A ground surface deformation monitoring method and system based on deep learning
By combining deep learning with InSAR, GNSS, and deep underground displacement data, a set of constraint equations is constructed to solve the initial value of three-dimensional deformation. This solves the problem of high-precision surface deformation monitoring in complex terrain using InSAR technology, and realizes efficient and automated identification of geological hazard risks and assessment of surface stability.
Patent Information
- Application Number
- CN202610987412.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-03
- Publication Date
- 2026-11-10
AI Technical Summary
Existing InSAR technology has high requirements for SAR data quality and is significantly affected by factors such as spatiotemporal baseline decoherence and atmospheric delay, which limits its applicability. It is difficult to achieve real-time dynamic surface deformation monitoring under complex terrain and it is difficult to fully characterize the integrated deformation features of the surface and subsurface and their evolution.
Using deep learning methods, combined with InSAR, GNSS, inclinometer and deep underground displacement data, a set of constraint equations is constructed to solve the initial value of three-dimensional deformation. Temporal evolution features are extracted through spatiotemporal graph neural network, and high-resolution deformation field is recovered by combining deconvolution graph network, and uncertainty is quantified.
Maintaining high accuracy and robustness in complex terrain, providing reliable technical support for geological disaster early warning and surface stability assessment, fully leveraging the complementary advantages of multi-source data, and achieving efficient and automated identification of potential geological hazards.
Smart Images

Figure CN123069170A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of surface deformation monitoring technology, specifically relating to a surface deformation monitoring method and system based on deep learning. Background Technology
[0002] Currently, interferometric synthetic aperture radar (InSAR) technology has become an important means of monitoring surface deformation, mainly including three types: differential interferometric synthetic aperture radar (D-InSAR), permanent scatterer interferometric synthetic aperture radar (PS-InSAR), and small baseline set interferometric synthetic aperture radar (SBAS-InSAR). However, these methods have high requirements for SAR data quality, are significantly affected by factors such as spatiotemporal baseline decoherence and atmospheric delay, and are highly dependent on the number and distribution of stable ground features or scatterers, thus limiting their applicability. Furthermore, InSAR technology is constrained by satellite revisit cycles, making it difficult to achieve real-time dynamic monitoring of the study area. In addition, existing monitoring results mostly focus on two-dimensional or line-of-sight surface deformation, making it difficult to fully characterize the integrated surface-subsurface deformation features and their evolution. In complex terrain environments, there is still a lack of effective technical means to efficiently and automatically identify potential geological hazards based on multi-source surface deformation data. Summary of the Invention
[0003] The purpose of this invention is to provide a method and system for monitoring land surface deformation based on deep learning, so as to improve the above-mentioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:
[0004] Firstly, this application provides a deep learning-based method for monitoring land surface deformation, comprising:
[0005] Acquire SBAS-InSAR data, identify anomalous areas based on SBAS-InSAR data, and remove data from anomalous areas based on data reliability to obtain usable areas;
[0006] Collect GNSS observation data from available areas and calculate GNSS time-series deformation data and corresponding uncertainties.
[0007] Time-series data of surface dip angle in available areas are collected using inclinometers, and the uncertainty of the data is calculated;
[0008] Obtain deep underground displacement data and corresponding uncertainties for the available area;
[0009] Using dip angle data as a constraint on the direction of surface deformation, an LOS projection equation is established for each InSAR pixel containing GNSS monitoring points. GNSS celestial deformation is used as an absolute constraint to construct a set of constraint equations. Weighted least squares is used to solve the equations to obtain the initial three-dimensional deformation value of the InSAR pixel and the corresponding uncertainty.
[0010] Heterogeneous maps are constructed using GNSS, surface dip, deep underground displacement monitoring points and InSAR pixels as nodes, and feature vectors of nodes are constructed based on InSAR pixels, monitoring data of monitoring points and uncertainties.
[0011] The heterogeneous graph is input into the spatiotemporal graph neural network in a temporal sequence to obtain temporal evolution features. The temporal evolution features are then input into the deconvolution graph network to obtain the deformation rate field and uncertainty distribution.
[0012] Secondly, this application provides a deep learning-based surface deformation monitoring system, comprising:
[0013] The first module is used to acquire SBAS-InSAR data, identify anomalous areas based on SBAS-InSAR data, and remove data from anomalous areas based on data reliability to obtain usable areas.
[0014] The second module is used to collect GNSS observation data from the available area and calculate the GNSS time-series deformation data and the corresponding uncertainties.
[0015] The third module is used to collect time-series data of surface dip angle in the available area using an inclinometer and to calculate the uncertainty of the data;
[0016] The fourth module is used to obtain deep underground displacement data and corresponding uncertainties in the available area;
[0017] The fifth module is used to use dip angle data as a constraint on the direction of surface deformation. For each InSAR pixel containing GNSS monitoring points, an LOS projection equation is established, and GNSS celestial deformation is used as an absolute constraint to construct a set of constraint equations. Weighted least squares is used to solve the equations to obtain the initial value of the three-dimensional deformation of the InSAR pixel and the corresponding uncertainty.
[0018] The sixth module is used to construct a heterogeneous map using GNSS, surface dip, deep underground displacement monitoring points and InSAR pixels as nodes, and to construct the feature vectors of the nodes based on the InSAR pixels, monitoring data of the monitoring points and uncertainties.
[0019] The seventh module is used to input heterogeneous graphs into a spatiotemporal graph neural network in a temporal sequence to obtain temporal evolution features, and then input the temporal evolution features into a deconvolutional graph network to obtain the deformation rate field and uncertainty distribution.
[0020] The beneficial effects of this invention are as follows:
[0021] This invention combines InSAR, GNSS, inclinometer, and deep subsurface displacement data to construct a set of constraint equations to solve the initial values of three-dimensional deformation, thus resolving the underdetermined problem of single-image InSAR. Furthermore, heterogeneous maps are used to characterize the spatiotemporal correlation between multi-source monitoring points and pixels. Temporal evolution features are extracted through a spatiotemporal graph neural network, and a deconvolutional graph network is combined to recover the high-resolution deformation field and quantify uncertainties. This method fully leverages the complementary advantages of multi-source data, maintaining high accuracy and robustness even in complex terrain, providing reliable technical support for geological disaster early warning and surface stability assessment.
[0022] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. Attached Figure Description
[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0024] Figure 1 This is a flowchart of a deep learning-based method for monitoring land deformation, as described in an embodiment of this application.
[0025] Figure 2 This is a structural diagram of a deep learning-based surface deformation monitoring device according to an embodiment of this application.
[0026] Symbol explanation: 800 - Deep learning-based surface deformation monitoring equipment; 801 - Processor; 802 - Memory; 803 - Multimedia component; 804 - I / O interface; 805 - Communication component. Detailed Implementation
[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0028] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0029] Example 1:
[0030] See Figure 1 This embodiment provides a deep learning-based method for monitoring land deformation, including steps S100, S200, S300, S400, S500, S600 and S700.
[0031] S100. Acquire SBAS-InSAR data, identify anomalous areas based on SBAS-InSAR data, and remove data from anomalous areas based on data reliability to obtain usable areas;
[0032] Multiple Sentinel-1 and other SAR image data covering the monitoring area are acquired, with the time span set according to monitoring needs (generally using data from the past 5 years). Data processing is performed using the SBAS-InSAR standard workflow, including image registration, interferogram generation, phase unwrapping, orbit refinement, spatiotemporal baseline optimization, and geocoding, to obtain the surface deformation rate map and time-series deformation sequence of the monitoring area, forming the SBAS-InSAR deformation dataset.
[0033] Based on the SBAS-InSAR deformation rate map, a deformation rate threshold (e.g., 10 mm / a) is set to identify areas of abnormal rate. The uncertainty of each area of abnormal rate (e.g., areas traversed by railway lines) is simultaneously estimated, as follows:
[0034] Based on SBAS-InSAR data, within a sliding window, the coherence coefficient of the complex values of the main and auxiliary images is calculated, and the phase noise component is calculated based on the coherence coefficient.
[0035] Atmospheric delay components are calculated using GNSS tropospheric models or InSAR atmospheric correction algorithms.
[0036] Obtain precise track parameters and calculate track error components based on the precise track parameters;
[0037] The uncertainty is calculated based on the phase noise component, atmospheric delay component, and orbital error component. Data with uncertainty exceeding a preset threshold are then removed; the preset threshold can be 0.3.
[0038] As another optional implementation, the reliability of the data can be evaluated by combining the coherence coefficient and uncertainty, and data in which both the coherence coefficient and uncertainty meet preset conditions can be classified as usable data.
[0039] The excluded InSAR data is unreliable and will not be included in the subsequent numerical inversion process. It will only be used to mark the spatial location and topographic background for manual geological inspection or supplementary ground monitoring.
[0040] S200: Collect GNSS observation data from the available area and calculate the GNSS time-series deformation data and the corresponding uncertainties;
[0041] By rationally deploying BeiDou GNSS monitoring points within the monitoring area, collecting raw BeiDou GNSS observation data, and obtaining three-dimensional time-series deformation data (longitude, latitude, and elevation deformation) of each monitoring point through data processing (including ionospheric delay correction, tropospheric delay correction, and multipath effect suppression), a high-precision GNSS deformation dataset is formed.
[0042] GNSS uncertainty calibration: The fixed solution accuracy (E, N, U directions) in each direction is directly obtained from the PPK / RTK solution results and used as the uncertainty of the monitoring point to participate in subsequent calculations.
[0043] S300: Collect time-series data of surface dip angle in the available area using an inclinometer, and calculate the uncertainty of the data;
[0044] Inclinometers are deployed around BeiDou GNSS monitoring points and in deformation-sensitive areas (such as slopes and areas prone to collapse) to collect time-series data on the tilt angle of the ground. Outlier removal (based on the 3σ criterion) and smoothing filtering are performed on the data to obtain a standardized inclinometer dataset that reflects the tilt deformation characteristics of the local area.
[0045] Inclinometer uncertainty calibration: Taking into account the instrument's nominal accuracy, temperature drift coefficient, and installation error, calculate the direction measurement uncertainty at each measuring point.
[0046] S400: Obtain deep underground displacement data and corresponding uncertainties for the available area;
[0047] Automatic monitoring equipment for deep underground displacement is deployed in the monitoring area by drilling to collect time-series data of underground soil or rock displacement at different depths (generally 40m, with one inclinometer every 2m, and the depth and number of inclinometers can be adjusted as needed). The data is then zero-point corrected and trend terms are extracted to form a deep underground displacement dataset, revealing the evolution law of underground deformation.
[0048] The uncertainty of deep underground displacement accumulates nonlinearly with depth. A depth-dependent uncertainty estimate is established: considering the inherent accuracy of the inclinometer, borehole tilt projection error, and depth accumulation error, the uncertainty of each of the 20 measuring points is calculated. The uncertainty is low at shallow measuring points and high at deep measuring points.
[0049] S500. Using dip angle data as a constraint on the direction of surface deformation, an LOS projection equation is established for each InSAR pixel containing GNSS monitoring points. The GNSS celestial deformation is used as an absolute constraint to construct a set of constraint equations. Weighted least squares is used to solve the equations to obtain the initial value of the three-dimensional deformation of the InSAR pixel and the corresponding uncertainty.
[0050] Traditional methods typically assume that surface deformation is predominantly vertical, directly projecting the LOS-direction deformation from InSAR as vertical deformation. This assumption introduces significant errors in complex terrain areas. This approach does not presuppose the deformation direction; instead, it constructs an underdetermined set of equations for the three-dimensional deformation field and solves them jointly using multi-source data. The three-dimensional surface deformation field is categorized into east, north, and celestial components. The LOS-direction deformation measured by InSAR has a projection relationship with the three-dimensional deformation. A single InSAR image provides only one observation equation, making the three-dimensional solution an underdetermined problem.
[0051] For each InSAR pixel, the unknown to be solved is: d = [d E d N d U ] T ;
[0052] The horizontal dominance direction (azimuth angle, measured clockwise from north) provided by the inclinometer corresponds to the unit vector = [sinα,cosα,0] (assuming that the surface deformation is mainly along this horizontal direction, and does not necessarily include the vertical component).
[0053] Construct the InSAR LOS observation equation:
[0054] u E ×d E +u N ×d N +u U ×d U =d LOS ;
[0055] Where, d E For the eastward deformation component, d N For the northward deformation component, d U For the celestial deformation component, u E u N u U These represent the components of the LOS unit vector in the east, north, and sky directions, respectively; d LOS Line-of-sight deformation obtained from InSAR observations;
[0056] Weights of the InSAR LOS observation equation: w LOS =1 / ; To observe the noise variance;
[0057] Based on the horizontal dominant direction of surface deformation given by the inclinometer, the constraint equations are constructed as follows:
[0058] -cos ×d E +sin ×d N =0;
[0059] in, The horizontal dominant azimuth angle measured by the inclinometer; d E For the eastward deformation component, d N This represents the northward deformation component;
[0060] The weights of the constraint equation: w dir =1 / ; The accuracy of the inclinometer measurement (azimuth error) and the smoothness of the terrain are determined by the inclinometer's measurement accuracy and the terrain's smoothness.
[0061] GNSS astronomical absolute constraint equations:
[0062] d U =d U GNSS ;
[0063] Where, d U For the axial deformation component, d U GNSS The absolute celestial deformation value is obtained from GNSS observations;
[0064] The weights of the constraint equation: w U =1 / ; The variance of GNSS astronomical deformation observations;
[0065] Each pixel forms:
[0066]
[0067] Where, d E For the eastward deformation component, d N For the northward deformation component, d U For the celestial deformation component, u E u N u U These represent the components of the LOS unit vector in the east, north, and sky directions, respectively; d U GNSS d represents the absolute celestial deformation value obtained from GNSS observations. LOS Line-of-sight deformation obtained from InSAR observations; The horizontal dominant azimuth angle measured by the inclinometer;
[0068] The above expression can be written as Ad=b;
[0069] The weight matrix is: W = diag(w LOS w dir w U );
[0070] The normal equation is: A T WA× =A T Wb;
[0071] The weighted least squares solution is: = (A T WA) -1 A T Wb;
[0072] Covariance matrix: C = (A T WA) -1 The final covariance matrix is used to reflect the uncertainty after combining all the data.
[0073] S600 uses GNSS, surface dip angle, deep underground displacement monitoring points and InSAR pixels as nodes to construct a heterogeneous map, and constructs the feature vector of the node based on the InSAR pixels, monitoring data of the monitoring points and uncertainties.
[0074] S610. Connect adjacent InSAR pixel nodes to construct spatial adjacency edges, and determine the edge weights based on the uncertainties of the two nodes.
[0075] S620. Connect the monitoring point node with its nearest pixel node to construct the monitoring association edge, and determine the edge weight based on the monitoring data accuracy, spatial distance and uncertainty;
[0076] Among them, the characteristics of GNSS monitoring point nodes include the true value of three-dimensional deformation, the accuracy of three-dimensional direction calculation, and the status of calculation; the characteristics of tilt monitoring point nodes include tilt angle, tilt azimuth, and direction measurement uncertainty.
[0077] S630. For deep underground displacement monitoring data, multiple nodes are taken in each borehole according to the depth to obtain deep displacement nodes. By connecting deep displacement nodes at adjacent depths in the same borehole, depth-related edges are obtained, and the edge weights are determined according to the depth of the node and the shear modulus of the formation. These deep displacement nodes are also connected to the nearest pixel node.
[0078] S700. Input the heterogeneous graph into the spatiotemporal graph neural network in time sequence to obtain the temporal evolution features. Input the temporal evolution features into the deconvolution graph network to obtain the deformation rate field and uncertainty distribution.
[0079] The uncertainty of neighboring nodes accumulates according to the error propagation law. When updating nodes, a heteroscedastic attention mechanism is adopted, and the message passing weight is inversely proportional to the uncertainty of neighboring nodes. High-precision neighbor messages contribute more to node updates, while low-precision neighbor information is automatically suppressed, thus achieving adaptive confidence fusion.
[0080] S710. Calculate the attention weight of the node based on the uncertainty of the node;
[0081] Attention weight Modulation by uncertainty:
[0082] ;
[0083] in, Let the attention weights be those between node i and node j. This represents vector concatenation. For learnable attention vectors, for transpose, For the uncertainty of node j, The smaller, The larger the item value, the higher its weight after attention scoring; The weight matrix is learnable; z i For the features of node i, z j Features of node j; For activation function, To normalize neighbor j;
[0084] S720: Based on node attention weights, node features are encoded into low-dimensional embedding vectors through heteroscedasticity graph convolution to obtain node embeddings at each time step.
[0085] For the l-th layer, the feature update formula for node i is:
[0086] ;
[0087] in, For the embedding of node i at layer l+1, For the embedding of node i in layer l, Let the attention weights be those between node i and node j. Here, σ represents the learnable weights, and σ is the activation function. N(i) Let i be the neighborhood of node i; is the bias vector of the l-th layer.
[0088] This step also introduces node type gating, dynamically adjusting the feature path based on different types such as pixels, GNSS, inclinometers, and deep displacement.
[0089] Generate node type embeddings based on node type ;
[0090] The node embeddings from the convolutional output of the heteroscedasticity graph are concatenated with the node type embeddings, and then gating coefficients are generated through a fully connected layer and a sigmoid activation function.
[0091]
[0092] in, The gating coefficient, Embedded for nodes, Embedded for node type, These are learnable parameters.
[0093] The feature channels of the node are dynamically adjusted based on the gating coefficient.
[0094] S730. Embed the nodes into a temporal convolutional network with temporal input to obtain temporal evolution features;
[0095] By inputting temporal evolution features into a deconvolutional graph network, the deformation rate field and uncertainty distribution are obtained, including:
[0096] Spatial resolution is gradually restored by deconvolutional graph network to generate preliminary prediction results of three-dimensional deformation field; then, mechanical consistency correction of deformation field is performed by adaptive physical residual layer, and finally the three-dimensional deformation rate field of key area and corresponding uncertainty distribution map are output.
[0097] For the aforementioned neural network, an uncertainty-driven hybrid loss function is used for training and optimization. The loss function is constructed as follows:
[0098] Based on the uncertainty of monitoring points and the local structural similarity of the deformation rate field, a data fidelity loss term is constructed.
[0099] Data fidelity loss includes MSE loss (pixel-level accuracy, weighted by the inverse of the uncertainty of GNSS and high-confidence monitoring points) and SSIM loss (structural consistency, calculated using an 11×11 Gaussian window to determine local structural similarity). Its constraint ultimately determines the deformation. It is closer to the true observation value.
[0100] Based on the deformation rate field, the continuous residual, the compatibility residual, and the boundary condition residual are calculated, and a physical constraint loss term is constructed.
[0101] An uncertainty calibration loss term is constructed based on the predicted deformation rate field and the observed true value of deformation.
[0102] A loss function is constructed based on a data fidelity loss term, a physical constraint loss term, and an uncertainty calibration loss term. The spatiotemporal graph neural network and the deconvolution graph network are trained based on the loss function to obtain the prediction model.
[0103] Example 2:
[0104] This embodiment provides a deep learning-based surface deformation monitoring system, including:
[0105] The first module is used to acquire SBAS-InSAR data, identify anomalous areas based on SBAS-InSAR data, and remove data from anomalous areas based on data reliability to obtain usable areas.
[0106] The second module is used to collect GNSS observation data from the available area and calculate the GNSS time-series deformation data and the corresponding uncertainties.
[0107] The third module is used to collect time-series data of surface dip angle in the available area using an inclinometer and to calculate the uncertainty of the data;
[0108] The fourth module is used to obtain deep underground displacement data and corresponding uncertainties in the available area;
[0109] The fifth module is used to use dip angle data as a constraint on the direction of surface deformation. For each InSAR pixel containing GNSS monitoring points, an LOS projection equation is established, and GNSS celestial deformation is used as an absolute constraint to construct a set of constraint equations. Weighted least squares is used to solve the equations to obtain the initial value of the three-dimensional deformation of the InSAR pixel and the corresponding uncertainty.
[0110] The sixth module is used to construct a heterogeneous map using GNSS, surface dip, deep underground displacement monitoring points and InSAR pixels as nodes, and to construct the feature vectors of the nodes based on the InSAR pixels, monitoring data of the monitoring points and uncertainties.
[0111] The seventh module is used to input heterogeneous graphs into a spatiotemporal graph neural network in a temporal sequence to obtain temporal evolution features, and then input the temporal evolution features into a deconvolutional graph network to obtain the deformation rate field and uncertainty distribution.
[0112] As an optional implementation, the first module includes:
[0113] The first unit is used to calculate the coherence coefficient of the complex values of the main and auxiliary images within a sliding window based on SBAS-InSAR data, and to calculate the phase noise component based on the coherence coefficient.
[0114] The second unit is used to calculate the atmospheric delay component using the GNSS tropospheric model or the InSAR atmospheric correction algorithm.
[0115] The third unit is used to obtain precise track parameters and calculate track error components based on the precise track parameters.
[0116] The fourth unit is used to calculate the uncertainty based on the phase noise component, atmospheric delay component, and orbital error component, and to remove data with uncertainty exceeding a preset threshold.
[0117] As an optional implementation, the sixth module includes:
[0118] The fifth unit is used to connect adjacent InSAR pixel nodes, construct spatial adjacency edges, and determine edge weights based on the uncertainty of the two nodes.
[0119] The sixth unit is used to connect the monitoring point node with its nearest pixel node, construct the monitoring association edge, and determine the edge weight based on the monitoring data accuracy, spatial distance and uncertainty;
[0120] The seventh unit is used to monitor deep underground displacement data. Multiple nodes are taken in each borehole according to the depth to obtain deep displacement nodes. By connecting deep displacement nodes at adjacent depths in the same borehole, depth-related edges are obtained, and the edge weights are determined according to the depth of the node and the shear modulus of the formation.
[0121] As an optional implementation, the seventh module includes:
[0122] The eighth unit is used to calculate the attention weights of nodes based on their uncertainty.
[0123] The ninth unit is used for node-based attention weights. It encodes node features into low-dimensional embedding vectors through heteroscedasticity graph convolution to obtain node embeddings at each time step.
[0124] The tenth unit is used to embed nodes into a temporally input temporal convolutional network to obtain temporal evolution features.
[0125] Example 3:
[0126] Corresponding to the above method embodiments, this embodiment also provides a surface deformation monitoring device based on deep learning. The surface deformation monitoring device based on deep learning described below and the surface deformation monitoring method based on deep learning described above can be referred to each other.
[0127] Figure 2 This is a block diagram illustrating a deep learning-based surface deformation monitoring device 800 according to an exemplary embodiment. Figure 2As shown, the deep learning-based land deformation monitoring device 800 includes a processor 801 and a memory 802. The deep learning-based land deformation monitoring device 800 may also include one or more of the following: a multimedia component 803, an input / output (I / O) interface 804, and a communication component 805. The processor 801 controls the overall operation of the deep learning-based land deformation monitoring device 800 to complete all or part of the steps in the aforementioned deep learning-based land deformation monitoring method. The memory 802 stores various types of data to support the operation of the deep learning-based land deformation monitoring device 800. This data may include, for example, commands for any application or method operating on the deep learning-based land deformation monitoring device 800, and application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.
[0128] Multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals.
[0129] The received audio signals can be further stored in memory 802 or transmitted via communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as a keyboard, mouse, buttons, etc. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the deep learning-based surface deformation monitoring device 800 and other devices. Wireless communication includes, for example, Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof; therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, or an NFC module.
[0130] Example 4:
[0131] Corresponding to the above embodiment of the deep learning-based land deformation monitoring method, this embodiment also provides a readable storage medium. The readable storage medium described below can be referred to in correspondence with the deep learning-based land deformation monitoring method described above.
[0132] A readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described embodiment of the deep learning-based land deformation monitoring method.
[0133] Specifically, the readable storage medium can be a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or any other readable storage medium capable of storing program code.
[0134] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0135] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for monitoring land surface deformation based on deep learning, characterized in that, include: Acquire SBAS-InSAR data, identify anomalous areas based on SBAS-InSAR data, and remove data from anomalous areas based on data reliability to obtain usable areas; Collect GNSS observation data from available areas and calculate GNSS time-series deformation data and corresponding uncertainties. Time-series data of surface dip angle in available areas are collected using inclinometers, and the uncertainty of the data is calculated; Obtain deep underground displacement data and corresponding uncertainties for the available area; Using dip angle data as a constraint on the direction of surface deformation, an LOS projection equation is established for each InSAR pixel containing GNSS monitoring points. GNSS celestial deformation is used as an absolute constraint to construct a set of constraint equations. Weighted least squares is used to solve the equations to obtain the initial three-dimensional deformation value of the InSAR pixel and the corresponding uncertainty. Heterogeneous maps are constructed using GNSS, surface dip, deep underground displacement monitoring points and InSAR pixels as nodes, and feature vectors of nodes are constructed based on InSAR pixels, monitoring data of monitoring points and uncertainties. The heterogeneous graph is input into the spatiotemporal graph neural network in a temporal sequence to obtain temporal evolution features. The temporal evolution features are then input into the deconvolution graph network to obtain the deformation rate field and uncertainty distribution.
2. The deep learning-based surface deformation monitoring method according to claim 1, characterized in that, Data in outlier regions is removed based on data reliability, including: Based on SBAS-InSAR data, within a sliding window, the coherence coefficient of the complex values of the main and auxiliary images is calculated, and the phase noise component is calculated based on the coherence coefficient. Atmospheric delay components are calculated using GNSS tropospheric models or InSAR atmospheric correction algorithms. Obtain precise track parameters and calculate track error components based on the precise track parameters; Uncertainty is calculated based on phase noise component, atmospheric delay component and orbital error component, and data with uncertainty higher than a preset threshold are removed.
3. The deep learning-based surface deformation monitoring method according to claim 1, characterized in that, A heterogeneous map is constructed using GNSS, surface dip, deep subsurface displacement monitoring points, and InSAR pixels as nodes, including: Connect adjacent InSAR pixel nodes to construct spatial adjacency edges, and determine the edge weights based on the uncertainties between the two nodes; Connect the monitoring point node with its nearest pixel node to construct the monitoring association edge, and determine the edge weight based on the monitoring data accuracy, spatial distance and uncertainty; For deep underground displacement monitoring data, multiple nodes are taken in each borehole according to the depth to obtain deep displacement nodes. By connecting deep displacement nodes at adjacent depths in the same borehole, depth-related edges are obtained, and the edge weights are determined according to the depth of the node and the shear modulus of the formation.
4. The deep learning-based surface deformation monitoring method according to claim 1, characterized in that, The heterogeneous graph is input into a spatiotemporal graph neural network in a temporal sequence to obtain temporal evolution features, including: Calculate the attention weight of a node based on its uncertainty. Based on the attention weights of nodes, node features are encoded into low-dimensional embedding vectors through heteroscedasticity graph convolution to obtain node embeddings at each time step. By embedding nodes into a temporal convolutional network with temporal input, temporal evolution features are obtained.
5. The deep learning-based surface deformation monitoring method according to claim 4, characterized in that, The method includes: Generate node type embeddings based on node type; The node embeddings from the convolutional output of the heteroscedasticity graph are concatenated with the node type embeddings, and then gating coefficients are generated through a fully connected layer and a sigmoid activation function. The feature channels of the node are dynamically adjusted based on the gating coefficient.
6. The deep learning-based surface deformation monitoring method according to claim 1, characterized in that, The method includes: Based on the uncertainty of monitoring points and the local structural similarity of the deformation rate field, a data fidelity loss term is constructed. Based on the deformation rate field, the continuous residual, the compatibility residual, and the boundary condition residual are calculated, and a physical constraint loss term is constructed. An uncertainty calibration loss term is constructed based on the predicted deformation rate field and the observed true value of deformation. A loss function is constructed based on a data fidelity loss term, a physical constraint loss term, and an uncertainty calibration loss term. The spatiotemporal graph neural network and the deconvolution graph network are trained based on the loss function to obtain the prediction model.
7. A surface deformation monitoring system based on deep learning, characterized in that, include: The first module is used to acquire SBAS-InSAR data, identify anomalous areas based on SBAS-InSAR data, and remove data from anomalous areas based on data reliability to obtain usable areas. The second module is used to collect GNSS observation data from the available area and calculate the GNSS time-series deformation data and the corresponding uncertainties. The third module is used to collect time-series data of surface dip angle in the available area using an inclinometer and to calculate the uncertainty of the data; The fourth module is used to obtain deep underground displacement data and corresponding uncertainties in the available area; The fifth module is used to use dip angle data as a constraint on the direction of surface deformation. For each InSAR pixel containing GNSS monitoring points, an LOS projection equation is established, and GNSS celestial deformation is used as an absolute constraint to construct a set of constraint equations. Weighted least squares is used to solve the equations to obtain the initial value of the three-dimensional deformation of the InSAR pixel and the corresponding uncertainty. The sixth module is used to construct a heterogeneous map using GNSS, surface dip, deep underground displacement monitoring points and InSAR pixels as nodes, and to construct the feature vectors of the nodes based on the InSAR pixels, monitoring data of the monitoring points and uncertainties. The seventh module is used to input heterogeneous graphs into a spatiotemporal graph neural network in a temporal sequence to obtain temporal evolution features, and then input the temporal evolution features into a deconvolutional graph network to obtain the deformation rate field and uncertainty distribution.
8. The deep learning-based surface deformation monitoring system according to claim 7, characterized in that, The first module includes: The first unit is used to calculate the coherence coefficient of the complex values of the main and auxiliary images within a sliding window based on SBAS-InSAR data, and to calculate the phase noise component based on the coherence coefficient. The second unit is used to calculate the atmospheric delay component using the GNSS tropospheric model or the InSAR atmospheric correction algorithm. The third unit is used to obtain precise track parameters and calculate track error components based on the precise track parameters. The fourth unit is used to calculate the uncertainty based on the phase noise component, atmospheric delay component, and orbital error component, and to remove data with uncertainty exceeding a preset threshold.
9. The deep learning-based surface deformation monitoring system according to claim 7, characterized in that, The sixth module includes: The fifth unit is used to connect adjacent InSAR pixel nodes, construct spatial adjacency edges, and determine edge weights based on the uncertainty of the two nodes. The sixth unit is used to connect the monitoring point node with its nearest pixel node, construct the monitoring association edge, and determine the edge weight based on the monitoring data accuracy, spatial distance and uncertainty; The seventh unit is used to monitor deep underground displacement data. Multiple nodes are taken in each borehole according to the depth to obtain deep displacement nodes. By connecting deep displacement nodes at adjacent depths in the same borehole, depth-related edges are obtained, and the edge weights are determined according to the depth of the node and the shear modulus of the formation.
10. The deep learning-based surface deformation monitoring system according to claim 7, characterized in that, The seventh module includes: The eighth unit is used to calculate the attention weights of nodes based on their uncertainty. The ninth unit is used for node-based attention weights. It encodes node features into low-dimensional embedding vectors through heteroscedasticity graph convolution to obtain node embeddings at each time step. The tenth unit is used to embed nodes into a temporally input temporal convolutional network to obtain temporal evolution features.
