Beidou and insar adaptive weighted fusion method and system for coal mine subsidence

Abstract

The application discloses a Beidou and InSAR adaptive weighted fusion method for coal mine subsidence, comprising the following steps: extracting Beidou subsidence time series and InSAR subsidence time series from obtained Beidou GNSS observation data and SAR image data, and performing space-time registration; calculating a Beidou observation value precision index of each Beidou station and an InSAR observation value precision index of each InSAR pixel; performing space interpolation on the Beidou observation value precision index of each Beidou station by using a preset space interpolation method to obtain the Beidou observation value precision index of each InSAR pixel; adaptively determining the fusion weight of the Beidou subsidence time series and the InSAR subsidence time series based on the Beidou observation value precision index and the InSAR observation value precision index of each InSAR pixel; and performing weighted fusion on the Beidou subsidence time series and the InSAR subsidence time series by using the fusion weight to obtain an adaptively weighted fusion subsidence value. Based on this, the application generates a coal mine subsidence monitoring product with high precision, high reliability and high space-time resolution.

Description

Technical Field

[0001] This invention relates to the field of surveying and mapping science and technology, specifically to an adaptive weighted fusion method and system for BeiDou and InSAR for coal mine subsidence. Background Technology

[0002] Surface movement and subsidence caused by coal mining pose serious environmental and engineering geological problems to mining areas. This subsidence not only damages surface buildings, infrastructure, and farmland, but may also trigger geological disasters such as landslides and collapses, threatening the lives and property of people in mining areas and the stability of the ecological environment. Therefore, accurate, reliable, and continuous dynamic monitoring of coal mine subsidence areas is of paramount importance for safe production, disaster early warning, and environmental protection.

[0003] Currently, the mainstream technologies for surface deformation monitoring mainly include BeiDou / GNSS and InSAR technologies. BeiDou / GNSS technology, by deploying monitoring stations on the ground, can provide absolute three-dimensional coordinate time series with millimeter-level to centimeter-level accuracy, offering extremely high temporal resolution, reaching hourly or even higher levels, and capable of capturing rapid changes in subsidence. However, the fundamental drawback of this technology lies in its low spatial resolution; monitoring results are limited to discrete station locations, making it difficult to comprehensively reflect the spatial distribution characteristics of the entire subsidence basin. Furthermore, in mining areas with complex terrain, the deployment and maintenance costs of these stations are prohibitively high.

[0004] Interferometric Synthetic Aperture Radar (InSAR) technology, especially time-series InSAR technology based on long-term series analysis (such as SBAS-InSAR), compensates for the low spatial resolution of BeiDou / GNSS technology. It can acquire large-scale, high-density (meter-level resolution) surface deformation fields, enabling "area" monitoring of subsidence basins. However, InSAR technology also has inherent limitations. First, its measurement results are one-dimensional relative deformation along the line of sight (LOS), requiring absolute reference correction and three-dimensional deformation decomposition using ground control points or model assumptions. Second, its temporal resolution is limited by the satellite revisit period (e.g., 6-12 days for Sentinel-1), making it difficult to capture sudden deformation events. Furthermore, InSAR measurements are highly susceptible to atmospheric delay (especially tropospheric water vapor changes) and incoherence caused by factors such as changes in surface vegetation and rapid deformation; these factors are particularly prominent in mining areas with variable ecological environments.

[0005] To combine the advantages of both technologies and achieve synergy, researchers have proposed a monitoring method that fuses BeiDou / GNSS and InSAR data. Existing technologies typically employ simple weighted averaging or Kalman filtering methods for fusion. However, these methods have significant drawbacks:

[0006] Fixed-weight method: This method assigns fixed weights to different data sources, which cannot adapt to the dynamic changes in the accuracy of BeiDou and InSAR data over time. For example, in areas where InSAR is decoupling or during periods when BeiDou signals are interfered with by multipath effects, using fixed weights will severely reduce the reliability of the fusion results.

[0007] Kalman Filtering: While Kalman filtering can handle time-series data, the design of its state and observation equations heavily relies on a prior physical model of the settlement process. Settlement processes caused by coal mining are typically nonlinear and non-Gaussian, and the complex geological conditions and mining activities make it extremely difficult to establish accurate prior models. Model mismatch can lead to divergent filtering results, reduced accuracy, and even erroneous results. Summary of the Invention

[0008] To overcome the problems of static weight allocation, inability to adapt to spatiotemporal changes in data quality, and poor adaptability to complex deformation processes in existing technologies for monitoring surface deformation (especially coal mine subsidence) by integrating BeiDou Global Navigation Satellite System (GNSS) and InSAR data, this invention provides an adaptive weighted fusion method and system for BeiDou and InSAR for coal mine subsidence monitoring. It replaces the traditional spatial interpolation method with a spatiotemporal graph convolutional network model. By learning the spatiotemporal dependencies of the observation accuracy indicators of all stations in the BeiDou monitoring network, it predicts a high-resolution BeiDou accuracy field across the entire monitoring area. This enables the generation of high-precision, high-reliability, and high spatiotemporal resolution coal mine subsidence monitoring products to meet the urgent needs of modern mine safety and environmental monitoring.

[0009] According to one aspect of the present invention, an adaptive weighted fusion method for BeiDou and InSAR for coal mine subsidence is provided, comprising: step S1, extracting BeiDou subsidence time series and InSAR subsidence time series from acquired BeiDou GNSS observation data and SAR image data, and performing spatiotemporal registration; step S2, based on the spatiotemporally registered BeiDou subsidence time series and InSAR subsidence time series, dynamically calculating the BeiDou observation accuracy index of each BeiDou station at different times and the InSAR observation accuracy index of each InSAR pixel at different times; and employing a preset spatial interpolation method. The method involves spatial interpolation of the BeiDou observation accuracy index of each BeiDou station at different times to obtain the BeiDou observation accuracy index of each InSAR pixel at different times; step S3, based on the BeiDou observation accuracy index and InSAR observation accuracy index of each InSAR pixel at different times, adaptively determining the fusion weight of the BeiDou settling time series and InSAR settling time series at each spatiotemporal point; step S4, using the fusion weight, performing weighted fusion on the registered BeiDou settling time series and InSAR settling time series to obtain the final adaptive weighted fused settling value.

[0010] Furthermore, the preset spatial interpolation method in step S2 uses a pre-trained spatiotemporal graph convolutional network model to predict the accuracy index of each InSAR pixel at different times based on the accuracy index of BeiDou observation values ​​of all BeiDou stations in the monitoring area at the current time and historical time.

[0011] Furthermore, the BeiDou station network within the monitoring area is abstracted into a spatiotemporal map. Where: node set V represents all BeiDou stations within the monitoring area; edge set E represents the spatial proximity or geological structural correlation between the BeiDou stations; node feature matrix X represents the accuracy index of BeiDou observation values ​​for each BeiDou station at a continuous time step.

[0012] Furthermore, the spatiotemporal graph convolutional network model includes a graph convolutional network module and a gated recurrent unit module; the graph convolutional network module is used to extract the spatial dependency features of the BeiDou observation accuracy index at each time step; the gated recurrent unit module is used to extract the dynamic change features of the spatial dependency features over time.

[0013] Furthermore, the calculation basis for the BeiDou observation accuracy index in step S2 includes either the solution variance of the three-dimensional deformation time series or the noise level of the time series; the calculation basis for the InSAR observation accuracy index includes correlation, atmospheric phase screen estimation residual and deformation model fitting residual.

[0014] Further, in step S3, the fusion weights of the BeiDou and InSAR settling time series at each spatiotemporal point are adaptively determined based on the BeiDou observation accuracy index and the InSAR observation accuracy index of each InSAR pixel at different times. The corresponding formula is:

[0015] ,

[0016] ,

[0017] in, The fusion weights for the BeiDou settling time series at InSAR pixel j and time t are given. The fusion weights for the InSAR settlement time series at InSAR pixel j and time t are given. This refers to the accuracy index of the BeiDou observation value at InSAR pixel j and time t. Let be the accuracy index of the InSAR observation at position j and time t in InSAR.

[0018] Furthermore, the fusion method also includes: step S5, outputting and visualizing the adaptive weighted fused settlement value.

[0019] According to one aspect of the present invention, a BeiDou and InSAR adaptive weighted fusion system for coal mine subsidence is provided, comprising: a data preprocessing and spatiotemporal registration module, used to extract BeiDou subsidence time series and InSAR subsidence time series from acquired BeiDou GNSS observation data and SAR image data, and perform spatiotemporal registration; an accuracy evaluation module, used to dynamically calculate the BeiDou observation accuracy index of each BeiDou station at different times and the InSAR observation accuracy index of each InSAR pixel at different times based on the spatiotemporally registered BeiDou subsidence time series and InSAR subsidence time series; and employing preset spatial interpolation. The method involves spatial interpolation of the BeiDou observation accuracy index of each BeiDou station at different times to obtain the BeiDou observation accuracy index of each InSAR pixel at different times; a fusion weight calculation module is used to adaptively determine the fusion weight of the BeiDou settling time series and the InSAR settling time series at each spatiotemporal point based on the BeiDou observation accuracy index and the InSAR observation accuracy index of each InSAR pixel at different times; an adaptive weighted fusion calculation module is used to use the fusion weight to perform weighted fusion of the registered BeiDou settling time series and the InSAR settling time series to obtain the final adaptive weighted fused settling value.

[0020] According to one aspect of the present invention, an electronic device is provided, including a memory and a processor, the memory storing program instructions executed by the processor, the processor invoking the program instructions to execute the aforementioned adaptive weighted fusion method of BeiDou and InSAR for coal mine subsidence.

[0021] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, the non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the aforementioned BeiDou and InSAR adaptive weighted fusion method for coal mine subsidence.

[0022] The present invention provides a more advanced, reliable, and intelligent BeiDou and InSAR data fusion solution for coal mine settlement monitoring. Specifically, it replaces the traditional spatial interpolation method with a spatiotemporal graph convolutional network model. By learning the spatiotemporal dependencies of the observation accuracy indicators of all stations in the BeiDou monitoring network, it predicts a high-resolution BeiDou accuracy field across the entire monitoring area. Based on this predicted accuracy field and the InSAR's own accuracy indicators, dynamic fusion weights are calculated using the variance component estimation principle, and BeiDou and InSAR settlement data are weighted and fused. Based on this, the present invention overcomes the limitations of traditional interpolation methods for sparse data and non-stationary fields, significantly improving the reliability of the fusion weights and the accuracy and spatiotemporal consistency of the final settlement product, providing a more precise technical means for safe production, disaster prevention, and environmental protection in mining areas.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] (1) This invention abandons static, empirical fixed weights and adopts a physically-based, multi-component composite accuracy index to dynamically determine the weights. This ensures that high-quality observation data dominates the fusion result at any spatiotemporal point, while the weights of low-quality or gross data (such as InSAR images with severe atmospheric interference or GNSS epochs with significant multipath effects) are automatically suppressed. This fundamentally improves the accuracy and reliability of the final fused subsidence field;

[0025] (2) The data-driven fusion strategy of the present invention does not depend on any specific settlement physics model. Therefore, it can realistically reflect the highly nonlinear and non-stationary settlement process caused by complex geological conditions and mining activities, such as step, accelerated or decelerated settlement, and overcome the problem of decreased accuracy or even divergent results that may be caused by model-based Kalman filtering when the model is mismatched;

[0026] (3) This invention innovatively introduces the STGCN model to predict and interpolate the GNSS observation accuracy field. This model can learn and utilize the inherent spatiotemporal correlation of GNSS measurement errors, and its prediction results are more accurate and more in line with physical reality than traditional pure spatial interpolation methods. This ensures that reliable fusion weights can be obtained throughout the entire monitoring area (including areas far from GNSS stations), thereby obtaining spatially continuous and high-quality fusion results;

[0027] (4) The accuracy indicators of BeiDou observations and InSAR observations calculated during the fusion process are valuable byproducts in themselves. They provide users with a "map" of the uncertainty of the final fusion result, clearly indicating in which areas and time periods the reliability of the monitoring results is high or low, providing important quality control basis for data interpretation and decision-making. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a technical schematic diagram of an adaptive weighted fusion method for BeiDou and InSAR in coal mine subsidence, provided as an embodiment of the present invention.

[0030] Figure 2 The flowchart illustrates an adaptive weighted fusion method for BeiDou and InSAR based on coal mine subsidence, as provided in this embodiment of the invention.

[0031] Figure 3 This is an architecture diagram of the STGCN model provided in an embodiment of the present invention. Detailed Implementation

[0032] It should be noted that:

[0033] The terms “comprising” and “having”, and any variations thereof, in the specification, claims, and accompanying drawings of this invention are intended to cover a non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0034] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be decomposed, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0035] 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 embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form new technical solutions. Such combinations are not bound by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0036] Please refer to Figure 1 and Figure 2 This invention discloses an adaptive weighted fusion method for BeiDou and InSAR data regarding coal mine subsidence. Its core idea is to overcome the limitations of traditional fusion methods (such as fixed-weight methods or Kalman filtering methods relying on prior physical models) and establish a dynamic, objective, and physically-based observation accuracy evaluation system. This system can independently quantify the instantaneous quality of BeiDou and InSAR observation data at each spatiotemporal point. Based on this accuracy evaluation result, the fusion weights are adaptively determined using a weighting principle inversely proportional to variance. This gives higher-quality data sources greater weight in the fusion calculation, suppresses the noise impact of low-quality data sources, and ultimately generates a high-precision, high-reliability, and high spatiotemporal resolution surface subsidence field.

[0037] This paradigm shift from model-driven to data-driven approaches represents a key technological leap in this invention. Traditional Kalman filtering methods attempt to predict settlement behavior by establishing a prior physical model. However, in complex and nonlinear scenarios such as coal mining, establishing accurate prior models is extremely difficult, and model mismatch is the main reason for the divergence and reduced accuracy of filtering results. The method of this invention, however, does not rely on prior assumptions about the settlement process. Its intelligence lies in the system's ability to dynamically evaluate and trust the data itself, rather than depending on a potentially inaccurate model. This makes the method of this invention more robust and adaptable to unexpected or complex deformation patterns (such as step settlement, accelerated settlement, etc.), making it particularly suitable for monitoring mining areas with variable geological conditions and complex mining activities.

[0038] Example 1:

[0039] The following will combine Figure 1 and Figure 2The flowchart shown provides a detailed breakdown and explanation of the fusion method proposed in this invention. This embodiment takes monitoring the surface subsidence field of a large coal mine as an example; the specific steps are as follows:

[0040] Step S1: Acquire BeiDou GNSS observation data and SAR image data and preprocess them to obtain BeiDou settling time series and InSAR settling time series under absolute reference, and perform spatiotemporal registration on the BeiDou settling time series and InSAR settling time series.

[0041] Step S1 aims to extract the deformation time series of the two techniques from the raw observation data and unify them under the same spatiotemporal reference.

[0042] (1) BeiDou GNSS observation data processing

[0043] 1) Data source acquisition: Several BeiDou / GNSS continuously operating reference stations (CORS) or periodically re-measured monitoring points are deployed in and around the stable area of ​​the monitoring mining area to acquire their raw observation data, which is usually in RINEX format.

[0044] 2) High-precision data processing software: The internationally recognized high-precision academic software GAMIT / GLOBK is used for data processing to ensure the acquisition of millimeter-level coordinate time series. Specific processing flow:

[0045] 1. Baseline Calculation (GAMIT): The GAMIT module is used to calculate the baseline based on double-difference carrier phase observations. Double-difference technology effectively eliminates clock errors between satellites and receivers and significantly reduces the impact of atmospheric delay and orbital errors. During this process, a series of sophisticated geophysical models are applied for corrections, including: Atmospheric Delay Correction: Tropospheric zenith delay models such as the Saastamoinen model are used, and gradient parameters are estimated to correct for path delay caused by the signal passing through the troposphere. Earth Tidal Correction: Based on the International Earth Rotation Service (IERS) specifications, models such as solid Earth tides, ocean tidal loads, and polar tides are applied to correct for periodic station displacement caused by celestial gravity. Antenna Phase Center Correction: The Absolute Antenna Phase Center (PCV) model published by the International GNSS Service (IGS) is used to correct for the influence of antenna phase center variations with elevation and azimuth angles.

[0046] 2. Multi-period, multi-station joint adjustment (GLOBK): The GLOBK module, which employs a Kalman filter algorithm, performs joint adjustment on the daily or time-period single-day relaxation solutions (h-files) generated by GAMIT. By introducing global or regional stable reference stations, the station coordinates are unified to a high-precision reference frame such as ITRF (International Earth Reference Frame), ultimately generating a three-dimensional coordinate time series of each monitoring station i under an absolute reference.

[0047] 3) Output: Obtain the BeiDou settling time series, i.e., the three-dimensional displacement vectors of a series of discrete monitoring points i at each time t. .

[0048] (2) SAR image data processing

[0049] 1) Data Source Acquisition: Long-term C-band Sentinel-1 satellite single-view complex (SLC) imagery data covering the monitored mining area was acquired from the European Space Agency (ESA) Copernicus Data Center. To achieve three-dimensional deformation decomposition, it is preferable to acquire data from both ascending and descending orbit observation geometries simultaneously.

[0050] 2) Temporal InSAR Processing Technology: Small Baseline Set (SBAS)-InSAR technology is employed. This technology effectively suppresses decoherence effects by combining interferometric pairs with short spatiotemporal baselines, making it particularly suitable for distributed deformation monitoring of non-urban surfaces such as mining areas. Specific processing flow (using ESA SNAP software as an example):

[0051] 1. Data stacking and interferometric pair generation: Select a high-quality image as the main image, and automatically combine it into several interferometric pairs according to the set spatiotemporal baseline thresholds (e.g., spatial baseline < 200 meters, temporal baseline < 48 days) to form a processing network diagram.

[0052] 2. Precise registration: Using the S-1 TOPS Coregistration module, load precise orbital ephemeris (POE) data to precisely register all auxiliary images onto the geometric space of the main image.

[0053] 3. Interferogram generation and TOPS deburst: Perform complex conjugate multiplication on each interference pair to generate a differential interferogram, and perform TOPS Deburst operation to merge multiple sub-stripes into a complete interferogram.

[0054] 4. Terrain phase removal: Using external digital elevation models (DEMs), such as SRTM 1 arcsecond (approximately 30-meter resolution) data, the phase components caused by terrain undulations are removed from the differential interferogram.

[0055] 5. Multi-view and filtering: Multi-view processing (e.g., 4 views in the azimuth direction and 1 view in the range direction) is performed to suppress speckle noise, and the Goldstein phase filtering algorithm is applied to further improve the clarity and signal-to-noise ratio of the interference fringes.

[0056] 6. Phase unwrapping: Using algorithms based on minimum cost flow or statistical cost network flow, the filtered interferogram is unwrapped to recover the true phase change.

[0057] 7. Time series inversion: Perform matrix inversion on all unwrapped interferograms (usually using methods such as singular value decomposition (SVD)) to solve for the average deformation rate and deformation time series of each coherent pixel j in the line of sight (LOS).

[0058] 3) Output: Obtain the InSAR settlement time series, which is the LOS-direction one-dimensional relative deformation value of each SAR image acquisition time t for high spatial density InSAR pixels j covering the entire monitoring area. .

[0059] (3) Spatiotemporal reference unification and registration

[0060] 1) Spatial benchmark unification: The InSAR solution results (i.e., InSAR settlement time series) are geocoded to a geographic coordinate system (such as WGS84) and elevation benchmark consistent with the BeiDou results (BeiDou settlement time series). At the same time, the discrete BeiDou monitoring station coordinates are projected onto the InSAR grid coordinate system to align the two spatially.

[0061] 2) Time Standardization: The time resolution (e.g., days) of BeiDou settling time series is usually higher than that of InSAR settling time series (e.g., 6-12 days). Linear interpolation or cubic spline interpolation methods need to be used to interpolate the InSAR settling time series to match the time of the BeiDou settling time series, thus achieving time alignment.

[0062] Step S2: Based on the spatiotemporally registered BeiDou settling time series and InSAR settling time series, dynamically calculate the BeiDou observation accuracy index of each BeiDou station at different times and the InSAR observation accuracy index of each InSAR pixel at different times; use a preset spatial interpolation method to spatially interpolate the BeiDou observation accuracy index of each BeiDou station at different times to obtain the BeiDou observation accuracy index of each InSAR pixel at different times.

[0063] It should be noted that step S2 is one of the core aspects of the method of this invention, which aims to assign an objective and quantitative quality evaluation index to each data point.

[0064] (1) Accuracy indicators of BeiDou observations

[0065] The accuracy index of the BeiDou observation value is directly derived from the high-precision GNSS data processing in step S1, reflecting the internal consistency accuracy of the coordinates calculated by the BeiDou monitoring station (i.e., BeiDou site) at position i and time t.

[0066] The source of the indicator can be the variance of the coordinate components output by the GLOBK Kalman filter (i.e., the variance of the three-dimensional deformation time series), or the noise level obtained by noise analysis of the final coordinate time series (e.g., using Hector software).

[0067] Physical meaning: This index comprehensively reflects the impact of various factors on positioning accuracy at a given moment, including satellite geometry (PDOP value), multipath effects, and residual atmospheric noise. A higher PDOP value or more severe multipath effects at a given moment indicates lower positioning accuracy. The value increases accordingly, indicating that the uncertainty of the observed value is high.

[0068] It should be noted that only the accuracy indicators of BeiDou observations at position i and time t from the BeiDou monitoring station are included. It is known, but for non-BeiDou site locations (e.g., at each InSAR pixel), a pre-defined spatial interpolation method is needed to spatially interpolate the BeiDou observation accuracy index of each BeiDou site at different times, in order to obtain the BeiDou observation accuracy index of each InSAR pixel at different times. .

[0069] Unlike traditional spatial interpolation methods, such as Kriging or Inverse Distance Weighted (IDW) interpolation, which interpolate the accuracy indices of non-site locations based on the accuracy indices of surrounding BeiDou stations, this invention innovatively designs a spatial interpolation method. This method uses a pre-trained Spatiotemporal Graph Convolutional Network (STGCN) model to predict the BeiDou observation accuracy indices of each InSAR pixel at different times based on the BeiDou observation accuracy indices of all BeiDou stations within the monitoring area at the current and historical times (details of the STGCN model can be found in Example 2 below, and will not be elaborated here).

[0070] (2) InSAR observation accuracy indicators

[0071] Unlike existing technologies that rely solely on coherence as a single quality indicator, this invention innovatively constructs a composite accuracy indicator that more comprehensively reflects the physical causes of InSAR measurement errors. This composite indicator represents a significant advancement over single indicators (such as those using only coherence). While a single coherence indicator primarily reflects the stability and temporal variations of surface scattering characteristics (e.g., vegetation growth, agricultural activities), it fails to effectively capture the two most significant error sources in InSAR measurements: atmospheric delay and nonlinear deformation modeling errors. The spatiotemporal heterogeneity of atmospheric water vapor can introduce centimeter-level phase errors, and the complex nonlinear characteristics of mining subsidence are difficult to fully describe with simple models. This invention incorporates the atmospheric phase screen estimation residual and the deformation model fitting residual into the accuracy indicator, directly quantifying the uncertainties of these two major error sources, thus forming an evaluation system with more complete physical meaning and a better reflection of true measurement quality.

[0072] The accuracy index of InSAR observations It is determined by the following three components:

[0073] 1) Component 1: Coherence :

[0074] Calculation: During the interferogram generation process, the cross-correlation coefficient of the complex signals of the primary and secondary images is obtained by calculating it within a local window (e.g., 5x5 pixels). The calculation formula is as follows:

[0075] ,

[0076] in, It refers to coherence, where k represents the index variable of the pixels within the local window. This indicates the number of pixels within the window. and These are the complex values ​​of the primary and secondary images, respectively. Indicates complex conjugation. It is the number of pixels within the window.

[0077] Convert to standard deviation: coherence A higher value indicates better phase quality. It can be converted into the phase standard deviation using the following approximate formula, which serves as the... One component:

[0078] ,

[0079] in, For phase standard deviation, This represents the number of views processed in multi-view processing.

[0080] 2) Component Two: Atmospheric Phase Screen Estimation Residuals:

[0081] Correction: Atmospheric delay, particularly tropospheric water vapor variation, is a major source of error in InSAR monitoring. This embodiment uses external auxiliary data for correction, such as the GACOS product. GACOS combines high-resolution numerical weather models and global GNSS network data to provide near real-time zenith tropospheric delay (ZTD) products.

[0082] Residual Quantization: The tropospheric delayed phase screen generated by the GACOS product is subtracted from the unwrapped interferogram. The large-scale phase components with spatial correlation that still exist in the interferogram after correction are the atmospheric correction residuals. The spatial standard deviation of this residual phase within a local window around pixel j is calculated, i.e., the atmospheric phase screen estimation residual, and used as a measure of atmospheric correction uncertainty, contributing to... .

[0083] 3) Component 3: Deformation model fitting residuals:

[0084] Modeling: After obtaining the initial InSAR deformation time series, the time series of each pixel is fitted with a simple function model (such as a linear rate model or a linear + periodic term model).

[0085] Residual quantization: Calculate the root mean square error (RMSE) between the observed deformation time series and the fitted model. This residual reflects uncertainties such as nonlinear deformation and noise not captured by the model, and its magnitude also contributes to the model's performance. .

[0086] 4) Combine the three components:

[0087] Final InSAR observation accuracy indicators It is a synthesis of the standard deviations derived from the above three components. For example, it can be obtained by taking the square root of the sum of the squares of the three components, thus comprehensively quantifying the observation uncertainty of InSAR pixel j at time t.

[0088] Step S3: Based on the accuracy index of BeiDou observations and the accuracy index of InSAR observations for each InSAR pixel at different times, adaptively determine the fusion weight of BeiDou settling time series and InSAR settling time series at each spatiotemporal point.

[0089] Step S3 aims to determine the data allocation fusion weights for the BeiDou settling time series and the InSAR settling time series at each spatiotemporal point.

[0090] Weighting principle: The classic variance component weighting method (i.e., inverse variance weighting) in metrology is adopted. The basic idea is that the smaller the variance of the observation (i.e., the higher the accuracy), the greater its weight should be in the weighted average.

[0091] Weight calculation formula:

[0092] ,

[0093] ,

[0094] in, The fusion weights for the BeiDou settling time series at InSAR pixel j and time t are given. Let be the fusion weight of the InSAR settling time series at InSAR pixel j and time t, and satisfy that the sum of the two is 1. This refers to the accuracy index of the BeiDou observation value at InSAR pixel j and time t. Let be the accuracy index of the InSAR observation at position j and time t in InSAR.

[0095] Step S4: Using fusion weights, the registered BeiDou settlement time series and InSAR settlement time series are weighted and fused to obtain the final adaptive weighted fused settlement value.

[0096] Step S4 aims to use the adaptive fusion weights calculated in step S3 to perform weighted fusion of the registered BeiDou settling time series and InSAR settling time series at each spatiotemporal point.

[0097] Fusion calculation: The final fused settlement value at each pixel j and time t. Calculated using the following formula:

[0098] ,

[0099] in, For the final adaptive weighted fused settlement value, This is an InSAR settlement time series. This refers to the deformation value at InSAR pixel j, obtained by spatial interpolation of the BeiDou monitoring station's deformation value (which can be obtained using ordinary spatial interpolation methods such as Kriging). Simultaneously, to ensure the physical meaning of the calculation, the BeiDou 3D deformation vector and the InSAR LOS-direction deformation value must first be aligned to the same direction, typically the vertical direction. This can be achieved by fusing ascending and descending orbit InSAR data or by combining the east-west and north-direction components of BeiDou data to perform 3D deformation decomposition.

[0100] Example 2:

[0101] Please refer to the following: Figure 3 The following section will elaborate on the indicators used to predict the location accuracy of non-BeiDou sites. The Spatiotemporal Graph Convolutional Network (STGCN) model is described. The STGCN model consists of a Graph Convolutional Network (GCN) module and a Gated Recurrent Unit (GRU) module. The Graph Convolutional Network module is used to extract the spatial dependency features of the accuracy index of BeiDou station observations at each time step. The Gated Recurrent Unit module is used to extract the dynamic change features of the spatial dependency features over time, which is the key to achieving high-precision adaptive weights.

[0102] Traditional spatial interpolation methods (such as Kriging) are purely static interpolations in the spatial domain. They estimate the value of an unknown point using only observations from surrounding stations at the same time, ignoring the evolution of GNSS measurement errors over time. However, GNSS error sources, such as satellite orbital errors, atmospheric delay, and multipath effects, all exhibit strong spatiotemporal correlations. For example, a tropospheric air mass in an area of ​​influence may sweep across different stations over time, causing station accuracy indicators to change with spatiotemporal correlations; changes in satellite geometry (PDOP) also exhibit temporal and spatial continuity.

[0103] The STGCN model's architecture is perfectly suited to capture this complex spatiotemporal dependency. It abstracts the GNSS station network as a graph, using graph convolutional network modules to learn the spatial propagation and correlation of errors, while simultaneously using gated recurrent unit modules to learn the dynamic evolution of errors over time. Therefore, the STGCN model is not merely an interpolator, but a predictive model that has learned the inherent physical propagation laws of GNSS measurement uncertainties. Based on the accuracy-related characteristics (such as PDOP and SNR) of the entire station network over a past period, it can more accurately and realistically predict the accuracy indicators of any location at the current moment.

[0104] (1) Graph construction

[0105] To apply the STGCN model, it is first necessary to abstract the BeiDou / GNSS station network within the monitoring area into a spatiotemporal map. Where V is the set of nodes, E is the set of edges, and X is the node feature matrix. The construction of this spatiotemporal graph is fundamental to model performance, and its specific definition is as follows:

[0106] 1) Node set V

[0107] Node set V( The diagram represents the set of all (N) BeiDou stations within the monitoring area. Each node in the diagram... Each physical monitoring station uniquely corresponds to a geographically located station with coordinates (such as longitude, latitude, and elevation). It serves as the basic unit that carries the temporal characteristics and spatial relationships of the station.

[0108] 2) Edge set E

[0109] The edge set E defines the connection relationships and connection strengths (i.e., the topology of the graph) between N nodes. This invention preferably uses a weighted adjacency matrix. To represent. Weighted adjacency matrix. elements in Represents a node and nodes The connection weights between the nodes are defined by incorporating various prior knowledge, primarily based on the following combination:

[0110] 1. Based on spatial proximity (distance weighting): This is the most basic method, based on the assumption that "the closer the geographical locations of the stations, the stronger the correlation of their observation errors." This invention preferably uses a Gaussian kernel function to define smooth spatial relationship weights, rather than simple 0 / 1 connections.

[0111] ,

[0112] in: It is a spatial relation weight matrix. It is a Beidou measurement station and The Euclidean geographical distance between them; It is the bandwidth parameter of the Gaussian kernel, used to control the rate at which the weights decay with distance; This is a preset distance threshold (e.g., 50 km) used to maintain the sparsity of the graph. If the distance between two stations exceeds this threshold, then... .

[0113] 2. Based on geological structural correlation (prior knowledge encoding): In coal mine subsidence areas, geological structures (such as faults and goaf boundaries) have a strong constraining effect on the spatial distribution of errors and deformations. This invention innovatively introduces a geological prior matrix. To encode this association:

[0114]

[0115]

[0116] in, It is a geological prior matrix. The same geological unit can be defined as: located in the same coal mining face influence zone, located in the same fault block, or not separated by large active faults.

[0117] Final weighted adjacency matrix It can be a combination of the two matrices mentioned above, for example, through element-wise multiplication. This means that only stations that are both geographically close and geologically related have a strong connection.

[0118] 3) Node feature matrix X

[0119] The node feature matrix X is the input data of the STGCN model, which is defined as a three-dimensional tensor. .in This represents the number of BeiDou tracking stations in the diagram. It is the historical time step (e.g., T=12, which means using data from the past 12 epochs). This refers to the number of features (feature dimension) possessed by each node at each time step. The key to the success of this invention lies in the selection and quantization of features F. These features F are not arbitrarily selected, but rather are chosen to improve the final solution accuracy. (The model's prediction target) has the strongest physical relevance and causality among the covariates.

[0120] The following are the detailed definitions, data sources, and processing methods for the F features (F=8) in this embodiment:

[0121] 1. Feature origin: Satellite geometry (F=2)

[0122] Position Precision Factor (PDOP). Physical Meaning: Reflects the impact of the geometric distribution of satellites in space on positioning accuracy. One of the most direct contributing factors. Data source: Calculated from GNSS receiver NMEA output (such as GPGSA statements) or RINEX navigation and observation files. Processing method: PDOP is a scalar value, in time The PDOP value of the station can be directly used as this characteristic value. ,in, Indicates in time The first characteristic value of the station, for time PDOP value of the station.

[0123] Number of satellites involved in the solution. Physical meaning: Reflects the redundancy of observations. The more satellites involved, the more robust the solution. Generally, the smaller the value, the better. Data source: GNSS receiver NMEA output (such as GPGSA statements) or log files of high-precision calculation software (such as GAMIT, PRIDE PPP-AR). Processing method: This is an integer scalar value, directly used as this characteristic value. ,in, Indicates in time The second characteristic value of the station, Indicates at time The number of satellites participating in the calculation at the station.

[0124] 2. Source of characteristics: Signal quality (F=3)

[0125] Problem: Signal-to-noise ratio (SNR) and multipath effect are properties of each satellite (SV), and they must be aggregated into properties of each station.

[0126] Average signal-to-noise ratio. Physical meaning: reflects... The overall quality of all signals received by the station at that time. Data source: S1, S2 (or S5, etc.) observations in the RINEX observation file (O-file). Processing method: ... Always involved in the solution The signal-to-noise ratio of each satellite (e.g.) Find the arithmetic mean. ,in, Indicates in time The third characteristic value of the station, In order to be in time The total number of satellites participating in the calculation at the station Indicates the index or sequence number of the satellite participating in the solution. In order to be in time The signal-to-noise ratio of the s-th satellite received by the station.

[0127] Signal-to-noise ratio standard deviation. Physical meaning: reflects... This indicates the dispersion of signal quality at that station at any given time. A large value suggests that some satellite signals are good while others are poor, potentially indicating obstruction or interference, which can lead to decreased solution stability. Data source: Same as above. Processing method: Calculation time The standard deviation of the signal-to-noise ratio of each satellite. ,in, In order to be in time The fourth characteristic value of the station.

[0128] Average multipath effect. Physical meaning: The multipath effect is... The main sources of error. Data sources: intermediate output files from high-precision calculation software (such as GAMIT / GLOBK, Bernese), or multi-path combinations (such as MP1, MP2) calculated from RINEX files using tools such as TEQC. Processing method: aggregation. time The root mean square (RMS) value of the multipath index (e.g., MP1) of a satellite. ,in, In order to be in time The fifth characteristic value of the station, It is the root mean square. In order to be in time The set of multipath indicators for k satellites corresponding to a given station. In order to be in time The multipath index of the station corresponding to the s-th satellite.

[0129] 3. Feature source: Model goodness of fit (F=1)

[0130] Posterior observation residuals. Physical meaning: This is an extremely crucial feature. It reflects the... At any given time, the Kalman filter (or least squares) model fits the actual observations (such as carrier phase). Large residuals indicate unmodeled errors (such as atmospheric anomalies or multipath propagation). It will inevitably increase. Data source: Kalman filter output logs from high-precision calculation software (such as GLOBK, PRIDE PPP-AR), extracting the "posterior residual" of the carrier phase ionosphere-free combination (L0). Processing method: Aggregation. time The root mean square (RMS) value of the posterior residuals of each satellite. ,in, In order to be in time The 6th characteristic value of the station, In order to be in time The set of posterior residuals for k satellites corresponding to the station.

[0131] 4. Feature source: Atmospheric error proxy (F=2)

[0132] Zenith Tropospheric Wet Delay (ZWD). Physical Significance: Tropospheric delay is the largest source of error in GNSS positioning (especially elevation). Drastic spatiotemporal variations in ZWD are highly correlated with uncertainties in coordinate calculation. Data Source: Calculated using ZTD (Total Zenith Delay) and ZHD (Dry Zenith Delay) estimated by high-precision calculation software (such as GAMIT, PPP). Processing method: This is a scalar value, so it is directly used as the feature value. ,in, In order to be in time The 7th characteristic value of the station, In order to be in time The wet delay of the zenith troposphere at the station.

[0133] Tropospheric gradient. Physical meaning: The simple ZWD model assumes the atmosphere is isotropic. The tropospheric gradient (usually with eastward and northward components) reflects atmospheric asymmetry. A large gradient value indicates atmospheric instability. It will be affected. Data source: Tropospheric parameters estimated by high-precision calculation software. Processing method: Calculation. Eastward gradient at time ( ) and northward gradient ( The Euclidean norm (i.e. the magnitude of the gradient) of ). ,in, In order to be in time The 8th characteristic value of the station, for The Euclidean norm of the eastward gradient at time step 1 for The Euclidean norm of the northward gradient at time step.

[0134] 5. Key Step: Feature Normalization

[0135] The above F=8 features have completely different physical units and dimensions (e.g., PDOP has a dimension of 1, ZWD has a dimension of mm, and SNR has a dimension of dB-Hz). To prevent gradient explosion or dominance by a feature with a large dimension during neural network training, normalization is necessary. This invention uses Z-score normalization. The processing method is: (1) Calculate each feature on the training set. Global mean of (f=1...8) and global standard deviation (2) Apply this transformation to all data (training set, validation set, and test set): ,in, Let f represent the f-th eigenvalue of node i at time t in the node feature tensor after Z-score normalization. To represent the original node feature tensor before normalization, node i (i.e., The f-th eigenvalue of the station at time t. Features The global mean, Features The global standard deviation.

[0136] Missing value handling: If in time If a certain feature of the monitoring station (such as MP_Avg) is missing for some reason, the present invention preferably uses the global mean of that feature. Interpolation is performed using (or median) to ensure the integrity of tensor X.

[0137] Through the detailed steps 1.) to 5.) above, this invention constructs a node feature matrix X optimized for neural network input.

[0138] (2) STGCN model architecture

[0139] like Figure 3 As shown, the STGCN model used in this invention is composed of several spatiotemporal convolutional blocks (ST-ConvBlocks) stacked together. Each ST-Conv block adopts a "sandwich" structure, that is, a spatial graph convolutional layer is sandwiched between two temporal convolutional layers.

[0140] 1) Spatial Graph Convolution (GCN) module

[0141] Objective: At each time step, aggregate the feature information of a node and its neighboring nodes in the graph to capture the spatial dependence of accuracy metrics.

[0142] Mathematical Principle: The core of GCN is graph convolution operation. This embodiment adopts the GCN layer propagation rule proposed by Kipf & Welling. Its mathematical form is:

[0143] ,

[0144] in, It is the first The node feature matrix of the layer, It is the first The node feature matrix of the layer, It is an adjacency matrix with self-loops added. It is a weighted adjacency matrix. It is the identity matrix. yes The degree matrix, It is the first The weight parameter matrix that the layer needs to learn. It is a nonlinear activation function (such as ReLU). The essence of this operation is to use a first-order approximation of the graph Laplacian operator to perform a weighted summation of the features of each node and the features of its first-order neighbors, thereby realizing the spatial propagation and aggregation of information.

[0145] 2) Gated Sequence Convolution (GRU) Module

[0146] Objective: After extracting spatial features through the GCN layer, the feature time series of each node is processed to capture the dynamic evolution of accuracy indicators and their long-term and short-term dependencies in the time dimension.

[0147] Mathematical Principle: This embodiment employs a Gated Recurrent Unit (GRU), a more concise and efficient recurrent neural network unit than LSTM. Its core is the update gate. and reset door The update formula is as follows:

[0148] ,

[0149] in, It is the input at the current moment. It is the hidden state from the previous moment. It is the hidden state at the current moment. It is the candidate hidden state at the current moment. , , It is a reset door Parameters to be learned , , It's an update gate. Parameters to be learned , , yes Parameters to be learned This represents the Hadamard product. Through these two gating mechanisms, GRU can learn which historical information in a time series needs to be retained and which needs to be forgotten, thus effectively modeling time dependencies.

[0150] (3) Model training and prediction

[0151] 1) Training Phase: A training set is constructed using historical BeiDou monitoring data. The input is the feature matrix X (including PDOP, SNR, etc.) of each station over the past T time points, and the label is the true accuracy index of each station at the current time. The network parameters of the STGCN model are trained using the backpropagation algorithm and an optimizer (such as Adam) with the goal of minimizing the mean squared error (MSE) between the predicted and actual values.

[0152] 2) Prediction Phase: In practical fusion applications, the latest feature matrices X from the T time points are input into the trained STGCN model. The STGCN model outputs predicted accuracy indicators for all N stations at the current time. Since the GCN layer learns universal spatial aggregation weights, these weights can be used to spatially interpolate the prediction results, thereby obtaining the predicted BeiDou observation accuracy indicator for any InSAR pixel location j within the monitoring area. .

[0153] To ensure that those skilled in the art can reproduce this, the following table provides a set of exemplary hyperparameters for the STGCN model used in this embodiment.

[0154] Table 1 Examples of hyperparameters for the STGCN model

[0155]

[0156] Example 3:

[0157] After step S4, the fusion method further includes: step S5: outputting and visualizing the adaptive weighted fused settlement value.

[0158] Output Results: The final output is a set of time series products of three-dimensional (or vertical) subsidence fields in the mining area with high spatiotemporal resolution, high precision, and high reliability.

[0159] Storage and Visualization: Results can be stored in geospatial databases (such as PostGIS) for easy querying and management. Using Python libraries such as GeoPandas and Matplotlib, or professional GIS software, the fused subsidence field can be visualized in various forms, including 2D color renderings, 3D surface maps, deformation time series curves, and cross-sectional views, providing intuitive data support for mine safety production and disaster early warning.

[0160] Based on the same technical concept as the foregoing embodiments, the present invention also provides a BeiDou and InSAR adaptive weighted fusion system for coal mine subsidence. This fusion system is designed as a modular and scalable geospatial data processing architecture, capable of handling large-scale, long-term monitoring tasks.

[0161] A key feature is that the system architecture is designed for operability and scalability, fully leveraging modern big data and cloud computing paradigms. Compared to monolithic applications running on a single server, this embodiment describes an implementation based on Apache Spark for distributed computing, with Python and its rich geospatial ecosystem at its core. This architecture effectively addresses the challenges of generating massive amounts of data (terabytes of SAR data, continuous GNSS data streams) in regional monitoring projects, demonstrating that this invention is not merely a theoretical concept, but a practical and scalable solution designed for large-scale real-world deployment, thus enhancing its industrial applicability.

[0162] The fusion system includes the following modules:

[0163] 1. The data preprocessing and spatiotemporal registration module is used to extract BeiDou subsidence time series and InSAR subsidence time series from the acquired BeiDou GNSS observation data and SAR image data, and perform spatiotemporal registration. The data preprocessing and spatiotemporal registration module includes a data acquisition and preprocessing module and a benchmark unification and registration module.

[0164] 1.1 Data Acquisition and Preprocessing Module:

[0165] It consists of a series of automated scripts that are responsible for periodically pulling raw data from GNSS data centers (such as IGS) and SAR data centers.

[0166] This module automatically calls the professional processing software in the background (such as the command-line version of GAMIT / GLOBK, or the batch processing script based on ESASNAP-GPT) to execute the BeiDou GNSS observation data and SAR image data preprocessing process described in step S1.

[0167] 1.2 Benchmark Unification and Registration Module:

[0168] A core software component, developed using Python and libraries such as GeoPandas and Rasterio.

[0169] It is responsible for performing operations such as geocoding, coordinate projection transformation, and time interpolation to ensure that the two types of data are aligned under a unified spatiotemporal reference.

[0170] 2. Accuracy assessment module, used to dynamically calculate the accuracy index of BeiDou observations at different times and the accuracy index of InSAR observations at different times for each BeiDou station based on the spatiotemporally registered BeiDou settling time series and InSAR settling time series; and to spatially interpolate the accuracy index of BeiDou observations at different times for each BeiDou station using a preset spatial interpolation method to obtain the accuracy index of BeiDou observations at different times for each InSAR pixel.

[0171] This is the computational core of the system, responsible for performing the dynamic accuracy calculation of step S2.

[0172] It includes an STGCN model training submodule: this is an offline component that uses the PyTorch or TensorFlow deep learning framework to train and validate STGCN models on historical datasets.

[0173] It also includes a precision metric prediction submodule: this is an online component that, when processing real-time data streams, loads a pre-trained STGCN model to perform predictions for the current time step. It performs rapid predictions. Simultaneously, this module is also responsible for calculating composite... index.

[0174] 3. An adaptive weighted fusion calculation module is used to adaptively determine the fusion weights of the BeiDou settling time series and the InSAR settling time series at each spatiotemporal point based on the BeiDou observation accuracy index and the InSAR observation accuracy index of each InSAR pixel at different times.

[0175] To handle the point-by-point computation of massive pixels, this module was implemented as a distributed computing task.

[0176] The Apache Spark framework can be used, leveraging its PySpark interface to write custom fusion computing logic. Data is loaded in the form of distributed datasets (RDDs) or DataFrames, and meta-level weight calculations and weighted averages are executed in parallel across multiple worker nodes in the Spark cluster, greatly improving processing efficiency.

[0177] 4. Results Output and Visualization Module:

[0178] A user interface can be a web-based dashboard or a plugin for desktop GIS software.

[0179] This module queries the database for the final fusion results and visualizes them in various forms such as interactive maps, time series charts, and deformation profiles to provide decision support for users.

[0180] Based on the same technical concept as the foregoing embodiments, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores program instructions that are executed by the processor, and the processor calls the program instructions to execute the aforementioned adaptive weighted fusion method of BeiDou and InSAR for coal mine subsidence.

[0181] Based on the same technical concept as the foregoing embodiments, the present invention also provides a non-transitory computer-readable storage medium that stores computer instructions that cause the computer to execute the aforementioned adaptive weighted fusion method of BeiDou and InSAR for coal mine subsidence.

[0182] It should be noted that the aforementioned storage devices or media employ a hybrid storage strategy. Raw and intermediate large file data (such as SLC imagery and interferograms) can be stored in cloud object storage or a distributed file system. The calculated time-series results, metadata, and the final fusion product are stored in a database that supports spatial data types and queries, such as PostgreSQL and its PostGIS extensions, to enable efficient spatial and temporal queries.

[0183] The table below visually compares the key technical features of the two mainstream fusion methods mentioned in the background art with those of the present invention.

[0184] Table 2 Comparison between the present invention and the prior art

[0185]

[0186] In summary, this invention provides a more advanced, reliable, and intelligent BeiDou and InSAR data fusion solution for coal mine settlement monitoring. Specifically, it replaces the traditional spatial interpolation method with a spatiotemporal graph convolutional network model. By learning the spatiotemporal dependencies of the observation accuracy indicators of all stations in the BeiDou monitoring network, it predicts a high-resolution BeiDou accuracy field across the entire monitoring area. Based on this predicted accuracy field and the InSAR's own accuracy indicators, dynamic fusion weights are calculated using the variance component estimation principle, and BeiDou and InSAR settlement data are weighted and fused. Therefore, this invention overcomes the limitations of traditional interpolation methods for sparse data and non-stationary fields, significantly improving the reliability of the fusion weights and the accuracy and spatiotemporal consistency of the final settlement product, providing a more precise technical means for safe production, disaster prevention, and environmental protection in mining areas.

[0187] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. An adaptive weighted fusion method of BeiDou and InSAR for coal mine subsidence, characterized in that, include: Step S1: Extract the BeiDou subsidence time series and InSAR subsidence time series from the acquired BeiDou GNSS observation data and SAR image data, and perform spatiotemporal registration. Step S2: Based on the spatiotemporally registered BeiDou settling time series and InSAR settling time series, dynamically calculate the BeiDou observation accuracy index of each BeiDou station at different times and the InSAR observation accuracy index of each InSAR pixel at different times; use a preset spatial interpolation method to spatially interpolate the BeiDou observation accuracy index of each BeiDou station at different times to obtain the BeiDou observation accuracy index of each InSAR pixel at different times; wherein, the preset spatial interpolation method includes: using a pre-trained spatiotemporal graph convolutional network model, predicting the BeiDou observation accuracy index of each InSAR pixel at different times based on the BeiDou observation accuracy index of all BeiDou stations in the monitoring area at the current time and historical time. Step S3: Based on the accuracy index of BeiDou observations and the accuracy index of InSAR observations for each InSAR pixel at different times, adaptively determine the fusion weight of BeiDou settling time series and InSAR settling time series at each spatiotemporal point. Step S4: Using the fusion weights, the registered BeiDou settlement time series and InSAR settlement time series are weighted and fused to obtain the final adaptive weighted fused settlement value.

2. The adaptive weighted fusion method of BeiDou and InSAR for coal mine subsidence as described in claim 1, characterized in that, Abstracting the BeiDou network of monitoring stations within the monitoring area into a spatiotemporal map Where: node set V represents all BeiDou stations within the monitoring area; edge set E represents the spatial proximity or geological structural correlation between the BeiDou stations; node feature matrix X represents the accuracy index of BeiDou observation values ​​for each BeiDou station at a continuous time step.

3. The adaptive weighted fusion method of BeiDou and InSAR for coal mine subsidence as described in claim 1, characterized in that, The spatiotemporal graph convolutional network model includes a graph convolutional network module and a gated recurrent unit module; the graph convolutional network module is used to extract the spatial dependency features of the BeiDou observation accuracy index at each time step; the gated recurrent unit module is used to extract the dynamic change features of the spatial dependency features over time.

4. The adaptive weighted fusion method of BeiDou and InSAR for coal mine subsidence as described in claim 1, characterized in that, The calculation basis for the BeiDou observation accuracy index in step S2 includes either the solution variance of the three-dimensional deformation time series or the noise level of the time series; the calculation basis for the InSAR observation accuracy index includes correlation, atmospheric phase screen estimation residual and deformation model fitting residual.

5. The adaptive weighted fusion method of BeiDou and InSAR for coal mine subsidence as described in claim 1, characterized in that, In step S3, the fusion weights of the BeiDou and InSAR settling time series at each spatiotemporal point are adaptively determined based on the accuracy indices of the BeiDou and InSAR observations for each InSAR pixel at different times. The corresponding formula is as follows: ， ， in, The fusion weights for the BeiDou settling time series at InSAR pixel j and time t are given. The fusion weights for the InSAR settlement time series at InSAR pixel j and time t are given. This refers to the accuracy index of the BeiDou observation value at InSAR pixel j and time t. Let be the accuracy index of the InSAR observation at position j and time t in InSAR.

6. The adaptive weighted fusion method of BeiDou and InSAR for coal mine subsidence as described in claim 1, characterized in that, The fusion method further includes: Step S5: Output and visualize the adaptive weighted fused settlement value.

7. A BeiDou and InSAR adaptive weighted fusion system for coal mine subsidence, characterized in that, include: The data preprocessing and spatiotemporal registration module is used to extract BeiDou subsidence time series and InSAR subsidence time series from the acquired BeiDou GNSS observation data and SAR image data, and perform spatiotemporal registration. The accuracy assessment module is used to dynamically calculate the accuracy index of BeiDou observations at different times for each BeiDou station and the accuracy index of InSAR observations at different times for each InSAR pixel, based on the spatiotemporally registered BeiDou settling time series and InSAR settling time series. A preset spatial interpolation method is used to spatially interpolate the accuracy index of BeiDou observations at different times for each BeiDou station, so as to obtain the accuracy index of BeiDou observations for each InSAR pixel at different times. The preset spatial interpolation method includes: using a pre-trained spatiotemporal graph convolutional network model, based on the accuracy index of BeiDou observations of all BeiDou stations in the monitoring area at the current time and historical times, predicting the accuracy index of BeiDou observations for each InSAR pixel at different times. The fusion weight calculation module is used to adaptively determine the fusion weight of the BeiDou settling time series and the InSAR settling time series at each spatiotemporal point based on the BeiDou observation accuracy index and the InSAR observation accuracy index of each InSAR pixel at different times. The adaptive weighted fusion calculation module is used to perform weighted fusion of the registered BeiDou settlement time series and InSAR settlement time series using the fusion weights to obtain the final adaptive weighted fusion settlement value.

8. An electronic device, characterized in that, The system includes a memory and a processor, wherein the memory stores program instructions that are executed by the processor, and the processor invokes the program instructions to execute the BeiDou and InSAR adaptive weighted fusion method for coal mine subsidence as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions that cause the computer to execute the BeiDou and InSAR adaptive weighted fusion method for coal mine subsidence as described in any one of claims 1 to 6.

Images