A method, system, device and medium for identifying surface deformation driven by groundwater

Abstract

The application discloses a kind of underground water driving ground surface deformation identification method, system, equipment and medium, it is related to artificial intelligence and geodesy cross field, this method includes: obtaining multi-source spatio-temporal data cube;Groundwater-driven probability soft label containing expert knowledge, statistical correlation and forward simulation fusion and candidate groundwater-driven deformation reference constraint are constructed;Construct spatio-temporal deep learning network based on multi-head attention mechanism, output groundwater-driven ground surface deformation probability graph and candidate groundwater-driven deformation sequence;Iterative training is carried out to the spatio-temporal deep learning network by constructing joint loss function, the joint loss function includes the physical constraint term adaptively modulated by groundwater-driven probability gate and the probability and deformation space consistency constraint term;Underground water driving deformation component is output using the network that training is completed.The application realizes the accurate identification of groundwater-driven ground surface deformation under weak supervision condition, and improves the physical consistency of result.

Description

Technical Field

[0001] This invention relates to the fields of geodesy, hydrology, and artificial intelligence, and in particular to a method, system, device, and medium for identifying groundwater-driven surface deformation based on probabilistic gating and spatial consistency constraints. Background Technology

[0002] Groundwater is a vital freshwater resource, accounting for approximately one-third of the world's freshwater reserves. With the increasing impact of global climate change and human activities, the over-exploitation of groundwater resources has become a prominent issue, triggering significant land subsidence in regions such as the North China Plain, causing substantial damage to urban infrastructure and the ecological environment. Therefore, accurately identifying groundwater-driven surface deformation and rationally assessing its intensity and spatial distribution characteristics are key technological requirements for achieving precise control of land subsidence.

[0003] The mechanisms of surface deformation caused by groundwater changes are highly complex, involving elastic deformation due to hydrological loads, tectonic deformation, porosity elastic deformation of confined aquifers, and irreversible inelastic deformation. Currently, methods for identifying and assessing groundwater-driven surface deformation mainly fall into three categories: statistical analysis, geodetic observation, and hydrogeological modeling. However, statistical analysis methods struggle to reveal physical mechanisms and quantify contributions; single geodetic observation methods cannot comprehensively reflect the spatiotemporal characteristics of groundwater-driven deformation; and hydrogeological modeling heavily relies on the difficult-to-obtain precise aquifer structure and mechanical parameters. Overall, these methods are ineffective at separating mixed signals, and the mechanisms for identifying deformation causes are unclear and lack sufficient intelligence.

[0004] In recent years, deep learning technology has brought breakthroughs to the accurate identification and assessment of groundwater-driven surface deformation. However, existing methods for identifying and assessing groundwater-driven surface deformation mainly have the following technical problems that need to be solved: (1) The groundwater-driven deformation components cannot be directly obtained through existing observation methods, resulting in a lack of real supervision labels for groundwater-driven deformation that can be directly observed, which restricts the applicability of the model; (2) Although some studies have introduced physical regularization terms to constrain the model, they are usually applied in a global or fixed weight manner, without distinguishing between groundwater-dominant and non-dominant areas, which can easily lead to unstable model training or biased results due to excessive physical constraints; (3) Existing methods do not explicitly model the spatial coordination relationship between groundwater-driven probability and deformation response, and do not constrain the consistency between the "spatial structure of probability change" and the "spatial gradient structure of deformation response", which leads to spatial inconsistency and discontinuity between probability identification and deformation separation results, reducing the physical interpretability of the results. Summary of the Invention

[0005] To address the technical problems in existing technologies, such as the lack of direct real-world supervision labels, model instability caused by global physical constraints, and low interpretability due to the lack of spatial consistency in the separation results of probability and deformation, this invention provides a groundwater-driven surface deformation identification method, system, device, and medium based on probabilistic gating and spatial consistency constraints.

[0006] In a first aspect, the present invention provides a method for identifying groundwater-driven surface deformation, comprising the following steps: S1: Obtain a multi-source spatiotemporal data cube containing deformation feature data, driving factor data, and auxiliary feature data; perform spatial resampling and temporal interpolation alignment processing on the multi-source spatiotemporal data cube; and output the multi-source spatiotemporal data cube formatted as a multi-dimensional tensor. S2: Based on a weakly supervised learning strategy, three probabilities are calculated respectively: the spatial prior probability driven by expert knowledge, the statistical temporal correlation probability, and the forward modeling probability of the physical model. The three probabilities are dynamically weighted and fused to output a soft label for groundwater driving probability. The soft label for groundwater driving probability is combined with the deformation feature data to output a candidate groundwater driving deformation reference constraint. S3: Construct a spatiotemporal deep learning network that includes a convolutional neural network module, a temporal convolutional network module, an attention module, and a dual-output branch structure; input the formatted multi-source spatiotemporal data cube into the spatiotemporal deep learning network to extract spatiotemporal fusion features; predict the probability map of groundwater-driven surface deformation by outputting the first branch of the dual-output branch structure; and predict candidate groundwater-driven deformation sequences by outputting the second branch of the dual-output branch structure. S4: Construct a joint loss function, which includes a probabilistic identification loss term based on the groundwater-driven probability soft label, a physical consistency regularization term gated and adaptively modulated by the predicted groundwater-driven surface deformation probability map, and a probabilistic deformation space consistency constraint term that constrains the collaboration between the probability gradient and the deformation gradient; iteratively train the spatiotemporal deep learning network based on the joint loss function and the candidate groundwater-driven deformation reference constraints, input the test spatiotemporal data cube of the region to be identified into the trained spatiotemporal deep learning network, and multiply the output target groundwater-driven surface deformation probability map with the target candidate groundwater-driven deformation sequence pixel by pixel to output the final groundwater-driven deformation component.

[0007] As an optional implementation of the first aspect of this application, step S1 specifically includes: acquiring the measured deformation sequence of the Global Navigation Satellite System and the interferometric synthetic aperture radar surface deformation time series as the deformation feature data; acquiring groundwater storage change data retrieved by gravity satellites, groundwater measured water level monitoring data, and climate and hydrological data as the driving factor data; acquiring digital elevation models, groundwater aquifer structure zoning maps, and geological structure data as the auxiliary feature data; performing grid alignment and time scale unification on the deformation feature data, the driving factor data, and the auxiliary feature data, performing dimensionless processing using the extreme value normalization method, and splicing them according to the time dimension, spatial grid width dimension, spatial grid height dimension, and feature channel dimension to output the multi-dimensional tensor multi-source spatiotemporal data cube.

[0008] As an optional implementation of the first aspect of this application, step S2, outputting a groundwater-driven probability soft label, specifically includes: extracting groundwater over-extraction indexes, surface subsidence rate, and Quaternary sedimentary thickness; dividing the spatial grid according to a preset response level and assigning initial probability values; and outputting the expert knowledge-driven spatial prior probability. Calculate the correlation coefficient between the interferometric synthetic aperture radar surface deformation time series and the measured groundwater level monitoring data; extract the target areas that pass the significance test and map their correlation coefficients to a preset probability interval; output the statistical time series correlation probability. A linear physical model of surface subsidence and groundwater head variation is constructed based on Terzaghi's consolidation theory. The calculation formula of the linear physical model is Δh = S × ΔH × b, where Δh represents the surface subsidence, S represents the water storage coefficient or compressibility coefficient, ΔH represents the groundwater head variation, and b represents the thickness of the compressible layer. Geological parameters corresponding to each spatial grid cell are extracted for forward modeling calculation. The fitting determination coefficient between the simulated deformation and the measured deformation sequence is mapped to the preset probability interval, and the forward modeling probability of the physical model is output. ; prior probabilities of the expert knowledge-driven space The statistical time series correlation probability and the forward modeling probability of the physical model Pixel-by-pixel dynamic weighted fusion is performed to output the groundwater-driven probability soft label.

[0009] As an optional implementation of the first aspect of this application, the prior probability of the expert knowledge-driven space is... The statistical time series correlation probability and the forward modeling probability of the physical model Perform pixel-by-pixel dynamic weighted fusion, specifically including: according to the formula Computing the prior confidence of expert knowledge space According to the formula Calculate the statistical correlation confidence level ,in The correlation significance test value is used; based on the goodness of fit between the simulated deformation and the measured deformation sequences. Determine the forward modeling confidence level of the physical model ; the prior confidence of the expert knowledge space The statistical correlation confidence level and the forward modeling confidence level of the physical model Normalization processing is performed to obtain the corresponding first normalized fusion weight, second normalized fusion weight, and third normalized fusion weight; the three probabilities are weighted and summed using each normalized fusion weight to output the groundwater driving probability soft label corresponding to each pixel spatial grid unit.

[0010] As an optional implementation of the first aspect of this application, step S2, outputting candidate groundwater-driven deformation reference constraints, specifically includes: setting a probability threshold based on the numerical value of the groundwater-driven probability soft label, dividing the target area into a low-probability zone, a medium-probability zone, and a high-probability zone; extracting structural deformation with a long-term linear trend in the low-probability zone as a first reference deformation sequence; extracting porosity elastic deformation in the high-probability zone by comparing deformation amplitude and time lag characteristics as a second reference deformation sequence; obtaining mixed-driven deformation in the medium-probability zone as a third reference deformation sequence; combining and splicing the first reference deformation sequence, the second reference deformation sequence, and the third reference deformation sequence according to spatial distribution, and outputting the candidate groundwater-driven deformation reference constraints for weak constraint training of the model.

[0011] As an optional implementation of the first aspect of this application, step S3, outputting the predicted candidate groundwater-driven deformation sequence, includes: extracting local spatial features from the multi-source spatiotemporal data cube using multi-layer two-dimensional spatial convolution operations and max pooling layers in the convolutional neural network module; capturing long-term dependent sequence features using causal dilated convolution operations within the temporal convolutional network module; inputting spatial and temporal features into the attention module, calculating the multi-head self-attention weights of the query feature matrix, key feature matrix, and value feature matrix, and outputting the spatiotemporal fusion features using dimensionality reduction and aggregation with a linear transformation matrix; inputting the spatiotemporal fusion features into the dual-output branch structure, wherein the first branch uses a Sigmoid activation function to output the predicted groundwater-driven surface deformation probability map, and the second branch uses a linear mapping layer to output the predicted candidate groundwater-driven deformation sequence.

[0012] As an optional implementation of the first aspect of this application, step S4, constructing the joint loss function, specifically includes: constructing the probability identification loss term using the binary classification cross-entropy formula. The formula for calculating the probability identification loss term is as follows: ,in, For the sample size, It is the first i The groundwater-driven probability soft label corresponding to each grid cell. It is the first i The probability results output by the CNN-TCN-Attention model of each grid cell; construct the physical consistency regularization term. The formula for calculating the physical consistency regularization term is as follows: ,in It is a nonlinear gated function. This represents the expectation operation on the sample set. , For the Laplace operator, The model outputs candidate groundwater-driven deformation results. It is a reference deformation obtained based on forward modeling of the physical model or inversion of the GRACE model; constructing the consistency constraint term of the probabilistic deformation space. The formula for calculating the probability deformation space consistency constraint term is as follows: ,in Let P represent the spatial gradient operator, where P is the probability label value; the probability identification loss term is... The physical consistency regularization term and the aforementioned probability deformation space consistency constraint term The joint loss function, used to control the iterative optimization of the network, is output by weighting and summing the results using preset weight coefficients.

[0013] Secondly, embodiments of this application provide a groundwater-driven surface deformation identification system, comprising: The training sample input data processing module is used to acquire a multi-source spatiotemporal data cube containing deformation feature data, driving factor data and auxiliary feature data, perform spatial resampling and temporal interpolation alignment processing on the multi-source spatiotemporal data cube, and output the multi-source spatiotemporal data cube formatted as a multi-dimensional tensor. The target label construction module for training samples is used to calculate three probabilities based on a weakly supervised learning strategy: expert knowledge-driven spatial prior probability, statistical temporal correlation probability, and physical model forward modeling probability. The three probabilities are dynamically weighted and fused to output a soft label for groundwater driving probability. The soft label for groundwater driving probability is combined with the deformation feature data to output a candidate groundwater-driven deformation reference constraint. The spatiotemporal deep network model construction module is used to construct a spatiotemporal deep learning network that includes a convolutional neural network module, a temporal convolutional network module, an attention module, and a dual-output branch structure. The formatted multi-source spatiotemporal data cube is input into the spatiotemporal deep learning network to extract spatiotemporal fusion features. The first branch of the dual-output branch structure is used to output a probability map of groundwater-driven surface deformation, and the second branch of the dual-output branch structure is used to output a candidate groundwater-driven deformation sequence. The model training and result output module is used to construct a joint loss function, which includes a probabilistic identification loss term constructed based on the groundwater-driven probability soft label, a physical consistency regularization term gated and adaptively modulated by the predicted groundwater-driven surface deformation probability map, and a probabilistic deformation space consistency constraint term that constrains the collaboration between the probability gradient and the deformation gradient. The spatiotemporal deep learning network is iteratively trained based on the joint loss function and the candidate groundwater-driven deformation reference constraints. The test spatiotemporal data cube of the region to be identified is input into the trained spatiotemporal deep learning network. The output target groundwater-driven surface deformation probability map is multiplied pixel by pixel with the target candidate groundwater-driven deformation sequence to output the final groundwater-driven deformation component.

[0014] Thirdly, embodiments of this application provide an electronic device, which includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of the method described in the first aspect.

[0015] Fourthly, embodiments of this application provide a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. By introducing a probabilistic gating mechanism, physical constraints are adaptively applied in space, avoiding the introduction of unreasonable constraints in non-groundwater-dominated areas, thereby improving the stability of model training and the reliability of results; 2. By constructing a spatial consistency constraint between the groundwater-driven probability gradient and the candidate groundwater-driven deformation gradient, the spatial mismatch between the identification results and the deformation response is effectively avoided, and the physical interpretability of the results is significantly improved. 3. Effectively alleviates the problem of missing true labels for groundwater-driven deformation under weak supervision conditions, and can achieve stable identification of groundwater-driven areas and deformation assessment without relying on precise pixel-by-pixel annotation; 4. By using the CNN–TCN–Attention combined model, end-to-end modeling is achieved from multi-source spatiotemporal data input to groundwater-driven surface deformation identification and probability assessment. It can automatically learn the complex nonlinear relationship between groundwater changes and surface deformation, get rid of the dependence on manual experience interpretation and physical models, and intelligently identify and assess groundwater-driven surface deformation, improving identification efficiency and objectivity. 5. Achieve collaborative modeling for groundwater-driven probability assessment and deformation component identification. The model can simultaneously output the spatial probability distribution of groundwater-driven surface deformation and the corresponding groundwater-driven deformation components, providing a direct and usable quantitative reference for groundwater over-extraction area identification and water resource management. Attached Figure Description

[0017] Figure 1 This is a flowchart of a groundwater-driven surface deformation identification method proposed in an embodiment of the present invention; Figure 2 This is a flowchart of the target label construction process using the weakly supervised learning method in an embodiment of the present invention; Figure 3 This is a flowchart of the physical attribution judgment of groundwater-driven deformation probability results in an embodiment of the present invention; Figure 4 This is a structural diagram of the CNN-TCN-Attention model in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a groundwater-driven surface deformation identification system provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] Example 1 Please see Figure 1 This is a flowchart illustrating groundwater-driven surface deformation identification, provided by an embodiment of the present invention. The method may include the following steps: Step S1: Obtain a multi-source spatiotemporal data cube containing deformation feature data, driving factor data, and auxiliary feature data. Perform spatial resampling and temporal interpolation alignment processing on the multi-source spatiotemporal data cube and output the multi-source spatiotemporal data cube formatted as a multi-dimensional tensor.

[0021] First, deformation feature data, driving factor data, and auxiliary feature data within the study area are collected and processed as input variables for model training. Among them: (1) Deformation characteristic data consists of two types of observations: First, measured deformation sequences of the study area are obtained through InSAR (Global Navigation Satellite System) technology. First, Sentinel-1 images of different time periods in the study area are collected and organized; then, the collected images are registered and interferometrically processed to obtain interferograms and remove flat-land effects and topographic phases. Subsequently, filtering and coherence estimation are performed; finally, the unwrapped interferograms are used to construct a set of equations through the phase stacking time series inversion method, and the surface deformation of the study area at different time periods is calculated using the singular value decomposition method. Second, the time series of surface deformation in the study area is obtained through GNSS (Interferometric Synthetic Aperture Radar) technology. First, continuous GNSS observation data within the study area were collected and organized. Then, baseline calculations were performed on the GNSS observation data using GAMIT / GLOBK version 10.61 software to obtain the daily regional relaxation solutions for each station. The specific process included satellite orbit integration calculation, observation model construction, cycle slip detection and gross error correction of phase observations, and least squares estimation of baseline vectors. Based on this, GNSS stations in the IGS network under the ITRF framework were used as constraints, and Kalman filters were employed for network adjustment to obtain the three-dimensional deformation of GNSS stations within the study area.

[0022] (2) The driving factor data consists of water storage change data and hydro-climatic factor data. Groundwater storage change data is obtained by inversion through the GRACE / GRACE-FO joint GLDAS model or by using water level data monitored by actual groundwater wells. Hydro-climatic factor data is extracted based on the GLDAS NOAH v2.1 dataset and includes evapotranspiration, average temperature, runoff, and precipitation.

[0023] (3) Auxiliary feature data include groundwater aquifer structure zoning map, geological map and digital elevation model (DEM).

[0024] Further, data preprocessing and fusion are performed, including: Step S11: Data Preprocessing and Spatiotemporal Consistency Construction. Deformation feature data and driving factor data are spatially resampled to unify them to the same spatial grid resolution. Temporal interpolation is then used to unify various time series to a monthly resolution, thus constructing a spatiotemporally consistent multidimensional dataset. For auxiliary feature data, vector data is converted to raster data and extended to the same time dimension as the deformation feature data to ensure spatiotemporal alignment in the model input.

[0025] Step S12: Data Normalization and Construction of Multi-Source Spatiotemporal Data Cube. The constructed multidimensional dataset is normalized using the Min-Max method to eliminate the impact of differences in the scales of different variables on model training, ultimately forming a multi-dimensional tensor multi-source spatiotemporal data cube. For each sample (a region or the entire study area), the data organization is as follows: [ T, C, W, H ].in, T In terms of time dimension, W , H These are the width and height of the spatial grid, respectively. C This represents the number of feature channels. Feature channels include deformation feature data, driving factor data, and auxiliary feature data, etc.

[0026] Step S2: Based on the weakly supervised learning strategy, calculate three probabilities respectively: expert knowledge-driven spatial prior probability, statistical temporal correlation probability, and physical model forward modeling probability. Dynamically weight and fuse the three probabilities to output a soft label for groundwater-driven probability. Combine the soft label for groundwater-driven probability with the deformation feature data to output candidate groundwater-driven deformation reference constraints.

[0027] like Figure 2 As shown, this invention employs weakly supervised learning to construct target labels for model training, consisting of two parts: first, a probability label map of groundwater-driven deformation (probabilistic soft labels); and second, candidate groundwater-driven deformation reference constraints (not pixel-by-pixel strongly supervised ground truth, but reference deformation constraints used for weakly supervised training). The key to this design is that groundwater-driven deformation components are difficult to obtain directly in existing observation systems. Therefore, this invention does not construct pixel-by-pixel precisely labeled strongly supervised ground truth, but instead achieves stable model training in the absence of ground truth through a combination of "probabilistic soft labels + reference deformation constraints."

[0028] Step S21: The specific content of constructing groundwater-driven probability soft labels using weakly supervised learning includes: Step S21-1: Spatial Prior Constraints Based on Expert Knowledge. First, collect hydrogeological data, active tectonic data, human activity information, and surface deformation data related to this invention. Specifically, this includes deformation monitoring data (GNSS / leveling / InSAR), hydrogeological survey data (aquifer distribution map, Quaternary sedimentary thickness map, groundwater over-extraction zone delineation map), active tectonic data (active fault distribution map, seismic activity analysis), and human activity information (distribution of deep and shallow groundwater wells, distribution of oil and gas and other mineral resource extraction, and distribution of urban infrastructure buildings). Based on this, considering the influence of groundwater over-extraction degree, surface subsidence rate, groundwater level change trend, and other driving factors, a classification standard is constructed using expert knowledge and experience constraints. The study area is divided into three groundwater-driven response levels (high, medium, and low, with corresponding probability label values ​​P: P ≥ 0.7, 0.4 ≤ P < 0.7, and P < 0.4, respectively). The high-level criteria are as follows: high-level areas are characterized by: groundwater level decline rates greater than 0.5 m / year; significant subsidence (rate > 10 mm / year); Quaternary sedimentary thickness greater than 80 m; and no disturbance from faults or major engineering projects. Medium-level areas are characterized by: not located in over-extraction cones; existing groundwater activity structures; moderate subsidence (5 mm / year ≤ rate ≤ 10 mm / year) and groundwater variation (0.2 m / year ≤ rate ≤ 0.5 m / year); and Quaternary sedimentary thickness (20 ≤ thickness < 80). Low-level areas are characterized by: bedrock zones, tectonic activity zones, resource extraction zones, and geological disaster zones; subsidence (rate < 5 mm / year) and groundwater variation (rate < 0.2 m / year); and Quaternary sedimentary thickness (thickness < 20). Finally, expert-driven spatial prior probabilities are generated through vector clipping and rasterization.

[0029] Step S21-2: Data-driven statistical correlation analysis. For each spatial network unit, the correlation between the InSAR deformation sequence (obtained in step S1) and the measured groundwater change sequence is calculated. The statistical correlation results are then filtered based on the significance test. For regions that pass the significance test, the correlation coefficient is normalized and mapped to the [0,1] interval as the statistical time-series correlation probability reflecting the statistical consistency between groundwater changes and surface deformation.

[0030] Step S21-3: Simplified physical model forward modeling. Under simplified assumptions, based on Terzaghi consolidation theory, surface subsidence can be expressed as a linear function of groundwater head variation and formation compressibility characteristics: (1) Among them, △ h Indicates the amount of land subsidence. S Δ is the water storage coefficient or compressibility coefficient.H For changes in groundwater head, b The thickness is the compressible layer thickness.

[0031] The parameter spatial distribution field is generated by combining borehole lithology information and hydrogeological maps: First, parameters such as lithological stratification and permeability coefficient at each depth are extracted from the borehole lithology information to establish a discrete point dataset; at the same time, the aquifer structure, groundwater flow direction, and boundary conditions are defined and constrained according to the hydrogeological map data; finally, a continuous parameter spatial distribution field is constructed using the Kriging interpolation method. Based on Equation (1), a forward modeling simulation is performed on each pixel spatial grid cell to calculate the pseudo-determination coefficient between the simulated deformation and the InSAR measured deformation, and it is mapped to the probability value in the interval [0,1] to form the forward modeling probability of the physical model.

[0032] Step S21-4: Pixel-by-pixel dynamic weighted fusion of probability labels. Given the differences in the reliability of expert knowledge space priors, statistical correlation analysis, and forward modeling of physical models across different regions, a pixel-by-pixel dynamic weighted fusion strategy is adopted to construct the final groundwater-driven deformation probability label. In each pixel spatial grid cell ( i, j At point ), let the spatial prior probability, the statistical temporal correlation probability, and the physical model forward modeling probability be denoted as […]. , and The basic weights are respectively The values ​​of all three methods are in the range of [0,1]. During the fusion process, the completeness of the three methods is first checked. When a method is missing in a local area, a weighted average is calculated using formula (2) to avoid the missing values ​​interfering with the fusion result. (2) in, This represents the fusion probability result; ; .

[0033] When all three types of tags are unavailable, set The value is set to 0.5 to avoid human bias. When all three methods are effective, their weights are dynamically adjusted based on the confidence index of each method. The confidence of the spatial prior constraint method based on expert knowledge is measured by the degree to which its probability value deviates from the neutral value (0.5), and its confidence is defined as: (3) The statistical correlation confidence level is controlled by the significance test results, with significantly correlated regions receiving higher weights. The confidence level of the correlation significance test value is defined as: (4) The weight labels for the simplified physical model forward modeling are determined by the goodness of fit between the simulation results and the observed deformation. Its confidence level is defined as: (5) Based on this, the unnormalized weights of the three labeling methods are expressed as follows: , , (6) After normalizing the above weights, the final fusion weights for the three labels are as follows: (7) After weight normalization, the three types are weighted and fused (Equation (8)) to obtain the final spatial grid cell. i, j ) groundwater-driven probability soft tag (8) Step S22: Constructing candidate groundwater-driven deformation reference constraints. (e.g.) Figure 3 As shown. First, based on the numerical value P of the soft label of groundwater driving probability output by the model, the study area is divided into three probability categories: low probability zone (P < 0.4), medium-low probability zone (0.4 ≤ P < 0.7), and high-low probability zone (P ≥ 0.7). In the low probability zone, by comparing the consistency between InSAR deformation and the hydrological load elastic deformation inverted by GRACE / GRACE-FO joint GLDAS, and considering whether the deformation shows a long-term linear trend, tectonic deformation with a long-term linear trend is extracted as the first reference deformation sequence to distinguish between hydrological load elastic deformation and tectonic deformation. In the high probability zone, the consistency between deformation and groundwater level changes in amplitude, phase, and time delay characteristics is compared to distinguish between porosity elastic deformation and inelastic deformation, and porosity elastic deformation is extracted as the second reference deformation sequence. Based on this, candidate groundwater-driven deformation reference constraints for weakly supervised training are formed: First, reference deformation time series are selected based on physical consistency criteria, including seasonal deformation (porosity elasticity) obtained by InSAR combined with water level changes, GNSS deformation with a significant linear trend (tectonic trend), and hydrological load elastic deformation obtained by GRACE inversion; Second, a hybrid driving reference deformation obtained by GNSS or InSAR in the medium probability zone is used as a third reference deformation sequence to weakly constrain the model output. The first, second, and third reference deformation sequences are combined and spliced ​​according to their spatial distribution to output the candidate groundwater-driven deformation reference constraints for weakly constrained model training.

[0034] Step S3: Construct a spatiotemporal deep learning network including a convolutional neural network module, a temporal convolutional network module, an attention module, and a dual-output branch structure; input the formatted multi-source spatiotemporal data cube into the spatiotemporal deep learning network to extract spatiotemporal fusion features; predict the probability map of groundwater-driven surface deformation through the output of the first branch of the dual-output branch structure; and predict the candidate groundwater-driven deformation sequence through the output of the second branch of the dual-output branch structure.

[0035] like Figure 4 As shown, the spatiotemporal deep learning network consists of a convolutional neural network (CNN) module, a temporal convolutional network (TCN) module, and an attention module connected sequentially. It employs a probability-deformation decoupled dual-output structure to output a probability map of groundwater-driven deformation and a candidate groundwater-driven deformation sequence, respectively. The CNN module effectively characterizes the spatial distribution characteristics and spatial heterogeneity of surface deformation and groundwater changes; the TCN module effectively characterizes the time-delay response and long-term dependency between groundwater changes and surface deformation; and the attention module is used to weightedly fuse key temporal sequences and key spatial regions, achieving collaborative modeling of probability assessment and deformation recognition. Specific content includes: Step S31: Spatial feature extraction of the CNN module. The CNN module adopts a multi-layer convolution and pooling structure. The specific process is as follows: (1) The input multi-source spatiotemporal data cube first passes through the first convolutional layer, performs two-dimensional spatial convolution operation to extract local spatial features, and introduces non-linear expressive ability through the ReLU activation function; (2) The features after ReLU activation are input into the first max pooling layer, and the feature map is downsampled to reduce the data dimension and enhance the robustness of spatial features; (3) The output after pooling passes through the second convolutional layer and ReLU activation in sequence, and enters the second max pooling layer; (4) As the convolutional layer and pooling layer are stacked layer by layer, the model gradually extracts spatial structure information from shallow to deep layers, and finally forms a stable spatial feature output after convolution and pooling in the nth layer.

[0036] Step S32: Temporal feature extraction of the TCN module. The TCN module consists of multiple concatenated residual blocks, each of which adopts a causal dilated convolution structure. Each residual block includes: (1) a first-layer causal dilated convolution, used to capture long-term dependencies in the time series; (2) weight normalization operation to accelerate model convergence and improve training stability; (3) ReLU activation function to enhance nonlinear expressive power; (4) Dropout operation to reduce the risk of overfitting; (5) a second-layer causal dilated convolution and corresponding weight normalization, ReLU and Dropout operations; (6) the residual block input and the above convolution output are matched in dimension by 1×1 convolution and then added to form the residual connection output, and finally the long-term dependency sequence features are output in the nth layer residual block.

[0037] Step S33: Spatiotemporal feature weighted fusion and result output based on the Attention module. The specific process is as follows: (1) Sequencing the spatiotemporal features output by CNN-TCN, denoted as (2) Perform layer normalization on the serialization and calculate multi-head self-attention, including the construction of query (Q), key (K) and value (V) and the calculation of attention weights, as shown in formula (9). (3) Concatenate the outputs of each attention and obtain spatiotemporal fusion features through linear transformation, as shown in formulas (10) and (11). (4) Perform 1×1 convolution on the spatiotemporal fusion features to achieve feature dimensionality reduction and obtain a one-dimensional feature vector. (5) The final module adopts a dual-output branch structure. One branch outputs the probability map of predicted groundwater-driven surface deformation through the Sigmoid activation function, and the other branch outputs the corresponding predicted candidate groundwater-driven deformation sequence through the Linear linear mapping. Finally, the groundwater-driven surface deformation and the probability results are gated and weighted to obtain the groundwater-driven deformation component, thereby realizing the collaborative modeling of probability assessment and deformation recognition.

[0038] (9) (10) (11) in, , , Each represents each Queries, keys, and values; This is the scaling factor; This represents the attention calculation for the i-th attention head; For layer normalization operation; This is a multi-head self-attention mechanism used to perform multi-head attention calculations on the input sequence. The function normalizes the input values ​​and outputs attention weights; It is a linear transformation matrix used for changes in the characteristic space. This indicates the output result after MSA and residual connection.

[0039] Step S4: Construct a joint loss function, which includes a probabilistic identification loss term based on the groundwater-driven probability soft label, a physical consistency regularization term gated and adaptively modulated by the predicted groundwater-driven surface deformation probability map, and a probabilistic deformation space consistency constraint term that constrains the collaboration between the probability gradient and the deformation gradient; iteratively train the spatiotemporal deep learning network based on the joint loss function and the candidate groundwater-driven deformation reference constraints, input the test spatiotemporal data cube of the region to be identified into the trained spatiotemporal deep learning network, and multiply the output target groundwater-driven surface deformation probability map with the target candidate groundwater-driven deformation sequence pixel by pixel to output the final groundwater-driven deformation component.

[0040] First, the constructed training samples (multi-source spatiotemporal data cubes and their corresponding target labels) are divided into training and testing sets in a 7:3 ratio according to the time dimension. During model training, the multi-source spatiotemporal data cubes are input into the CNN module to extract spatial features, and the output spatial feature tensor […]. T, C', W', H' ],in, C' This represents the number of spatial feature channels extracted by the CNN; W', H' The width and height of the spatial features after CNN convolution and pooling are represented respectively. These are then spatially aggregated to form a [C',T] sequence, which is input into the TCN module. Modeling is performed using a multi-scale causal dilation convolutional structure. Based on this, an Attention module is introduced to adaptively weight and enhance key temporal and spatial regions.

[0041] Step S41: Separate model output from probabilistic gated deformation. The model adopts a dual-output structure, and its output includes two parts: one is the probability map of predicted groundwater-driven surface deformation. The first purpose is to characterize the dominance of groundwater changes in surface deformation; the second is to predict candidate groundwater-driven deformation outcomes. This is used to characterize the potential deformation response patterns learned by the model under the dominant assumption of groundwater variation, and ultimately the groundwater-driven deformation component. It is obtained through a probability gating mechanism, and its expression is as follows: (12) Where ⊙ represents pixel-by-pixel and time-by-time multiplication.

[0042] Step S42: Weakly supervised training and loss function construction.

[0043] During model training, the groundwater-driven probability soft label constructed before training is used as a weak supervision target, and the probability prediction result is optimized by the binary classification cross-entropy loss function, as shown in formula (13).

[0044] (13) in For probabilistic identification of loss terms, It is the first i The groundwater driving probability soft label corresponding to each grid cell has a value range of [0,1], which is used to characterize the dominant probability of groundwater in the formation of surface deformation at that location. It is the first i The probability values ​​output by the CNN-TCN-Attention model for each grid cell; This represents the number of samples.

[0045] Physical consistency regularization term for probabilistic gating. In order to guide the candidate groundwater-driven deformation output of the model to conform to the physical mechanism of groundwater change, a physical consistency regularization term based on reference deformation is introduced, as shown in formula (14).

[0046] (14) in, The candidate groundwater-driven deformation results output by the model are used to characterize the deformation response mode under the dominant groundwater change conditions. It is a reference deformation obtained based on forward modeling of the physical model or inversion of the GRACE model. It is used to characterize the spatial deformation characteristics that should be produced by changes in groundwater or hydrological load in a physical sense. This reference deformation is not used as a strongly supervised truth label, but is used to construct physical consistency constraints. For the Laplace operator, it describes the spatial curvature and smoothness of the deformed field.

[0047] To highlight the innovation of this invention, a probabilistic gating adaptive modulation mechanism is introduced into the physical consistency regularization term. This allows it to exert a restraining effect only in areas where groundwater is more likely to dominate, thereby avoiding the introduction of deviations by forcibly imposing physical constraints on groundwater in non-groundwater-dominated areas, as shown in formula (15).

[0048] (15) in, For nonlinear gated functions (such as Sigmoid), when P When the value is less than 0.7, the physical constraint is automatically weakened or turned off, while when... P As the value approaches 1, the physical constraints gradually increase. This represents the expectation operation on the sample set, used to calculate the average constraint effect of the probability-gated physical regularization term on the overall sample. Probability-deformation space consistency constraint. To further address the potential inconsistency between the groundwater-driven probability identification results and the deformation response space structure, this invention introduces a spatial consistency constraint between the probability gradient and the deformation gradient, as shown in formula (16), so that regions with significant changes in groundwater-driven probability correspond to reasonable deformation space gradient responses, thereby achieving collaborative optimization of groundwater-driven region identification and deformation response structure learning.

[0049] (16) in This is a consistency constraint term for the probabilistic deformation space. This represents the spatial gradient operator.

[0050] Step S43: Composition of the total loss function. Combining the groundwater-driven probabilistic identification constraint, the probabilistic gating adaptive modulation deformation physical consistency constraint, and the probability-deformation space consistency constraint, the total loss function of the model is shown in Equation (17).

[0051] (17) in, , , These represent the groundwater-driven probabilistic identification loss term, the probabilistic gating physical consistency regularization term, and the probabilistic-deformation space consistency constraint term, respectively. , , For the corresponding weight coefficients, satisfying Its value is determined through validation set performance tuning to achieve a balance between identification accuracy and physical consistency. Through the above-mentioned "probabilistic gating physical constraint" mechanism, the introduction of bias by forcibly imposing groundwater physical constraints in non-groundwater-dominated areas is avoided, thereby enhancing the physical consistency and interpretability of groundwater-driven deformation prediction results.

[0052] Model Validation and Accuracy Evaluation The trained model was evaluated using a test dataset. Thresholding was applied to the output groundwater-driven probability results (e.g., P ≥ 0.7, indicating a groundwater-dominated driving region), and prediction performance was evaluated using metrics such as AUC, F1 score, and precision / recall. Simultaneously, the consistency between the groundwater-driven deformation component and groundwater level change data, as well as terrestrial water storage changes retrieved from the GRACE gravity satellite, was assessed through comparative analysis to evaluate their temporal trends, periodic characteristics, and hysteresis responses. Furthermore, the groundwater-driven deformation component was incorporated into the surface deformation reconstruction process and compared with the total deformation observed by InSAR to quantitatively evaluate the explanatory power of the groundwater-driven deformation component for the total deformation.

[0053] In summary, this invention provides a groundwater-driven surface deformation identification method based on probabilistic gating and spatial consistency constraints, comprising: (1) Construction of a multi-source spatiotemporal data cube (input layer). Specifically, it includes: deformation feature data, driving factor data, and auxiliary feature data. Among them, deformation feature data includes GNSS measured deformation sequences and InSAR surface deformation time series; driving factor data includes groundwater storage change data obtained by inversion of the GRACE / GRACE-FO gravity satellite and GLDAS model, groundwater measured water level monitoring data, and climate data such as rainfall, evapotranspiration, runoff, and temperature; auxiliary feature data includes digital elevation model, groundwater aquifer structure zoning map, and geological structure data. The above multi-source data are spatially resampled, temporally interpolated, and normalized to generate a multi-dimensional spatiotemporal data cube with consistent spatiotemporal reference.

[0054] (2) Construction of target labels for training samples (weakly supervised layer). Addressing the issue that groundwater-driven deformation cannot be directly labeled, this invention does not construct pixel-by-pixel strongly supervised labels. Instead, it constructs groundwater-driven probability soft labels using the following three types of information: expert knowledge-driven spatial prior constraints to characterize the long-term spatial background of groundwater activity; statistical correlation analysis between deformation and groundwater changes to reflect the consistency of observational data; and forward simulation results based on simplified physical models or GRACE inversion to introduce explicit physical mechanism constraints. By dynamically weighting and fusing the above three types of information pixel by pixel, a groundwater-driven probability label map with values ​​ranging from [0,1] is generated to characterize the dominant probability of groundwater changes in surface deformation formation. Combined with datasets such as InSAR, GNSS, groundwater level, and GRACE gravity satellite, candidate groundwater-driven deformation reference constraints are formed.

[0055] (3) CNN–TCN–Attention Model Construction. The model constructed in this invention consists of a CNN module, a TCN module, and an Attention module, and introduces a probabilistic gating mechanism. The model adopts a dual-output structure, with one branch outputting the probability map of predicted groundwater-driven surface deformation and the other branch outputting the predicted candidate groundwater-driven deformation sequence under groundwater-dominant conditions.

[0056] (4) Probability-gated physical regularization adaptive modulation mechanism. Unlike the existing technology that uses fixed weights or global application to the physical regularization term, the groundwater driving probability is introduced into the physical regularization term as a weight gating factor, and combined with a nonlinear gating function, to realize the spatial adaptive modulation of the physical regularization term. This avoids forcibly introducing groundwater physical constraints in non-groundwater-dominated areas, thereby improving the stability of model training and the physical rationality of the results.

[0057] (5) Probability-deformation spatial consistency constraint mechanism. By introducing a joint constraint term of probability gradient and deformation gradient, the model learns the corresponding spatial deformation response structure while identifying the groundwater-driven region, ensuring the consistency and continuity of the probability identification result and the deformation separation result in spatial structure.

[0058] (6) Joint loss function optimization. A joint loss function is constructed based on probabilistic soft label probabilistic identification loss, deformation physical constraint driven by groundwater probabilistic gating adaptive modulation, and probability-deformation space consistency constraint. The model is trained and finally outputs the groundwater-driven surface deformation probability map and the corresponding groundwater-driven deformation components.

[0059] This invention does not simply superimpose physical constraints onto a deep learning model. Instead, it uses the groundwater-driven probability as the core control variable and employs a probabilistic gating mechanism to adaptively modulate the physical constraints, ensuring that they are activated only in regions where groundwater is highly likely to dominate. Furthermore, by introducing a spatial consistency constraint between the probability gradient and the candidate groundwater-driven deformation gradient, it achieves collaborative optimization of groundwater-driven causal identification and deformation response modeling under weak supervision. This avoids the spatial inconsistency between probability identification and deformation separation results in existing methods, improving the physical interpretability and spatial consistency of groundwater-driven deformation identification results.

[0060] Example 2 Please see Figure 5 The diagram shown is a schematic representation of a groundwater-driven surface deformation identification system according to a second embodiment of this application. The system includes the following key modules: The training sample input data processing module 100 is used to acquire a multi-source spatiotemporal data cube containing deformation feature data, driving factor data and auxiliary feature data, perform spatial resampling and temporal interpolation alignment processing on the multi-source spatiotemporal data cube, and output the multi-source spatiotemporal data cube formatted as a multi-dimensional tensor. The target label construction module 200 for training samples is used to calculate three probabilities based on a weakly supervised learning strategy: expert knowledge-driven spatial prior probability, statistical time-series correlation probability, and physical model forward modeling probability. The three probabilities are dynamically weighted and fused to output a soft label for groundwater-driven probability. The soft label for groundwater-driven probability is combined with the deformation feature data to output a candidate groundwater-driven deformation reference constraint. The spatiotemporal deep network model construction module 300 is used to construct a spatiotemporal deep learning network including a convolutional neural network module, a temporal convolutional network module, an attention module, and a dual-output branch structure; the formatted multi-source spatiotemporal data cube is input into the spatiotemporal deep learning network to extract spatiotemporal fusion features; the first branch of the dual-output branch structure outputs a predicted probability map of groundwater-driven surface deformation; and the second branch of the dual-output branch structure outputs a predicted candidate groundwater-driven deformation sequence. The model training and result output module 400 is used to construct a joint loss function, which includes a probabilistic identification loss term constructed based on the groundwater-driven probability soft label, a physical consistency regularization term gated and adaptively modulated by the predicted groundwater-driven surface deformation probability map, and a probabilistic deformation space consistency constraint term that constrains the collaboration between the probability gradient and the deformation gradient. The spatiotemporal deep learning network is iteratively trained based on the joint loss function and the candidate groundwater-driven deformation reference constraints. The test spatiotemporal data cube of the region to be identified is input into the trained spatiotemporal deep learning network. The output target groundwater-driven surface deformation probability map is multiplied pixel by pixel with the target candidate groundwater-driven deformation sequence to output the final groundwater-driven deformation component.

[0061] The groundwater-driven surface deformation identification system in this application embodiment can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, in-vehicle electronic devices, wearable devices, ultra-mobile personal computers (UMPCs), netbooks, or personal digital assistants (PDAs), etc., while non-mobile electronic devices can be servers, network-attached storage (NAS), personal computers (PCs), etc. This application embodiment does not impose specific limitations.

[0062] One embodiment of the groundwater-driven surface deformation identification system in this application can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application does not specifically limit the specific operating system used.

[0063] This application provides a groundwater-driven surface deformation identification system that can achieve... Figure 1The various processes implemented by the groundwater-driven surface deformation identification method in the method embodiment are not described in detail here to avoid repetition.

[0064] Optionally, this application embodiment also provides an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the various processes of the above-described embodiment of the groundwater-driven surface deformation identification method and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0065] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described embodiment of the groundwater-driven surface deformation identification method and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0066] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0067] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0068] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0069] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for identifying groundwater-driven surface deformation, characterized in that, Includes the following steps: S1: Obtain a multi-source spatiotemporal data cube containing deformation feature data, driving factor data, and auxiliary feature data; perform spatial resampling and temporal interpolation alignment processing on the multi-source spatiotemporal data cube; and output the multi-source spatiotemporal data cube formatted as a multi-dimensional tensor. S2: Based on a weakly supervised learning strategy, three probabilities are calculated respectively: the spatial prior probability driven by expert knowledge, the statistical temporal correlation probability, and the forward modeling probability of the physical model. The three probabilities are dynamically weighted and fused to output a soft label for groundwater driving probability. The soft label for groundwater driving probability is combined with the deformation feature data to output a candidate groundwater driving deformation reference constraint. S3: Construct a spatiotemporal deep learning network that includes a convolutional neural network module, a temporal convolutional network module, an attention module, and a dual-output branch structure; input the formatted multi-source spatiotemporal data cube into the spatiotemporal deep learning network to extract spatiotemporal fusion features; predict the probability map of groundwater-driven surface deformation by outputting the first branch of the dual-output branch structure; and predict candidate groundwater-driven deformation sequences by outputting the second branch of the dual-output branch structure. S4: Construct a joint loss function, which includes a probabilistic identification loss term based on the groundwater-driven probability soft label, a physical consistency regularization term gated and adaptively modulated by the predicted groundwater-driven surface deformation probability map, and a probabilistic deformation space consistency constraint term that constrains the collaboration between the probability gradient and the deformation gradient; iteratively train the spatiotemporal deep learning network based on the joint loss function and the candidate groundwater-driven deformation reference constraints, input the test spatiotemporal data cube of the region to be identified into the trained spatiotemporal deep learning network, and multiply the output target groundwater-driven surface deformation probability map with the target candidate groundwater-driven deformation sequence pixel by pixel to output the final groundwater-driven deformation component.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: The measured deformation sequence from the Global Navigation Satellite System and the temporal sequence of surface deformation from the Interferometric Synthetic Aperture Radar were obtained as the deformation feature data. The data on groundwater storage changes retrieved from gravity satellites, groundwater level monitoring data, and climate and hydrological data are used as the driving factor data. Digital elevation models, groundwater aquifer structure zoning maps, and geological structure data are acquired as auxiliary feature data. The deformation feature data, the driving factor data, and the auxiliary feature data are subjected to grid alignment and time scale unification. The extremum normalization method is used for dimensionless processing. The data is then spliced ​​according to the time dimension, the spatial grid width dimension, the spatial grid height dimension, and the feature channel dimension to output the multi-dimensional tensor multi-source spatiotemporal data cube.

3. The method according to claim 1, characterized in that, In step S2, the groundwater-driven probability soft label is output, specifically including: The system extracts indicators of groundwater over-extraction, surface subsidence rate, and Quaternary sedimentary thickness; divides the space into grids based on preset response levels and assigns initial probability values; and outputs the expert knowledge-driven spatial prior probability. ; The correlation coefficient between the interferometric synthetic aperture radar (ISAR) surface deformation time series and the measured groundwater level monitoring data is calculated. Target areas that pass the significance test are extracted, and their correlation coefficients are mapped to a preset probability interval. The statistical time series correlation probability is then output. ; A linear physical model of surface subsidence and groundwater head variation is constructed based on Terzaghi's consolidation theory. The calculation formula of the linear physical model is Δh = S × ΔH × b, where Δh represents the surface subsidence, S represents the water storage coefficient or compressibility coefficient, ΔH represents the groundwater head variation, and b represents the thickness of the compressible layer. Geological parameters corresponding to each spatial grid cell are extracted for forward modeling calculations. The fitting determination coefficient between the simulated deformation and the measured deformation sequence is mapped to the preset probability interval, and the forward modeling probability of the physical model is output. ; Prior probabilities of the expert knowledge-driven space The statistical time series correlation probability and the forward modeling probability of the physical model Pixel-by-pixel dynamic weighted fusion is performed to output the groundwater-driven probability soft label.

4. The method according to claim 3, characterized in that, Prior probabilities of the expert knowledge-driven space The statistical time series correlation probability and the forward modeling probability of the physical model Perform pixel-by-pixel dynamic weighted fusion, specifically including: According to the formula Computing the prior confidence of expert knowledge space ; According to the formula Calculate the statistical correlation confidence level ,in This is the value for the significance test of the correlation. Based on the goodness of fit between simulated deformation and measured deformation sequences Determine the forward modeling confidence level of the physical model ; The prior confidence of the expert knowledge space The statistical correlation confidence level and the forward modeling confidence level of the physical model Normalization is performed to obtain the corresponding first normalized fusion weight, second normalized fusion weight, and third normalized fusion weight; The three probabilities are weighted and summed using normalized fusion weights to output the groundwater-driven probability soft label corresponding to each pixel spatial grid cell.

5. The method according to claim 1, characterized in that, In step S2, candidate groundwater-driven deformation reference constraints are output, specifically including: Based on the numerical value of the groundwater-driven probability soft tag, a probability threshold is set to divide the target area into low-probability, medium-probability, and high-probability zones. In the low probability region, structural deformation with a long-term linear trend is extracted as the first reference deformation sequence. In the high probability region, porosity elastic deformation is extracted by comparing deformation amplitude and time delay characteristics as the second reference deformation sequence. In the medium probability region, hybrid driven deformation is obtained as the third reference deformation sequence. The first reference deformation sequence, the second reference deformation sequence, and the third reference deformation sequence are combined and spliced ​​according to their spatial distribution to output the candidate groundwater-driven deformation reference constraints for weak constraint training of the model.

6. The method according to claim 1, characterized in that, In step S3, the predicted candidate groundwater-driven deformation sequence is output, including: The multi-layer two-dimensional spatial convolution operation and max pooling layer in the convolutional neural network module are used to extract local spatial features from the multi-source spatiotemporal data cube. The causal dilated convolution operation within the temporal convolutional network module is used to capture long-term dependent sequence features; The spatial and temporal features are input into the attention module, the multi-head self-attention weights of the query feature matrix, key feature matrix and value feature matrix are calculated, and the spatiotemporal fusion features are output by dimensionality reduction and aggregation using a linear transformation matrix. The spatiotemporal fusion features are input into the dual-output branch structure. The first branch uses the Sigmoid activation function to output the predicted groundwater-driven surface deformation probability map, and the second branch uses a linear mapping layer to output the predicted candidate groundwater-driven deformation sequence.

7. The method according to claim 1, characterized in that, In step S4, the joint loss function is constructed, specifically including: The probability discrimination loss term is constructed using the binary cross-entropy formula. The formula for calculating the probability identification loss term is as follows: ,in, For the sample size, It is the first i The groundwater-driven probability soft label corresponding to each grid cell. It is the first i The probability values ​​output by the CNN-TCN-Attention model for each grid cell; Construct the physical consistency regularization term The formula for calculating the physical consistency regularization term is as follows: ,in It is a nonlinear gated function. This represents the expectation operation on the sample set. , For the Laplace operator, The model outputs candidate groundwater-driven deformation results. It is a reference deformation obtained based on forward modeling of the physical model or inversion of the GRACE model; Construct the consistency constraint term of the probabilistic deformation space The formula for calculating the probability deformation space consistency constraint term is as follows: ,in Let P represent the spatial gradient operator, where P is the probability label value; The probability identification loss term The physical consistency regularization term and the aforementioned probability deformation space consistency constraint term The joint loss function, used to control the iterative optimization of the network, is output by weighting and summing the results using preset weight coefficients.

8. A groundwater-driven surface deformation identification system, characterized in that, include: The training sample input data processing module is used to acquire a multi-source spatiotemporal data cube containing deformation feature data, driving factor data and auxiliary feature data, perform spatial resampling and temporal interpolation alignment processing on the multi-source spatiotemporal data cube, and output the multi-source spatiotemporal data cube formatted as a multi-dimensional tensor. The target label construction module for training samples is used to calculate three probabilities based on a weakly supervised learning strategy: expert knowledge-driven spatial prior probability, statistical temporal correlation probability, and physical model forward modeling probability. The three probabilities are dynamically weighted and fused to output a soft label for groundwater driving probability. The soft label for groundwater driving probability is combined with the deformation feature data to output a candidate groundwater-driven deformation reference constraint. The spatiotemporal deep network model construction module is used to construct a spatiotemporal deep learning network that includes a convolutional neural network module, a temporal convolutional network module, an attention module, and a dual-output branch structure. The formatted multi-source spatiotemporal data cube is input into the spatiotemporal deep learning network to extract spatiotemporal fusion features. The first branch of the dual-output branch structure is used to output a probability map of groundwater-driven surface deformation, and the second branch of the dual-output branch structure is used to output a candidate groundwater-driven deformation sequence. The model training and result output module is used to construct a joint loss function, which includes a probabilistic identification loss term constructed based on the groundwater-driven probability soft label, a physical consistency regularization term gated and adaptively modulated by the predicted groundwater-driven surface deformation probability map, and a probabilistic deformation space consistency constraint term that constrains the collaboration between the probability gradient and the deformation gradient. The spatiotemporal deep learning network is iteratively trained based on the joint loss function and the candidate groundwater-driven deformation reference constraints. The test spatiotemporal data cube of the region to be identified is input into the trained spatiotemporal deep learning network. The output target groundwater-driven surface deformation probability map is multiplied pixel by pixel with the target candidate groundwater-driven deformation sequence to output the final groundwater-driven deformation component.

9. An electronic device, characterized in that, The method includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of a groundwater-driven surface deformation identification method as described in any one of claims 1-7.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions, which, when executed by a processor, implement the steps of a groundwater-driven surface deformation identification method as described in any one of claims 1-7.

Images