Surface deformation prediction method of goaf area by integrating convolutional attention and multiple influencing factors

Abstract

The present invention belongs to the field of surface deformation prediction, and provides a method for predicting surface deformation of goaf areas that integrates convolutional attention and multiple influencing factors, including S1: obtaining relevant data covering the goaf area, the relevant data including SAR image data, auxiliary data and external influencing factor data; S2: constructing a time series data set based on temperature, precipitation and InSAR deformation of the goaf according to the relevant data; S3: preprocessing the time series data set and dividing it into a training set and a test set, constructing and training a deformation prediction model based on a convolutional attention mechanism and multivariate input; S4: predicting the future deformation of the goaf according to the deformation prediction model. The present invention can achieve high-precision deformation prediction under complex deformation modes; the deformation prediction model based on the Transformer encoder and Bi‑LSTM decoder architecture can improve the accuracy and robustness of the deformation prediction model.

Description

Technical Field

[0001] The present invention relates to the technical field of surface deformation prediction, and in particular to a method for predicting surface deformation of a goaf area by integrating convolutional attention and multiple influencing factors. Background Art

[0002] Underground coal mineral resources are an important energy pillar for national and social development, and also play a vital role in the prosperity of regional economy. However, the goaf formed during the mining process and after the closure of underground coal mines will lead to serious surface deformation, which in turn will cause many geological disasters, such as ground subsidence, ground fissures and building damage, which poses a potential threat to important local ground buildings (railways, highways and residential buildings, etc.) and ecological environment in the mining area. Therefore, predicting the surface deformation caused by goaf is of great practical significance for ensuring safe production in mining areas, ecological protection and restoration, and even the sustainable development of mineral resource-based cities.

[0003] InSAR (Interferometric synthetic aperture radar) is an emerging space geodetic technology developed in the past four decades. It has all-weather and all-day earth observation capabilities, large-scale spatial coverage, high spatial resolution and certain surface penetration capabilities. It can interpret the temporal and spatial evolution of the surface deformation of the entire goaf from a macroscopic perspective of remote sensing. It has been widely used in urban subsidence, mining area deformation monitoring, landslide monitoring and other fields. InSAR methods include D-InSAR (Differential Interferometric Synthetic Aperture Radar) method and time series InSAR method. Among them, the time series InSAR method includes SBAS-InSAR (Small baseline subset InSAR) method, DS-InSAR (Distributed Scatterer InSAR) method and MSBAS-InSAR (Multi-dimensional Small Baseline Subset InSAR) method.

[0004] At present, prediction models based on deep learning methods are widely used in the prediction of surface deformation in mining areas. The invention patent with publication number CN117633494A discloses a coal mine surface deformation prediction method based on the AWC-LSTM model. The time series deformation of the Shilausu coal mining area in Inner Mongolia Autonomous Region was obtained based on the time series InSAR method. The autoregressive integrated moving average model (ARIMA) was used to obtain the linear deformation prediction results, and then the convolutional neural network (CNN)-long short-term memory network (LSTM) model was constructed to obtain the nonlinear deformation prediction results. By fusing the two deformation prediction results, the complete coal mine area surface deformation prediction results were finally obtained. However, the prediction model in this patent lacks sufficient attention to key time steps or features, and has limited effect in capturing the time dependence of longer deformation time series and complex deformation characteristics. In addition, the patent also ignores the potential relationship between the surface deformation of the mining area and other influencing factors (such as temperature, precipitation, etc.), and only uses the single variable of InSAR deformation as the input of the prediction model. According to relevant research, when predicting deformation, using multiple factors that affect surface deformation as variable inputs can effectively improve the final prediction accuracy. In addition, the above patent mainly focuses on the prediction of settlement during mining, and does not consider the uplift deformation caused by groundwater upwelling when the mining area is closed. That is, the above patent only considers the single deformation mode of settlement caused by mining in the mining area, but does not consider the deformation prediction problem under complex deformation modes (mining and closing mining areas coexist). Summary of the invention

[0005] In view of this, the present invention aims to provide a method for predicting surface deformation of goaf areas by integrating convolutional attention with multiple influencing factors, so as to solve the problem that the existing prediction models based on deep learning methods have limitations in predicting surface deformation.

[0006] To achieve the above object, the technical solution created by the present invention is implemented as follows:

[0007] A method for predicting surface deformation of goaf areas by integrating convolutional attention and multiple influencing factors includes the following steps:

[0008] S1: Acquire relevant data covering the goaf area; wherein the relevant data include SAR image data, auxiliary data and external influencing factor data, SAR image data include ascending SAR image and descending SAR image, auxiliary data include digital elevation model data and precise orbit data, external influencing factor data include surface temperature data and precipitation data;

[0009] S2: Based on the acquired relevant data covering the goaf area, a time series dataset based on temperature, precipitation and InSAR deformation of the goaf area is constructed;

[0010] S3: Preprocess the time series data set of the goaf to obtain a sample sequence data set including three variables: surface temperature, precipitation and deformation, and divide the sample sequence data set into a training set and a test set. According to the training set and the test set, a deformation prediction model based on the convolutional attention mechanism and multi-variable input is constructed and trained; wherein the deformation prediction model includes an input module, a feature extraction module, an encoder module based on the Transformer encoder structure, a decoder module based on the Bi-LSTM structure and an output module; the input module is used to input the sample sequence data set into the deformation prediction model as the input of the feature extraction module; the feature extraction module is used to extract features of the input sample sequence data set in the time dimension; the encoder module is used to perform embedded encoding on the extracted features at different time series positions and capture the correlation between the features; the decoder module is used to perform asymmetric decoding and prediction on the features output by the encoder module; the output module is used to output the predicted deformation time series graph;

[0011] S4: Predict the future deformation of the goaf based on the deformation prediction model.

[0012] Furthermore, step S1 specifically includes the following steps:

[0013] S11: vector range for delineating the location of the goaf;

[0014] S12: Acquire ascending SAR images, descending SAR images, digital elevation model data, precise orbit data, surface temperature data, and precipitation data according to the vector range of the goaf location.

[0015] Furthermore, the digital elevation model data are ALOS WORLD 3D-30m DEM data, the precise orbit data are POD precise orbit determination ephemeris data, and the surface temperature data and precipitation data are 2m temperature hourly data and total precipitation hourly data in the ERA5 hourly dataset, respectively.

[0016] Furthermore, step S2 specifically includes the following steps:

[0017] S21: The ascending SAR image and the descending SAR image are processed by the distributed scatterer InSAR method to obtain the ascending unwrapped phase map and the descending unwrapped phase map of the goaf area;

[0018] S22: The ascending and descending unwrapped phase images are processed using the multidimensional small baseline InSAR method to obtain the vertical deformation time series of the goaf.

[0019] Furthermore, step S21 specifically includes the following steps:

[0020] S211: Select a super master image, establish a master-slave relationship for all images according to a set spatiotemporal baseline threshold, align and resample all slave images to the super master image, and then perform interference processing to generate multiple interference images;

[0021] S212: Removal of terrain phase in interferograms using ALOS WORLD 3D-30m DEM data;

[0022] S213: Based on the amplitude information of the ascending SAR image and the amplitude information of the descending SAR image, a fast statistical homogeneous pixel selection method is used to identify the number of homogeneous pixels around each pixel, and pixels with a surrounding homogeneous pixel number higher than a threshold are used as DS preliminary candidate points, and pixels with a surrounding homogeneous pixel number lower than the threshold are used as PS preliminary candidate points;

[0023] S214: establishing a coherence matrix on each DS preliminary candidate point and PS preliminary candidate point, and using a coherence matrix eigenvalue decomposition method to select the phase of the eigenvector corresponding to the maximum eigenvalue as the optimized phase to replace the original phase of each DS preliminary candidate point and PS preliminary candidate point;

[0024] S215: First, the GACOS tool is used to correct the atmospheric phase delay of the optimized phase. Then, the best fitting surface is used to remove the residual linear trend in the optimized phase. After that, the optimized phase is three-dimensionally unwrapped. Finally, the ascending orbit unwrapped phase map and the descending orbit unwrapped phase map are obtained respectively.

[0025] Furthermore, the optimization quality of the optimized phase is evaluated by the following formula:

[0026]

[0027] In the formula, represents the temporal coherence of the DS preliminary candidate points or the PS preliminary candidate points; N represents the number of ascending SAR images or the number of descending SAR images; represents the phase difference between the optimized phase and the original phase; represents the original interferometric phase of the interferograms acquired by the nth and mth SAR; and They represent the phases after phase optimization of the interferograms acquired by the nth and mth SAR respectively; An imaginary unit representing a complex number;

[0028] The temporal coherence of the DS preliminary candidate points and the PS preliminary candidate points after phase optimization is calculated, and the pixels whose temporal coherence of the DS preliminary candidate points is higher than the threshold are determined as the DS final candidate points, and the pixels whose temporal coherence of the PS preliminary candidate points is higher than the threshold are determined as the PS final candidate points.

[0029] Further, step S22 specifically includes the following steps:

[0030] S221: geocoding the ascending orbit unwrapped phase image and the descending orbit unwrapped phase image, and selecting high coherence pixels to be resampled to the common area of ​​the ascending orbit unwrapped phase image and the descending orbit unwrapped phase image;

[0031] S222: unify the time base and create a time matrix about the time intervals between consecutive SAR acquisitions;

[0032] S223: Use the singular value decomposition method to solve each pixel in the common area to obtain the two-dimensional vertical deformation rate V of each pixel U ;

[0033] S224: Two-dimensional vertical deformation rate V U By performing time domain integration, the vertical deformation time series of the goaf is obtained.

[0034] Furthermore, the time series data set of the goaf is preprocessed to obtain a sample sequence data set including three variables: surface temperature, precipitation, and deformation. The specific process of dividing the sample sequence data set into a training set and a test set is as follows:

[0035] The multi-dimensional small baseline InSAR method is used to fill the monitoring gaps in the deformation time series in the time dimension;

[0036] The data of all hours of each day in the 2m temperature hourly data are averaged to obtain the 2m temperature daily data set, and the data of all hours of each day in the total precipitation hourly data are averaged to obtain the total precipitation daily data set;

[0037] The 2m temperature daily dataset, the total precipitation daily dataset, and the vertical deformation time series of a single pixel are combined and divided into training and test sets in a ratio of 9:1.

[0038] Furthermore, the feature extraction module includes a Conv1D layer, a BatchNorm layer, an ELU layer, and a Softmax layer. The Conv1D layer is used to perform a one-dimensional convolution operation in the time dimension to extract features from the time series information carried by each historical moment in the training set; the BatchNorm layer performs a batch normalization operation on the output result of the Conv1D layer; the ELU layer is used to perform an ELU activation function operation on the output of the BatchNorm layer; and the Softmax layer is used to perform a Softmax activation function operation on the output result of the ELU layer to obtain the attention weight corresponding to the time series information carried by each historical moment in the training set;

[0039] The encoder module mainly includes a position embedding layer, a multi-head attention layer, two residual connections and normalization layers, and a feedforward neural network layer. The encoder module is used to encode and embed the multi-dimensional features output by the feature extraction module, embed them through the position embedding layer, and use the multi-head attention mechanism to capture the correlation between features, and combine the feedforward neural network to further enhance the feature expression ability. First, the position embedding layer embeds the position information in the input multi-dimensional feature sequence into each feature vector. Then the output of the position embedding layer is added to the multi-dimensional features of the input to generate the position enhanced features; the multi-head attention layer is used to divide the enhanced features after position embedding into multiple heads, and perform self-attention calculation on each head separately, and calculate the weighted average by generating Query, Key and Value vectors. The results of each head are concatenated and fused through linear transformation to capture the multi-level dependencies in the feature sequence; the first residual connection and normalization layer is used to perform residual connection and layer normalization, and the output of the multi-head attention layer is added to the enhanced features after position embedding through residual connection to form new features, and then layer normalization is performed on the new features; the feedforward neural network layer is used to perform nonlinear transformation and mapping on the features of each position, so as to capture more complex deep patterns and relationships; the second residual connection and normalization layer is used to perform residual connection and layer normalization again, and the output of the feedforward neural network layer is added to the input of the feedforward neural network layer to form new features, and then layer normalization is performed on the new features;

[0040] The decoder module includes two Bi-LSTM networks. A single Bi-LSTM network consists of a forward LSTM layer and a reverse LSTM layer. The calculation process of the Bi-LSTM network is as follows:

[0041] ;

[0042] In the formula, Represents the output of the forward LSTM layer; Represents the output of the reverse LSTM layer; h t express and The concatenation result is the output of the Bi-LSTM network;

[0043] The calculation process of the forward LSTM layer and the reverse LSTM layer is as follows:

[0044] ;

[0045] In the formula, , and Respectively The input gate, forget gate and output gate at each moment; express The input feature matrix at time t; express The hidden state of the moment; h t-1 express The hidden state of the moment; express Candidate cell states at time instant; , , , Respectively The corresponding weight matrix; , , , Respectively The corresponding weight matrix; , , , All represent bias matrices; Indicates the current cell status; represents the sigmoid activation function, and tanh represents the hyperbolic tangent activation function; Represents the Hadamard product operator symbol;

[0046] The output module includes a flattening layer and a linear layer. The flattening layer is used to flatten the time series features output by the decoder module into a one-dimensional vector, and the linear layer is used to output the predicted time series deformation map of the future time step.

[0047] Furthermore, the interference graph generated in step S211 is subjected to multi-view processing with a range direction of 10 and an azimuth direction of 2, and the interference graph generated in step S211 is subjected to Goldstein filtering processing.

[0048] Compared with the prior art, the present invention can achieve the following technical effects:

[0049] (1) The deformation prediction model takes into account external influencing factors (temperature and precipitation), and expands the single deformation variable into three variables: temperature, precipitation, and deformation. The deformation prediction model with multi-variable input can achieve high-precision deformation prediction under complex deformation modes (where both mining areas and closed mining areas exist at the same time);

[0050] (2) The deformation prediction model uses an innovative feature extraction module that takes into account convolutional attention proposed in this paper. By combining convolution operations with the attention mechanism, it effectively captures the multi-scale spatial features and local correlations in the deformation time series. This module enhances the focus on key time steps and important features, significantly improves the model's feature expression capabilities, and provides a more accurate and rich foundation for subsequent predictions;

[0051] (3) The deformation prediction model uses the encoder module of the Transformer encoder structure combined with the decoder module of the Bi-LSTM structure to form an encoder-decoder architecture, which can better capture the long-term dependencies in the deformation time series. The powerful sequence correlation feature extraction capability of the Transformer encoder and the excellent long time series processing capability of the Bi-LSTM decoder complement each other to improve the accuracy and robustness of the deformation prediction model.

[0052] (4) The deformation prediction model provides early warning decision-making information for geological disaster prevention and safe production in coal mining areas. It also provides important scientific support for ecological protection and restoration in coal mining areas and even for the sustainable development of coal resource-based cities. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 A schematic diagram of the overall process of the method for predicting surface deformation of goaf areas by integrating convolutional attention and multiple influencing factors as described in an embodiment of the present invention;

[0054] Figure 2 A schematic diagram for verifying the vertical deformation results obtained by the MSBAS-InSAR method described in the embodiment of the present invention;

[0055] Figure 3 A schematic diagram of the overall architecture of the deformation prediction model described in the embodiment of the present invention;

[0056] Figure 4a It is a schematic diagram of the observed deformation of the last phase described in the embodiment of the present invention;

[0057] Figure 4b A schematic diagram of predicted deformation of the deformation prediction model described in the embodiment of the present invention;

[0058] Figure 4c A schematic diagram of the residual between the observed deformation and the predicted deformation according to the embodiment of the present invention;

[0059] Figure 4d The correlation and scatter density diagram between the observed deformation and the predicted deformation described in the embodiment of the present invention;

[0060] Figure 4e A histogram of deformation differences between the observed deformation and the predicted deformation described in the embodiment of the present invention;

[0061] Figure 5a-5d It is a schematic diagram comparing the observed deformation curve of the maximum deformation point in the four main deformation areas within 6 months described in the embodiment of the present invention and the predicted deformation curve based on different prediction models. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solution and advantages of the invention more clear, the invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described here are only used to explain the invention and do not constitute a limitation of the invention.

[0063] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0064] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments, and the Hunchun coal mining area (including multiple goaf areas in the mining area) is selected as an implementation case of surface deformation prediction.

[0065] like Figure 1 As shown, the present invention provides a method for predicting surface deformation of goaf areas by integrating convolutional attention and multiple influencing factors, including four steps: data preparation, deformation time series modeling, deformation prediction model construction, and deformation prediction. The four steps are described in detail below.

[0066] 1. Data Preparation

[0067] The data sets that need to be prepared include SAR image data of the goaf location, auxiliary data and external influencing factor data. The SAR image data include ascending SAR images and descending SAR images. The auxiliary data include digital elevation model data and precise orbit data. The external influencing factor data include surface temperature data and precipitation data.

[0068] First, a vector range covering the goaf needs to be preliminarily defined, which is used to determine the range of the SAR image data and external influencing factor data that will be downloaded later.

[0069] SAR image data: According to the vector range of the goaf, the ascending SAR images (Sentinel-1A SAR images) and descending SAR images (Sentinel-1B SAR images) are downloaded from the ASF Data Search website. Since the time series surface deformation results need to be obtained, multiple periods of data must be downloaded for both the ascending SAR images and the descending SAR images. The present invention downloaded 131 ascending Sentinel-1A SAR images from 2017 / 7 / 8 to 2021 / 12 / 14 and 134 descending Sentinel-1B SAR images from 2017 / 7 / 7 to 2021 / 12 / 13. The specific parameters of these two SAR images are: ① File type: single look complex (SLC) image; ② Sensor mode: interferometric wideband (IW) mode; ③ Polarization mode: vertical polarization (VV); ④ Image orbit number: 156 for ascending and 61 for descending.

[0070] Auxiliary data: The digital elevation model data is the ALOS WORLD 3D-30m DEM data, which can be manually downloaded from this website (https: / / www.eorc.jaxa.jp / ALOS / en / aw3d30 / data / index.htm); the precise orbit data is the POD precise orbit determination ephemeris data released by the European Space Agency (ESA) 21 days after the SAR data are distributed, which can be manually downloaded from the ASF website.

[0071] External influencing factor data: The surface temperature data and precipitation data are the hourly data of “2m temperature (air temperature 2 meters above the land)” and “total precipitation” in the ERA5 hourly dataset, respectively. The hourly data of “2m temperature” and “total precipitation” in the corresponding vector range in the ERA5 hourly dataset were downloaded from the Climate Data Store (CDS) website.

[0072] 2. Deformation Time Series Modeling

[0073] Based on the data prepared above (ascending and descending SAR images, ALOS WORD 3D-30m DEM data and POD precise orbit determination ephemeris data), firstly, the DS-InSAR method (i.e. distributed scatterer InSAR method) is used to obtain the unwrapped phase map of the ascending SAR image (referred to as ascending unwrapped phase map) and the unwrapped phase map of the descending SAR image (referred to as descending unwrapped phase map) of the goaf; then, the MSBAS-InSAR method (i.e. multidimensional small baseline InSAR method) is used to solve the vertical deformation time series covering all 136 observation periods; finally, the leveling data of the goaf is used to verify the vertical deformation results to prove its reliability, which is the basis for deformation time series prediction.

[0074] For DS-InSAR method

[0075] The DS-InSAR method is based on SBAS-InSAR processing and is implemented using the ISCE+StaMPS platform.

[0076] The steps of processing the ascending SAR image and the descending SAR image to obtain the ascending unwrapped phase map and the descending unwrapped phase map are as follows:

[0077] (1) SBAS-InSAR processing: The present invention determines a suitable spatiotemporal baseline threshold (63 days for time baseline and 200 m for space baseline) through a trial-and-error strategy to construct an interferometric pair set. A super master image is selected, and a master-slave relationship is established for all images according to the determined spatiotemporal baseline threshold. All slave images are registered and resampled to the super master image, and then all interferometric pairs are interferometrically processed to generate multiple interferograms. Considering the relationship between effective signals and noise, the generated interferograms are subjected to multi-view processing with a range of 10 and an azimuth of 2 to improve the signal-to-noise ratio of the interferogram. The ALOS WORLD 3D-30mDEM data is used to remove the terrain phase in the interferogram. Considering the presence of high noise phases in the interferogram, the interferogram is filtered by using Goldstein filtering to improve the visibility of the fringes.

[0078] (2) Identification of statistically homogeneous pixels SHPs: In the present invention, based on the amplitude information of the ascending SAR image and the amplitude information of the descending SAR image, the fast statistical homogeneous pixel selection algorithm (FaSHPS) is used to identify homogeneous pixels SHPs. By identifying the number of homogeneous pixel SHPs within a rectangular window of a certain size (15×15) around each pixel, the pixels with more than 20 surrounding homogeneous pixels are selected as DS preliminary candidate points, and the pixels with less than 20 surrounding homogeneous pixels are selected as PS preliminary candidate points.

[0079] Based on the amplitude deviation index method, the PS preliminary candidate points are further screened by setting the amplitude deviation index threshold (the thresholds for both ascending and descending orbits are 0.6) among the PS preliminary candidate points.

[0080] (3) Optimizing the phase. The present invention uses a coherent matrix eigenvalue decomposition method to optimize the phase. A coherent matrix is ​​established on each DS preliminary candidate point and PS preliminary candidate point, and the phase of the eigenvector corresponding to the maximum eigenvalue selected by the coherent matrix eigenvalue decomposition method is used as the optimized phase to replace the original phase of each DS preliminary candidate point and PS preliminary candidate point.

[0081] (4) Calculate the temporal coherence. The quality of phase optimization is evaluated by the following formula (1). The temporal coherence of the DS preliminary candidate points and the PS preliminary candidate points after phase optimization are calculated by formula (1). The pixels whose temporal coherence of the DS preliminary candidate points is higher than the threshold (the ascending and descending thresholds are 0.65 and 0.7, respectively) are determined as the DS final candidate points. At the same time, the pixels whose temporal coherence of the PS preliminary candidate points is higher than the threshold (the ascending and descending thresholds are 0.65 and 0.7, respectively) are determined as the PS final candidate points.

[0082] Temporal coherence of DS preliminary candidate points or PS preliminary candidate points γ T The calculation formula is:

[0083] (1);

[0084] Where N represents the number of ascending or descending SAR images; represents the phase difference between the optimized phase and the original phase; represents the original interferometric phase of the interferograms acquired by the nth and mth SAR; and They represent the phases after phase optimization of the interferograms acquired by the nth and mth SAR respectively; An imaginary unit representing a complex number;

[0085] (5) Generate InSAR unwrapped phase map. First, use the GACOS tool to perform atmospheric phase delay correction on the optimized phase to reduce atmospheric noise; then, use the best fit surface to remove the residual linear trend in the optimized phase; finally, perform three-dimensional phase unwrapping on the optimized phase to obtain the ascending unwrapped phase map and the descending unwrapped phase map.

[0086] For the MSBAS-InSAR method

[0087] The present invention uses the MSBAS-InSAR method to obtain the vertical deformation time series of the goaf. The MSBAS-InSAR method includes the following steps:

[0088] (1) Geocode the ascending and descending unwrapped phase images, and select highly coherent pixels to be resampled to the common area of ​​the ascending and descending unwrapped phase images.

[0089] (2) Unify the time base and create a time matrix about the time intervals between consecutive SAR acquisitions.

[0090] (3) Apply the singular value decomposition (SVD) method to solve each pixel in the common area and obtain the two-dimensional vertical deformation rate V of each pixel. U .

[0091] (4) By calculating the two-dimensional vertical deformation rate V U By integrating in the time domain, the vertical deformation time series of the goaf is obtained.

[0092] The main principle formula of the MSBAS-InSAR method is:

[0093] For the ascending SAR image set and the descending SAR image set, the two-dimensional deformation time series can be reconstructed by the following equations (2) and (3):

[0094] (2);

[0095] (3);

[0096] In the formula, ; represents the matrix composed of the time intervals of continuous SAR acquisition; α and θ represent the azimuth and incident angle of the SAR satellite respectively; V E 、V U They represent the two-dimensional east-west and vertical deformation rates respectively; λ represents the regularization parameter; I represents the unit matrix; represents the observed interferometric phase; Represents the vertical deformation in the i-th period.

[0097] Verification of vertical deformation results

[0098] The reliability of the vertical deformation results is the basis for subsequent deformation time series prediction. Therefore, in order to verify the reliability of the vertical deformation obtained by the MSBAS-InSAR method in the present invention, the leveling results of the previous monitoring in the goaf were used for verification. The P2, P3, P4, P8, P9, P15, and P16 leveling points that coincide with the effective deformation positions in the vertical deformation time series diagram were used, as shown in Figure 2. Figure 2 As shown, within almost the same time range, the cumulative vertical deformation obtained by the MSBAS-InSAR method is highly consistent with the deformation measured by leveling, which proves the reliability of the vertical deformation obtained by the MSBAS-InSAR method of the present invention.

[0099] 3. Construction of deformation prediction model

[0100] In order to carry out deformation time series prediction, the acquired time series data (2m temperature, total precipitation and vertical deformation data) must first be preprocessed and the data set must be prepared. The purpose of data preprocessing is to collect and preprocess the original time series data in advance (for example, data interpolation) so that the data set can be prepared. The purpose of data set preparation is to divide the preprocessed time series data into training sets and test sets so as to train and verify the subsequent deformation prediction model.

[0101] Data preprocessing: Due to SAR acquisition gaps, there are often monitoring gaps in the InSAR deformation time series, which is not conducive to deep learning prediction. Therefore, the existing technology usually uses different interpolation methods to fill in, including linear interpolation, cubic spline interpolation, etc., but these interpolation methods usually introduce new errors. In the present invention, in order to use the vertical deformation time series for accurate terrain deformation prediction, the MSBAS-InSAR method is used to fill the monitoring gaps in the time dimension. This method is an InSAR time series solution method, not an interpolation method. Therefore, it can better retain the nonlinear time series characteristics of the deformation of the monitoring gap. Finally, based on this method, a vertical deformation time series covering all 136 observation periods is obtained.

[0102] In addition, the external influencing factor data collected above are the hourly data of "2m temperature" and "total precipitation" of the ERA5 hourly data set in the same period as the 136 observation periods of the vertical deformation time series. The hourly data of "2m temperature" and "total precipitation" are averaged over all hours (00:00-23:00) of each period to obtain the daily data set of "2m temperature" and "total precipitation" corresponding to the date of the vertical deformation time series. It is worth noting that due to the low resolution of the ERA5 data (0.25°×0.25°), one pixel range just covers the range of this goaf, so it is regarded as a time series input of one pixel point.

[0103] Dataset preparation: In the present invention, the "2m temperature", "total precipitation" and the vertical deformation time series of a single pixel are combined as the training set and test set of the deformation time series prediction task in the present invention, so that the single deformation variable input is expanded to three variable inputs of temperature, precipitation and deformation. The deformation prediction model with multi-variable input can achieve high-precision deformation prediction under complex deformation modes (coexistence of mining areas and closed mining areas).

[0104] During the training process, each valid deformed pixel point in the vertical deformation map is input and time series prediction is performed separately. For the time series data set of a single deformed pixel point input, in order to ensure a sufficient amount of training data, the time series data set is segmented using a sliding window and an overlapping segmentation method. The present invention adopts a single-step prediction method and uses multiple time steps to predict one time step. Based on a trial-and-error strategy, considering the final training accuracy, all odd numbers between 3 and 17 are tried as input time steps. When the step size is 7, the accuracy on the test set is the highest, so 7 time steps are finally used to predict one time step. That is, a single training sample sequence is (x t ,y t ,z t ), (x t+1 ,yt+1 ,z t+1 ),…,(x t+6 ,y t+6 ,z t+6 ), labelled x t+7 (1≤t≤129). Finally, 129 sample sequences were constructed according to the time step for all 136 periods, and divided into training set and test set according to 90% and 10% respectively. Therefore, for each pixel point, 116 training sample sequences and 13 test sample sequences can be obtained.

[0105] Based on the training set and the test set, a deformation prediction model integrating the convolutional attention mechanism and multi-source variable input is constructed and trained. The overall architecture of the deformation prediction model is as follows: Figure 3 The deformation prediction model is named "FeTLnet" and consists of an input module, a feature extraction module, an encoder module based on the Transformer encoder structure, a decoding module based on the Bi-LSTM structure, and an output module.

[0106] The input module is used to input the sample sequence data set including three variables: surface temperature, precipitation and deformation into the model as the input of the feature extraction module; the feature extraction module is used to extract the features in the time dimension of the input sample sequence; the encoder module is used to perform embedded encoding on the extracted features at different time positions and capture the correlation between the features; the decoder module is used to perform asymmetric decoding and prediction on the output features of the encoder; the output module is used to output the predicted deformation time series diagram.

[0107] Feature extraction module: It is used to assign attention weights to the time series information carried by each historical moment of the input sample sequence to select and emphasize the important features in the input time series and realize the multi-feature extraction function. The feature extraction module includes three information flows, among which the attention calculation part includes a one-dimensional convolution layer (Conv1D), a batch normalization layer (BatchNorm), an ELU layer, and a Softmax layer. The output of the attention calculation part is Hadamard producted with the output of a one-dimensional convolution layer to calculate the scaling, and then the scaled result is added to the residual connection and output. The one-dimensional convolution layer uses a convolution kernel of size 3, the number of output channels is 32, and the stride is 1. The one-dimensional convolution layer is used for data embedding and feature extraction. By performing 1D convolution on the time dimension, different features are extracted from the input time series information to explore the impact of the previous and next historical moments on the current moment. The BatchNorm layer batch normalizes the output of the convolution layer to stabilize the training process. Then a nonlinear activation function is introduced. After experimental attempts, the ELU activation function of the ELU layer will obtain better accuracy and stability. Next, the Softmax activation function of the Softmax layer is used to convert the output into a probability distribution, and the corresponding time attention weights of each historical moment are obtained to enhance the information expression of key moments and better capture the complex time dependencies in the time series. The specific calculation process of the above structure is as follows:

[0108] (4);

[0109] (5);

[0110] (6);

[0111] (7);

[0112] In the formula, Represents convolution operation; Softmax() represents Softmax activation function operation; ELU() represents ELU activation function operation; BN() represents batch normalization calculation; represents the calculation of Hadamard product; and Represents the weight matrix of the convolution calculation. represents transpose; X Represents the input data of the model; and Represents the bias term of convolution calculation; and They represent the weight matrix and bias term of the convolution calculation in the convolutional attention mechanism respectively; Represents convolutional attention calculation The result after convolution; H2 and H3 represent the intermediate representation and output representation of the model, where H2 is the intermediate feature after convolution attention calculation and other operations, and H3 is the final output feature representation.

[0113] Encoder module: The encoder adopts a Transformer structure encoder, and its structure mainly includes a position embedding layer, a multi-head attention layer, two residual connections and a normalization layer (Add&Norm) and a feedforward neural network layer; among them, the number of attention heads in the multi-head attention layer is set to 4, and the number of channels of Key and Query is set to 128. Specifically, the encoder module is used to encode and embed the multi-dimensional features output by the feature extraction module, embed them through the position embedding layer, and use the multi-head attention mechanism to capture the correlation between features, and combine the feedforward neural network to further enhance the feature expression ability. The position embedding layer embeds the position information in the input multi-dimensional feature sequence into each feature vector. Then, the output of the position embedding layer is added to the multi-dimensional features of the input to generate the position-enhanced features (which is also the first layer output of the encoder module); the multi-head self-attention mechanism of the multi-head attention layer allows the deformation prediction model to pay attention to different contextual information at different positions in the sequence. Therefore, the multi-head attention layer divides the enhanced features after the input position embedding into multiple heads, and performs self-attention calculation on each head separately. By generating Query, Key and Value vectors and calculating the weighted average, the results of each head are concatenated and fused through linear transformation to capture the multi-level dependencies in the feature sequence; the first residual connection performs a residual connection with the normalization layer (Residual Connection Connection) and layer normalization (LayerNormalization) processing, the output of the multi-head attention layer is added to the input entering the layer (that is, the enhanced features after position embedding) through residual connection to form new features, and then layer normalization is performed on the new features to stabilize the numerical distribution of the features and improve the stability and speed during training; the feedforward neural network layer is used to receive the output of the first residual connection and normalization layer, and is used to perform nonlinear transformation and mapping on the features of each position to capture more complex deep patterns and relationships; the second residual connection and normalization layer is used to perform residual connection and layer normalization again, adding the output of the feedforward neural network layer to the input entering the feedforward neural network layer (residual connection) to form new features, and then layer normalization is performed on the new features. The output of the second residual connection and normalization layer will be passed to the decoder module.

[0114] The Transformer encoder can process multiple input variables at the same time, so it can encode multiple variables and capture the correlation between them. In the encoding stage, the multi-head attention mechanism based on the Transformer is used to capture the temporal dependency of the input feature sequence, which enhances the ability of the deformation prediction model to simulate time series data. The calculation of the multi-head attention mechanism can be described as:

[0115] (8);

[0116] (9);

[0117] (10);

[0118] In the formula, Q, K and V represent Query, Key and Value vectors respectively. K T represents the transpose of the key vector; d k represents the dimension of the key vector, represents the scaling factor, which helps stabilize the gradient; , , and represents the learnable parameter matrix, head i represents the current attention head, and n represents the number of parallel self-attention mechanisms.

[0119] The multi-head attention mechanism improves the attention ability of the deformation prediction model by linearly mapping the Q, K, and V vectors to different representation spaces, calculating the self-attention in parallel, and then concatenating and mapping the outputs again.

[0120] Decoder module: The decoder reconstructs and outputs multiple features by asymmetric decoding of the input features from the encoder. The present invention uses a two-layer Bi-LSTM structure to form a decoder module, instead of a Transformer decoder structure. This simplifies the parameters while also comprehensively mining the bidirectional time dependencies in the time series based on the bidirectional time feature extraction capability of the Bi-LSTM structure, thereby capturing and predicting nonlinear and complex time series patterns. The first Bi-LSTM structure uses 16 hidden units, and the second Bi-LSTM structure uses 32 hidden units. A single Bi-LSTM network consists of a forward LSTM layer and a reverse LSTM layer. Compared with the unidirectional state transfer method in the LSTM layer, the Bi-LSTM structure generally has superior performance than the LSTM structure through bidirectional time series feature extraction. Usually, the calculation process of a standard LSTM layer (the same applies to the forward LSTM layer and the reverse LSTM layer) is as follows:

[0121] (11);

[0122] In the formula, , and Respectively The input gate, forget gate and output gate at each moment; express The input feature matrix at time t; express The hidden state of the moment; h t-1 express The hidden state of the moment; express Candidate cell states at time instant; , , , Respectively The corresponding weight matrix; , , , Respectively The corresponding weight matrix; , , , All represent bias matrices; Indicates the current cell status; represents the sigmoid activation function, and tanh represents the hyperbolic tangent activation function; Represents the Hadamard product operator.

[0123] The calculation formula of the Bi-LSTM structure can be expressed as follows:

[0124] (12);

[0125] In the formula, Represents the output of the forward LSTM layer; Represents the output of the reverse LSTM layer; h t express and The concatenation result is the output of the Bi-LSTM network.

[0126] Output module: The output module includes a flattening layer (Flatten) and a linear layer (Dense). The flattening layer is used to flatten the temporal features output by the decoder module into a one-dimensional vector, and the linear layer is used to output the predicted deformation time series of a single pixel in the future time step. Finally, each valid pixel in the deformation time series graph is trained and predicted one by one to obtain a time series prediction deformation graph.

[0127] In summary, the present invention can better capture the long-term dependencies in the sequence by combining the Transformer structure with the Bi-LSTM structure, and utilize the powerful sequence association feature extraction capability of the Transformer and the excellent long time series processing capability of the Bi-LSTM to complement each other and improve the accuracy and robustness of the deformation prediction model.

[0128] Model training and loss evaluation: The FeTLnet model proposed in the present invention is based on pixel-by-pixel processing, and the computing resources and storage resources are relatively low. The experimental platform parameters of the present invention are shown in Table 1. The Pytorch framework is used for FeTLnet model training and reasoning. In the specific model parameter setting, the batch size batch_size during model training is 32, and the number of training rounds epochs is 150. The loss function during the training process is set to the root mean square error (RMSE). The weights of the neural network are iteratively updated by the Adam optimizer. The learning rate of the initial training is 1e-3. In addition, an adaptive learning rate strategy is adopted. When the loss of the test set has not decreased for 10 consecutive epochs, the learning rate drops to 0.5 of the current value. During reasoning, the trained FeTLnet model is used for prediction, and the rolling prediction mode is used for the future deformation prediction process. Finally, the output is the predicted time series deformation (x8, x9, ..., x136) of each deformed pixel point, and the time series prediction deformation map containing all deformed pixels is further obtained.

[0129]

[0130] In addition, in order to quantitatively evaluate the prediction performance of the model, a test set containing 13 test sample sequences for each pixel was input, and the following indicators were used, including root mean square error (RMSE), mean absolute error (MAE), determination coefficient (R 2 ) and symmetric mean absolute percentage error (SMAPE) to evaluate the model's prediction results on the test set.

[0131] (13);

[0132] (14);

[0133] (15);

[0134] (16);

[0135] In the formula, N represents the number of samples in the test set; and They represent the InSAR observed deformation and the deformation predicted by the FeTLnet model at period t, respectively.

[0136] 4. Deformation Prediction

[0137] In order to evaluate the prediction accuracy of the FeTLnet model proposed in this invention, support vector machine regression (SVR), LSTM, and Bi-LSTM were selected as benchmark methods to compare with the FeTLnet model proposed in this invention, and four evaluation indicators were used (see Table 2). Considering the impact of single variable input and multi-variable input on model performance, the same model added single variable and multi-variable comparisons and used the same model parameters. For the comparison between Bi-LSTM and LSTM, the Bi-LSTM model with multi-variable input was adopted, and the same parameters as LSTM were used. The setting parameters of the SVR and LSTM models are shown in Table 3.

[0138] Table 2 shows the prediction accuracy of each model. Overall, the FeTLnet model proposed in the present invention is the best in all evaluation indicators, which shows that the FeTLnet model proposed in the present invention has the smallest error and better prediction performance. In addition, in all models, the univariate prediction evaluation index is worse than the multivariate prediction, which shows that more variable inputs are conducive to the model to better capture the nonlinear relationship in long time series data.

[0139]

[0140] The present invention converts the predicted deformation of the FeTLnet model (such as Figure 4b ) and the observed vertical deformation of the last period (2021 / 12 / 13) (as shown in Figure 4a First, there are no large outliers or significant abnormal areas in the residual graph between the two (such as Figure 4c The difference between the two is mainly between -0.02m and 0.02m (as shown in Figure 4e The mean and standard deviation of the difference between the two are -0.0014m and 0.006m, respectively, which are both very small. Secondly, the Pearson correlation was used to evaluate the correlation between the observed deformation and the predicted deformation, and the r value was 0.9992 (as shown in Figure 4dAs shown), this indicates that there is a very high correlation between the two. Therefore, the deformation predicted by the FeTLnet model proposed in the present invention has good prediction accuracy. In addition, the present invention selected the points with the highest cumulative deformation in the C1 area of ​​Chengxi Coal Mine, the L1 area of ​​Baliancheng Coal Mine, the S1 area of ​​Banshi Coal Mine, and the Y1 area of ​​Ying'an Coal Mine, respectively, and extracted the time series of observed deformation and predicted deformation corresponding to these points in the 6-month period from 2021 / 7 / 1 to 2021 / 12 / 13, and used several benchmark methods in Table 2 for comparison with the FeTLnet model (such as Figure 5a-5d The results show that the predicted values ​​of the FeTLnet model proposed in the present invention are in good agreement with the observed deformations, and the fluctuation is small compared with other models. In general, the above analysis verifies that the FeTLnet model proposed in the present invention has very good prediction performance.

[0141] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the disclosure of the present invention can be performed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document does not limit this.

[0142] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting surface deformation of goaf areas by integrating convolutional attention and multiple influencing factors, characterized in that: The steps include: S1: Acquire relevant data covering the goaf area; wherein the relevant data include SAR image data, auxiliary data and external influencing factor data, SAR image data include ascending SAR image and descending SAR image, auxiliary data include digital elevation model data and precise orbit data, external influencing factor data include surface temperature data and precipitation data; S2: Based on the acquired relevant data covering the goaf area, a time series dataset based on temperature, precipitation and InSAR deformation of the goaf area is constructed; S3: Preprocess the time series data set of the goaf to obtain a sample sequence data set including three variables: surface temperature, precipitation and deformation, and divide the sample sequence data set into a training set and a test set. According to the training set and the test set, a deformation prediction model based on the convolutional attention mechanism and multi-variable input is constructed and trained; wherein the deformation prediction model includes an input module, a feature extraction module, an encoder module based on the Transformer encoder structure, a decoder module based on the Bi-LSTM structure and an output module; the input module is used to input the sample sequence data set into the deformation prediction model as the input of the feature extraction module; the feature extraction module is used to extract features of the input sample sequence data set in the time dimension; the encoder module is used to perform embedded encoding on the extracted features at different time series positions and capture the correlation between the features; the decoder module is used to perform asymmetric decoding and prediction on the features output by the encoder module; the output module is used to output the predicted deformation time series graph; The feature extraction module includes Conv1D layer, BatchNorm layer, ELU layer and Softmax layer. The Conv1D layer is used to perform one-dimensional convolution operation in the time dimension to extract features from the time series information carried by each historical moment in the training set; the BatchNorm layer performs batch normalization operation on the output result of the Conv1D layer; the ELU layer is used to perform ELU activation function operation on the output of the BatchNorm layer; the Softmax layer is used to perform Softmax activation function operation on the output result of the ELU layer to obtain the attention weight corresponding to the time series information carried by each historical moment in the training set; The encoder module includes a position embedding layer, a multi-head attention layer, two residual connections and normalization layers, and a feedforward neural network layer. The position embedding layer is used to embed the position information in the multi-dimensional features output by the feature extraction module into each feature vector. Then, the output of the position embedding layer is added to the input multi-dimensional features to generate position-enhanced features; the multi-head attention layer is used to divide the position-enhanced features into multiple heads and perform self-attention calculations on each head. By generating Query, Key, and Value vectors, the weighted average is calculated, and the results of each head are spliced ​​and fused through linear transformation to capture the features. Multi-level dependencies in the sequence; the first residual connection and normalization layer is used to perform residual connection and layer normalization, and the output of the multi-head attention layer is added to the enhanced features after position embedding through residual connection to form new features, and then layer normalization is performed on the new features; the feedforward neural network layer is used to perform nonlinear transformation and mapping on the features of each position to capture more complex deep patterns and relationships; the second residual connection and normalization layer is used to perform residual connection and layer normalization again, and the output of the feedforward neural network layer is added to the input of the feedforward neural network layer to form new features, and then layer normalization is performed on the new features; The decoder module includes two Bi-LSTM networks. Both Bi-LSTM networks consist of a forward LSTM layer and a reverse LSTM layer. The calculation process of the Bi-LSTM network is as follows: ； In the formula, Represents the output of the forward LSTM layer; Represents the output of the reverse LSTM layer; h t express and The concatenation result is the output of the Bi-LSTM network; The calculation process of the forward LSTM layer and the reverse LSTM layer is as follows: ； In the formula, , and Respectively The input gate, forget gate and output gate at each moment; express The input feature matrix at time t; express The hidden state of the moment; h t-1 express The hidden state of the moment; express Candidate cell states at time instant; , , , Respectively The corresponding weight matrix; , , , Respectively The corresponding weight matrix; , , , All represent bias matrices; Indicates the current cell status; represents the sigmoid activation function, and tanh represents the hyperbolic tangent activation function; Represents the Hadamard product operator symbol; The output module includes a flattening layer and a linear layer. The flattening layer is used to flatten the time series features output by the decoder module into a one-dimensional vector, and the linear layer is used to output the predicted time series deformation map of the future time step; S4: Predict the future deformation of the goaf based on the deformation prediction model.

2. The method for predicting surface deformation of goaf areas by integrating convolutional attention and multiple influencing factors according to claim 1 is characterized in that: Step S1 specifically includes the following steps: S11: vector range for delineating the location of the goaf; S12: Acquire ascending SAR images, descending SAR images, digital elevation model data, precise orbit data, surface temperature data, and precipitation data according to the vector range of the goaf location.

3. The method for predicting surface deformation of goaf areas by integrating convolutional attention and multiple influencing factors according to claim 1 is characterized in that: The digital elevation model data is the ALOS WORLD 3D-30m DEM data, the precise orbit data is the POD precise orbit determination ephemeris data, and the surface temperature data and precipitation data are the 2m temperature hourly data and total precipitation hourly data in the ERA5 hourly dataset, respectively.

4. The method for predicting surface deformation of goaf areas by integrating convolutional attention and multiple influencing factors according to claim 3 is characterized in that: Step S2 specifically includes the following steps: S21: The ascending SAR image and the descending SAR image are processed by the distributed scatterer InSAR method to obtain the ascending unwrapped phase map and the descending unwrapped phase map of the goaf area; S22: The ascending and descending unwrapped phase images are processed using the multidimensional small baseline InSAR method to obtain the vertical deformation time series of the goaf.

5. The method for predicting surface deformation of goaf areas by integrating convolutional attention and multiple influencing factors according to claim 4 is characterized in that: Step S21 specifically includes the following steps: S211: Select a super master image, establish a master-slave relationship for all images according to a set spatiotemporal baseline threshold, align and resample all slave images to the super master image, and then perform interference processing to generate multiple interference images; S212: Removal of terrain phase in interferograms using ALOS WORLD 3D-30m DEM data; S213: Based on the amplitude information of the ascending SAR image and the amplitude information of the descending SAR image, a fast statistical homogeneous pixel selection method is used to identify the number of homogeneous pixels around each pixel, and pixels with a surrounding homogeneous pixel number higher than a threshold are used as DS preliminary candidate points, and pixels with a surrounding homogeneous pixel number lower than the threshold are used as PS preliminary candidate points; S214: establishing a coherence matrix on each DS preliminary candidate point and PS preliminary candidate point, and using a coherence matrix eigenvalue decomposition method to select the phase of the eigenvector corresponding to the maximum eigenvalue as the optimized phase to replace the original phase of each DS preliminary candidate point and PS preliminary candidate point; S215: First, the GACOS tool is used to correct the atmospheric phase delay of the optimized phase. Then, the best fitting surface is used to remove the residual linear trend in the optimized phase. After that, the optimized phase is three-dimensionally unwrapped. Finally, the ascending orbit unwrapped phase map and the descending orbit unwrapped phase map are obtained respectively.

6. The method for predicting surface deformation of goaf areas by integrating convolutional attention and multiple influencing factors according to claim 5 is characterized in that: The optimization quality of the optimized phase is evaluated by the following formula: In the formula, represents the temporal coherence of the DS preliminary candidate points or the PS preliminary candidate points; N represents the number of ascending SAR images or the number of descending SAR images; represents the phase difference between the optimized phase and the original phase; represents the original interferometric phase of the interferograms acquired by the nth and mth SAR; and They represent the phases after phase optimization of the interferograms acquired by the nth and mth SAR respectively; An imaginary unit representing a complex number; The temporal coherence of the DS preliminary candidate points and the PS preliminary candidate points after phase optimization is calculated, and the pixels whose temporal coherence of the DS preliminary candidate points is higher than the threshold are determined as the DS final candidate points, and the pixels whose temporal coherence of the PS preliminary candidate points is higher than the threshold are determined as the PS final candidate points.

7. The method for predicting surface deformation of goaf areas by integrating convolutional attention and multiple influencing factors according to claim 4 or 5, characterized in that: Step S22 specifically includes the following steps: S221: geocoding the ascending orbit unwrapped phase image and the descending orbit unwrapped phase image, and selecting high coherence pixels to be resampled to the common area of ​​the ascending orbit unwrapped phase image and the descending orbit unwrapped phase image; S222: unify the time base and create a time matrix about the time intervals between consecutive SAR acquisitions; S223: Use the singular value decomposition method to solve each pixel in the common area to obtain the two-dimensional vertical deformation rate V of each pixel U ; S224: Two-dimensional vertical deformation rate V U By performing time domain integration, the vertical deformation time series of the goaf is obtained.

8. The method for predicting surface deformation of goaf areas by integrating convolutional attention and multiple influencing factors according to claim 1 is characterized in that: The specific process of preprocessing the time series data set of the goaf to obtain the sample sequence data set including three variables: surface temperature, precipitation and deformation, and dividing the sample sequence data set into training set and test set is as follows: The multi-dimensional small baseline InSAR method is used to fill the monitoring gaps in the deformation time series in the time dimension; The data of all hours of each day in the 2m temperature hourly data are averaged to obtain the 2m temperature daily data set, and the data of all hours of each day in the total precipitation hourly data are averaged to obtain the total precipitation daily data set; The 2m temperature daily dataset, the total precipitation daily dataset, and the vertical deformation time series of a single pixel are combined and divided into training and test sets in a ratio of 9:

1.

9. The method for predicting surface deformation of goaf areas by integrating convolutional attention and multiple influencing factors according to claim 5 is characterized in that: The interference graph generated in step S211 is subjected to multi-view processing with a range direction of 10 and an azimuth direction of 2, and the interference graph generated in step S211 is subjected to Goldstein filtering processing.

Images