of the range of all image data through cropping;

[0010] Step 2: extract parameters from the pre-processed Landsat8 OLI and Sentinel-2A / B optical data, the extracted parameters include normalized difference bare soil and built-up land index, soil-adjusted vegetation index, improved normalized water index, bare soil index, enhanced bare soil index, and enhanced built-up land index, the extraction formula is as follows:

[0011]

[0012]

[0013]

[0014]

[0015]

[0016]

[0017] In the above formula, NDBBI is the normalized difference bare soil and built-up land index, SAVI is the soil-adjusted vegetation index, MNDWI is the improved normalized water index, BSI is the bare soil index, EBSI is the enhanced bare soil index, EIBI is the enhanced built-up land index, B, G, R, NIR, SWIR1 / SWIR2 are the blue band, green band, red band, near-infrared band and short-wave infrared band in the Landsat8 OLI or Sentinel-2 remote sensing image respectively; L is the soil adjustment factor, its value is between 0-1, which is set according to the vegetation coverage of the study area, the higher the vegetation coverage, the smaller the value of L;

[0018] Step 3: The SLC format Sentinel-1A SAR data after preprocessing is processed by SBAS-InSAR technology to obtain the ground subsidence information of the prediction area. First, select one of the time series Sentinel-1A SAR data as the main image, and perform high-precision registration with the other SAR images. After setting the time baseline and spatial baseline threshold critical value, all SAR images and the main image will form an interference pair. Remove the flat ground effect by introducing 30m precision DEM data, select Goldstein filtering method, use the least cost flow method for phase unwrapping, select ground control points in the flat terrain area without deformation stripes and phase jumps, estimate and remove residual constant phase and slope phase by using cubic polynomial method, establish a linear model of deformation rate and elevation coefficient of coherent points to form an equation group, and use matrix singular value decomposition method to preliminarily estimate the deformation rate and terrain error. After the first inversion of the deformation rate, estimate and remove the atmospheric phase by using atmospheric time domain high-pass and spatial domain low-pass filtering;

[0019] Step 4: Through correlation analysis, the correlation between ground subsidence and building area is estimated. The distance correlation coefficient is used to measure the correlation between time series subsidence data and building area, and the calculation formula is as follows:

[0020]

[0021] Wherein,

[0022]

[0023]

[0024]

[0025] Above, DC represents distance correlation, n is the length of random variables x and y, and || || mean represents taking the average of the sum of all elements in the matrix, A x and A y represent the centering distance matrix, and the size of the DC coefficient is between [0-1]. The closer the DC coefficient is to 1, the stronger the correlation between the two random variables, and the DC coefficient is 0, which means that the two random variables are completely independent.

[0026] Step 5: On the basis of correlation analysis, an LSTM long short-term neural network is constructed based on the deep learning designer tool. First, the overall framework of the ground subsidence model is established, and then the parameters of the ground subsidence model are configured,

[0027] Establish the overall framework of the ground subsidence model:

[0028] An overall framework for determining a land subsidence model based on an LSTM long short-term memory neural network,

[0029] The calculation of a single neuron in an LSTM hidden layer includes cell state updating and output value calculation, and there are three gate functions in the neuron: a forgetting gate, an input gate, and an output gate, which control the input value, memory value, and output value through the gate functions. The forgetting gate, input gate, and output gate calculation formulas are as follows:

[0030] f t =σ(W f h t-1 +W f x t +b f )(11)

[0031] i t =σ(W i h t-1 +W i x t +b i )(12)

[0032] o t =σ(W o h t-1 +W o x t +b o )(13)

[0033] where h t-1 is the hidden state at the previous time; σ is a sigmoid function; f t , i t , o t are the forgetting gate, input gate, and output gate state calculation results, respectively; W f , W i , W o are the weight matrices of the forgetting gate, input gate, and output gate, respectively; b f , b i , b o are the forgetting gate, input gate, and output gate bias terms, respectively; the final output of the LSTM long short-term neural network is determined by the output gate and the cell state, is a candidate value vector, and the product of the input value and the candidate value vector is used to update the cell state, and the calculation process is as follows:

[0034]

[0035]

[0036] h t =o t tanh(C t )(16)

[0037]

[0038] f(x)=tanh(x)(18)

[0039] where W c is the input unit state weight matrix; b c is the input unit state bias term; tanh is the activation function tanh, the forget gate controls how much information is discarded from the cell state at the current time, o t is the neuron output value, h t is the current time hidden state;

[0040] Parameter configuration of land subsidence model: in the training, the input layer neuron format of the LSTM neural network, the number of network iterations, the gradient threshold value, the specified initial training rate need to be set, the loss function is RMSE, the activation function is tanh, the network training adopts the Adam optimization algorithm, the gradient threshold value is set to 1, the specified initial training rate is 0.003, the input of the network is composed of the current time SBAS-InSAR ground feature point subsidence and the building area, the network output is the predicted feature point subsidence at the current time, the time series subsidence data and the time series building area data are divided into two time periods T1 and T2, all the samples of the two time period data are divided into a training set and a test set according to the ratio of 8:2, the model is trained by using the training set, and the predicted value is compared with the known value, and the absolute error and the root mean square error are used as the evaluation index of the prediction accuracy;

[0041] Step 6, after the land subsidence model is trained and tested, the future land subsidence of the predicted area can be predicted, and the specific steps are as follows: after the land subsidence model is trained and tested, the predictAndUpdateState function in the deep learning designer tool of the Matlab software platform is used to carry out multi-step prediction on the future subsidence, and the network state is updated at each prediction, since the future monitoring value is lacking, the predicted value is used as the true value to continuously construct the input of the new prediction sequence, and thus the future land subsidence of the feature point is predicted and analyzed.

[0042] As a preferred technical scheme of the application: in step 1, the same name control points are selected on SAR data and optical data for spatial registration, and the range of all image data is ensured to be consistent through cropping.

[0043] As a preferred technical scheme of the present application: in step 2, the normalized difference bare land and building land index, the soil-adjusted vegetation index, the improved normalized water index and the bare soil index are extracted from the Landsat8 OLI and Sentinel-2A / B time series optical data by using the grid statistical method, then the enhanced bare soil index for enhancing the bare land information is calculated according to the bare soil index and the improved normalized water index, then the enhanced building land index for reflecting the time series change of building load is calculated according to the normalized difference bare land and building land index, the improved normalized water index, the soil-adjusted vegetation index and the enhanced bare soil index, and finally the research area time series building land area is obtained by normalization processing and grid calculation.

[0044] The research area time series building land area is obtained by normalization processing and grid calculation.

[0045] As a preferred technical scheme of the present application: in step 3, the registration error range of the main image and other SAR images is 0.5 pixels.

[0046] As a preferred technical scheme of the present application: in step 5, the LSTM long short-term neural network is constructed based on the deep learning designer tool by using the Matlab software platform, the input of the LSTM long short-term neural network is composed of the current time SBAS-InSAR ground feature point subsidence and the building area, and the network output is the current time predicted feature point subsidence.

[0047] As a preferred technical scheme of the present application: in step 6, the prediction time of the predicted area future ground subsidence is two years.

[0048] The application provides an SBAS-InSAR ground subsidence prediction method considering ground load, which comprises the following steps: pre-processing all Sentinel-1A SLC format SAR time series data covering the same region within a certain period of time and time series optical data composed of all Landsat8 OLI images and Sentinel-2A / B images; extracting normalized difference bare land and building land index, soil adjusted vegetation index, improved normalized water body index and bare soil index of the time series optical data, calculating an enhanced building land index according to the above indexes, and obtaining the time series building land area of the research region through normalization processing and grid calculation; extracting feature point subsidence information and average subsidence rate based on the SBAS-InSAR technology of the time series Sentinel-1A SAR data; calculating the correlation between the ground subsidence and the ground load by using the distance correlation coefficient; determining the overall framework of the ground subsidence model and the configuration parameters of the ground subsidence model based on the LSTM long short-term memory neural network, selecting training samples and verification samples, training the ground subsidence model, and evaluating the accuracy of the ground subsidence model; and predicting and analyzing the subsequent subsidence of the feature points by using the trained ground subsidence model, so that the influence of the ground subsidence on the buildings can be better evaluated, the ground subsidence trend can be more accurately predicted, and scientific basis can be provided for city planning, land development and infrastructure maintenance.

[0049] Compared with the prior art, the application has the following beneficial effects:

[0050] The application has the advantages that the SBAS-InSAR ground subsidence prediction model considering the ground building load is constructed based on the LSTM long short-term memory neural network in deep learning, so that the influence of the building load on the ground subsidence can be better evaluated, the ground subsidence trend can be more accurately predicted, and scientific basis can be provided for city planning, land development and infrastructure maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 The flowchart of the application is shown in the figure;

[0052] Figure 2 The SBAS-InSAR ground subsidence monitoring result in the embodiment is shown in the figure;

[0053] Figure 3 The LSTM neuron structure in the embodiment is shown in the figure. DETAILED DESCRIPTION

[0054] The application will be further described in detail in combination with the drawings and specific embodiments:

[0055] As Figure 1As shown, the present application proposes a SBAS-InSAR land subsidence prediction method considering ground load, comprising the following steps:

[0056] Step 1: download all Sentinel-1A SLC format SAR data and all Landsat8 OLI and Sentinel-2A / B optical data covering the prediction area, then pre-process the Sentinel-1A to obtain time series SAR data covering the study area; pre-process the Landsat8 OLI and Sentinel-2A / B to obtain accurate optical data, then select the same name control points on SAR data and optical data respectively for spatial registration, and ensure the range of all image data through cropping;

[0057] Step 2: parameter extraction is performed on the pre-processed Landsat8 OLI and Sentinel-2A / B optical data, and the extracted parameters include: normalized difference bare and built-up index, soil-adjusted vegetation index, improved normalized water index, bare soil index, enhanced bare soil index, and enhanced built-up index, and the extraction formula is as follows:

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] In the above formula, NDBBI is the normalized difference bare and built-up index, SAVI is the soil-adjusted vegetation index, MNDWI is the improved normalized water index, BSI is the bare soil index, EBSI is the enhanced bare soil index, EIBI is the enhanced built-up index, B, G, R, NIR, SWIR1 / SWIR2 are the blue band, green band, red band, near-infrared band and short-wave infrared band in Landsat8 OLI or Sentinel-2 remote sensing image respectively; L is the soil adjustment factor, its value is between 0-1, which is set according to the vegetation coverage of the study area, the higher the vegetation coverage, the smaller the value of L;

[0065] Step 3: The SLC format Sentinel-1A SAR data after preprocessing is processed by SBAS-InSAR technology to obtain the ground subsidence information of the prediction area. First, select one of the time series Sentinel-1A SAR data as the main image, and perform high-precision registration with the other SAR images. After setting the time baseline and spatial baseline threshold critical value, all SAR images and the main image will form an interference pair. Remove the flat ground effect by introducing 30m precision DEM data, select Goldstein filtering method, use the least cost flow method for phase unwrapping, select ground control points in the flat terrain area without deformation stripes and phase jumps, estimate and remove residual constant phase and slope phase by using cubic polynomial method, establish a linear model of deformation rate and elevation coefficient of coherent points to form an equation group, and use matrix singular value decomposition method to preliminarily estimate the deformation rate and terrain error. After the first inversion of the deformation rate, estimate and remove the atmospheric phase by using atmospheric time domain high-pass and spatial domain low-pass filtering;

[0066] Step 4: Through correlation analysis, the correlation between the ground subsidence and the area of the ground building is estimated. The distance correlation coefficient is used to measure the correlation between the time series subsidence data and the building area, and the calculation formula is as follows:

[0067]

[0068] Wherein,

[0069]

[0070]

[0071]

[0072] Above, DC represents distance correlation, n is the length of random variables x and y, and || || mean represents taking the average of the sum of all elements in the matrix, A x and A y represent the centering distance matrix, and the size of the DC coefficient is between [0-1]. The closer the DC coefficient is to 1, the stronger the correlation between the two random variables. If the DC coefficient is 0, it means that the two random variables are completely independent.

[0073] Step 5: On the basis of correlation analysis, an LSTM long short-term neural network is constructed based on the deep learning designer tool. First, the overall framework of the ground subsidence model is established, and then the parameters of the ground subsidence model are configured,

[0074] Establish the overall framework of the ground subsidence model:

[0075] An overall framework for determining a land subsidence model based on an LSTM long short-term memory neural network,

[0076] The calculation of a single neuron in an LSTM hidden layer includes cell state updating and output value calculation, and there are three gate functions in the neuron: a forgetting gate, an input gate, and an output gate, which control the input value, memory value, and output value through the gate functions, and the forgetting gate, input gate, and output gate calculation formulas are as follows:

[0077] f t =σ(W f h t-1 +W f x t +b f )(11)

[0078] i t =σ(W i h t-1 +W i x t +b i )(12)

[0079] o t =σ(W o h t-1 +W o x t +b o )(13)

[0080] where h t-1 is the hidden state at the previous time; σ is a sigmoid function; f t , i t , o t are the forgetting gate, input gate, and output gate state calculation results, respectively; W f , W i , W o are the weight matrices of the forgetting gate, input gate, and output gate, respectively; b f , b i , b o are the forgetting gate, input gate, and output gate bias terms, respectively; the final output of the LSTM long short-term neural network is determined by the output gate and the cell state, is a candidate value vector, and the product of the input value and the candidate value vector is used to update the cell state, and the calculation process is as follows:

[0081]

[0082]

[0083] h t =o t tanh(C t )(16)

[0084]

[0085] f(x) = tanh(x) (18)

[0086] where W c is the input unit state weight matrix; b c is the input unit state bias term; tanh is the activation function tanh, the forget gate controls how much information is discarded from the cell state at the current time step, o t is the neuron output value, h t is the current hidden state;

[0087] Parameter configuration of land subsidence model: In the training, the input layer neuron format of LSTM neural network, network iteration number, gradient threshold value, specified initial training rate need to be set, the loss function is RMSE, the activation function is tanh, the network training adopts Adam optimization algorithm, the gradient threshold value is set to 1, the specified initial training rate is 0.003, the input of the network is composed of the current time SBAS-InSAR ground feature point subsidence and building area, the network output is the predicted feature point subsidence at the current time, the time series subsidence data and time series building area data are divided into two time periods T1 and T2, all samples of the two time period data are divided into training set and test set according to the proportion of 8:2, the model is trained by using the training set, and the predicted value is compared with the known value, and the absolute error and the root mean square error are used as the evaluation index of the prediction accuracy;

[0088] Step 6, after the land subsidence model is trained and tested, the future land subsidence of the predicted area can be predicted, and the specific steps are as follows: after the land subsidence model is trained and tested, the predictAndUpdateState function in the deep learning designer tool of the Matlab software platform is used to make multi-step prediction on the future subsidence, and the network state is updated at each prediction, since the future monitoring value is lacking, the predicted value is used as the true value to construct the input of the new prediction sequence, so as to realize the prediction and analysis of the land subsidence of the feature point in the future time.

[0089] In step 1, the same name control points are selected on SAR data and optical data for spatial registration, and the range of all image data is ensured to be consistent through cropping.

[0090] In step 2, the normalized difference built-up and land index, the soil-adjusted vegetation index, the improved normalized water index and the bare soil index are extracted from the Landsat8 OLI and Sentinel-2A / B time series optical data by using the grid statistical method, then the enhanced bare soil index for enhancing the bare soil information is calculated according to the bare soil index and the improved normalized water index, then the enhanced built-up index for reflecting the time series change of building load is calculated according to the normalized difference built-up and land index, the improved normalized water index, the soil-adjusted vegetation index and the enhanced bare soil index, and finally the time series building area of the study area is obtained by normalization processing and grid calculation.

[0091] In step 3, the registration error range of the main image and other SAR images is 0.5 pixels.

[0092] In step 5, the LSTM long short-term neural network is constructed by using the Matlab software platform based on the deep learning designer tool, the input of the network is composed of the current time SBAS-InSAR ground feature point settlement and the building area, and the output of the network is the predicted feature point settlement at the current time.

[0093] In step 6, the prediction time of the predicted area future ground settlement is two years.

[0094] As shown in Figure 2 In this embodiment, Nanjing is taken as the predicted area for illustration and demonstration, firstly, all Sentinel-1A SLC format SAR data and all Landsat8 OLI and Sentinel-2A / B optical data covering Nanjing area since 2016 need to be downloaded, the preprocessing of Sentinel-1A mainly includes cropping to obtain time series SAR data covering the study area; the Landsat8 OLI and Sentinel-2A / B data are subjected to radiometric correction and geometric precise correction to eliminate the solar electromagnetic radiation distortion and geometric distortion caused by atmospheric influence, then the same name control points are selected on the SAR data and optical data respectively for spatial registration, and the range of all image data is ensured to be consistent by cropping;

[0095] Then, the pre-processed Landsat8 OLI and Sentinel-2A / B optical data are extracted for parameters, mainly including: normalized difference built-up and bare index (NDBBI), soil-adjusted vegetation index (SAVI), modified normalized water index (MNDWI), bare soil index (BSI), enhanced bare soil index (EBSI), and enhanced built-up index (EIBI). The normalized difference built-up and bare index (NDBBI), soil-adjusted vegetation index (SAVI), modified normalized water index (MNDWI), and bare soil index (BSI) are extracted from the time series of Landsat8 OLI and Sentinel-2A / B optical images by using the grid statistical method. The enhanced bare soil index (EBSI) for enhancing the bare soil information is calculated according to the BSI index and the MNDWI index. The enhanced built-up index (EIBI) for reflecting the time series change of building load is calculated according to the NDBBI index, the MNDWI index, the SAVI index, and the EBSI index. The time series building area of the study area is obtained by normalization processing and grid calculation.

[0096] Then, the pre-processed SLC format Sentinel-1A SAR data is processed by SBAS-InSAR technology to obtain the ground subsidence information of Nanjing. First, one of the time series Sentinel-1A SAR data is selected as the master image, and the other SAR images are registered with it with high precision. The registration error is controlled within 0.5 pixels. After setting the time baseline and spatial baseline threshold critical value, all SAR images and the master image will form an interference pair. The flat earth effect is removed by introducing 30m precision DEM data. The Goldstein filtering method is selected, and the Minimum Cost Flow (MCF) method is used for phase unwrapping. Ground control points are selected in the terrain flat area without deformation stripes and phase jumps. The residual constant phase and slope phase are estimated and removed by using the cubic polynomial method. The linear model of the deformation rate and the elevation coefficient of the coherent points is established to form the equation group. The matrix singular value decomposition method is used to preliminarily estimate the deformation rate and the terrain error. After the first inversion of the deformation rate, the atmospheric time domain high-pass and spatial domain low-pass filtering is used to estimate and remove the atmospheric phase to obtain a more pure time series deformation variable.

[0097] Then, the correlation between the ground subsidence and the ground building area is estimated by correlation analysis to provide a basis for subsequent subsidence prediction using deep learning modeling. The Distance Correlation (DC) coefficient is used to measure the correlation between the time series subsidence data and the building area.

[0098] Next, based on the correlation analysis, the overall framework of the ground settlement model was first established, and then the parameters of the ground settlement model were configured.

[0099] The overall framework of the land subsidence model is established based on an LSTM (Long Short-Term Memory) neural network. The computation of a single neuron in the LSTM hidden layer includes cell state updates and output value calculation. Each neuron has three gate functions: a forget gate, an input gate, and an output gate. These gate functions control the input, memory, and output values. During training, the parameters of the land subsidence model need to be configured, including the input layer neuron format, the number of network iterations, the gradient threshold, and the initial training rate. The loss function is RMSE, the activation function is tanh, and Adam optimization is used for network training. To prevent gradient explosion, the gradient threshold is set to 1, and the initial training rate is specified as 0.003. The network input consists of the current SBAS-InSAR ground feature point subsidence and building area, and the network output is the predicted feature point subsidence at the current time. Here, the time series subsidence data and time series building area data are divided into two time periods, T1 and T2. All samples of the data in the two time periods are divided into training set and test set in a ratio of 8:2. The model is trained using the training set, and the predicted values ​​are compared and analyzed with the known values. The absolute error and root mean square error are used as evaluation indicators of prediction accuracy.

[0100] Finally, based on the training and testing sets, the ground subsidence model outputs training results to predict the ground subsidence in Nanjing over the next two years. Specifically, after training and testing, the `predictAndUpdateState` function in the deep learning designer tool of the Matlab software platform is used to perform multi-step predictions of the subsidence over the next two years, updating the network state at each prediction. Since future monitoring values ​​are lacking, the predicted values ​​are continuously used as real values ​​to construct new prediction sequences, thereby achieving the prediction and analysis of ground subsidence at feature points over a future period.

[0101] The advantage of this invention lies in its construction of an SBAS-InSAR ground settlement prediction model that takes into account ground building loads, based on the LSTM (Long Short-Term Memory) neural network in deep learning. This method allows for a better assessment of the impact of building loads on ground settlement and a more accurate prediction of ground settlement trends, providing a scientific basis for urban planning, land development, and infrastructure maintenance.

[0102] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A SBAS-InSAR ground settlement prediction method that takes into account ground load, characterized in that: Comprising the following steps: Step 1: download all Sentinel-1A SLC format SAR data and all Landsat8 OLI and Sentinel-2A / B optical data covering the predicted area, then pre-process the Sentinel-1A to obtain time series SAR data covering the study area, pre-process the Landsat8 OLI and Sentinel-2A / B to obtain accurate optical data, then select the same name control points on the SAR data and optical data respectively for spatial registration, and ensure that the range of all image data is consistent through cropping; Step 2: extract parameters from the pre-processed Landsat8 OLI and Sentinel-2A / B optical data, including normalized difference bare and built-up index, soil-adjusted vegetation index, improved normalized water index, bare soil index, enhanced bare soil index, and enhanced built-up index, the extraction formula is as follows: (1)； (2)； （3）； (4)； (5)； (6)； In the above formula, NDBBI is the normalized difference bare and built-up index, SAVI is the soil-adjusted vegetation index, MNDWI is the improved normalized water index, BSI is the bare soil index, EBSI is the enhanced bare soil index, EIBI is the enhanced built-up index, B, G, R, NIR, SWIR1 / SWIR2 are the blue band, green band, red band, near-infrared band and short-wave infrared band of Landsat8 OLI or Sentinel-2 remote sensing image respectively; L is the soil adjustment factor, its value is between 0~1, which is set according to the vegetation coverage of the study area, the higher the vegetation coverage, the smaller the value of L; In step 2, the raster statistical method is used to extract the normalized difference bare and built-up index, soil-adjusted vegetation index, improved normalized water index, and bare soil index from the Landsat8 OLI and Sentinel-2A / B time series optical data, then the enhanced bare soil index for enhancing bare land information is calculated according to the bare soil index and the improved normalized water index, then the enhanced built-up index for reflecting the time series change of building load is calculated according to the normalized difference bare and built-up index, the improved normalized water index, the soil-adjusted vegetation index, and the enhanced bare soil index, and the time series building area of the study area is obtained through normalization processing and raster calculation; Step 3: The SLC format Sentinel-1A SAR data after preprocessing is processed by SBAS-InSAR technology to obtain ground subsidence information in the prediction area. First, select one of the time series Sentinel-1A SAR data as the main image, and perform high-precision registration with the other SAR images. After setting the time baseline and spatial baseline threshold critical value, all SAR images and the main image will form an interference pair. Remove the flat ground effect by introducing 30m precision DEM data, select Goldstein filtering method, use the least cost flow method for phase unwrapping, select ground control points in the flat terrain area without deformation stripes and phase jumps, estimate and remove residual constant phase and slope phase by using cubic polynomial method, establish a linear model of deformation rate and elevation coefficient of coherent points to form an equation system, and use matrix singular value decomposition method to estimate the deformation rate and terrain error. After the first inversion of the deformation rate, estimate and remove the atmospheric phase by using atmospheric time domain high-pass and spatial domain low-pass filtering; Step 4: Through correlation analysis, estimate the correlation between ground subsidence and building area. Use distance correlation coefficient to measure the correlation between time series subsidence data and building area, and the calculation formula is as follows: (7)； Wherein, (8)； (9)； (10)； DC denotes distance correlation, n is the length of random variables x and y, denotes the average of all elements in the matrix, and denotes the centering distance matrix, the size of the DC coefficient is between [0-1], the closer the DC coefficient is to 1, the stronger the correlation between the two random variables, and the DC coefficient is 0, which means that the two random variables are completely independent; Step 5: Based on the correlation analysis, build LSTM long short-term neural network based on deep learning designer tool. First, establish the overall framework of the ground subsidence model, then configure the parameters of the ground subsidence model, Establish the overall framework of the ground subsidence model: Determine the overall framework of the ground subsidence model based on LSTM long short-term memory neural network, The calculation of a single neuron in the LSTM hidden layer includes cell state update and output value calculation. There are three gate functions in the neuron: forget gate, input gate and output gate. The input value, memory value and output value are controlled by the gate functions. The calculation formulas of the forget gate, input gate and output gate are as follows: (11)； (12)； (13)； wherein is the hidden state of the previous time; is a sigmoid function; , , are the state settlement results of the forget gate, the input gate and the output gate respectively; , , are the weight matrices of the forget gate, the input gate and the output gate respectively; , , are the bias terms of the forget gate, the input gate and the output gate respectively; the final output of the LSTM long short-term neural network is determined by the output gate and the cell state, is a candidate value vector, and the product of the input value and the candidate value vector is used to update the cell state, and the calculation process is as follows: (14)； (15)； (16)； (17)； (18)； wherein, is an input unit state weight matrix; is an input unit state bias term; is an activation function , the forget gate controls how much information of the current time cell state is discarded, is a neuron output value, is a current time hidden state; Parameter configuration of ground subsidence model: In training, the input layer neuron format of LSTM neural network, network iteration number, gradient threshold, specified initial training rate need to be set. The loss function is RMSE, the activation function is tanh, the network training uses Adam optimization algorithm, the gradient threshold is set to 1, the specified initial training rate is 0.003, the network input is composed of the current time SBAS-InSAR ground feature point subsidence and building area, and the network output is the predicted feature point subsidence at the current time. Time series subsidence data and time series building area data are divided into two time periods T1 and T2. All samples in the two time periods are divided into training set and test set according to the ratio of 8:

2. Train the model with the training set, compare the predicted value with the known value, and use absolute error and root mean square error as the evaluation index of prediction accuracy. After the ground subsidence model is trained and tested, the future ground subsidence of the predicted area can be predicted. The specific steps are as follows: after the ground subsidence model is trained and tested, the predictAndUpdateState function in the deep learning designer tool of the Matlab software platform is used to make multi-step prediction of the future subsidence, and the network state is updated at each prediction. Since there is a lack of future monitoring values, the predicted values are used as the true values to construct new input prediction sequences, thereby realizing the prediction and analysis of the ground subsidence of the feature points in the future time. 2.The SBAS-InSAR land subsidence prediction method considering ground load according to claim 1, wherein: In step 1, the same control points are selected on the SAR data and the optical data for spatial registration, and the range of all image data is ensured to be consistent through cropping. 3.The SBAS-InSAR land subsidence prediction method considering ground load according to claim 1, wherein: In step 3, the registration error range of the main image and other SAR images is 0.5 pixels.

4. The SBAS-InSAR land subsidence prediction method considering ground load according to claim 1, characterized in that: In step 5, the LSTM long short-term neural network is constructed based on the deep learning designer tool of the Matlab software platform. The input of the LSTM long short-term neural network is composed of the SBAS-InSAR ground feature point subsidence and the building area at the current time, and the network output is the predicted feature point subsidence at the current time.

5. The SBAS-InSAR land subsidence prediction method considering ground load according to claim 1, characterized in that: In step 6, the prediction time of the future ground subsidence of the predicted area is two years.

Images