Geological subsidence monitoring method and system fusing deep learning and multi-source data

Abstract

The application discloses a geological subsidence monitoring method and system fusing deep learning and multi-source data, relates to the technical field of geological subsidence, and comprises the following steps: collecting multi-source subsidence observation and driving data, generating a subsidence risk area mask based on historical records and adaptively dividing a grid; generating a multi-source subsidence field through Kriging interpolation and mapping a driving factor grid field; performing spatial pyramid and time multi-scale decomposition on the subsidence field and the driving factor to obtain a time-space subband; performing Bayesian fusion based on a subband confidence weight to obtain a fused subsidence field; inputting the fused field and the driving factor into a deep learning model to train a multi-scale prediction submodel, and reconstructing a global continuous subsidence prediction field through cross-scale consistency; and dynamically optimizing a risk mask and a grid according to a prediction result to realize iterative monitoring. The application solves the problems of multi-source data fusion difficulty, spatial scale heterogeneity and difficulty in accurately predicting the subsidence of a local high-risk area in geological subsidence monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geological subsidence, more particularly, the present application relates to a geological subsidence monitoring method and system fusing deep learning and multi-source data. BACKGROUND

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

[0003] However, the geological subsidence field usually shows a highly heterogeneous spatial scale, that is, the local high-risk area subsides rapidly and changes dramatically, while the surrounding area subsides relatively gently. The existing methods model and interpolate at a uniform spatial resolution, which can easily lead to smoothing or submerging of local abnormal subsidence characteristics, and cannot accurately reflect the local details of subsidence. At the same time, the data fusion of different observation sources also has the problem of scale mismatch, which affects the monitoring accuracy.

[0004] In the field of remote sensing image processing, methods such as spatial pyramid and multi-resolution wavelet decomposition have been proposed to solve the problem of multi-resolution image fusion. By decomposing image data into subbands of different spatial resolutions and time scales, local details can be highlighted while preserving global trends, effectively alleviating the information aliasing problem caused by inconsistent scales of original data, and providing a reference for joint modeling of multi-source data.

[0005] However, directly applying analog methods to the process of geological subsidence monitoring will introduce new problems. The subbands formed by multi-scale decomposition of the subsidence field and the driving factor grid field have differences in observation accuracy and reliability. High-precision subbands can accurately reflect the subsidence characteristics, while low-precision subbands may contain noise. If all subbands are directly fused, the noise in low-precision subbands may interfere with the overall subsidence reconstruction, affecting the local and global prediction accuracy, and causing instability of the monitoring results.

[0006] In summary, the existing geological subsidence monitoring technology cannot simultaneously consider global subsidence trend and local high-risk characteristics. On the other hand, in the process of multi-source data fusion, scale differences and accuracy differences can easily lead to information aliasing and reconstruction errors. These problems restrict the high-precision reconstruction of subsidence field, the identification of abnormal areas, and the reliable prediction of future subsidence. SUMMARY

[0007] In order to overcome the above-mentioned defects of the prior art, embodiments of the present application provide a geological subsidence monitoring method and system fusing deep learning and multi-source data, which solves the problems of multi-source data fusion difficulty, spatial scale heterogeneity and local high-risk area subsidence difficult to accurately predict in geological subsidence monitoring by carrying out multi-resolution spatio-temporal decomposition and Bayesian fusion on multi-source subsidence data, and combining deep learning multi-scale prediction and dynamic grid optimization.

[0008] To achieve the above object, the present application provides the following technical scheme:

[0009] The geological subsidence monitoring method fusing 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 area mask based on the subsidence observation data and historical records, and adaptively dividing a grid according to the mask; projecting the subsidence observation data to the grid based on Kriging interpolation to generate a multi-source subsidence field, including a first subsidence field and a second subsidence field; mapping the subsidence driving data to the grid through spatial superposition to generate a driving factor grid field; carrying out spatial pyramid and time multi-scale decomposition on the multi-source subsidence field and the driving factor grid field to generate a plurality of time and space resolution subbands; carrying out Bayesian fusion on the first subsidence field and the second subsidence field based on a multi-resolution weighting strategy of the subbands to obtain a fused subsidence field; inputting the fused subsidence field and the driving factor grid field into a deep learning model on each scale subband to train multi-scale prediction submodels, and reconstructing a global continuous subsidence prediction field through cross-scale consistency constraint; dynamically updating the risk mask and the grid division according to the prediction field to realize iterative optimization of geological subsidence monitoring.

[0010] In a preferred embodiment, the step of generating a subsidence risk area mask based on the subsidence observation data and historical records, and adaptively dividing a grid according to the mask specifically comprises: constructing a subsidence time series curve; carrying out least squares regression fitting on the subsidence time series curve to calculate the average subsidence rate and the acceleration component of each observation point; marking the rate abnormal value and the acceleration mutation value as potential risk points to form a preliminary subsidence risk point set; applying a density-based clustering algorithm to the preliminary risk point set to aggregate adjacent high-risk points into a risk area; generating a rasterized mask layer in the spatial range of the monitoring area with the risk area as the polygon boundary, and encoding and assigning values to the risk area according to the risk level; adaptively adjusting the grid division density according to the risk level of the risk area mask, and merging the division results to generate a grid framework covering the whole area.

[0011] In a preferred implementation, the projecting the subsidence observation data to the grid based on Kriging interpolation comprises: under the framework of the grid, establishing a grid node index and recording the spatial coordinates and corresponding risk levels of each node; estimating a spatial trend surface based on linear regression with driving factors for the overall time series to obtain a trend field; detrending the original observation based on the trend field to obtain residual observation at each time, the residual observation including InSAR residual observation and GNSS residual observation; projecting the InSAR residual observation to the grid based on collaborative Kriging interpolation to obtain a first subsidence field; and projecting the GNSS residual observation to the grid based on partition Kriging interpolation to obtain a second subsidence field.

[0012] In a preferred implementation, the mapping the subsidence driving data to the grid by spatial superposition to generate a driving factor grid field comprises: classifying the subsidence driving data to obtain data categories, the data categories including point time series, raster data and vector polygon; projecting the subsidence driving data of the point time series category to the grid based on point-to-surface interpolation; projecting the subsidence driving data of the raster data category to the grid based on resampling and alignment interpolation; projecting the subsidence driving data of the vector polygon category to the grid based on area-weighted allocation interpolation; and performing difference mapping of the subsidence driving data to the grid one by one for each data category to obtain the driving factor grid field.

[0013] In a preferred implementation, the spatial pyramid and time multiscale decomposition of the multi-source subsidence field and the driving factor grid field generates a plurality of time-space resolution sub-bands, and the spatial pyramid decomposition comprises: setting a preset first number of spatial pyramid resolution layers based on the spatial resolution of the grid, and each resolution layer is half of the previous layer; downsampling the first subsidence field, the second subsidence field and the driving factor grid field to each resolution layer to obtain copies of the subsidence field and the driving factor field at different spatial resolutions, and retaining mapping indexes; performing multiscale decomposition based on wavelet transform on the grid time series of each spatial resolution layer to generate multi-time scale components of the multi-source subsidence field and the driving factor grid field; and combining the resolution layers of the spatial pyramid and the time components to form a first number of time-space resolution sub-bands, the first number being the product of the first number and the number of multi-time scale components.

[0014] In a preferred implementation, the Bayesian fusion of the first subsidence field and the second subsidence field based on the sub-band multi-resolution weighting strategy generates a fused subsidence field, and the Bayesian fusion comprises: establishing a sub-band observation error model to output a sub-band observation error variance; and fusing the time-space resolution sub-bands of the first subsidence field and the second subsidence field based on Bayesian update and sub-band confidence weight, and performing multi-resolution reconstruction to obtain the fused subsidence field, the sub-band confidence weight being obtained by taking the inverse of the sub-band observation error variance and normalizing.

[0015] In a preferred embodiment, the fusion precipitation field and the driving factor grid field are combined on each scale sub-band to input a deep learning model, and a multi-scale prediction sub-model is trained, specifically: the fusion precipitation field is inversely split according to a defined space-time resolution sub-band to obtain multi-scale input features; the multi-scale input features are used to synchronously extract node values of the driving factor grid field on the corresponding sub-band to form a driving factor multi-scale feature matrix; a deep learning model is constructed based on a space-time hybrid neural network; the deep learning model includes a network input layer and an intermediate layer; in the network input layer, each sub-band resolution is independently coded to extract local spatial features and temporal trend features; in the intermediate layer, a cross-scale feature interaction module is introduced to fuse high-resolution sub-band features and low-resolution sub-band features through a feature pyramid to realize multi-scale perception; the deep learning model is trained based on the driving factor multi-scale feature matrix to obtain the multi-scale prediction sub-model.

[0016] In a preferred embodiment, the global continuous precipitation prediction field is reconstructed through cross-scale consistency constraints, specifically: a cross-scale loss function is introduced in the training process to constrain the prediction results of different resolution sub-bands to remain consistent after upsampling and downsampling; the prediction results of the multi-scale prediction sub-model are reconstructed one by one to the original grid resolution, and a multi-scale deconvolution reconstruction method is used to smooth the boundaries; the prediction results of the multi-scale prediction sub-model are fused to generate a global future precipitation prediction field, and a confidence distribution is output.

[0017] In a preferred embodiment, the risk mask and grid division are dynamically updated according to the prediction field to realize iterative optimization of geological settlement monitoring, specifically: the settlement value distribution and the confidence distribution of the global future precipitation prediction field are obtained; a settlement risk area mask is regenerated based on the settlement value threshold area of the prediction field, and the grid is locally refined or merged; the updated risk mask and grid are used as the basis for projection and driving factor mapping of the next round of settlement observation data to realize adaptive optimization of the grid in the risk key area; after several cycles of iterative updating, the final converged settlement risk area division and continuous settlement prediction field are output to realize real-time dynamic monitoring and early warning of geological settlement.

[0018] The geological subsidence monitoring system fusing deep learning and multi-source data comprises a multi-source subsidence data acquisition module, an adaptive grid division 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 is used for collecting subsidence observation data and subsidence driving data; the adaptive grid division module is used for generating a subsidence risk area mask based on the subsidence observation data and performing adaptive grid division; the grid projection module is used for projecting the subsidence observation data and the subsidence driving data to the grid respectively to generate a multi-source subsidence field and a driving factor grid field; the fused subsidence field module is used for performing time-space multi-scale decomposition on the multi-source subsidence field and the driving factor grid field to obtain a time-space resolution subband, and performing Bayesian fusion on the multi-source subsidence field based on a multi-resolution weighting strategy of the subband to obtain a fused subsidence field; the prediction field module is used for reconstructing a global prediction field through a deep learning model from the fused subsidence field and the driving factor grid field; and the monitoring module is used for dynamically updating a risk mask and grid division according to the prediction field.

[0019] The technical effects and advantages of the geological subsidence monitoring method and system fusing deep learning and multi-source data are as follows:

[0020] 1. The spatial pyramid and time multi-scale decomposition are performed on the multi-source subsidence field and the driving factor grid field, the originally inconsistent spatial resolution and complex time dynamics of the subsidence information are decomposed into a plurality of hierarchical and time scale subbands, the feature aliasing problem caused by the high spatial scale heterogeneity in the geological subsidence monitoring is effectively relieved, the overall trend of large-scale subsidence is retained, and the detail features of the local high-risk area are highlighted, so that the subsequent analysis can simultaneously perceive the global and local information under the multi-resolution, the subsidence field reconstruction accuracy and the sensitivity to abnormal subsidence are improved, and the problem that different observation sources are difficult to be directly and uniformly modeled in the prior art is solved.

[0021] 2. The multi-scale subbands of the multi-source subsidence field and the driving factor grid field differ in accuracy and reliability, and if directly fused, the noise of the low-precision subband may affect the overall reconstruction effect; the subband observation error model is established, and the reciprocal of the error variance is taken as the confidence weight for Bayesian fusion, so that the high-precision subband dominates the fusion result, the influence of the low-precision subband is weakened, and the integrity of the multi-resolution information is retained. The fused subsidence field obtained through the multi-resolution reconstruction not only maintains the local high-resolution features, but also takes into account the global trend, thereby improving the reconstruction accuracy and reliability of the subsidence field and providing a high-credibility input for subsequent multi-scale deep learning prediction. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 It is a flowchart of the geological subsidence monitoring method fusing deep learning and multi-source data.

[0023] Figure 2A structural schematic diagram of a geological settlement monitoring system fusing deep learning and multi-source data according to the present application. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative effort belong to the protection scope of the present application.

[0025] Embodiment 1, Figure 1 The geological settlement monitoring method fusing deep learning and multi-source data according to the present application is given, including the following steps:

[0026] S1, collecting multi-source settlement data, including settlement observation data and settlement driving data.

[0027] In this embodiment, the multi-source settlement data, including settlement observation data and settlement driving data, is specifically:

[0028] The settlement observation data includes InSAR time series settlement observation data and GNSS continuous observation data;

[0029] InSAR time series settlement observation data and GNSS continuous observation data are obtained;

[0030] The InSAR time series settlement observation data and GNSS settlement sequence data are time-synchronized based on timestamps;

[0031] Based on the nearest neighbor, the missing observation time is filled in, and geographic registration is performed according to a unified spatial coordinate system to form settlement observation data;

[0032] Driving factor data is obtained from a weather monitoring system, a geological survey system and a groundwater monitoring system, each driving factor data is spatio-temporally registered, and is converted to the same spatial reference system and time step as the settlement observation data set to form settlement driving factor data.

[0033] In this embodiment, the InSAR time series settlement observation data is obtained by collecting multi-orbit SAR image data covering the target area, performing differential interference processing on each orbit image sequence to generate a pixel-level displacement sequence under the same time reference, and removing atmospheric delay and orbit error through time series analysis method 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.

[0034] The GNSS continuous observation data of the embodiment is obtained, specifically: GNSS continuous observation stations are laid out, three-dimensional coordinate data of a target area is collected in real time, a precise point positioning and baseline solution algorithm is used to calculate a displacement sequence in a subsidence direction, and abnormal jumps are filtered and smoothed to generate GNSS subsidence sequence data; the GNSS subsidence sequence data includes vertical displacement values of each station at each observation time and measurement error covariances.

[0035] The driving factor data of the embodiment includes rainfall, groundwater level change, soil thickness, stratum structure, underground mining activity record and the like.

[0036] S2, a subsidence risk area mask is generated based on subsidence observation data and historical records, and a grid is adaptively divided based on the mask.

[0037] In the embodiment, the subsidence risk area mask is generated based on subsidence observation data and historical records, and a grid is adaptively divided based on the mask, specifically:

[0038] A subsidence time sequence curve is constructed.

[0039] Least square regression fitting is performed on the subsidence time sequence curve, and average subsidence rates and acceleration components of each observation point are calculated.

[0040] Rate abnormal values and acceleration mutation values are marked as potential risk points to form a preliminary subsidence risk point set.

[0041] A density-based clustering algorithm is applied to the preliminary risk point set to aggregate adjacent high-risk points into risk areas.

[0042] A raster mask layer is generated in the spatial range of the monitoring area with the risk areas as polygon boundaries, and mask values are encoded and assigned to the risk areas according to risk levels.

[0043] The grid division density is adaptively adjusted according to the risk levels of the risk area mask, and the division results are combined to generate a grid framework covering the whole area.

[0044] The subsidence time sequence curve of the embodiment is constructed, specifically: based on the formed subsidence observation data, the subsidence displacement sequences of each pixel or station point are extracted according to the spatial position, a global subsidence time sequence curve set is constructed, and historical subsidence records are imported to fill in the missing sequences through time interpolation to obtain continuous subsidence time sequence curves.

[0045] The mask values are encoded and assigned to the risk areas according to risk levels, for example: low-risk areas are assigned a value of 0, medium-risk areas are assigned a value of 1, and high-risk areas are assigned a value of 2 to form a subsidence risk area mask.

[0046] The feasible implementation of the risk level adaptive adjustment of grid division density according to the risk area mask of the embodiment is as follows: a small grid size is used in a high-risk area to improve spatial resolution, a medium grid size is used in a medium-risk area, and a large grid size is used in a low-risk area to reduce the calculation burden; and the division results are combined to generate a framework of grids covering the whole area.

[0047] The least square regression fitting of the embodiment is to, for each settlement time series curve, use the least square method to fit a trend curve with time as the independent variable and settlement displacement as the dependent variable, so that the residual sum of squares of the curve to each observation point is minimized. The average settlement rate of the point is obtained by the slope of the fitted curve, and the settlement acceleration component is obtained by the second-order change trend of the curve. This method can filter out random noise, making the estimation of the rate and acceleration more stable.

[0048] The rate abnormal value and acceleration mutation value of the embodiment are observation points that are obviously deviated from the global average or quantile threshold in the distribution of average settlement rate and acceleration component of all observation points. For example, points with a rate much higher than the overall average rate are marked as rate abnormalities, and points with a significant positive jump or negative jump in acceleration are marked as acceleration mutation points. Such points usually represent potential ground subsidence, ground settlement acceleration, and other risks, and are key inputs for risk area identification.

[0049] The preliminary settlement risk point set of the embodiment is a set composed of all observation points marked as rate abnormalities or acceleration mutations. The set is a sparse point set that spatially covers possible locations where settlement risks may occur. It is the input data for subsequent spatial clustering, used to identify continuous risk areas from discrete high-risk points.

[0050] The density-based clustering algorithm of the embodiment takes the spatial coordinates of observation points as input and, according to the neighborhood density of local points, classifies high-risk points that are spatially adjacent and densely distributed 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 area clusters of any shape and avoid boundary misjudgments that may be caused by using simple circular buffers.

[0051] The risk area of the embodiment is a spatially continuous high-risk point cluster identified by the clustering algorithm, with the minimum enclosing polygon or convex hull forming the area boundary. Each risk area corresponds to a local settlement active area, and the settlement rate and acceleration characteristics in the area are significantly higher than those in the surrounding background area, which is the focus of subsequent grid subdivision.

[0052] The gridded mask layer of the embodiment is referenced to the polygon boundary of the risk area, a regular grid is established in the entire monitoring area, each grid unit corresponds to a spatial pixel or a grid node, and a mask value is assigned according to whether the center point of the grid falls within the risk area and the corresponding risk level. The layer is a direct basis for subsequent adaptive grid division and monitoring weight allocation.

[0053] The grid framework of the embodiment is a global grid system obtained after completing adaptive grid division based on the risk mask, including all grid nodes and units covering the monitoring area. The node refers to the center point or corner point of the grid, representing the spatial position of the deposition field calculation or interpolation; the unit is a polygon area composed of adjacent nodes, representing the spatial unit of calculation. The grid framework contains both large-size grids in low-risk areas and subdivided grids in high-risk areas, and is a unified spatial reference for subsequent multi-source deposition field projection, driving factor mapping and deep learning modeling.

[0054] S3, projecting the deposition observation data to the grid based on Kriging interpolation to generate a multi-source deposition field, including a first deposition field and a second deposition field.

[0055] In the embodiment, the projecting the deposition observation data to the grid based on Kriging interpolation is specifically:

[0056] Under the framework of the grid, a grid node index is established and the spatial coordinates and corresponding risk level of each node are recorded;

[0057] Based on linear regression with driving factors, the spatial trend surface of the overall time series is estimated to obtain a trend field;

[0058] Based on the trend field, the original observation is detrended to obtain residual observation at each time, the residual observation including InSAR residual observation and GNSS residual observation;

[0059] The InSAR residual observation is interpolated to the grid based on collaborative Kriging to obtain the first deposition field.

[0060] The GNSS residual observation is interpolated to the grid based on partitioned Kriging to obtain the second deposition field.

[0061] It should be noted that the grid node index refers to establishing a unique number for the geometric center or vertex of each grid unit in the grid framework, and recording the two-dimensional spatial coordinates of the node in the monitoring area and the risk level it is in. The significance of this is that all subsequent deposition field calculation results can be accurately associated to specific spatial positions and risk categories through the node index, so that the interpolation results and risk classification can be one-to-one corresponding.

[0062] The spatial trend surface of the embodiment refers to a smooth spatial field estimated by linear regression with driving factors, using the overall time series of subsidence observations, to depict the overall trend of subsidence at a large scale. The trend surface can reflect the long-term background subsidence, such as regional geological subsidence or overall stratum compression, and provide a benchmark for subsequent identification of local abnormal subsidence.

[0063] The detrending process of the embodiment refers to subtracting the trend surface value at the corresponding position from the original subsidence observation value, thereby obtaining residual observations with more obvious local changes. The residual data after detrending can eliminate the influence of regional background subsidence, so that subsequent spatial interpolation highlights local anomalies, thereby improving the sensitivity of risk identification.

[0064] The residual observation of the embodiment includes InSAR residual observation and GNSS residual observation, which respectively refer to the results obtained by detrending the subsidence observation data acquired by synthetic aperture radar interferometry (InSAR) and global navigation satellite system (GNSS). InSAR residual observation generally has large coverage and high spatial resolution, but the single-point precision is slightly lower; GNSS residual observation has sparse points but high precision, and the combination of the two can complement each other.

[0065] The collaborative kriging interpolation of the embodiment refers to using InSAR residual observation as the main variable, while introducing driving factors related to subsidence (such as groundwater level, stratum structure, etc.) as covariates, to realize collaborative interpolation through a spatial covariance model, thereby estimating the subsidence value of each node on the grid. This can integrate multi-source information and improve the fitting degree of the interpolation result to the actual subsidence distribution.

[0066] The first subsidence field of the embodiment refers to the gridded subsidence distribution obtained by collaborative kriging interpolation of InSAR residual observation at grid nodes, with each node storing a time series of subsidence residual values. This subsidence field focuses on depicting the subsidence pattern in a large range of space, and is suitable for discovering subsidence risk areas with wide distribution.

[0067] The partitioned kriging interpolation of the embodiment refers to dividing the monitoring area into several spatial sub-areas based on the risk area mask, and performing kriging interpolation on GNSS residual observations in each sub-area, thereby obtaining subsidence estimates with higher local precision. Partitioned interpolation can avoid error propagation caused by cross-regional interpolation, making the interpolation results of high-risk areas more consistent with reality.

[0068] The second subsidence field of the embodiment refers to the subsidence distribution obtained by partitioned kriging interpolation of GNSS residual observations at grid nodes, with each node also corresponding to a subsidence time series. The results of this subsidence field generally have higher point precision and can be used to check and correct abnormal areas of the first subsidence field, thereby improving the reliability of overall subsidence risk assessment.

[0069] S4, mapping the subsidence driving data to the grid through spatial superposition to generate a driving factor grid field.

[0070] In the embodiment, the step of mapping the subsidence driving data to the grid through spatial superposition to generate a driving factor grid field comprises:

[0071] The subsidence driving data is classified to obtain data categories, including point time series, raster data, and vector polygons.

[0072] The subsidence driving data of the point time series category is interpolated to the grid based on point-to-surface interpolation.

[0073] The subsidence driving data of the raster data category is interpolated to the grid based on resampling and alignment.

[0074] The subsidence driving data of the vector polygon category is interpolated to the grid based on area-weighted distribution.

[0075] The subsidence driving data is mapped to the grid by difference one by one to obtain a driving factor grid field.

[0076] The subsidence driving data of the embodiment is a data set of external or internal factors that can explain or affect the occurrence and evolution of land subsidence, including underground water level monitoring, pumping volume, soil thickness, building load, seismic activity, precipitation, etc. These data are usually diverse in source and different in format, and are important inputs for building subsidence cause analysis and risk prediction models.

[0077] The data categories of the embodiment are classified according to the spatial form and time attribute of the subsidence driving data, and common categories include point time series, raster data, and vector polygons. Point time series data describe the time variation of local points, such as water well monitoring sequences; raster data describe continuous spatial distribution, such as precipitation raster; and vector polygon data describe discrete regional characteristics, such as land use zoning or geological fault zones. Through classification, appropriate spatial mapping methods can be selected to ensure the consistency of expression of different data on a unified grid.

[0078] The point-to-surface interpolation of the embodiment is to distribute discrete point driving data to each node or element of the grid through spatial interpolation algorithms, so that each grid node can obtain a corresponding driving factor value. The essence is to infer the representative value of the node according to the spatial distance or proximity relationship between the point and the node, so as to extend the point information to a continuous spatial field.

[0079] The resampling and alignment interpolation of the embodiment is performed on the driving factor data originally in the form of a grid. The spatial resolution of the driving factor data is first adjusted to be consistent with the grid, and then the grid points are spatially aligned to ensure that each grid node can accurately match a driving factor value. This process ensures that grid data of different resolutions can be seamlessly integrated in the same spatial framework.

[0080] The area-weighted distribution interpolation of the embodiment is performed when the driving factor is represented by a polygon. The polygon attribute is weighted and distributed to the grid nodes according to the area ratio by calculating the overlapping area of the polygon and the grid cell. This method can reflect the spatial influence range of the polygon driving factor on the grid, so that the grid-based factor value is representative.

[0081] The driving factor grid field of the embodiment is a set of values of one or more driving factors at each node of the grid after interpolation mapping, forming a multi-dimensional spatial data field. The grid field can unify the originally dispersed 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.

[0082] S5, spatial pyramid and time multi-scale decomposition are performed on the multi-source settlement field and the driving factor grid field to generate a plurality of time and space resolution sub-bands.

[0083] In the embodiment, the spatial pyramid and time multi-scale decomposition are performed on the multi-source settlement field and the driving factor grid field to generate a plurality of time and space resolution sub-bands, and specifically:

[0084] Based on the spatial resolution of the grid, a preset first number of spatial pyramid resolution layers is set, and each layer has a resolution that is half of the previous layer;

[0085] The first settlement field, the second settlement field, and the driving factor grid field are down-sampled to each resolution layer to obtain copies of the settlement field and the driving factor field with different spatial resolutions, and the mapping index is retained;

[0086] The grid time series of each spatial resolution layer is multi-scale decomposed based on wavelet transform to generate multi-time scale components of the multi-source settlement field and the driving factor grid field;

[0087] The resolution layers of the spatial pyramid and the time components are combined to form a first number of time and space resolution sub-bands, and the first number is the product of the first number of levels and the number of multi-time scale components.

[0088] The following is an arithmetic expression for the feasible spatial pyramid resolution layer setting of the embodiment:

[0089]

[0090] In the formula, is the spatial pyramid resolution of the yth layer, is the original grid spatial resolution, is the first level number.

[0091] The following is an arithmetic expression of the feasible spatial pyramid resolution layer setting of the embodiment:

[0092]

[0093]

[0094]

[0095] In the formula, is the first settling field after spatial down-sampling of the yth layer, is the first settling field, is the second settling field after spatial down-sampling of the yth layer, is the second settling field, is the driving factor grid field after spatial down-sampling of the yth layer, is the driving factor grid field, is the spatial down-sampling operation performed according to the resolution .

[0096] The embodiment performs multi-scale decomposition based on wavelet transform, specifically, wavelet multi-scale decomposition is performed on the grid time sequence of each layer of spatial resolution. It is assumed that the grid point (x, z) has an observation value in the time sequence, k represents the settling field number first and second, and t represents the time step. The following is a specific arithmetic example:

[0097]

[0098]

[0099] In the formula, is the settling residual of the yth layer and the jth time scale sub-band, is the time scale sub-band corresponding to the driving factor, is the number of layers of time multi-scale decomposition, is the wavelet decomposition operation.

[0100] The embodiment combines the spatial pyramid resolution layer and the time scale sub-band to generate a specific arithmetic example of the space-time resolution sub-band as follows:

[0101]

[0102] In the formula, is the space-time resolution sub-band, and the total number of sub-bands is Y*J. ​​​​​​

[0103] It is to be explained 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 usually half of the previous level, forming a spatial scale system from coarse to fine. The internal function of this method is to enable the settlement field and the driving factor field to be observed and analyzed at different spatial scales, so as to capture both regional overall trends and small-range settlement characteristics with local high gradients.

[0104] The spatial resolution layer of the embodiment is the grid fineness corresponding to each level of the spatial pyramid. The node number and grid cell size of each layer are different. Higher resolution layers can more accurately depict local details, but the calculation amount is larger; lower resolution layers can capture large-scale overall deformation trends, and the calculation burden is smaller. Through the combination of multiple levels, the overall and local settlement feature analysis can be considered.

[0105] The down-sampling of the embodiment is to map the high-resolution settlement field and the driving factor field to the low-resolution layer through dimension reduction to reduce the number of nodes and aggregate local features. Down-sampling not only compresses the data amount, but also smooths out local noise, so that the low-resolution layer can better reflect the macroscopic change trend, and ensure that the data between the layers of the spatial pyramid have consistent spatial coverage.

[0106] The mapping index of the embodiment is the corresponding relationship between the high-resolution node and its corresponding low-resolution unit recorded in the down-sampling process. This index relationship ensures that information transmission or inversion recovery can be performed between different resolution layers in the future, such as redistributing the prediction results of the low-resolution layer to the high-resolution layer to realize cross-scale result fusion.

[0107] The multi-scale decomposition of wavelet transform of the embodiment is a signal decomposition of the settlement time series of each grid node, which decomposes the original time series into components of different time scales, such as long-term trend, seasonal component and short-term fluctuation component. Its internal function is to separate information of different time frequencies, so that the model can identify the difference between long-term slow settlement and short-term abnormal settlement, and improve the accuracy and interpretability of subsequent prediction.

[0108] The multi-time scale component of the embodiment is a plurality of time frequency components obtained by wavelet decomposition, each component representing a settlement change pattern at an independent time scale. For example, the low-frequency component describes multi-year scale stable settlement, and the high-frequency component captures sudden abnormal settlement or seasonal fluctuations. Through the combination analysis of different components, the driving mechanism of settlement can be studied in more detail.

[0109] The spatio-temporal resolution subband of the embodiment is a decomposition unit formed by the Cartesian product combination of the spatial pyramid resolution layer and the time scale component, each subband having a spatial scale and a time scale. The role of the subband is to divide the deposition field and the driving factor field into multiple independent spatio-temporal blocks, each block describing the deposition characteristics at a specific scale, so that the subsequent model can model and predict changes at different scales.

[0110] In the embodiment, by performing spatial pyramid and time multi-scale decomposition on the multi-source deposition field and the driving factor grid field, the deposition information originally having inconsistent spatial resolution and complex time dynamics is decomposed into multiple levels and multiple time scale subbands, effectively alleviating the feature aliasing problem caused by high spatial scale heterogeneity in geological subsidence monitoring. This method not only retains the overall trend of large-scale subsidence, but also highlights the detailed features of local high-risk areas, so that the subsequent Bayesian fusion and deep learning prediction can simultaneously perceive global and local information at multiple resolutions, improving the reconstruction accuracy of the deposition field and the sensitivity to abnormal subsidence. Therefore, it indeed expands and optimizes the idea of analog remote sensing multi-resolution image fusion in the prior art, and specifically solves the problem of large spatial scale difference of geological deposition field and difficulty in directly unified modeling of different observation sources.

[0111] S6, Bayesian fusion of the first deposition field and the second deposition field based on a multi-resolution weighting strategy of the subband, to obtain a fused deposition field.

[0112] In the embodiment, the Bayesian fusion of the first deposition field and the second deposition field based on the multi-resolution weighting strategy of the subband to obtain the fused deposition field is specifically:

[0113] A subband observation error model is established to output a subband observation error variance;

[0114] The spatio-temporal resolution subbands of the first deposition field and the second deposition field are fused based on Bayesian update and subband confidence weight, and multi-resolution reconstruction is performed to obtain a fused deposition field, wherein the subband confidence weight is obtained by taking the inverse of the subband observation error variance and normalizing.

[0115] The following is a calculation example of a feasible subband observation error model of the embodiment:

[0116] ;

[0117] ;

[0118] In the formula, is the observation error variance of the subband of the first deposition field, is the number of grid points contained in the jth subband, is the deposition value of the first deposition field at the i th point of the jth subband. is the average deposition value of the sub-band of the first deposition field, is the observation error variance of the sub-band of the second deposition field, is the deposition value of the i-th point in the j-th sub-band of the second deposition field, is the average deposition value of the sub-band of the second deposition field.

[0119] The arithmetic expression of the Bayesian update and sub-band confidence weight-based fusion of the spatio-temporal resolution sub-bands of the first deposition field and the second deposition field in the embodiment is as follows:

[0120] ;

[0121] In the formula, is the deposition value of the i-th point in the j-th sub-band of the fused deposition field, and is the confidence weight of the first deposition field and the second deposition field in the j-th sub-band.

[0122] The arithmetic expression of the multi-resolution reconstruction of the fusion results of each sub-band to obtain the final fused deposition field in the embodiment is as follows:

[0123] ;

[0124] In the formula, is the fused deposition field, is a multi-resolution reconstruction operator that integrates the results of the sub-band level into a deposition field of the original resolution.

[0125] The sub-band observation error model in the embodiment is a statistical model that describes the error distribution of the interpolation results of the first deposition field and the second deposition field on each spatio-temporal resolution sub-band, and is used to quantify the deviation between the observation value and the true deposition value. Its internal role is to help the subsequent Bayesian fusion to realize differentiated weighting on different sub-bands by estimating the error size of each sub-band, so that the sub-band with smaller error contributes more in the fusion result, and the sub-band with larger error contributes less, thereby improving the overall fusion precision.

[0126] The observation error variance in the embodiment is a key indicator output by the sub-band observation error model, which is used to measure the fluctuation degree of the observation residual in the sub-band. The smaller the observation error variance, the more reliable the interpolation result of the sub-band; the larger the variance, the higher the uncertainty of the sub-band, which needs to reduce its weight in fusion to avoid introducing too much noise.

[0127] The sub-band confidence weight of the embodiment is a set of weight values obtained by taking the inverse of the observation error variance and normalizing, which is used to express the contribution degree of each sub-band to the final fused subsidence field. The distribution of weights follows the principle that the sub-band with high confidence contributes more and the sub-band with low confidence contributes less, thereby dynamically balancing the reliability and coverage of data in multi-source information fusion.

[0128] The Bayesian update of the embodiment is a process of establishing a conditional probability relationship between prior information and observation information, using observation data to correct the prior subsidence estimate, and obtaining the posterior estimate. In this step, the prior estimate usually comes from the first subsidence field, and the observation likelihood comes from the second subsidence field. Through Bayesian update, the subsidence value that conforms to the real geological state is obtained, so that the fusion result takes into account the advantages of both types of data sources.

[0129] The multi-resolution weighting strategy of the embodiment is to apply sub-band confidence weight to weighted average for different resolution layers and time components of the spatial pyramid, so that the subsidence information of different resolution layers can participate in the fusion in a coordinated manner. This strategy ensures that the overall trend of the low-resolution layer and the local details of the high-resolution layer can be considered in the final result, avoiding the problem of deviation dominated by a single scale.

[0130] The multi-resolution reconstruction of the embodiment is to reassemble the sub-band results under different spatial resolutions and time scales according to the pyramid index and time sequence order after completing the Bayesian fusion of each sub-band, and restore the complete subsidence field under the original spatial resolution. The process of reconstruction ensures that the final fused subsidence field can seamlessly cover the entire study area and has multi-scale feature expression capability, providing consistent input data for subsequent deep learning prediction.

[0131] In the embodiment, the multi-source subsidence field and the driving factor grid field are decomposed into time-space sub-bands of different spatial resolutions and time scales. Although this can alleviate the problem of high spatial scale heterogeneity of geological subsidence, it may introduce new problems, i.e., the observation data accuracy and reliability of different sub-bands are inconsistent, and direct fusion may lead to noise of low-precision sub-bands affecting the overall subsidence reconstruction effect.

[0132] This step calculates the observation error variance of each sub-band by establishing a sub-band observation error model, and uses its inverse as the confidence weight for Bayesian fusion, effectively realizing the dominant role of high-precision sub-bands in the fusion result and weakening the influence of low-precision sub-bands, while preserving the integrity of multi-resolution information. Through multi-resolution reconstruction, the fused subsidence field is obtained, which can maintain the fine subsidence characteristics of local high-resolution areas and take into account the global low-resolution trend, thereby improving the reconstruction accuracy and reliability of the subsidence field.

[0133] In short, this step provides a solution to the problem of low-precision sub-band affecting the overall fusion quality in the process of analog remote sensing multi-resolution image fusion, realizes the optimal fusion of multi-source and multi-resolution deposition information, and ensures that the input of subsequent deep learning prediction has higher reliability and accuracy.

[0134] S7, jointly input the fused deposition field and the driving factor grid field on each scale sub-band into a deep learning model, train a multi-scale prediction sub-model, and reconstruct a global continuous deposition prediction field through cross-scale consistency constraints.

[0135] In this embodiment, the joint input of the fused deposition field and the driving factor grid field on each scale sub-band into the deep learning model to train the multi-scale prediction sub-model is specifically:

[0136] The fused deposition field is inversely split according to the defined space-time resolution sub-band to obtain multi-scale input features;

[0137] The nodes of the driving factor grid field on the corresponding sub-band are extracted synchronously to form a driving factor multi-scale feature matrix;

[0138] A deep learning model is constructed based on a space-time hybrid neural network;

[0139] The deep learning model includes a network input layer and an intermediate layer;

[0140] The network input layer independently encodes each sub-band resolution to extract local spatial features and temporal trend features;

[0141] 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 realize multi-scale perception;

[0142] The deep learning model is trained based on the driving factor multi-scale feature matrix to obtain a multi-scale prediction sub-model.

[0143] In this embodiment, the reconstruction of the global continuous deposition prediction field through cross-scale consistency constraints is specifically:

[0144] A cross-scale loss function is introduced in the training process to constrain the prediction results of different resolution sub-bands to remain consistent after upsampling and downsampling;

[0145] The prediction results of the multi-scale prediction sub-model are reconstructed to the original grid resolution one by one, and a multi-scale deconvolution reconstruction method is used to smooth the boundaries;

[0146] The prediction results of the multi-scale prediction sub-model are fused to generate a global future deposition prediction field, and a confidence distribution is output.

[0147] The multiscale input feature of the embodiment is to re-split the fused settlement field according to the defined spatiotemporal resolution subband, forming a feature set of multiple levels and multiple time scales. Each feature set corresponds to a combination of a spatial resolution layer and a time component, and can represent slow settlement trends on a large scale and local rapid settlement anomalies, respectively, thereby providing multi-angle input information for the deep learning model to perceive settlement dynamics at different scales.

[0148] The driving factor multiscale feature matrix of the embodiment is to extract driving factor values corresponding to each subband spatial position and time resolution from the driving factor grid field, and form a matrix input by combining subband levels. The matrix not only retains the spatial distribution characteristics of geological, hydrological, and load driving factors, but also synchronously aligns the time series, so that the deep learning model can correlate and model the settlement evolution and external driving force, and improve the causal explanation and generalization ability of the prediction.

[0149] The space-time hybrid neural network of the embodiment is a deep learning network structure that simultaneously has spatial feature extraction and time series modeling capabilities. In the spatial dimension, convolution or graph neural networks are usually used to capture the spatial relationship between grid nodes, and in the time dimension, recurrent units or attention mechanisms are used to describe the time sequence dependence of settlement, so that the network output can reflect both spatial morphology and temporal evolution rules.

[0150] The network input layer independent coding of the embodiment is to encode features for each subband resolution separately at the input stage of the model, converting the original settlement values and driving factor values into high-dimensional feature vectors. This can avoid interference between features of different resolutions in the early stage, so that local details of high-resolution subbands and overall trends of low-resolution subbands can be learned separately.

[0151] The cross-scale feature interaction module of the embodiment is a feature fusion unit designed in the middle layer of the model, which is used to transfer and fuse feature information between different resolution subbands. By gradually integrating high-resolution and low-resolution features in a feature pyramid manner, information is aggregated from bottom to top and details are enhanced from top to bottom, so that the model has both global trend grasping ability and local anomaly recognition ability.

[0152] The multiscale prediction submodel of the embodiment is a prediction subnetwork trained for each resolution layer, and each subnetwork focuses on settlement change prediction at its corresponding scale, and can output a settlement distribution field at that resolution. The outputs of all submodels are jointly reconstructed into a global continuous settlement prediction field through consistency constraints, which not only ensures the coordination and consistency of the prediction results at each scale, but also avoids the bias that may be caused by single-scale prediction.

[0153] The cross-scale consistency constraint of the embodiment is a regularization strategy introduced in the training process, which is used to force the prediction results of different resolution layers to remain coordinated in the overlapping area. By limiting the high-resolution prediction to be close to the low-resolution prediction after downsampling, a smooth transition across scales is achieved, and the final global prediction field can capture local details without losing global trends, meeting the continuity and stability requirements of geological subsidence monitoring.

[0154] S8, dynamically updating the risk mask and grid division according to the prediction field, realizing iterative optimization of geological subsidence monitoring.

[0155] In the embodiment, the dynamic updating of the risk mask and grid division according to the prediction field realizes iterative optimization of geological subsidence monitoring, specifically:

[0156] Obtain the settlement value distribution and confidence distribution of the global future settlement prediction field;

[0157] Based on the settlement value threshold area of the prediction field, the settlement risk area mask is regenerated, and the grid is locally refined or merged;

[0158] The updated risk mask and grid are used as the basis for the next round of settlement observation data projection and driving factor mapping, realizing adaptive optimization of the grid in the risk key area;

[0159] After several cycles of iterative updating, the final converged settlement risk area division and continuous settlement prediction field are output, realizing real-time dynamic monitoring and early warning of geological subsidence.

[0160] The settlement risk area division of the embodiment has many research and application (existing technology common achievements) on the grading division of the settlement risk area based on threshold, statistical modeling, machine learning, etc. in geological disaster monitoring and urban infrastructure subsidence monitoring. For example: the area with settlement rate exceeding a certain threshold is marked as a high-risk area, which is a mature method.

[0161] The continuous settlement prediction field of the embodiment has application scenarios of generating continuous settlement field (or future prediction field) based on Kriging interpolation, Bayesian update, deep learning, etc. in water conservancy, subway settlement monitoring and surface deformation research. It is a spatially continuous prediction result used to assist risk judgment and also a common output, not a novel point.

[0162] The embodiment realizes global continuous settlement prediction by deep learning modeling of multi-source settlement data on multi-scale subbands, combined with joint training of driving factors, and introduction of cross-scale consistency constraints, effectively alleviates the prediction error caused by settlement spatial heterogeneity and nonlinear evolution, compared with the low prediction accuracy of single-scale or single-source data in the prior art, and the difficulty of taking into account high and low resolution spatial information, and simultaneously realizes adaptive refinement and risk monitoring of settlement key areas through iterative optimization based on dynamic updating of the prediction field risk mask and grid division, compared with the existing static grid method, solves the problem of insufficient real-time dynamic monitoring and early warning capability, thereby providing an effective technical means for high-precision, continuous and iterative monitoring of geological settlement.

[0163] Embodiment 2, Figure 2 The geological settlement monitoring system fusing deep learning and multi-source data provided by the present application comprises 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 grid division module is used to generate a settlement risk area mask based on the settlement observation data and perform adaptive grid division. The grid projection module is used to project the settlement observation data and the settlement driving data to the grid respectively to generate a multi-source settlement field and a driving factor grid field. The fused settlement field module is used to perform temporal and spatial multi-scale decomposition on the multi-source settlement field and the driving factor grid field to obtain a time-space resolution subband, and perform Bayesian fusion on the multi-source settlement field based on a multi-resolution weighting strategy of the subband to obtain a fused settlement field. The prediction field module is used to reconstruct a global prediction field by a deep learning model from the fused settlement field and the driving factor grid field. The monitoring module is used to dynamically update the risk mask and the grid division according to the prediction field.

[0164] In the embodiments provided by the present application, it should be understood that the disclosed system can be implemented in other manners. For example, the system embodiments described above are merely schematic; for example, the division of the modules is only a logical function division; and there can be another division manner in actual implementation.

[0165] In addition, each function module in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware, or in the form of hardware plus software function module.

[0166] Therefore, the embodiments should be considered in all respects as illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the description given above, therefore all changes which come within the meaning and range of equivalency of the claims are intended to be embraced therein. Any reference signs in the claims should not be construed as limiting the scope of the claims.

[0167] It is apparent for a person skilled in the art that the present application is not limited to the details of the above-described exemplary embodiments, but can be implemented in other concrete forms without departing from the spirit or essential characteristics of the application.

[0168] A person of ordinary skill in the art can understand that all or part of the processes in the above-described embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when executed, can include the processes of the above-described embodiments. Any reference to memory, storage, database, or other medium used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct RAM bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0169] It can be clearly understood by a person skilled in the art that, for the convenience and brevity of description, only the division of the above-described functional units and modules is taken as an example for illustration, and in actual applications, the above-described functions can be completed by different functional units and modules according to needs, that is, the internal structure of the system is divided into different functional units or modules to complete all or part of the functions described above.

[0170] In the embodiments provided in the present disclosure, it should be understood that the disclosed system and method can also be implemented in other manners. The above described system embodiments are merely illustrative, for example, the flowcharts and block diagrams in the accompanying drawings show possible implementation architectures, functions and operation of the system, method and computer program product according to the embodiments of the present disclosure. In this regard, each block in the flowcharts or block diagrams can represent a module, a segment or a portion of code which comprises one or more executable instructions for implementing the specified logic function. It should also be noted that in some alternative implementations, the functions shown in the blocks can occur in a different order than that shown in the accompanying drawings. For example, two blocks shown in succession can actually be executed substantially concurrently or in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented by dedicated hardware-based systems which perform the specified functions or actions, or can be implemented by a combination of dedicated hardware-based systems and computer instructions.

[0171] In the present disclosure, the term "comprising" or "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or apparatus including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent in such process, method, article or apparatus. Without more limitations, the element limited by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus including the element.

[0172] The above described embodiments are merely used to illustrate the technical solutions of the present application, rather than limiting them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

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 predicted field. Spatiotemporal multi-scale decomposition of the multi-source settlement field and the driving factor grid field yields spatiotemporally resolved subbands, 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.

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 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.

7. The geological subsidence monitoring method integrating deep learning and multi-source data according to claim 6, 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.

8. The geological subsidence monitoring method integrating deep learning and multi-source data according to claim 7, 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.

9. A system for monitoring geological subsidence using the method for integrating deep learning and multi-source data as described in any one of claims 1-8, 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