A displacement prediction space domain construction method based on landslide scale mapping

Abstract

The application discloses a displacement prediction space domain construction method based on landslide scale mapping, belongs to the technical field of geological disaster monitoring, and comprises the following steps: obtaining multi-temporal remote sensing observation data of a research area, and inversely obtaining a displacement time sequence; constructing a quantitative scale mapping relationship between a prediction space domain and landslide scale based on landslide space scale parameters, introducing a displacement time sequence space correlation as a physical constraint for adaptive correction, and determining an optimal prediction space domain; selecting a monitoring point in the prediction space domain and preprocessing the displacement time sequence; fusing the displacement time sequence and external driving factors to construct an input factor set; using a deep learning model to perform displacement prediction; after quantitatively evaluating the prediction result, constructing a space domain quality evaluation index, and if a threshold value is not reached, triggering a feedback optimization mechanism, reversely adjusting mapping parameters and iterating until the quality requirement is met. The application effectively reduces the subjectivity of space range selection, and improves the stability and applicability of landslide displacement prediction.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring technology, and in particular to a method for constructing a spatial domain for displacement prediction based on landslide scale mapping. Background Technology

[0002] Landslides are a common geological hazard in mountainous and reservoir areas, and their deformation and evolution processes exhibit significant spatiotemporal nonlinear characteristics. With the development of Synthetic Aperture Radar Interferometry (InSAR) technology, landslide surface displacement monitoring based on time-series InSAR inversion has become an important means of identifying and analyzing long-term landslide deformation. Furthermore, combining this with deep learning models for landslide displacement prediction has gradually become a research hotspot in the field of landslide disaster prediction and forecasting.

[0003] In existing technologies, landslide displacement prediction typically uses single-point displacement time series or displacement information from multiple monitoring points within a certain spatial range as model input. In practical applications, the selection of the prediction input spatial range often relies on empirical judgment or manually set fixed buffer radii, lacking a unified and transferable determination criterion across different research objects. If the selected spatial range is too small, it is difficult to fully reflect the overall deformation characteristics of the landslide; if the spatial range is too large, it easily introduces redundant information with weak correlation to the target prediction point, reducing the stability and generalization ability of the prediction model.

[0004] Existing methods generally neglect the impact of variations in the spatial scale of landslides on the rationality of the predicted input spatial range. Landslides of different scales exhibit significant differences in spatial structure, deformation coordination, and controlling factors. Directly adopting a uniform or empirically defined predicted spatial range is difficult to apply to landslides of different scales and types. Therefore, establishing the intrinsic correlation between the spatial scale of landslides and the predicted displacement input spatial range, and thereby rationally determining the predicted spatial domain, has become a crucial technical issue restricting the accuracy and applicability of landslide displacement prediction. Summary of the Invention

[0005] The purpose of this invention is to provide a displacement prediction spatial domain construction method based on landslide scale mapping, addressing the problems of strong subjectivity in the selection of the prediction input spatial range, lack of scale adaptability, and difficulty in promoting its application among landslides of different scales in the existing landslide displacement prediction process.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for constructing a spatial domain for displacement prediction based on landslide scale mapping, comprising the following steps: S1. Obtain multi-temporal relative ground observation data of the study area and invert to obtain the displacement time series of the landslide area; S2. Based on the landslide spatial scale parameters, construct a quantitative scale mapping relationship between the prediction spatial domain and the landslide scale, and introduce displacement temporal spatial correlation as a physical constraint to adaptively correct the quantitative scale mapping relationship and determine the optimal prediction spatial domain that matches the target landslide scale. S3. Select monitoring points within the optimal prediction spatial domain and preprocess the displacement time series of the monitoring points to obtain a standardized displacement time series. S4. The standardized displacement time series is fused with the external driving factors of landslide deformation to construct a set of input factors for landslide displacement prediction. S5. Based on the set of input factors, landslide displacement is predicted using a deep learning model to obtain the displacement prediction result; S6. The displacement prediction results are quantitatively evaluated using mean square error, root mean square error, mean absolute error and coefficient of determination to obtain prediction accuracy index. S7. Based on the prediction accuracy index and the deformation synergy coefficient in the physical constraints, construct a spatial domain quality evaluation index and determine whether the quality evaluation index meets the preset threshold: if it does, then take the current optimal prediction spatial domain as the final prediction spatial domain; if it does not, then trigger the feedback optimization mechanism, reverse the adjustment of the parameters in the quantitative scale mapping relationship and adaptive correction, and jump to step S2 to re-execute until the quality requirements are met.

[0007] A storage medium storing instructions and data for implementing a displacement prediction spatial domain construction method based on landslide scale mapping.

[0008] A displacement prediction spatial domain construction device based on landslide scale mapping includes: a processor and the storage medium; the processor loads and executes instructions and data in the storage medium to implement a displacement prediction spatial domain construction method based on landslide scale mapping.

[0009] The beneficial effects of this invention are as follows: This invention introduces a landslide scale mapping mechanism to establish a quantitative correlation between landslide spatial scale and displacement prediction spatial domain, effectively overcoming the problem of existing methods where the prediction input range depends on empirical selection. Furthermore, the scale mapping ratio of this mechanism is obtained based on known landslide examples, making it applicable to landslides of different scales and types. The results of this invention can provide a scientific and reusable basis for determining the spatial domain for deformation prediction and risk assessment of geological disasters such as reservoir landslides and mountain landslides, demonstrating promising engineering application prospects. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the technical route of the method of the present invention; Figure 2 This is a schematic diagram of the principle architecture of the multi-scale hierarchical attention BiGRU model; Figure 3 This is a schematic diagram of the displacement sequence, daily rainfall, and daily reservoir water level of the selected points; Figure 4 This is a comparison chart of the prediction results for PSI-01 points; Figure 5 This is a schematic diagram of the hardware device of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific examples described herein are merely illustrative and not intended to limit the scope of the invention.

[0012] Before formally describing the present invention, a general description of the solution of the present invention will be given first to facilitate understanding.

[0013] Example 1:

[0014] Please refer to Figure 1 The present invention provides a method for constructing a spatial domain for displacement prediction based on landslide scale mapping, comprising the following steps: S1. Obtain multi-temporal relative ground observation data of the study area and invert to obtain the displacement time series of the landslide area; Specifically, multi-temporal synthetic aperture radar image data covering the study area is acquired, and the radar images are processed using the small baseline set temporal interferometry method to invert and obtain the time series of surface displacement in the landslide area.

[0015] S2. Based on the landslide spatial scale parameters, construct a quantitative scale mapping relationship between the prediction spatial domain and the landslide scale, and introduce displacement temporal spatial correlation as a physical constraint to adaptively correct the quantitative scale mapping relationship and determine the optimal prediction spatial domain that matches the target landslide scale. It should be noted that step S2 specifically includes: S21. Taking the target prediction point as the center, the prediction spatial domain is divided into three layers: local scale spatial domain, regional scale spatial domain, and global scale spatial domain. The radius of each layer is determined by the basic radius R and the scale coefficients β1, β2, and β3, and satisfies 0 < β1 < β2 < β3 = 1. Specifically, with the target prediction point Centered on the base space domain, the prediction spatial domain is equivalent to a buffer zone, with the radius of the base spatial domain as the buffer zone. The spatial domain is divided into three layers:

[0016]

[0017] Local scale spatial domain: Regional-scale spatial domain: Global scale spatial domain: in: For the first Layer space domain radius; It is a scaling factor that satisfies .

[0018] S22. Using the known landslide area A of the landslide sample and the corresponding optimal prediction spatial domain area S, establish the initial scale mapping coefficient k = S / A; Specifically, through the known landslide area of ​​landslide samples Area of ​​the corresponding optimal prediction spatial domain Establish a proportional mapping relationship and construct an initial proportional mapping relationship.

[0019]

[0020] in, The landslide scale mapping coefficient.

[0021] S23. Within each scale spatial domain, select the displacement time series of all monitoring points and calculate the correlation coefficient r between any two points. ij Furthermore, the average correlation coefficient of all monitoring point pairs in the entire spatial domain is calculated as the deformation synergy coefficient ρ; Specifically, after the initial proportional relationships are determined, in each scale spatial domain Within the selected spatial domain, the deformation synergy coefficient among all monitoring points in that spatial domain is calculated. It is used to characterize the consistency and coordination of deformation within a landslide.

[0022] Suppose that the candidate spatial domain contains The displacement time series of each monitoring point are as follows: Then any two points and The correlation can be expressed as:

[0023] In the formula: For monitoring points and The correlation coefficient; The length of the time series; For point The time series mean.

[0024] The deformation compatibility coefficient of the spatial domain is:

[0025] S24. When the deformation coordination coefficient ρ is lower than the preset coordination threshold ρ0, the initial scale mapping coefficient k is corrected using the coordination adjustment function g(ρ) to obtain the corrected scale mapping coefficient. and based on The final optimal prediction spatial domain area is determined by the landslide area A. When ρ ≥ ρ0, the optimal predicted spatial domain area S = k·A is determined directly using the initial scale mapping coefficient k.

[0026] Specifically, based on a known landslide sample database, the overall deformation synergy coefficient of landslides of different sizes was statistically analyzed. The correlation with prediction accuracy determines the deformation synergy constraint threshold. .

[0027] (1) When This indicates that the overall deformation of the landslide is well coordinated, and the deformation trends at each monitoring point are consistent. The current spatial domain can effectively characterize the overall deformation features of the landslide, and the basic proportional relationship can be directly used. ; (2) When This indicates that there is significant spatial differentiation in the deformation within the landslide, with poor overall coordination. The basic proportional relationship cannot accurately depict the effective influence range, requiring correction of the proportional mapping relationship. The corrected proportional coefficient is... ,in The formula for calculating the synergistic adjustment function is as follows:

[0028]

[0029] In the formula: This is an adjustment coefficient, and its value range is... , The smaller, The smaller, the better. The smaller the value, the smaller the predicted spatial domain, in order to avoid introducing too much spatial heterogeneous information into areas with low coherence and improve prediction accuracy.

[0030] Based on the corrected scaling factor With landslide area The final predicted spatial domain area at different scales was calculated. Based on this, the optimal prediction spatial domain matching different landslide scales is determined.

[0031]

[0032] S3. Select monitoring points within the optimal prediction spatial domain and preprocess the displacement time series of the monitoring points to obtain a standardized displacement time series. It should be noted that the preprocessing of the displacement time series in step S3 specifically includes: S31. Construct a spatial buffer zone centered on the target prediction point, based on the optimal prediction spatial domain radius determined in step S2; specifically, construct a spatial buffer zone centered on the target prediction point. Centered on the radius of the spatial domain at different scales determined in step S2 Construct the corresponding spatial buffer area.

[0033] S32. Select monitoring points with a valid data ratio greater than the preset threshold η in each scale spatial domain to form a monitoring point set; In spatial domains of various scales Internal screening selects valid monitoring points with continuous observation records to form a monitoring point set. ,satisfy:

[0034] in: It is a monitoring point The number of effective monitoring data; It is the total time series length; This is the threshold for the proportion of valid data.

[0035] S33. Calculate the average deformation direction of all monitoring points in the set of monitoring points. and will satisfy The monitoring points were removed as anomalies with the opposite deformation direction; Specifically, considering the consistent directional deformation of the landslide as a whole, the average deformation direction within the calculation area is as follows:

[0036] If a monitoring point satisfies the following relationship, it indicates that the deformation direction of that point is opposite to the overall deformation trend, and it is judged as an abnormal monitoring point and removed:

[0037] S34. Perform linear interpolation on the missing values ​​in the displacement time series of the remaining monitoring points to obtain a continuous time series; Specifically, the displacement time series of the selected monitoring points is processed using a linear interpolation method to ensure the continuity of the time series:

[0038] in: , This represents the most recent valid observation time before and after the missing point.

[0039] S35. Resample the continuous time series to a uniform time interval Δt to form a uniform time axis; S36. The resampled time series is smoothed using a moving average method to eliminate high-frequency noise; as shown in the following formula:

[0040] in: This refers to the window size.

[0041] S37. Normalize all smoothed displacement time series to obtain standardized displacement time series, as shown in the following formula:

[0042] S38. Average and aggregate the standardized displacement time series of each monitoring point in different scale spatial domains to construct a hierarchical feature vector F(t)=[F local (t), F region (t), F global (t)].

[0043] Specifically, let the first Layer space domain ( (Corresponding to local scale, mesoscale, and global scale respectively) contains The displacement time series of the monitoring points is as follows: Then the feature at this scale is represented as

[0044] Features at different scales are combined to construct a hierarchical feature vector:

[0045] S4. The standardized displacement time series is fused with the external driving factors of landslide deformation to construct a set of input factors for landslide displacement prediction. It should be noted that the specific set of input factors for landslide displacement prediction constructed in step S4 includes: S41. Obtain external driving factors related to landslide deformation, including rainfall, reservoir water level and previous cumulative displacement. S42. Perform time scale unification and normalization processing on the external driving factor and the standardized displacement time series obtained in step S3. S43. Combine the normalized displacement time series with the external driving factors to form an initial input factor set; S44. Perform grey correlation analysis on each factor in the initial input factor set and the landslide displacement, and calculate the grey correlation degree. Specifically, grey correlation analysis is performed on the set of input factors to calculate the correlation between each input factor and the landslide displacement:

[0046]

[0047] in: The resolution coefficient (usually taken as 0.5); This represents the grey relational degree.

[0048] S45. Select the top-ranked key influencing factors based on the degree of grey relational analysis to form an optimized set of input factors.

[0049] S5. Based on the set of input factors, landslide displacement is predicted using a deep learning model to obtain the displacement prediction result. It should be noted that the deep learning model in step S5 is a multi-scale hierarchical attention bidirectional gated recurrent unit model, and the specific prediction steps include: S51. The optimized set of input factors obtained in step S4 is fused with the hierarchical feature vector F(t) obtained in step S3 to construct the model input vector Z(t). Specifically, the hierarchical feature vector constructed in step S3 The input vector is then fused with the filtered set of input factors to construct the model input vector. :

[0050]

[0051] S52. Input the model input vector Z(t) into a bidirectional gated recurrent unit (GRU). Extract forward and backward time-series features through forward and backward GRU branches respectively. Then, concatenate the hidden states in the two directions to obtain the bidirectional time-series feature h. t ; Please refer to Figure 2 Specifically, BiGRU consists of two independent gated recurrent units (GRUs), which process the input sequence in the forward and backward directions of time, respectively. The core computational unit of the GRU contains a reset gate. and Update Gate Its mathematical expression is:

[0052] in, The input vector at the current time step. This is the hidden state of the previous time step. This indicates element-wise multiplication. It is the Sigmoid activation function. and These are the weight matrix and bias term to be trained, respectively.

[0053] Forward GRU from arrive The forward hidden state sequence is obtained by calculating sequentially. ; Reverse GRU from arrive Reverse computation yields the backward hidden state sequence. By concatenating the hidden states of two directions at the same time step, a bidirectional temporal feature is obtained. : This concatenated vector contains contextual information from both historical and future moments, enabling a more comprehensive depiction of the dynamic patterns of landslide displacement evolution.

[0054] S53. Introduce a hierarchical attention mechanism to address local scale features F. local (t), Regional scale characteristics F region (t) and global scale features F global (t) Calculate the attention weight w l Multi-scale weighted fusion features are obtained by weighted summation. ; Specifically, to adaptively emphasize the importance of features at different spatial scales to the current prediction task, a hierarchical attention mechanism is introduced. For each time step... Attention weights are calculated for features at three scales: local, regional, and global.

[0055] First, each scale feature ( (Corresponding to local, regional, and global levels respectively) are mapped to the latent space through a single-layer perceptron:

[0056] in, , , These are learnable parameters.

[0057] Then, the Softmax function is used to measure the three scales. Normalization is performed to obtain the attention weights. :

[0058] Obviously, The larger the weight, the more important that scale feature is to the prediction result at the current time step.

[0059] Finally, the features at each scale are weighted and summed to obtain the multi-scale weighted fusion features. :

[0060] S54, the bidirectional timing feature h t With the multi-scale weighted fusion features The features are spliced ​​and fused to obtain the comprehensive feature representation H(t); Specifically, the bidirectional timing features obtained in step S52 The multi-scale weighted fusion features obtained in step S53 The features are concatenated along the feature dimension to form a comprehensive feature representation. :

[0061] in This represents a vector concatenation operation. The fused representation includes both the bidirectional dependency of landslide deformation on the time axis and multi-scale spatial information weighted by an attention mechanism.

[0062] S55. Input the comprehensive feature representation H(t) into the fully connected layer and output the predicted landslide displacement value. .

[0063] Specifically, the comprehensive feature representation One or more fully connected layers (also known as dense layers) are input and mapped to the target output dimension (in this embodiment, a single output value, i.e., the displacement prediction value for the next time step) through linear transformation and nonlinear activation:

[0064] in, This is the weight matrix of the fully connected layer. For bias terms, This is the predicted landslide displacement output by the model.

[0065] S6. The displacement prediction results are quantitatively evaluated using mean square error, root mean square error, mean absolute error and coefficient of determination to obtain prediction accuracy index. Specifically, step S6 includes: Step S61: Calculate the mean square error (MSE), using the following formula: ;in, This represents the true displacement value. This is the predicted displacement value. The number of samples; Step S62: Calculate the Mean Absolute Error (MAE), using the following formula: ;in, This represents the true displacement value. The sample mean. The number of samples; Step S63: Calculate the root mean square error (RMSE), using the following formula: ;in, This represents the true displacement value. This is the predicted displacement value. The number of samples; Step S64: Calculate the coefficient of determination R 2 The formula is ;in, This represents the true displacement value. This is the predicted displacement value. The sample mean. This represents the number of samples.

[0066] S7. Based on the prediction accuracy index and the deformation synergy coefficient in the physical constraints, construct a spatial domain quality evaluation index and determine whether the quality evaluation index meets the preset threshold: if it does, then take the current optimal prediction spatial domain as the final prediction spatial domain; if it does not, then trigger the feedback optimization mechanism, reverse the adjustment of the parameters in the quantitative scale mapping relationship and adaptive correction, and jump to step S2 to re-execute until the quality requirements are met.

[0067] It should be noted that step S7, which involves constructing spatial domain quality evaluation indicators and triggering a feedback optimization mechanism, specifically includes: S71. The coefficient of determination R in the prediction accuracy index obtained in step S6 2 Using the root mean square error (RMSE) and the deformation synergy coefficient (ρ) calculated in step S23, a comprehensive evaluation index for spatial domain quality, Q = ω1·[β1·R], is constructed. 2 + β2·(1 - RMSE / max(RMSE))] + ω2·ρ, where ω1+ω2=1, β1+β2=1; Specifically, based on the prediction error index obtained in step S6 and the deformation synergy coefficient calculated in step S2... Construct a unified spatial domain quality evaluation index:

[0068] in, , The predicted evaluation index obtained in step S6; The deformation synergy coefficient calculated in step S2; , For the weighting coefficients, satisfying ; , is the internal weighting coefficient of the accuracy index; max(RMSE) is the maximum RMSE value recorded in the historical iteration optimization process of the same landslide object (if there is no historical record, the RMSE value calculated under the initial spatial domain can be used as the benchmark). The physical meaning of the above fusion formula is that: when the prediction accuracy is high ( big, Small) and good deformation coordination within the spatial domain ( When (large), The value approaches 1; conversely, The value is relatively small. This indicator can be used to unify two evaluation indicators with different dimensions onto the same scale for comparison.

[0069] S72. Set the spatial domain quality threshold Q0 and determine whether Q is greater than or equal to Q0: if yes, determine that the current best predicted spatial domain is reasonable and use it as the final predicted spatial domain; if no, trigger the spatial domain reconstruction mechanism. In order to make an objective judgment on the rationality of the current spatial domain, a mass threshold needs to be set. This embodiment uses the statistical quantile method to determine... The specific steps are as follows: In the initial prediction space domain (i.e., using empirical initial values) , , Based on this, a set of prediction results are obtained through steps S2 to S6, and the corresponding... .

[0070] For key parameters ( , , To perform small-scale disturbances (e.g.) Multiply by 0.8 and 1.2 respectively. Adding or subtracting 0.1, etc., generates several candidate spatial domains.

[0071] Calculate the quality evaluation index for each candidate spatial domain. , and obtained a series value.

[0072] Take the above The 75th percentile (third quartile) of the value sequence is used as the quality threshold. .

[0073] The discrimination rule is: when At this time: it is determined that the current prediction spatial domain is reasonably selected and no further adjustment is needed. This spatial domain and its corresponding mapping parameters ( , , The iteration ends when the final prediction spatial domain for the target landslide is determined.

[0074] when When the current predicted spatial domain quality is determined to be unsatisfactory, the spatial domain reconstruction mechanism (i.e., feedback optimization) is triggered, and step S73 is entered.

[0075] S73. In the spatial domain reconstruction mechanism, a grid search method is used to adaptively adjust at least one of the following parameters: scale mapping coefficient k, adjustment coefficient α in the cooperative adjustment function, and scale coefficients β1, β2, and β3 of the multi-scale hierarchical spatial domain. Specifically, once the feedback optimization mechanism is triggered, a systematic adjustment needs to be made to the core parameters affecting the spatial domain construction. In this embodiment, the parameters to be adjusted include: Scale mapping coefficients : Directly controls the ratio between the predicted spatial area and the landslide area.

[0076] The adjustment coefficient in the cooperative adjustment function : The extent of expansion of the time-space domain that affects poor synergy ( The larger the size, the stronger the scalability.

[0077] Scale coefficients of multi-scale hierarchical spatial domains : Controls the relative size of the three spatial domains: local, regional, and global.

[0078] This embodiment uses a grid search method for adaptive parameter adjustment, and the specific implementation method is as follows: Search scope settings: Search scope: Step size ,in These are the scale mapping coefficient values ​​currently in use. Search scope: Step size . Search scope: , , (Fixed), and guaranteed Step size .

[0079] Search execution flow: Perform a Cartesian product combination on the above parameter space to generate a candidate parameter set. For each set of candidate parameters, repeat steps S2 (spatial domain partitioning and correction) and S3 (monitoring point selection and preprocessing). For the newly constructed spatial domain, repeat steps S4 to S6 (model training and prediction). Calculate the quality evaluation index obtained under this set of parameters. Records make The value is the largest and satisfies The first set of parameters is used as the output of this optimization; if all candidate parameters do not satisfy... Then select The set of parameters with the largest value is selected and proceeded to the next iteration.

[0080] In the first iteration of this embodiment, the initial parameters are: (Obtained from Example 2) , , , Calculated less than the threshold This triggers feedback optimization.

[0081] The optimal parameters found through grid search are: (Increase of approximately 15%) , , , After rerunning with these parameters, Value increased to ,satisfy Stop iterating.

[0082] S74. Based on the adjusted parameters, repeat steps S2 to S6 and recalculate the new quality evaluation index Q. Repeat this process until Q ≥ Q0.

[0083] Specifically, this feedback optimization mechanism adopts an iterative loop structure, and its process can be summarized as follows: Initialization: Set the current optimal parameter set as the initial parameters. Maximum number of iterations quality threshold .

[0084] Loop begins: For each iteration Based on current parameters Perform steps S2 to S6 to calculate the quality evaluation indicators. .like If the condition is met, the loop terminates and the current parameter is output as the final result. and Then, perform the grid search in step S73 to obtain a new set of parameters. ,make Then increment the iteration count by 1. If the condition is still not met after reaching the maximum number of iterations, then take the iteration count from the previous iteration. The set of parameters with the largest value is taken as the final result, and a warning message is recorded (indicating that this landslide object may require manual intervention or is determined to be a special case of extremely inconsistent deformation).

[0085] Example 2:

[0086] Step S1: SBAS-InSAR inversion displacement time series; specifically: Multi-temporal synthetic aperture radar imagery data covering the Muyubao landslide area was acquired, and the radar images were processed using Small Baseline Set Interferometry (SBAS-InSAR) technology to invert the time series of surface displacement in the Muyubao landslide area, demonstrating the temporal deformation characteristics of the landslide during the observation period, and providing basic data for subsequent spatial domain determination and displacement prediction.

[0087] It should be noted that the method of acquiring displacement time series data in this step is not limited to the specific InSAR inversion method, but can also use other remote sensing or ground monitoring methods that can obtain landslide time series displacement information.

[0088] Step S2: Based on the spatial morphology and boundary information of the Muyubao landslide, extract its spatial scale parameters, construct the mapping relationship between the landslide scale and the displacement prediction spatial domain, and introduce the deformation synergy coefficient as a physical constraint to achieve adaptive determination of the prediction spatial domain, thereby determining a prediction spatial domain that matches the target landslide scale. Specifically: The monitoring point PSI-01, which shows significant displacement changes within the landslide, was selected as the target prediction point. Centered on this target prediction point, the spatial domains at different scales were equivalently represented as multiple ring buffer zones, and the spatial domain area was used as the basis for prediction. Calculate the radius of the basic spatial domain Introducing a scaling factor The spatial domain is divided into three scales: local scale, regional scale, and global scale. Each candidate prediction spatial domain is used to cover the local deformation area of ​​the landslide.

[0089] Based on the landslide area Area of ​​the optimal prediction spatial domain Calculate the landslide scale mapping ratio The calculation method is as follows:

[0090] After obtaining the initial scale mapping relationship, corresponding InSAR monitoring points are selected in the spatial domain at each scale, their displacement time series are extracted, and the correlation between the monitoring points is calculated to characterize the consistency of deformation within the landslide. Assume the spatial domain contains... The displacement time series of each monitoring point are as follows: Then the correlation coefficient between any two points can be expressed as:

[0091] Calculate the overall deformation compatibility coefficient of the spatial domain Its expression is:

[0092] Based on the statistical results of existing landslide samples, the deformation synergy threshold was determined. .when This indicates that the internal deformation of the landslide exhibits good overall coordination, and the current spatial domain can effectively characterize the landslide deformation features. Therefore, the initial proportional mapping coefficients can be directly used. ;when This indicates that there is spatial differentiation in the deformation within the landslide, requiring correction of the proportional mapping relationship and the introduction of a synergistic adjustment function. The corrected proportionality coefficient is obtained. :

[0093]

[0094] in: This is an adjustment coefficient, and its value range is... By adjusting the scaling factor through the adjustment function, the predicted spatial domain range can be adaptively adjusted according to the synergistic effect of landslide deformation.

[0095] (6) Based on the corrected scaling factor With landslide area The final predicted spatial domain area at different scales was calculated. Based on this, a multi-scale prediction spatial domain matching the spatial scale of the Muyubao landslide was determined, providing spatial range constraints for subsequent selection of monitoring points and construction of displacement prediction models.

[0096]

[0097] Please refer to Figure 3 In this embodiment, the Muyubao landslide is taken as the research object. Through multiple rounds of iterative optimization, the optimal prediction spatial domain that matches the spatial scale of the Muyubao landslide is finally determined. Under the conditions of this embodiment, its area is S = 250πm. 2 The corresponding scale mapping ratio The scale mapping ratio, as a mapping parameter between the landslide spatial scale and the displacement prediction spatial domain, can be used to guide the construction of the prediction spatial domain for other landslide objects.

[0098] It should be noted that the numerical value is the experimental result obtained under specific landslide objects and data conditions in this embodiment, and its specific value is not intended to limit the scope of protection of this invention. Under different landslide types, different spatial morphologies, or different monitoring data conditions, the scale mapping ratio can vary within a reasonable range.

[0099] Step S3: Selection and preprocessing of monitoring points within the prediction spatial domain; specifically: Centered on the target prediction point PSI-01, based on the spatial domain radii of different scales determined in S2 The corresponding spatial buffer is constructed, and InSAR monitoring points with positive deformation rates and continuous observation records are selected within the buffer. The displacement time series of the selected monitoring points are processed by outlier removal and missing value interpolation. Then, wavelet transform is used to smooth the interpolated displacement time series to reduce the interference of high-frequency noise on displacement prediction and enhance the trend and stability of the displacement series.

[0100] It should be noted that during the spatial domain iterative optimization process in step S2, the buffer range changes synchronously with the dynamic adjustment of the predicted spatial domain scale, and correspondingly, the set of monitoring points participating in the modeling is also updated. If the spatial domain quality requirements are not met in step S2, the data selection and preprocessing process in this step needs to be repeated with the reselection of the spatial domain. After the above processing, standardized displacement time series data corresponding to the current predicted spatial domain are obtained, providing basic data support for subsequent input factor construction and model training.

[0101] Step S4: Constructing the prediction input factor set; specifically: The preprocessed displacement time series is fused with external driving factors related to landslide deformation. These external driving factors include 12 influencing factors, such as the average reservoir water level of the previous 12 days, the previous 24 days, the previous 36 days, the reservoir water level change, the cumulative rainfall, and the cumulative displacement, to construct an input factor set for landslide displacement prediction. Further grey correlation analysis was performed on the input factor set to quantify the correlation between each influencing factor and landslide displacement changes. Based on the correlation strength, key influencing factors with high correlation to landslide displacement changes were selected to form an optimized input dataset. This reduced the interference of redundant information on the prediction model and improved prediction accuracy and model stability. After all factors were constructed, the dataset was normalized to form standardized input data.

[0102] Step S5: Multi-scale hierarchical attention BiGRU model predicts displacement; specifically: The optimized input dataset obtained in step S4 is input into a Bidirectional Gated Recurrent Unit (BiGRU) model with added multi-scale hierarchical attention. The BiGRU model is a bidirectional extension of the GRU, and its core includes two major control modules: a reset gate and an update gate. The reset gate is used to control the degree of forgetting of historical information, and the update gate is used to adjust the fusion ratio of historical information and current input information, thereby effectively alleviating the gradient vanishing or gradient explosion problems in traditional recurrent neural networks and enhancing the ability to model long-term temporal dependencies. The temporal features of the input factors are extracted bidirectionally through the forward GRU branch and the backward GRU branch, and the bidirectional hidden states are spliced ​​and fused to obtain a temporal feature representation containing historical and future contextual information.

[0103] Building upon this foundation, a hierarchical attention mechanism is introduced to address the multi-scale feature differences in landslide displacement evolution. This mechanism adaptively weights features at different time scales. Specifically, the input features are divided into local-scale, meso-scale, and global-scale features. Importance weights for each scale are calculated, and normalization is applied to obtain the attention weight coefficients for each scale. These attention weights are then used to weight and fuse the multi-scale features, resulting in a comprehensive multi-scale feature representation. This representation is then fused with the temporal features extracted by BiGRU to form a final high-dimensional feature expression. Subsequently, the fused features are input into a fully connected layer, outputting landslide displacement prediction results. This enables refined prediction of the displacement changes of the Muyubao landslide. For details, please refer to [reference needed]. Figure 4 .

[0104] The specific model building process is as follows: (1) The fused input sequence Input the BiGRU model, obtain the hidden states, and concatenate them:

[0105] in: (2) For multi-scale hierarchical features, a hierarchical attention mechanism is introduced to calculate the weights at different scales:

[0106]

[0107] in: These correspond to local scale, mesoscale, and global scale, respectively. Corresponding scale weights.

[0108] Multi-scale features are weighted and fused using attention weights:

[0109] Further integration with BiGRU features: ( (Indicates splicing or merging operation) The prediction results are output through the fully connected layer: It should be noted that during the iterative optimization process from steps S2 to S6, as the prediction spatial domain and input dataset are dynamically updated, the BiGRU model and hierarchical attention mechanism are retrained and re-predicted based on the updated data in each iteration round to ensure that the model can adapt to the landslide deformation characteristics under different spatial scales and improve the accuracy and stability of the prediction results.

[0110] Step S6: Quantitatively verify the accuracy using multiple indicators; specifically: The mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) were used. 2 The accuracy of the landslide displacement prediction results was verified. A training set of 126 time steps from March 12, 2017 to April 20, 2021 was used, and a test set of 32 time steps from May 2, 2021 to May 9, 2022 was used. The prediction results show that, under the test set conditions, all accuracy evaluation indicators performed well, and the prediction results showed a high degree of consistency with the actual displacement change trend. This indicates that the method of the present invention has good prediction accuracy and stability under the determined prediction spatial domain conditions, and can meet the practical application requirements of landslide displacement prediction.

[0111] Step S7: Spatial domain quality evaluation and feedback optimization mechanism; specifically: A spatial domain quality evaluation index system is constructed based on the landslide displacement prediction results to quantitatively evaluate the rationality of the prediction spatial domain. Based on the evaluation results, the spatial domain is optimized through feedback to achieve dynamic adjustment of the prediction spatial domain.

[0112] Specifically, the steps include the following: (1) Construction of spatial domain quality evaluation index Based on the prediction accuracy evaluation index obtained in S6 and the deformation synergy coefficient ρ calculated in S2, a unified comprehensive evaluation index Q for spatial domain quality is constructed to measure the rationality and effectiveness of the current prediction spatial domain. Its expression is as follows: Where: R 2 RMSE is the prediction accuracy evaluation index obtained in S6; ρ is the deformation synergy coefficient calculated in S2; ω1 and ω2 are weight coefficients, satisfying ω1 + ω2 = 1; β1 and β2 are the internal weight coefficients of the prediction accuracy index, used to adjust the degree of influence of different evaluation indicators on the comprehensive evaluation result. By constructing the above indicators, the prediction accuracy and spatial deformation consistency are quantified in a unified manner, so as to achieve a comprehensive evaluation of the quality of the prediction spatial domain. (2) Construction of spatial domain quality discrimination and feedback mechanism Based on the Q-value distribution calculated from the candidate spatial domain set, a statistical method is used to determine the spatial domain quality evaluation threshold Q0. Specifically, according to the distribution characteristics of the Q-values, the quantile (75%) is selected as the threshold Q0. For the Muyubao landslide, the calculated threshold Q0 = 0.738. When the comprehensive evaluation index Q meets different conditions, corresponding strategies are implemented: When Q ≥ Q0, it indicates that the current prediction spatial domain is reasonably selected and no spatial domain adjustment is needed. This spatial domain and its corresponding scale mapping relationship are taken as the final prediction spatial domain of the target landslide.

[0113] When Q < Q0, it indicates that the current predicted spatial domain fails to effectively characterize the landslide deformation features, triggering a spatial domain reconstruction mechanism to adaptively update key parameters in S2, including: Scale mapping parameters: Spatial domain extension control parameters: Multi-scale weight parameters: in, , and To adjust the step size of the parameters, a grid search method is used to update the parameters. The parameter update step size is no longer set to a fixed value, but is adaptively determined by the search strategy within the parameter space.

[0114] (3) Spatial domain iterative optimization process After completing the parameter adjustment, S2 to S6 are executed again to select data, construct inputs, train the model and evaluate accuracy for the updated prediction spatial domain, and recalculate the spatial domain quality index Q. Repeat the above process until the spatial domain quality requirement (Q ≥ Q0) is met, then stop the iteration.

[0115] Example 3:

[0116] Please see Figure 5 , Figure 5 This is a schematic diagram of the hardware device in operation according to an embodiment of the present invention. The hardware device specifically includes: a displacement prediction spatial domain construction device 401 based on landslide scale mapping, a processor 402, and a storage medium 403.

[0117] A displacement prediction spatial domain construction device 401 based on landslide scale mapping: The displacement prediction spatial domain construction device 401 based on landslide scale mapping implements the displacement prediction spatial domain construction method based on landslide scale mapping.

[0118] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the displacement prediction spatial domain construction method based on landslide scale mapping.

[0119] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the displacement prediction spatial domain construction method based on landslide scale mapping.

[0120] This invention is not limited to the specific embodiments described above. Those skilled in the art can implement this invention using various other specific embodiments based on the content disclosed herein. Therefore, any design that adopts the design structure and concept of this invention and makes some simple changes or modifications falls within the scope of protection of this invention.

Claims

1. A method for constructing a spatial domain for displacement prediction based on landslide scale mapping, characterized in that: Includes the following steps: S1. Obtain multi-temporal relative ground observation data of the study area and invert to obtain the displacement time series of the landslide area; S2. Based on the landslide spatial scale parameters, construct a quantitative scale mapping relationship between the prediction spatial domain and the landslide scale, and introduce displacement temporal spatial correlation as a physical constraint to adaptively correct the quantitative scale mapping relationship and determine the optimal prediction spatial domain that matches the target landslide scale. S3. Select monitoring points within the optimal prediction spatial domain and preprocess the displacement time series of the monitoring points to obtain a standardized displacement time series. S4. The standardized displacement time series is fused with the external driving factors of landslide deformation to construct a set of input factors for landslide displacement prediction. S5. Based on the set of input factors, landslide displacement is predicted using a deep learning model to obtain the displacement prediction result. S6. The displacement prediction results are quantitatively evaluated using mean square error, root mean square error, mean absolute error and coefficient of determination to obtain prediction accuracy index. S7. Based on the prediction accuracy index and the deformation synergy coefficient in the physical constraints, construct a spatial domain quality evaluation index and determine whether the quality evaluation index meets the preset threshold: if it does, then take the current optimal prediction spatial domain as the final prediction spatial domain; if it does not, then trigger the feedback optimization mechanism, reverse the adjustment of the parameters in the quantitative scale mapping relationship and adaptive correction, and jump to step S2 to re-execute until the quality requirements are met.

2. The method for constructing a spatial domain for displacement prediction based on landslide scale mapping according to claim 1, characterized in that: Step S2 specifically includes: S21. Taking the target prediction point as the center, the prediction spatial domain is divided into three layers: local scale spatial domain, regional scale spatial domain, and global scale spatial domain. The radius of each layer is determined by the basic radius R and the scale coefficients β1, β2, and β3, and satisfies 0 < β1 < β2 < β3 = 1. S22. Using the known landslide area A of the landslide sample and the corresponding optimal prediction spatial domain area S, establish the initial scale mapping coefficient k = S / A; S23. Within each scale spatial domain, select the displacement time series of all monitoring points and calculate the correlation coefficient r between any two points. ij Furthermore, the average correlation coefficient of all monitoring point pairs in the entire spatial domain is calculated as the deformation synergy coefficient ρ; S24. When the deformation coordination coefficient ρ is lower than the preset coordination threshold ρ0, the initial scale mapping coefficient k is corrected using the coordination adjustment function g(ρ) to obtain the corrected scale mapping coefficient. and based on The final optimal prediction spatial domain area is determined by the landslide area A. When ρ ≥ ρ0, the optimal predicted spatial domain area S = k·A is determined directly using the initial scale mapping coefficient k.

3. The method for constructing a spatial domain for displacement prediction based on landslide scale mapping according to claim 2, characterized in that: The specific cooperative adjustment function in step S24 is: g(ρ) = 1 - α(ρ0 - ρ), where α is an adjustment coefficient ranging from 0 to 1; the smaller ρ is, the smaller g(ρ) is, the smaller the corrected k is, and the smaller the area S of the optimal prediction spatial domain is.

4. The method for constructing a spatial domain for displacement prediction based on landslide scale mapping according to claim 1, characterized in that: Step S3, the preprocessing of the displacement time series, specifically includes: S31. Construct a spatial buffer zone centered on the target prediction point and based on the optimal prediction spatial domain radius determined in step S2. S32. Select monitoring points with a valid data ratio greater than the preset threshold η in each scale spatial domain to form a monitoring point set; S33. Calculate the average deformation direction of all monitoring points in the set of monitoring points. and will satisfy The monitoring points were removed as anomalies with the opposite deformation direction; S34. Perform linear interpolation on the missing values ​​in the displacement time series of the remaining monitoring points to obtain a continuous time series; S35. Resample the continuous time series to a uniform time interval Δt to form a uniform time axis; S36. The resampled time series is smoothed using the moving average method to eliminate high-frequency noise; S37. Normalize all smoothed displacement time series to obtain standardized displacement time series. S38. Average and aggregate the standardized displacement time series of each monitoring point in different scale spatial domains to construct a hierarchical feature vector F(t)=[F local (t), F region (t), F global (t)].

5. The method for constructing a spatial domain for displacement prediction based on landslide scale mapping according to claim 1, characterized in that: Step S4, which involves constructing the set of input factors for landslide displacement prediction, specifically includes: S41. Obtain external driving factors related to landslide deformation, including rainfall, reservoir water level and previous cumulative displacement. S42. Perform time scale unification and normalization processing on the external driving factor and the standardized displacement time series obtained in step S3. S43. Combine the normalized displacement time series with the external driving factors to form an initial input factor set; S44. Perform grey correlation analysis on each factor in the initial input factor set and the landslide displacement, and calculate the grey correlation degree. S45. Select the top-ranked key influencing factors based on the degree of grey relational analysis to form an optimized set of input factors.

6. The method for constructing a spatial domain for displacement prediction based on landslide scale mapping according to claim 4, characterized in that, The deep learning model in step S5 is a multi-scale hierarchical attention bidirectional gated recurrent unit model.

7. The method for constructing a spatial domain for displacement prediction based on landslide scale mapping as described in claim 6, characterized in that, The prediction steps in step S5 include: S51. The optimized set of input factors obtained in step S4 is fused with the hierarchical feature vector F(t) obtained in step S3 to construct the model input vector Z(t). S52. Input the model input vector Z(t) into a bidirectional gated recurrent unit (GRU). Extract forward and backward time-series features through forward and backward GRU branches respectively. Then, concatenate the hidden states in the two directions to obtain the bidirectional time-series feature h. t ; S53. Introduce a hierarchical attention mechanism to address local scale features F. local (t), Regional scale characteristics F region (t) and global scale features F global (t) Calculate the attention weight w l Multi-scale weighted fusion features are obtained by weighted summation. ; S54, the bidirectional timing feature h t With the multi-scale weighted fusion features The features are spliced ​​and fused to obtain the comprehensive feature representation H(t); S55. Input the comprehensive feature representation H(t) into the fully connected layer and output the predicted landslide displacement value. .

8. The method for constructing a spatial domain for displacement prediction based on landslide scale mapping according to claim 1, characterized in that, Step S7, which involves constructing spatial domain quality evaluation indicators and triggering a feedback optimization mechanism, specifically includes: S71. The coefficient of determination R in the prediction accuracy index obtained in step S6 2 Using the root mean square error (RMSE) and the deformation synergy coefficient (ρ) calculated in step S23, a comprehensive evaluation index for spatial domain quality, Q = ω1·[β1·R], is constructed. 2 + β2·(1 -RMSE / max(RMSE))] + ω2·ρ, where ω1+ω2=1, β1+β2=1; S72. Set the spatial domain quality threshold Q0 and determine whether Q is greater than or equal to Q0: if yes, determine that the current best predicted spatial domain is reasonable and use it as the final predicted spatial domain; if no, trigger the spatial domain reconstruction mechanism. S73. In the spatial domain reconstruction mechanism, a grid search method is used to adaptively adjust at least one of the following parameters: scale mapping coefficient k, adjustment coefficient α in the cooperative adjustment function, and scale coefficients β1, β2, and β3 of the multi-scale hierarchical spatial domain. S74. Based on the adjusted parameters, repeat steps S2 to S6 and recalculate the new quality evaluation index Q. Repeat this process until Q ≥ Q0.

9. A storage medium, characterized in that: The storage medium stores instructions and data to implement the displacement prediction spatial domain construction method based on landslide scale mapping as described in any one of claims 1 to 8.

10. A displacement prediction spatial domain construction device based on landslide scale mapping, characterized in that: include: A processor and a storage medium; the processor loads and executes instructions and data in the storage medium to implement the displacement prediction spatial domain construction method based on landslide scale mapping as described in any one of claims 1 to 8.

Images