Geological settlement monitoring method and system integrating deep learning and multi-source data

Abstract

The invention discloses a geological settlement monitoring method and system fusing deep learning and multi-source data, and relates to the technical field of geological settlement, and the method comprises the steps: collecting multi-source settlement observation and driving data, generating a settlement risk area mask based on historical records, and carrying out the self-adaptive grid division; a multi-source settlement field is generated through Kriging interpolation, and a driving factor grid field is mapped; performing spatial pyramid and time multi-scale decomposition on the settlement field and the driving factor to obtain a time-space sub-band; performing Bayesian fusion based on the sub-band confidence weight to obtain a fusion settlement field; inputting the fusion field and the driving factor into a deep learning model to train a multi-scale prediction sub-model, and reconstructing a global continuous settlement prediction field through cross-scale consistency; and dynamically optimizing the risk mask and the grid according to a prediction result to realize iterative monitoring. The problems of difficulty in multi-source data fusion, spatial scale heterogeneity and difficulty in accurate prediction of local high-risk area settlement in geological settlement monitoring are solved.

Description

Technical Field

[0001] This invention relates to the field of geological subsidence technology, and more specifically, to a geological subsidence monitoring method and system that integrates deep learning and multi-source data. Background Technology

[0002] Existing methods for monitoring geological subsidence mainly rely on multi-source observation data, such as InSAR, GNSS, and subsidence observation stations, to estimate surface subsidence through time series analysis, spatial interpolation, or statistical models. These methods can, to some extent, obtain subsidence change trends, support geological disaster early warning and engineering safety assessment, and provide a data foundation for urban construction and underground engineering management.

[0003] However, geological subsidence fields typically exhibit highly heterogeneous spatial scales, meaning that local high-risk areas experience rapid and dramatic subsidence, while surrounding areas experience relatively gentle subsidence. Existing methods, which model and interpolate at a uniform spatial resolution, can easily lead to the smoothing or burying of local anomalous subsidence features, failing to accurately reflect local subsidence details. Furthermore, data fusion from different observation sources also suffers from scale mismatch issues, affecting monitoring accuracy.

[0004] In the field of remote sensing image processing, methods such as spatial pyramids and multi-resolution wavelet decomposition have been proposed to address the fusion problem of multi-resolution images. By decomposing image data into sub-bands with different spatial resolutions and temporal scales, local details can be highlighted while preserving global trends, thus effectively alleviating the information aliasing problem caused by inconsistent scales of the original data. This provides a valuable approach for joint modeling of multi-source data.

[0005] However, directly applying analogy methods to geological subsidence monitoring introduces new problems. The sub-bands formed after multi-scale decomposition of the subsidence field and the driving factor grid field differ in observation accuracy and reliability. High-precision sub-bands accurately reflect subsidence characteristics, while low-precision sub-bands may contain noise. If all sub-bands are directly merged, the noise from low-precision sub-bands may interfere with the overall subsidence reconstruction, affecting both local and global prediction accuracy and causing instability in the monitoring results.

[0006] In summary, existing geological subsidence monitoring technologies struggle to simultaneously capture both global subsidence trends and localized high-risk characteristics. Furthermore, during multi-source data fusion, scale and accuracy differences can easily lead to information aliasing and reconstruction errors. These issues limit the ability to reconstruct subsidence fields with high precision, identify anomalous areas, and reliably predict future subsidence. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a geological subsidence monitoring method and system that integrates deep learning and multi-source data. By performing multi-resolution spatiotemporal decomposition and Bayesian fusion on multi-source subsidence data, and combining deep learning multi-scale prediction and dynamic grid optimization, the method addresses the problems of difficulty in multi-source data fusion, spatial scale heterogeneity, and difficulty in accurately predicting subsidence in local high-risk areas during geological subsidence monitoring.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A geological subsidence monitoring method integrating deep learning and multi-source data comprises the following steps: Collecting multi-source subsidence data, including subsidence observation data and subsidence driving data; generating a subsidence risk zone mask based on the subsidence observation data and historical records, and adaptively dividing the grid accordingly; projecting the subsidence observation data onto the grid using Kriging interpolation to generate a multi-source subsidence field, including a first subsidence field and a second subsidence field; mapping the subsidence driving data onto the grid through spatial overlay to generate a driving factor grid field; performing spatial pyramid and temporal multi-scale decomposition on the multi-source subsidence field and the driving factor grid field to generate several spatiotemporally resolved sub-bands; performing Bayesian fusion of the first and second subsidence fields based on a multi-resolution weighting strategy for the sub-bands to obtain a fused subsidence field; jointly inputting the fused subsidence field and the driving factor grid field into a deep learning model at each scale sub-band to train a multi-scale prediction sub-model, and reconstructing a continuous subsidence prediction field across the entire domain through cross-scale consistency constraints; dynamically updating the risk mask and grid division according to the prediction field to achieve iterative optimization of geological subsidence monitoring.

[0009] In a preferred embodiment, the step of generating a settlement risk zone mask based on settlement observation data and historical records, and adaptively dividing the grid accordingly, specifically involves: constructing a settlement time series curve; using least squares regression to fit the settlement time series curve and calculating the average settlement rate and its acceleration components at each observation point; marking rate anomalies and acceleration mutations as potential risk points to form a preliminary set of settlement risk points; applying a density-based clustering algorithm to the preliminary risk point set to aggregate adjacent high-risk points into risk regions; generating a rasterized mask layer within the monitoring area using the risk regions as polygonal boundaries, and encoding and assigning mask values ​​to the risk regions according to their risk levels; adaptively adjusting the grid division density based on the risk level of the risk zone mask, and merging the division results to generate a grid framework covering the entire area.

[0010] In a preferred embodiment, the step of projecting settlement observation data onto a grid based on kriging interpolation specifically involves: establishing a grid node index within the grid framework and recording the spatial coordinates and corresponding risk level of each node; estimating the spatial trend surface of the overall time series based on linear regression with driving factors to obtain a trend field; detrending the original observations based on the trend field to obtain residual observations at each time step, wherein the residual observations include InSAR residual observations and GNSS residual observations; interpolating the InSAR residual observations onto the grid based on co-kriging to obtain a first settlement field; and interpolating the GNSS residual observations onto the grid based on partitioned kriging to obtain a second settlement field.

[0011] In a preferred embodiment, the step of mapping the settlement driving data to a grid through spatial overlay to generate a driving factor grid field specifically involves: classifying the settlement driving data to obtain data categories, which include point time series, raster data, and vector polygons; interpolating the settlement driving data of the point time series category to the grid based on point-to-surface interpolation; interpolating the settlement driving data of the raster data category to the grid based on resampling and alignment; interpolating the settlement driving data of the vector polygon category to the grid based on area-weighted allocation; and performing interpolation mapping on the settlement driving data to the grid for each data category to obtain the driving factor grid field.

[0012] In a preferred embodiment, the step of performing spatial pyramid and temporal multi-scale decomposition on the multi-source settlement field and the driving factor grid field to generate several spatiotemporally resolved sub-bands specifically involves: setting a preset first-level spatial pyramid resolution layer based on the spatial resolution of the grid, with each layer having a resolution that is half that of the previous layer; downsampling the first settlement field, the second settlement field, and the driving factor grid field to each resolution layer in turn to obtain copies of the settlement field and driving factor field with different spatial resolutions, and retaining the mapping index; performing multi-scale decomposition on the grid temporal sequence of each spatial resolution layer based on wavelet transform to generate multi-timescale components of the multi-source settlement field and the driving factor grid field; and combining the resolution layers of the spatial pyramid with the time components to form a first number of spatiotemporally resolved sub-bands, where the first number is the product of the first-level layer and the number of multi-timescale components.

[0013] In a preferred embodiment, the sub-band-based multi-resolution weighting strategy performs Bayesian fusion on the first and second settlement fields to obtain a fused settlement field. Specifically, this involves: establishing a sub-band observation error model and outputting the sub-band observation error variance; fusing the spatiotemporally resolved sub-bands of the first and second settlement fields based on Bayesian updates and sub-band confidence weights, and performing multi-resolution reconstruction to obtain the fused settlement field. The sub-band confidence weights are obtained by taking the reciprocal of the sub-band observation error variance and normalizing it.

[0014] In a preferred embodiment, the step of jointly inputting the fused settlement field and the driving factor grid field into the deep learning model at each scale sub-band to train the multi-scale prediction sub-model specifically involves: reversibly splitting the fused settlement field into spatiotemporally resolved sub-bands according to the defined parameters to obtain multi-scale input features; synchronously extracting the node values ​​of the driving factor grid field on the corresponding sub-band from the multi-scale input features to form a driving factor multi-scale feature matrix; constructing a deep learning model based on a spatial-temporal hybrid neural network; the deep learning model includes a network input layer and an intermediate layer; independently encoding each sub-band resolution at the network input layer to extract local spatial features and temporal trend features; introducing a cross-scale feature interaction module in the intermediate layer to fuse high-resolution sub-band features and low-resolution sub-band features through a feature pyramid to achieve multi-scale perception; and training the deep learning model based on the driving factor multi-scale feature matrix to obtain the multi-scale prediction sub-model.

[0015] In a preferred embodiment, the reconstruction of the global continuous settlement prediction field through cross-scale consistency constraints specifically involves: introducing a cross-scale loss function during training to constrain the prediction results of different resolution sub-bands to remain consistent after upsampling and downsampling; reconstructing the prediction results of the multi-scale prediction sub-models one by one to the original grid resolution, and smoothing the boundaries using a multi-scale deconvolution reconstruction method; fusing the prediction results of the multi-scale prediction sub-models to generate the global future settlement prediction field, and outputting the confidence distribution.

[0016] In a preferred embodiment, the step of dynamically updating the risk mask and grid division based on the prediction field to achieve iterative optimization of geological subsidence monitoring specifically involves: acquiring the subsidence value distribution and confidence distribution of the future subsidence prediction field across the entire region; regenerating the subsidence risk zone mask based on the subsidence value exceeding the threshold area in the prediction field, and locally refining or merging the grid; using the updated risk mask and grid as the basis for the next round of subsidence observation data projection and driving factor mapping to achieve adaptive optimization of the grid in key risk areas; and after several iterative update cycles, outputting the finally converged subsidence risk zone division and continuous subsidence prediction field to achieve real-time dynamic monitoring and early warning of geological subsidence.

[0017] A geological subsidence monitoring system integrating deep learning and multi-source data includes a multi-source subsidence data acquisition module, an adaptive gridding module, a grid projection module, a fused subsidence field module, a prediction field module, and a monitoring module. The multi-source subsidence data acquisition module collects subsidence observation data and subsidence driving data. The adaptive gridding module generates a subsidence risk zone mask based on the subsidence observation data and performs adaptive gridding. The grid projection module projects the subsidence observation data and subsidence driving data onto the grid, generating a multi-source subsidence field and a driving factor grid field. The fused subsidence field module performs spatiotemporal multi-scale decomposition of the multi-source subsidence field and the driving factor grid field to obtain spatiotemporally resolved sub-bands, and performs Bayesian fusion of the multi-source subsidence field based on a multi-resolution weighting strategy for the sub-bands to obtain a fused subsidence field. The prediction field module reconstructs a global prediction field from the fused subsidence field and the driving factor grid field using a deep learning model. The monitoring module dynamically updates the risk mask and grid division based on the prediction field.

[0018] The technical effects and advantages of this invention's geological subsidence monitoring method and system that integrates deep learning and multi-source data are as follows: 1. This invention decomposes multi-source subsidence fields and driving factor grid fields into spatial pyramids and temporal multi-scale decompositions. This process breaks down the originally inconsistent spatial resolution and complex temporal dynamics of subsidence information into multiple sub-bands at various levels and time scales, effectively mitigating the feature aliasing problem caused by high spatial scale heterogeneity in geological subsidence monitoring. This method preserves the overall trend of large-scale subsidence while highlighting the detailed features of local high-risk areas. This allows subsequent analysis to simultaneously perceive global and local information at multiple resolutions, improving the accuracy of subsidence field reconstruction and sensitivity to anomalous subsidence. It also solves the problem in existing technologies where different observation sources are difficult to directly and uniformly model.

[0019] 2. The multi-scale sub-bands of multi-source settlement fields and driving factor grid fields differ in accuracy and reliability. Direct fusion may lead to noise from low-precision sub-bands affecting the overall reconstruction effect. This invention establishes a sub-band observation error model and uses the inverse of the error variance as a confidence weight for Bayesian fusion, effectively allowing high-precision sub-bands to dominate the fusion result, weakening the influence of low-precision sub-bands, and preserving the integrity of multi-resolution information. The fused settlement field obtained through multi-resolution reconstruction maintains both local high-resolution features and global trends, improving the reconstruction accuracy and reliability of the settlement field and providing a highly reliable input for subsequent multi-scale deep learning predictions. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the geological subsidence monitoring method that integrates deep learning and multi-source data according to the present invention.

[0021] Figure 2This is a schematic diagram of the geological subsidence monitoring system that integrates deep learning and multi-source data according to the present invention. Detailed Implementation

[0022] 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, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1, Figure 1 The present invention provides a geological subsidence monitoring method integrating deep learning and multi-source data, comprising the following steps: S1 collects multi-source settlement data, including settlement observation data and settlement driving data.

[0024] In this embodiment, the collection of multi-source settlement data includes settlement observation data and settlement driving data, specifically as follows: The settlement observation data includes InSAR time-series settlement observation data and GNSS continuous observation data; Acquire InSAR time-series settlement observation data and GNSS continuous observation data; Time synchronization of InSAR time-series settlement observation data and GNSS settlement sequence data based on timestamps; The missing observation times were filled in based on the nearest neighbor and georegistered according to a unified spatial coordinate system to form settlement observation data; Driving factor data are obtained from meteorological monitoring, geological surveys and groundwater monitoring systems. The data of each driving factor are spatiotemporally registered and transformed into the same spatial reference frame and time step as the settlement observation dataset to form settlement driving factor data.

[0025] The acquisition of InSAR time-series settlement observation data in this embodiment is specifically as follows: acquiring multi-orbit SAR image data covering the target area, performing differential interferometry on each orbit image sequence to generate a pixel-level displacement sequence under the same time reference, and removing atmospheric delay and orbital error through time series analysis methods to obtain InSAR time-series settlement observation data; the InSAR time-series settlement observation data includes the settlement displacement value of each pixel at each observation time and the corresponding confidence index.

[0026] The acquisition of GNSS continuous observation data in this embodiment is specifically as follows: GNSS continuous observation stations are deployed to collect three-dimensional coordinate data of the target area in real time. The displacement sequence in the settlement direction is calculated using precise single-point positioning and baseline calculation algorithms, and abnormal jumps are filtered and smoothed to generate GNSS settlement sequence data. The GNSS settlement sequence data includes the vertical displacement value of each base station at each observation time and its measurement error covariance.

[0027] The driving factor data in this embodiment include rainfall, groundwater level changes, soil layer thickness, geological structure, and underground mining activity records.

[0028] S2 generates a subsidence risk zone mask based on subsidence observation data and historical records, and adaptively divides the grid accordingly.

[0029] In this embodiment, the step of generating a subsidence risk zone mask based on subsidence observation data and historical records, and adaptively dividing the grid accordingly, specifically involves: Construct settlement time-series curves; Least square regression was used to fit the settlement time series curve to calculate the average settlement rate and its acceleration components at each observation point. Rate anomalies and acceleration mutations are marked as potential risk points to form a preliminary set of settlement risk points. A density-based clustering algorithm is applied to the initial risk point set to aggregate adjacent high-risk points into risk regions; Using the risk area as the polygon boundary, a raster mask layer is generated within the monitoring area. The mask values ​​are encoded and assigned to the risk area according to the risk level. The mesh density is adaptively adjusted based on the risk level of the risk area mask, and the meshing results are merged to generate a mesh framework covering the entire area.

[0030] The construction of settlement time series curves in this embodiment is specifically as follows: based on the formed settlement observation data, the settlement displacement sequence of each pixel or base station is extracted according to the spatial location to construct a set of global settlement time series curves; at the same time, historical settlement records are imported, and the sequences of missing years are filled in by time interpolation to obtain continuous settlement time series curves.

[0031] The following is a feasible example of encoding and assigning risk area values ​​according to risk level in this embodiment: low risk area is assigned a value of 0, medium risk area is assigned a value of 1, and high risk area is assigned a value of 2, thus forming a subsidence risk area mask.

[0032] The following is a feasible example of how this embodiment adaptively adjusts the mesh density based on the risk level of the risk area mask: a smaller mesh size is used in high-risk areas to improve spatial resolution, a medium mesh size is used in medium-risk areas, and a larger mesh size is used in low-risk areas to reduce computational burden; the meshing results are then merged to generate a framework that covers the entire area.

[0033] In this embodiment, the least squares regression fitting involves fitting a trend curve for each settlement time series curve, using time as the independent variable and settlement displacement as the dependent variable, employing the least squares method to minimize the sum of squared residuals from this curve to each observation point. The average settlement rate at that point is obtained from the slope of the fitted curve, and the settlement acceleration component is obtained from the second-order trend of the curve. This method can filter out random noise, making the estimates of rate and acceleration more stable.

[0034] In this embodiment, rate anomalies and acceleration abrupt changes are identified by filtering out observation points that significantly deviate from the global average or quantile threshold from the average settlement rate and acceleration component distributions across all observation points. For example, points with rates much higher than the overall average rate are marked as rate anomalies, and points exhibiting significant positive or negative acceleration jumps are marked as acceleration abrupt changes. These points typically represent potential risks such as ground subsidence and accelerated ground settlement, and are key inputs for risk area identification.

[0035] The initial settlement risk point set in this embodiment consists of all observation points marked as having anomalies in rate or abrupt changes in acceleration. This set is sparsely distributed, spatially covering locations where settlement risk may occur. It serves as input data for subsequent spatial clustering, used to identify continuous risk areas from discrete high-risk points.

[0036] The density-based clustering algorithm in this embodiment takes the spatial coordinates of the observation points as input and, based on the neighborhood density of local points, groups high-risk points that are spatially close and densely packed into the same cluster, while discarding isolated points as noise. Feasible implementations of this type of algorithm include DBSCAN or OPTICS, which can identify risk region clusters of arbitrary shapes, avoiding boundary misjudgments that may occur when using only a simple circular buffer.

[0037] In this embodiment, the risk regions are spatially continuous clusters of high-risk points identified by a clustering algorithm, forming region boundaries in the form of minimum bounding polygons or convex hulls. Each risk region corresponds to a local subsidence active area, where the subsidence rate and acceleration characteristics are significantly higher than the surrounding background area, making it the focus of subsequent mesh subdivision.

[0038] In this embodiment, the rasterized mask layer uses the polygonal boundaries of the risk area as a reference to establish a regular grid within the entire monitoring area. Each grid cell corresponds to a spatial pixel or grid node. A mask value is assigned to the grid cell based on whether its center point falls within the risk area and the corresponding risk level. This layer serves as the direct basis for subsequent adaptive grid division and monitoring weight allocation.

[0039] The mesh framework in this embodiment is a global mesh system obtained after completing adaptive mesh generation based on a risk mask. It includes all mesh nodes and cells covering the monitoring area. A node is the center point or corner point of the mesh, representing the spatial location for settlement field calculation or interpolation; a cell is a polygonal region composed of adjacent nodes, representing the spatial unit of calculation. The mesh framework includes both large-size meshes in low-risk areas and subdivided meshes in high-risk areas, serving as a unified spatial benchmark for subsequent multi-source settlement field projection, driving factor mapping, and deep learning modeling.

[0040] S3 projects the settlement observation data onto the grid based on Kriging interpolation to generate a multi-source settlement field, including a first settlement field and a second settlement field.

[0041] In this embodiment, the process of projecting settlement observation data onto a grid based on kriging interpolation specifically involves: Within the grid framework, a grid node index is established, and the spatial coordinates of each node and its corresponding risk level are recorded. The spatial trend surface of the overall time series is estimated based on linear regression with driving factors, and the trend field is obtained. Detrending the original observations based on the trend field yields residual observations at each time step, including InSAR residual observations and GNSS residual observations. The first settlement field was obtained by interpolating the InSAR residual observations to the grid using co-kriging.

[0042] The second settlement field was obtained by interpolating the GNSS residual observations to the grid based on partitioned kriging.

[0043] It should be noted that the grid node index refers to assigning a unique number to the geometric center or vertex of each grid cell within the grid framework, and recording the two-dimensional spatial coordinates of that node in the monitoring area and its corresponding risk level. The significance of this is that all subsequent settlement field calculation results can be precisely correlated to specific spatial locations and risk categories through the node index, ensuring a one-to-one correspondence between interpolation results and risk classifications.

[0044] In this embodiment, the spatial trend surface refers to a smoothed spatial field estimated by linear regression with driving factors using the overall time series of subsidence observations. This field is used to characterize the overall trend of subsidence changes on a large scale. The trend surface can reflect long-term background subsidence, such as regional geological subsidence or overall stratigraphic compression, providing a benchmark for subsequent identification of localized anomalous subsidence.

[0045] The detrending process in this embodiment refers to subtracting the trend surface value at the corresponding location from the original settlement observation value, thereby obtaining residual observations with more obvious local changes. The detrended residual data can eliminate the influence of regional background settlement, making subsequent spatial interpolation more prominent for local anomalies, thus improving the sensitivity of risk identification.

[0046] The residual observations in this embodiment include InSAR residual observations and GNSS residual observations, which refer to the results of detrended settlement observation data obtained using Synthetic Aperture Radar Interferometry (InSAR) and Global Navigation Satellite System (GNSS), respectively. InSAR residual observations typically have a large coverage area and high spatial resolution, but slightly lower accuracy per point; GNSS residual observations have sparser points but higher accuracy, and the combination of the two can complement each other.

[0047] The co-kriging interpolation in this embodiment uses InSAR residual observations as the main variable and introduces settlement-related driving factors (such as groundwater level and geological structure) as covariates. Co-interpolation is achieved through a spatial covariance model, thereby estimating the settlement value of each node on the grid. This integrates multi-source information and improves the fit of the interpolation results to the actual settlement distribution.

[0048] The first settlement field in this embodiment refers to the gridded settlement distribution obtained by co-kriging interpolation from InSAR residual observations at the grid nodes, with each node storing a time-series settlement residual value. This settlement field focuses on characterizing settlement patterns over a large spatial area and is suitable for discovering widely distributed settlement risk areas.

[0049] The partitioned kriging interpolation in this embodiment refers to dividing the monitoring area into several spatial sub-regions based on the risk area mask, and performing kriging interpolation on the GNSS residual observations in each sub-region to obtain a settlement estimate with higher local accuracy. Partial interpolation can avoid the error diffusion caused by cross-regional interpolation, making the interpolation results in high-risk areas more consistent with reality.

[0050] The second settlement field in this embodiment refers to the settlement distribution obtained by partitioned kriging interpolation from GNSS residual observations at the grid nodes. Each node also corresponds to a settlement time series. The results of this settlement field typically have higher point accuracy and can be used to verify and correct anomalous areas in the first settlement field, thereby improving the reliability of the overall settlement risk assessment.

[0051] S4 maps the settlement driving data onto the grid through spatial overlay to generate a driving factor grid field.

[0052] In this embodiment, the step of mapping the settlement driving data onto a grid through spatial overlay to generate a driving factor grid field specifically involves: The settlement-driven data is classified to obtain data categories, which include point time series, raster data, and vector polygons. Settlement-driven data for point-to-surface interpolation to a grid are used; Settlement-driven data for raster data categories is based on resampling plus aligned interpolation to the grid; Settlement-driven data for vector polygon categories are interpolated to the grid based on area-weighted assignment; For each data category, interpolation mapping is performed on the settlement driving data to the grid to obtain the driving factor grid field.

[0053] The settlement-driving data in this embodiment is a collection of data on external or internal factors that can explain or influence the occurrence and evolution of ground settlement, including groundwater level monitoring, extraction volume, soil layer thickness, building load, seismic activity, and precipitation. These data typically come from diverse sources and are presented in various formats, serving as crucial inputs for constructing settlement causal analysis and risk prediction models.

[0054] In this embodiment, the data categories are based on the spatial form and temporal attributes of the settlement-driven data. Common categories include point time series, raster data, and vector polygons. Point time series data describes the temporal changes of local points, such as well water level monitoring sequences; raster data describes continuous spatial distributions, such as precipitation rasters; and vector polygon data describes the characteristics of discrete regions, such as land use zones or geological fault zones. This classification allows for the selection of appropriate spatial mapping methods, ensuring consistency in the representation of different data on a unified grid.

[0055] The point-to-surface interpolation in this embodiment distributes discrete point-based driving data to each node or cell of the grid using a spatial interpolation algorithm, enabling each grid node to obtain a corresponding driving factor value. Essentially, it infers the representative values ​​of nodes based on the spatial distance or proximity between points and nodes, thereby expanding point-based information into a continuous spatial field.

[0056] The resampling and alignment interpolation in this embodiment first adjusts the spatial resolution of the driving factor data, which already exists in raster form, to match the grid, and then spatially aligns the raster nodes to ensure that each grid node can accurately match a driving factor value. This process ensures that raster data of different resolutions can be seamlessly integrated within the same spatial framework.

[0057] The area-weighted allocation interpolation in this embodiment calculates the overlapping area between the polygon and the grid cells when the driving factor is represented by a polygon, and then allocates the polygon attributes to the grid nodes according to the area ratio. This method can reflect the spatial influence range of the polygon driving factor on the grid, making the gridded factor values ​​representative.

[0058] The driving factor grid field in this embodiment, after interpolation mapping, contains a set of values ​​for one or more driving factors at each node of the grid, forming a multidimensional spatial data field. This grid field can unify the originally scattered multi-source driving data into the same spatial reference system, providing an input basis for subsequent quantitative modeling of settlement risk and multi-factor correlation analysis.

[0059] S5 performs spatial pyramid and temporal multi-scale decomposition on the multi-source settlement field and the driving factor grid field to generate several spatiotemporally resolved sub-bands.

[0060] In this embodiment, the spatial pyramid and temporal multi-scale decomposition of the multi-source settlement field and the driving factor grid field to generate several spatiotemporally resolved sub-bands is specifically as follows: Based on the spatial resolution of the grid, a preset first-level spatial pyramid resolution layer is set, with each layer having a resolution that is half that of the previous layer. The first settlement field, the second settlement field, and the driving factor grid field are downsampled to each resolution layer one by one to obtain copies of the settlement field and driving factor field at different spatial resolutions, and the mapping index is retained. The time series of the grid at each spatial resolution layer is decomposed into multiple scales based on wavelet transform to generate multi-source settlement field and driving factor grid field with multiple time scale components. The resolution layers of the spatial pyramid are combined with time components to form a first number of spatiotemporally resolved sub-bands, where the first number is the product of the first level and the number of multiple time scale components.

[0061] The following is an arithmetic expression for a feasible spatial pyramid resolution layer setting in this embodiment:

[0062] In the formula, Let y be the resolution of the spatial pyramid at level y. For the original grid space resolution, This is the first level number.

[0063] The following is an arithmetic expression for a feasible spatial pyramid resolution layer setting in this embodiment: ; ; ; In the formula This represents the first settlement field after spatial downsampling at the y-th layer. As the first settlement area, This represents the second settlement field after spatial downsampling of the y-th layer. As the second settlement area, Let y be the driving factor grid field after spatial downsampling at the y-th layer. For the driving factor grid field, For resolution Perform spatial downsampling operation.

[0064] This embodiment uses wavelet transform for multi-scale decomposition, specifically performing wavelet multi-scale decomposition on the time series of each spatial resolution grid. Let grid point (x, z) have observations in the time series. Here, k represents the first and second subsidence field numbers, and t represents the time step. The following is a specific example of the arithmetic expression: ; ; In the formula, Let y be the settlement residual of the sub-band at the j-th time scale. The timescale sub-band corresponding to the driving factor. The number of layers in the time-scale decomposition. This is a wavelet decomposition operation.

[0065] The specific arithmetic expression for combining the spatial pyramid resolution layer with the temporal scale sub-band in this embodiment to generate a spatiotemporally resolved sub-band is shown below: ; In the formula, There are Y*J spatiotemporally resolved subbands.

[0066] It should be noted that the spatial pyramid is a hierarchical representation method that divides the original grid into multiple levels according to spatial resolution. The resolution of each level is typically half that of the level above, thus forming a spatial scale system from coarse to fine. The inherent function of this method is to allow the settlement field and driving factor field to be observed and analyzed at different spatial scales, thereby capturing both existing regional overall trends and identifying local high-gradient small-scale settlement characteristics.

[0067] In this embodiment, the spatial resolution layer represents the mesh fineness corresponding to each level of the spatial pyramid, with different numbers of nodes and mesh cell sizes in each layer. Higher resolution layers can more accurately depict local details but require more computation; lower resolution layers can capture large-scale overall deformation trends with less computational burden. By combining multiple levels, both overall and local settlement characteristics can be analyzed.

[0068] In this embodiment, downsampling maps high-resolution settlement fields and driving factor fields to a low-resolution layer through dimensionality reduction, thereby reducing the number of nodes and aggregating local features. Downsampling not only compresses the amount of data but also smooths out local noise, enabling the low-resolution layer to better reflect macroscopic trends and ensuring consistent spatial coverage across all layers of the spatial pyramid.

[0069] The mapping index in this embodiment records the correspondence between high-resolution nodes and their corresponding low-resolution units during the downsampling process. This index relationship ensures that information can be transferred or inverted between different resolution layers, such as redistributing the prediction results of the low-resolution layer to the high-resolution layer to achieve cross-scale result fusion.

[0070] The multi-scale decomposition of wavelet transform in this embodiment decomposes the settlement time series of each grid node into components at different time scales, such as long-term trend, seasonal components, and short-term fluctuation components. Its inherent function is to separate information at different time frequencies, enabling the model to identify the differences between long-term slow settlement and short-term abnormal settlement, thereby improving the accuracy and interpretability of subsequent predictions.

[0071] The multi-timescale components in this embodiment are multiple time-frequency components obtained through wavelet decomposition, each representing a settlement change pattern at an independent time scale. For example, low-frequency components describe stable settlement over multi-year timescales, while high-frequency components capture sudden abnormal settlement or seasonal fluctuations. By combining and analyzing different components, the driving mechanisms of settlement can be studied in more detail.

[0072] In this embodiment, the spatiotemporally resolved subbands are decomposition units formed by combining the spatial pyramid resolution layer with the Cartesian product of the temporal scale component. Each subband has both a spatial scale and a temporal scale. The function of the subbands is to divide the subsidence field and driving factor field into multiple independent spatiotemporal blocks. Each block focuses on describing the subsidence characteristics at a specific scale, enabling subsequent models to perform targeted modeling and prediction of changes at different scales.

[0073] In this embodiment, by performing spatial pyramid and temporal multi-scale decomposition on the multi-source subsidence field and the driving factor grid field, the originally inconsistent spatial resolution and complex temporal dynamics of subsidence information are decomposed into sub-bands of multiple levels and multiple time scales, effectively alleviating the feature aliasing problem caused by the high heterogeneity of spatial scale in geological subsidence monitoring. This method preserves the overall trend of large-scale subsidence while highlighting the detailed features of local high-risk areas, enabling subsequent Bayesian fusion and deep learning prediction to simultaneously perceive global and local information at multiple resolutions, improving the reconstruction accuracy of the subsidence field and its sensitivity to abnormal subsidence. Therefore, it truly expands and optimizes the existing approach of analogous remote sensing multi-resolution image fusion, specifically solving the problem of large spatial scale differences in geological subsidence fields and the difficulty in directly unifying modeling different observation sources.

[0074] S6. Based on the multi-resolution weighting strategy of sub-bands, Bayesian fusion is performed on the first and second settlement fields to obtain the fused settlement field.

[0075] In this embodiment, the sub-band-based multi-resolution weighting strategy performs Bayesian fusion on the first and second settlement fields to obtain a fused settlement field, specifically as follows: Establish a sub-band observation error model and output the sub-band observation error variance; Based on Bayesian update and subband confidence weights, the spatiotemporally resolved subbands of the first and second settlement fields are fused, and multi-resolution reconstruction is performed to obtain the fused settlement field. The subband confidence weights are obtained by taking the reciprocal of the subband observation error variance and normalizing it.

[0076] The following are calculation examples of feasible sub-band observation error models in this embodiment: ; ; In the formula, Let Variance be the observation error variance of the sub-zone of the first settlement field. Let j be the number of grid points contained in the j-th sub-band. Let be the settlement value at the i-th point of subzone j in the first settlement field. This represents the average settlement value of the first sub-zone of the settlement field. Let Variance be the observation error variance of the sub-zone of the second settlement field. Let be the settlement value at the i-th point of subzone j in the second settlement field. This represents the average settlement value of the second settlement field sub-zone.

[0077] The arithmetic expression for fusing the spatiotemporally resolved subbands of the first and second subsidence fields based on Bayesian update and subband confidence weights in this embodiment is exemplified below: ; In the formula, Let i be the settlement value of the sub-zone j at the i-th point after merging. and denoted as the confidence weight of subzone j for the first and second settlement fields.

[0078] In this embodiment, the fusion results of each sub-band are reconstructed at multiple resolutions to obtain the arithmetic representation of the final fused settlement field. An example is shown below: ; In the formula, In order to integrate the settlement field, The multi-resolution reconstruction operator synthesizes the sub-band level results into a settling field at the original resolution.

[0079] The sub-band observation error model in this embodiment establishes a statistical model describing the error distribution of the interpolation results for the first and second settlement fields in each spatiotemporally resolved sub-band. This model quantifies the degree of deviation between observed and actual settlement values. Its intrinsic function is to estimate the error magnitude of each sub-band, helping subsequent Bayesian fusion to achieve differentiated weighting across different sub-bands. This ensures that sub-bands with smaller errors contribute more to the fusion result, while sub-bands with larger errors contribute less, thereby improving the overall fusion accuracy.

[0080] The observation error variance in this embodiment is a key indicator output by the sub-band observation error model, used to measure the degree of fluctuation of the observation residuals in that sub-band. The smaller the observation error variance, the more reliable the interpolation results of that sub-band; the larger the variance, the higher the uncertainty of that sub-band, and its weight needs to be reduced during fusion to avoid introducing too much noise.

[0081] In this embodiment, the sub-band confidence weights are a set of weight values ​​obtained by taking the reciprocal of the observation error variance and normalizing it. These weights represent the contribution of each sub-band to the final fused settlement field. The weight allocation follows the principle that sub-bands with higher confidence contribute more and sub-bands with lower confidence contribute less, thereby dynamically balancing the reliability and coverage of data in multi-source information fusion.

[0082] The Bayesian update in this embodiment establishes a conditional probability relationship between prior information and observational information, and uses observational data to correct the prior settlement estimate to obtain the posterior estimate. In this step, the prior estimate is usually derived from the first settlement field, and the observational likelihood is derived from the second settlement field. The Bayesian update obtains a settlement value that better reflects the actual geological conditions, allowing the fusion result to take into account the advantages of both types of data sources.

[0083] The multi-resolution weighting strategy in this embodiment applies sub-band confidence weights to the weighted average for different resolution layers and time components of the spatial pyramid, ensuring that sedimentation information from different resolution layers can participate in the fusion in a coordinated manner. This strategy guarantees that the overall trend of low-resolution layers and the local details of high-resolution layers are both taken into account in the final result, avoiding the problem of single-scale bias dominating.

[0084] The multi-resolution reconstruction in this embodiment involves reassembling the sub-band results at different spatial resolutions and time scales according to the pyramid index and time series order after completing the Bayesian fusion of each sub-band, restoring the complete settlement field at the original spatial resolution. This reconstruction process ensures that the final fused settlement field seamlessly covers the entire study area and possesses multi-scale feature representation capabilities, providing consistent input data for subsequent deep learning predictions.

[0085] In this embodiment, the multi-source subsidence field and the driving factor grid field are decomposed into spatiotemporal sub-bands with different spatial resolutions and time scales. While this can alleviate the problem of high spatial heterogeneity in geological subsidence, it also introduces new problems—the accuracy and reliability of observation data from different sub-bands are inconsistent. Direct fusion may lead to noise from low-precision sub-bands affecting the overall subsidence reconstruction effect.

[0086] This step establishes a sub-band observation error model, calculates the observation error variance of each sub-band, and uses its reciprocal as a confidence weight for Bayesian fusion. This effectively realizes the dominant role of high-precision sub-bands in the fusion result, weakens the influence of low-precision sub-bands, and preserves the integrity of multi-resolution information. The fused settlement field obtained through multi-resolution reconstruction maintains the detailed settlement characteristics of local high-resolution areas while also taking into account global low-resolution trends, thereby improving the accuracy and reliability of the settlement field reconstruction.

[0087] In short, this step provides a solution to the problem of low-precision sub-bands affecting the overall fusion quality during the fusion of analog remote sensing multi-resolution images. It achieves optimal fusion of multi-source, multi-resolution settlement information, ensuring that the input for subsequent deep learning predictions has higher credibility and accuracy.

[0088] S7 integrates the settlement field and the driving factor grid field into the deep learning model at each scale sub-band, trains the multi-scale prediction sub-model, and reconstructs the global continuous settlement prediction field through cross-scale consistency constraints.

[0089] In this embodiment, the step of jointly inputting the fused settlement field and the driving factor grid field into the deep learning model at each scale sub-band to train the multi-scale prediction sub-model specifically involves: The fused settlement field is split in reverse according to the defined spatiotemporal resolution sub-bands to obtain multi-scale input features; The node values ​​of the driving factor grid field on the corresponding sub-band are extracted synchronously from the multi-scale input features to form the driving factor multi-scale feature matrix. Constructing deep learning models based on spatial-temporal hybrid neural networks; The deep learning model includes a network input layer and intermediate layers; Each sub-band resolution is independently encoded at the network input layer to extract local spatial features and temporal trend features; A cross-scale feature interaction module is introduced in the intermediate layer to fuse high-resolution sub-band features and low-resolution sub-band features through a feature pyramid to achieve multi-scale perception. The deep learning model is trained based on the multi-scale feature matrix of driving factors to obtain a multi-scale prediction sub-model.

[0090] In this embodiment, the reconstruction of the global continuous settlement prediction field through cross-scale consistency constraints specifically involves: A cross-scale loss function is introduced during training to ensure that the prediction results of different resolution sub-bands remain consistent after upsampling and downsampling. The prediction results of the multi-scale prediction sub-model are reconstructed to the original grid resolution one by one, and the boundary is smoothed by multi-scale deconvolution reconstruction method. The prediction results of the multi-scale prediction sub-models are fused to generate a global future settlement prediction field and output the confidence distribution.

[0091] The multi-scale input features in this embodiment involve re-splitting the fused settlement field according to predefined spatiotemporal resolution sub-bands, forming multi-level, multi-time-scale feature sets. Each feature set corresponds to a combination of spatial resolution layer and temporal components, which can respectively characterize large-scale slow settlement trends and local rapid settlement anomalies, thereby providing deep learning models with multi-angle input information, enabling them to perceive settlement dynamics at different scales.

[0092] The multi-scale feature matrix of driving factors in this embodiment is formed by extracting the driving factor values ​​corresponding to the spatial location and temporal resolution of each sub-zone from the driving factor grid field and combining them according to the sub-zone hierarchy to form a matrix input. This matrix not only preserves the spatial distribution characteristics of driving factors such as geology, hydrology, and load, but also synchronously aligns with the time series, enabling deep learning models to model the correlation between settlement evolution and external driving forces, thereby improving the causal interpretability and generalization ability of predictions.

[0093] The spatial-temporal hybrid neural network in this embodiment is a deep learning network structure that simultaneously possesses the ability to extract spatial features and model temporal series. In the spatial dimension, convolutional or graph neural networks are typically used to capture the spatial relationships between grid nodes, while in the temporal dimension, recurrent units or attention mechanisms are used to describe the temporal dependence of settlement, thereby ensuring that the network output can simultaneously reflect spatial morphology and temporal evolution patterns.

[0094] In this embodiment, the network input layer uses independent encoding, which encodes features separately for each sub-band resolution during the model's input stage, converting the original sedimentation values ​​and driving factor values ​​into high-dimensional feature vectors. This avoids interference between features of different resolutions in the early stages, allowing the local details of high-resolution sub-bands and the overall trends of low-resolution sub-bands to be fully learned separately.

[0095] The cross-scale feature interaction module in this embodiment is a feature fusion unit designed in the middle layer of the model, used to transfer and fuse feature information between sub-bands of different resolutions. By integrating high-resolution and low-resolution features step by step through a feature pyramid approach, information is converged from bottom to top and details are enhanced from top to bottom, thereby enabling the model to simultaneously grasp global trends and identify local anomalies.

[0096] The multi-scale prediction sub-model in this embodiment is a prediction sub-network trained separately for each resolution layer. Each sub-network focuses on predicting settlement changes at its corresponding scale and can output the settlement distribution field at that resolution. The outputs of all sub-models are jointly reconstructed into a continuous settlement prediction field across the entire domain through consistency constraints, which not only ensures the consistency of prediction results at each scale, but also avoids the deviations that may be caused by single-scale prediction.

[0097] The cross-scale consistency constraint in this embodiment is a regularization strategy introduced during training to force the prediction results of different resolution layers to remain consistent in overlapping areas. By restricting high-resolution predictions from approaching low-resolution predictions after downsampling, a smooth transition across scales is achieved. The final output global prediction field can capture local details without losing the global trend, meeting the continuity and stability requirements of geological subsidence monitoring.

[0098] S8 dynamically updates the risk mask and grid division based on the predicted field to achieve iterative optimization of geological subsidence monitoring.

[0099] In this embodiment, the step of dynamically updating the risk mask and grid division based on the predicted field to achieve iterative optimization of geological subsidence monitoring specifically involves: Obtain the settlement value distribution and confidence level distribution of the global future settlement prediction field; Based on the areas where the predicted settlement value exceeds the threshold, a settlement risk zone mask is regenerated, and the mesh is locally refined or merged. The updated risk mask and grid are used as the basis for the next round of settlement observation data projection and driving factor mapping to achieve adaptive grid optimization in key risk areas. After several iterations, the final converged subsidence risk zone division and continuous subsidence prediction field are output, enabling real-time dynamic monitoring and early warning of geological subsidence.

[0100] The subsidence risk zone classification in this embodiment is based on existing technologies and has been widely studied and applied in geological disaster monitoring and urban infrastructure subsidence monitoring. Methods such as thresholding, statistical modeling, and machine learning are commonly used for classifying subsidence risk zones. For example, areas with subsidence rates exceeding a certain threshold are marked as high-risk areas, a well-established practice.

[0101] The continuous settlement prediction field in this embodiment has existing applications in water conservancy, subway settlement monitoring, and surface deformation research, where continuous settlement fields (or future prediction fields) are generated using methods such as Kriging interpolation, Bayesian updates, and deep learning. It is a spatially continuous prediction result used to assist in risk assessment, and is a common output, not a novel point.

[0102] This embodiment achieves continuous subsidence prediction across the entire domain by performing deep learning modeling on multi-scale sub-bands using multi-source subsidence data and conducting joint training with driving factors. It also introduces cross-scale consistency constraints. Compared with the shortcomings of existing technologies, such as low prediction accuracy of single-scale or single-source data and difficulty in taking into account high and low resolution spatial information, this embodiment effectively alleviates the prediction errors caused by spatial heterogeneity and nonlinear evolution of subsidence. Furthermore, through iterative optimization based on dynamically updating the risk mask and grid division of the prediction field, it achieves adaptive refinement and risk monitoring of key subsidence areas. Compared with existing static grid methods, it solves the problem of insufficient real-time dynamic monitoring and early warning capabilities, thus providing an effective technical means for high-precision, continuous, and iterative monitoring of geological subsidence.

[0103] Example 2, Figure 2 This invention presents a geological subsidence monitoring system integrating deep learning and multi-source data, comprising a multi-source subsidence data acquisition module, an adaptive grid partitioning module, a grid projection module, a fused subsidence field module, a prediction field module, and a monitoring module. The multi-source subsidence data acquisition module collects subsidence observation data and subsidence driving data. The adaptive grid partitioning module generates a subsidence risk zone mask based on the subsidence observation data and performs adaptive grid partitioning. The grid projection module projects the subsidence observation data and subsidence driving data onto the grid, generating a multi-source subsidence field and a driving factor grid field. The fused subsidence field module performs spatiotemporal multi-scale decomposition of the multi-source subsidence field and the driving factor grid field to obtain spatiotemporally resolved sub-bands, and performs Bayesian fusion of the multi-source subsidence field based on a multi-resolution weighting strategy for the sub-bands to obtain a fused subsidence field. The prediction field module reconstructs a global prediction field from the fused subsidence field and the driving factor grid field using a deep learning model. The monitoring module dynamically updates the risk mask and grid partitioning based on the prediction field.

[0104] In the embodiments provided by this invention, it should be understood that the disclosed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0105] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0106] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.

[0107] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0108] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0109] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0110] In the embodiments provided in this disclosure, it should be understood that the disclosed systems and methods can also be implemented in other ways. The system embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0111] In this disclosure, the terms "comprising," "including," or any other variations thereof are intended to cover a 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 limitation, an element limited by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0112] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A geological subsidence monitoring method integrating deep learning and multi-source data, characterized in that, Includes the following steps: Collect settlement observation data and settlement driving data; A subsidence risk zone mask is generated based on subsidence observation data and adaptive mesh generation is performed. The settlement observation data and settlement driving data are projected onto the grid to generate a multi-source settlement field and a driving factor grid field. Spatiotemporal multi-scale decomposition of the multi-source settlement field and the driving factor grid field is performed to obtain spatiotemporally resolved sub-bands. Based on the multi-resolution weighting strategy of the sub-bands, Bayesian fusion of the multi-source settlement field is performed to obtain the fused settlement field. The global prediction field is reconstructed by fusing the settlement field and the driving factor grid field through a deep learning model. The risk mask and grid division are dynamically updated based on the prediction field.

2. The geological subsidence monitoring method integrating deep learning and multi-source data according to claim 1, characterized in that, The process of generating a subsidence risk zone mask based on subsidence observation data and performing adaptive mesh generation includes the construction of the mesh framework, and the specific steps are as follows: Construct settlement time-series curves and analyze their temporal characteristics. The time-series characteristics include settling rate and acceleration components; Potential risk points are identified based on temporal features, and risk area masks are generated through spatial clustering. The mesh density is adaptively adjusted based on the risk level of the risk area mask to generate a mesh framework covering the entire area.

3. The geological subsidence monitoring method integrating deep learning and multi-source data according to claim 2, characterized in that, The step of projecting settlement observation data and settlement driving data onto a grid to generate a multi-source settlement field and a driving factor grid field includes: A subsidence risk zone mask is generated based on subsidence observation data, and the mesh is adaptively divided. The settlement observation data is projected onto the grid based on interpolation to generate a multi-source settlement field that includes the first settlement field and the second settlement field; Settlement-driven data are mapped onto a grid to generate a driving factor grid field.

4. The geological subsidence monitoring method integrating deep learning and multi-source data according to claim 3, characterized in that, The process of projecting settlement observation data onto a grid based on interpolation to generate a multi-source settlement field containing a first settlement field and a second settlement field specifically involves: Establish a grid node index to record the spatial coordinates of each node and its corresponding risk level; The trend field is obtained by estimating the spatial trend surface based on linear regression with driving factors; Detrending the original observations based on the trend field yields InSAR residual observations and GNSS residual observations. The first settlement field is obtained by interpolating the InSAR residual observations to the grid through co-kriging. The second settlement field was obtained by interpolating the GNSS residual observations to the grid based on partitioned kriging.

5. The geological subsidence monitoring method integrating deep learning and multi-source data according to claim 4, characterized in that, The specific steps for generating the driving factor grid field are as follows: The settlement-driven data is classified to obtain data categories, which include point time series, raster data, and vector polygons. Based on the data category, the corresponding spatial interpolation method is used to map the settlement driving data of each category to a unified grid, and the driving factor grid field is generated by class-by-class mapping and integration.

6. The geological subsidence monitoring method integrating deep learning and multi-source data according to claim 5, characterized in that, The spatiotemporal multi-scale decomposition of the multi-source settlement field and the driving factor grid field yields spatiotemporally resolved sub-bands, including: Construct a spatial pyramid resolution layer to downsample the multi-source settlement field and driving factor field layer by layer; The temporal data of each spatial resolution layer is decomposed into multiple scales to generate multiple time scale components. The multi-scale decomposition of the temporal data is performed using the wavelet transform method. By combining spatial resolution layers with temporal scale components, multiple spatiotemporally resolved sub-bands are formed.

7. The geological subsidence monitoring method integrating deep learning and multi-source data according to claim 6, characterized in that, The specific steps for obtaining the fused settlement field are as follows: Establish a sub-band observation error model and output the sub-band observation error variance; The confidence weights of each sub-band are determined based on the observed error variance. A Bayesian update method is used to fuse the spatiotemporally resolved sub-bands of the multi-source settlement field in combination with the confidence weights, and the fused settlement field is obtained through multi-resolution reconstruction.

8. The geological subsidence monitoring method integrating deep learning and multi-source data according to claim 7, characterized in that, The process of reconstructing the global prediction field by fusing the settlement field and the driving factor grid field through a deep learning model includes the construction of a multi-scale prediction model, specifically: The fused settlement field is split into multi-scale input features, and the driving factor features of the corresponding scales are extracted to form a multi-scale feature matrix. A spatial-temporal hybrid neural network is used to construct a multi-scale prediction model by independently encoding the features of each sub-band and introducing cross-scale feature interactions. The multi-scale prediction model is trained based on the multi-scale feature matrix.

9. The geological subsidence monitoring method integrating deep learning and multi-source data according to claim 8, characterized in that, The method of reconstructing the global prediction field by fusing the settlement field and the driving factor grid field through a deep learning model also includes: A cross-scale loss function is introduced during training to ensure that the prediction results of different resolution sub-bands remain consistent after upsampling and downsampling. The prediction results at each scale are reconstructed to the original grid resolution, and boundary smoothing is achieved through multi-scale deconvolution. The prediction results at various scales are integrated to generate a global settlement prediction field and output the confidence distribution.

10. A system for monitoring geological subsidence using the method for integrating deep learning and multi-source data as described in any one of claims 1-9, characterized in that, It includes a multi-source settlement data acquisition module, an adaptive grid division module, a grid projection module, a fused settlement field module, a prediction field module, and a monitoring module; The multi-source settlement data acquisition module is used to collect settlement observation data and settlement driving data; The adaptive mesh generation module is used to generate a subsidence risk zone mask based on subsidence observation data and perform adaptive mesh generation. The grid projection module is used to project settlement observation data and settlement driving data onto the grid respectively to generate a multi-source settlement field and a driving factor grid field. The fusion settlement field module is used to perform spatiotemporal multi-scale decomposition of multi-source settlement fields and driving factor grid fields to obtain spatiotemporally resolved sub-bands, and to perform Bayesian fusion of multi-source settlement fields based on the multi-resolution weighting strategy of sub-bands to obtain the fusion settlement field. The prediction field module is used to reconstruct the global prediction field by fusing the settlement field and the driving factor grid field through a deep learning model; The monitoring module is used to dynamically update the risk mask and grid division based on the prediction field.

Images