A method and device for identifying and analyzing deformation of a bank landslide

Abstract

The application discloses a deformation identification analysis method and device for reservoir bank landslide, and belongs to the technical field of geological disaster monitoring and identification. Multi-source basic data of a reservoir bank research area are acquired, the multi-source basic data including synthetic aperture radar image data, deformation inversion is performed on the multi-source basic data, at least two initial deformation data are obtained, fusion deformation data are generated, a multi-dimensional feature vector is generated based on the fusion deformation data and the multi-source basic data, a pre-trained identification network model is used to obtain a landslide identification result, a dynamic early warning threshold is determined for a deformation landslide in the landslide identification result, a corresponding early warning level is determined according to the dynamic early warning threshold, the deformation monitoring precision and the spatial coverage range are considered, the spatio-temporal deformation law of the landslide can be more comprehensively captured, on this basis, the accuracy of the landslide identification is improved, the dynamic adjustment of the landslide early warning is realized, and reliable technical support is provided for disaster prevention and control of the reservoir bank landslide.

Description

Technical Field

[0001] This application belongs to the field of geological disaster monitoring and identification technology, specifically relating to a deformation identification and analysis method, device, electronic equipment, and storage medium for reservoir bank landslides. Background Technology

[0002] Reservoir bank landslides are one of the major geological hazards in the field of water conservancy and hydropower engineering. For example, since the Three Gorges Reservoir began impounding water, reservoir bank landslides have become frequent, especially "small deformation landslides." These landslides are characterized by long-term minute deformations or step displacements under specific conditions, and are often distributed in mountainous areas with complex terrain and high vegetation cover. Traditional monitoring and identification methods are insufficient to effectively detect them, seriously threatening the safety of the reservoir area. Moreover, the deformation of reservoir bank landslides is influenced by a combination of factors, including periodic fluctuations in reservoir water levels and rainfall, resulting in a complex deformation mechanism.

[0003] Existing methods for monitoring and identifying reservoir bank landslides mainly include deformation monitoring methods based on synthetic aperture radar interferometry and landslide identification methods based on neural networks. However, in practical applications, existing solutions still have certain limitations, and the accuracy of deformation identification of reservoir bank landslides needs to be further improved. Summary of the Invention

[0004] The purpose of this application is to provide a method and apparatus for deformation identification and analysis of reservoir bank landslides, which can solve the problem that existing reservoir bank landslide monitoring and identification methods still have certain limitations, and how to further improve the accuracy of deformation identification of reservoir bank landslides.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows: In a first aspect, embodiments of this application provide a deformation identification and analysis method for reservoir bank landslides, the method comprising: Acquire multi-source basic data of the reservoir shore research area; wherein, the multi-source basic data includes synthetic aperture radar image data; Deformation inversion is performed on the synthetic aperture radar image data to obtain at least two initial deformation data, and fused deformation data is generated based on the at least two initial deformation data; Based on the fused deformation data and the multi-source basic data, a multi-dimensional feature vector is generated; Based on the aforementioned multi-dimensional feature vectors, a pre-trained recognition network model is used to obtain landslide recognition results; For the deformed landslides in the landslide identification results, a dynamic early warning threshold is determined, and the corresponding early warning level is determined based on the dynamic early warning threshold.

[0006] Optionally, the multi-source basic data includes deformation verification data, and the generation of fused deformation data based on the at least two initial deformation data includes: Based on the deformation verification data, the accuracy of the at least two initial deformation data is evaluated, and the accuracy weights corresponding to the at least two initial deformation data are determined. Based on the precision weights, the at least two initial deformation data are weighted and fused to generate fused deformation data.

[0007] Optionally, the multi-source basic data also includes reservoir water data and rainfall data, and the generation of a multi-dimensional feature vector based on the fused deformation data and the multi-source fused data includes: Based on the reservoir water data, multiple reservoir water movement stages are determined. Based on the rainfall data, a heavy rainfall event is identified; Based on the fused deformation data, the multiple reservoir water movement stages, and the heavy rainfall event, a multi-dimensional feature vector is generated; wherein, the multi-dimensional feature vector includes any one or more of the following: basic temporal displacement features, deformation gradient features, spatiotemporal correlation features, and reservoir water-rainfall coupling features.

[0008] Optionally, for deformed landslides in the landslide identification results, a dynamic early warning threshold is determined, including: Extract deformation stage information, reservoir water change information, and rainfall information of the deformed landslide; Based on the deformation stage information, reservoir water change information, and rainfall information, the coupling influence coefficient is determined; The dynamic early warning threshold for the deformed landslide is determined based on the preset benchmark displacement increment threshold and the coupling influence coefficient.

[0009] Optionally, determining the corresponding warning level based on the dynamic warning threshold includes: Obtain the actual displacement increment of the deformed landslide; Determine the warning ratio between the actual displacement increment and the dynamic warning threshold, and determine the corresponding warning level based on the warning ratio.

[0010] Optionally, it also includes: Extract the spatial attention weights and temporal attention weights from the recognition network model; Based on the spatial attention weights, the key deformation areas of the landslide are determined; Based on the aforementioned time attention weights, the key triggering time points for landslides are determined; Based on the key deformation area of ​​the landslide and the key induction time node of the landslide, the dominant inducing factors and critical parameters of the landslide deformation are determined. Based on the dominant inducing factors and the critical parameters of landslide deformation, landslide prevention and control recommendations are generated.

[0011] Optionally, the pre-trained recognition network model is a two-branch attention long short-term memory network model; The bi-branch attention long short-term memory network model includes a spatial attention branch and a temporal attention branch. The spatial attention branch and the temporal attention branch are used to assign attention weights to spatially relevant features and temporally relevant features in the input features, respectively.

[0012] Optionally, the training steps of the recognition network model include: Select landslide samples and non-landslide samples, and define landslide type label and deformation stage label for each sample; wherein, the deformation stage label is based on the quantitative classification of landslide deformation trend in each sample; A stratified sampling method was adopted, and the samples were divided into training set, validation set and test set according to the landslide type label and the deformation stage label; The training set is optimized using cross-validation to obtain a target training set, which is then used to train the recognition network model.

[0013] Secondly, embodiments of this application provide a deformation identification and analysis device for reservoir bank landslides, the device comprising: The data acquisition module is used to acquire multi-source basic data of the reservoir shore research area; wherein, the multi-source basic data includes synthetic aperture radar image data; The data fusion module is used to perform deformation inversion on the synthetic aperture radar image data to obtain at least two initial deformation data, and generate fused deformation data based on the at least two initial deformation data; The feature generation module is used to generate multi-dimensional feature vectors based on the fused deformation data and the multi-source basic data; The identification module is used to obtain landslide identification results based on the multi-dimensional feature vector and a pre-trained identification network model. The early warning module is used to determine a dynamic early warning threshold for deformed landslides in the landslide identification results, and to determine the corresponding early warning level based on the dynamic early warning threshold.

[0014] Thirdly, embodiments of this application provide an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, 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] Fifthly, embodiments of this application provide a chip, the chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the method as described in the first aspect.

[0017] In this embodiment, multi-source basic data of the reservoir bank research area is acquired. This multi-source basic data includes synthetic aperture radar (SAR) image data. Deformation inversion is performed on the SAR image data to obtain at least two initial deformation data sets. Based on these initial deformation data sets, fused deformation data is generated. A multi-dimensional feature vector is generated based on the fused deformation data and the multi-source basic data. A pre-trained recognition network model is used based on the multi-dimensional feature vector to obtain landslide recognition results. For deformed landslides identified in the landslide recognition results, a dynamic early warning threshold is determined. Based on this dynamic early warning threshold, the corresponding early warning level is determined. This achieves the goal of generating fused deformation data by weighted fusion of at least two initial deformation data sets. This approach balances the accuracy and spatial coverage of deformation monitoring, enabling a more comprehensive capture of the spatiotemporal deformation patterns of landslides. Furthermore, it improves the accuracy of landslide recognition and enables dynamic adjustment of landslide early warning, providing reliable technical support for disaster prevention and control of reservoir bank landslides. Attached Figure Description

[0018] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the present invention will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the steps of a deformation identification and analysis method for reservoir bank landslides provided in some embodiments of the present invention; Figure 2 This is a structural block diagram of a deformation identification and analysis device for reservoir bank landslides provided in some embodiments of the present invention; Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in some embodiments of the present invention. Detailed Implementation

[0020] 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.

[0021] 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 use of 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.

[0022] Existing reservoir bank landslide monitoring and identification technologies mainly include the following categories: InSAR (Interferometric Synthetic Aperture Radar) deformation monitoring technology has been widely used in landslide deformation monitoring. The mainstream approach involves using a single time-series InSAR technology (such as SBAS-InSAR (Small Baseline Subset InSAR) or StaMPS-MTI (Stanford Method for Persistent Scatterers - Multi-Temporal InSAR)) or a combination of both. SBAS-InSAR offers advantages in area coverage, while StaMPS-MTI offers advantages in high-precision point monitoring. However, both are often compared after independent inversion, without achieving deep data fusion. Furthermore, data processing, such as phase unwrapping and atmospheric correction, lacks standardized parameters and precision weighting mechanisms.

[0023] Machine learning landslide identification technology: Machine learning algorithms (such as LSTM (Long Short-Term Memory), random forest, and support vector machine) have been used for landslide identification, but they have obvious limitations: ① Feature indicators mostly rely on optical images and topographic data (such as slope and aspect). A few studies involving SAR (Synthetic Aperture Radar) data have not made full use of InSAR temporal deformation information and have not constructed a multi-dimensional feature system that includes spatiotemporal correlation and reservoir water-rainfall coupling; ② The training set design is simple, mostly classified as landslide / non-landslide binary, without combining landslide deformation mechanism (such as floating reduction type, hydrodynamic pressure type) or deformation stage for classification, and has not considered the imbalance of positive and negative sample ratios and the mining of "difficult sample" landslides with small deformation.

[0024] Landslide early warning and deformation analysis technology: Traditional early warning methods mostly use static displacement threshold methods, which do not consider the dynamic changes of landslide deformation stages and reservoir water-rainfall coupling conditions; deformation feature analysis is mostly qualitative description, and does not accurately locate key deformation areas and triggering time nodes based on the combination of InSAR and machine learning results, and lacks quantitative analysis of the contribution rate of triggering factors and critical parameters.

[0025] However, existing reservoir bank landslide monitoring and identification technologies have the following core problems: Limitations of InSAR data applications: A single InSAR technology cannot balance accuracy and spatial coverage (SBAS-InSAR has full coverage but slightly lower accuracy, while StaMPS-MTI has high accuracy but limited coverage); combining two technologies is only for simple comparison and verification, without achieving deep fusion through quantification of accuracy weights, which can easily lead to misidentification due to technical errors.

[0026] The feature index system is incomplete: Most studies take optical images and topographic data as the core features. Even those that use InSAR data do not make full use of temporal deformation information and do not construct multi-dimensional features that include basic displacement, deformation gradient, spatiotemporal correlation, and reservoir water-rainfall coupling, making it difficult to fully capture the spatiotemporal deformation patterns of landslides.

[0027] Machine learning models lack specificity: Traditional LSTM and other models lack a focusing mechanism for key spatiotemporal features of landslides, and cannot adaptively enhance the focus on features of reservoir water-sensitive stages, key periods of heavy rainfall, and areas of strong landslide deformation, thus limiting the recognition accuracy; the model training set classification method is simple (not considering the landslide deformation stage), and the positive and negative sample ratio and difficult case mining are not optimized, resulting in weak recognition ability of small deformation landslides.

[0028] Early warning and deformation analysis lack quantitative support: early warning of small deformation landslides mostly uses static thresholds and does not dynamically adjust them in combination with deformation stages and reservoir water-rainfall coupling conditions, resulting in low accuracy of early warning; deformation analysis of typical landslides is mostly qualitative description and does not quantify the contribution rate of inducing factors and critical parameters (such as reservoir water threshold and dynamic water pressure lag time), making it difficult to support engineering prevention and control decisions.

[0029] This application utilizes standardized preprocessing of multi-source basic data, employs deep fusion of dual InSAR technology to obtain high-precision deformation data, constructs a multi-dimensional fusion feature index system, designs a landslide-specific dual-branch attention LSTM model, and optimizes model performance by combining a hierarchical training set with deformation stage labels. This enables accurate identification of reservoir bank landslide hazards and dynamic early warning of small-deformation landslides. Simultaneously, it completes in-depth analysis of the spatiotemporal deformation characteristics of typical landslides, providing technical support for landslide prevention and control. This approach can solve technical problems such as the contradiction between accuracy and coverage of single InSAR technology, insufficient attention to key features in traditional LSTM models, and difficulties in identifying and warning of small-deformation landslides.

[0030] The following description, in conjunction with the accompanying drawings, details a deformation identification and analysis method for reservoir bank landslides provided in this application through specific embodiments and application scenarios.

[0031] Reference Figure 1 The diagram illustrates a flowchart of a deformation identification and analysis method for reservoir bank landslides provided by some embodiments of the present invention, which may specifically include the following steps: Step 101: Obtain multi-source basic data of the reservoir shore research area; wherein, the multi-source basic data includes synthetic aperture radar image data.

[0032] In step 101, multi-source basic data related to landslide monitoring in the reservoir bank study area can be acquired. This multi-source basic data includes at least synthetic aperture radar imagery, topographic data, deformation verification data, and auxiliary data on triggering factors.

[0033] Specifically, the synthetic aperture radar imagery data was obtained using the Sentinel-1 satellite's ascending orbit interferometric wide-swath mode data, with a resolution of 5 meters × 20 meters. Vertical transmission and vertical reception polarization and vertical transmission and horizontal reception polarization were employed, acquiring a total of 158 images. The time span was from March 12, 2017 to December 23, 2022, and the radar wavelength was 5.55 centimeters.

[0034] The terrain data used was 30-meter resolution digital elevation model data provided by NASA, and the projected coordinate system was Global Reference System 1984 / Universal Transverse Mercator Projection Zone 50 (i.e., WGS84 / UTM Zone 50N).

[0035] The deformation verification data were obtained from surface line-of-sight deformation data from four global navigation satellite system monitoring stations in the study area, with a sampling frequency of once a day.

[0036] The auxiliary data for triggering factors include daily rainfall data of the study area from 2017 to 2022, with a spatial resolution of 1 km, and daily reservoir water level data of the Three Gorges Reservoir area, with water level fluctuations ranging from 145 meters to 175 meters.

[0037] In one example, after acquiring multi-source basic data, the multi-source basic data can be standardized to eliminate data errors, unify data format and coordinate system, complete missing data, and normalize all data to a unified numerical range, ensuring data consistency and availability, and providing a standardized data foundation for subsequent deformation inversion and feature analysis.

[0038] Specifically, synthetic aperture radar image data (i.e. SAR data) can be radiometrically calibrated, thermal noise removed, and orbitally corrected using SNAP software. Coherence enhancement is performed using Goldstein filtering (5×5 filter window, 0.25 filter coefficient), and the data is uniformly cropped to the study area (110°26′3″~110°39′23″ E, 30°56′32″~31°3′48″ N).

[0039] Topographic data can be filled in depressions and resampled using ArcGIS (a geographic information system software), and the coordinate system and range can be completely matched with SAR data with an error of ≤1 pixel.

[0040] GNSS (Global Navigation Satellite System), rainfall, and reservoir water level data: Data were timestamped to the SAR image imaging time (accurate to the minute), and missing values ​​were filled using linear interpolation. All data were normalized to [-1, 1]. The specific formula is as follows:

[0041] in, The normalized value, The original value, This represents the maximum / minimum value in the dataset.

[0042] Step 102: Perform deformation inversion on the synthetic aperture radar image data to obtain at least two initial deformation data, and generate fused deformation data based on the at least two initial deformation data.

[0043] In step 102, two different synthetic aperture radar interferometric deformation inversion techniques can be used to perform deformation inversion on the preprocessed synthetic aperture radar image data, resulting in two different types of initial deformation data. One type of initial deformation data has the advantage of area coverage, while the other type has the advantage of high-precision point monitoring. Then, based on deformation verification data, the two initial deformation data are weighted and fused to generate high-precision, full-coverage area fused deformation data.

[0044] In practical applications, the first deformation inversion technique can employ Small Baseline Assembled Aperture Radar Interferometry (SBAS-InSAR). Specifically, the inversion operation is performed using GMTSAR software, with the following parameter settings: a temporal baseline threshold of 200 days (defined as the maximum time interval between adjacent SAR images; images exceeding this value are discarded to prevent distortion of deformation information due to excessively long time intervals); a spatial baseline threshold of 180 meters (defined as the maximum spatial distance between adjacent SAR images; images exceeding this value are discarded to ensure inversion accuracy); and a coherence threshold of 0.4 (defined as an index measuring the correlation between SAR image pairs; pixels below this value are considered invalid and discarded, not participating in the inversion). Phase unwrapping can be performed using the minimum cost flow method, with an unwrapping accuracy of less than or equal to 0.1 radians. The GACOS (Generic Atmospheric Correction Online Service for InSAR) atmospheric correction method is used to eliminate the influence of atmospheric disturbances, with a correction error of less than or equal to 2 millimeters. The final output is the planar grid deformation rate. and time-series cumulative displacement ( For the first Each imaging moment, =1,2.....158), with a fixed grid resolution of 10 meters × 10 meters, and a time span from March 12, 2017 to December 23, 2022.

[0045] The second deformation inversion technique can employ the Stanford Permanent Scatterer Multi-Temporal Synthetic Aperture Radar Interferometry (StaMPS-MTI) technique. Specifically, the inversion operation is jointly performed using ISCE (InSAR Scientific Computing Environment) and StaMPS software, with the following parameter settings: amplitude deviation value 0.2 (defined as an indicator of SAR image amplitude stability; pixels exceeding this value are considered unstable and discarded from the inversion); standard deviation threshold 0.6 (defined as an indicator of pixel deformation dispersion; pixels exceeding this value are considered anomalous and discarded from the inversion); a dual-reference point correction method is used, with the Three Gorges Dam as the first reference point for regional stability, and the central ridge of the reservoir bank as the second reference point, with coordinates 110°32′15″E, 30°59′48″N; the GACOS atmospheric correction method is used to eliminate the influence of atmospheric disturbances, with a correction error of less than or equal to 2 mm. The final output vector point deformation rate is... and time-series cumulative displacement Kriging interpolation was used to generate raster data for the aforementioned vector points. The interpolation model was a spherical model with the following parameters: nugget value 0.01 and range 500 meters. The interpolated raster resolution and spatial extent were completely consistent with the small baseline ensemble aperture radar interferometry inversion results, yielding the raster deformation rate. and time-series cumulative displacement .

[0046] In some embodiments of this application, the multi-source basic data includes deformation verification data, and the generation of fused deformation data based on the at least two initial deformation data includes: Sub-step 11: Based on the deformation verification data, perform an accuracy evaluation on the at least two initial deformation data and determine the accuracy weights corresponding to the at least two initial deformation data.

[0047] In sub-step 11, GNSS monitoring data can be used as deformation verification data to calculate the grid-by-grid accuracy weighting factor (i.e., accuracy weight) for the raster data obtained from SBAS-InSAR and StaMPS-MTI techniques. The GNSS monitoring data can be the LOS (Line of Sight) deformation data from four GNSS monitoring stations (ZGX295, ZGX296, ZGX297, and ZGX298) within the reservoir study area, sampled once daily with an accuracy of ±0.3 mm. The GNSS monitoring data is precisely aligned with the SAR image imaging time, spanning from March 12, 2017 to December 23, 2022.

[0048] Specifically, the root mean square error (RMSE) can be calculated grid-by-grid first. For SBAS-InSAR and StaMPS-MTI inversion data, the RMSE is calculated using the following formulas. 、:

[0049] The sample size of the GNSS monitoring data is the daily monitoring data from March 12, 2017 to December 23, 2022. This represents the InSAR deformation data of the raster corresponding to the k-th time point, in millimeters. This represents the deformation data of the GNSS monitoring station at time k, in millimeters. It should be noted that when calculating the root mean square error per grid corresponding to the SBAS-InSAR inversion data... hour, Take D ASRS (x, y, t) k When calculating the root mean square error per grid corresponding to the StaMPS-MTI inversion data hour, Pick .

[0050] Then, the accuracy weighting factor is calculated based on the root mean square error. The formula for calculating the accuracy weighting factor of SBAS-InSAR raster data is as follows:

[0051] The formula for calculating the precision weighting factor of StaMPS-MTI raster data is:

[0052] in, This refers to the accuracy weighting factor (i.e., accuracy weight) for SBAS-InSAR raster data. For StaMPS-MTI raster data, the precision weighting factor is... The weights represent the planar coordinates of the grid, in meters, consistent with the WGS84 / UTM Zone 50N coordinate system. A larger weight value indicates higher deformation inversion accuracy of the corresponding InSAR data at that grid location; a weight value of 0 indicates that the corresponding InSAR data at that grid location is invalid and will not participate in subsequent fusion calculations.

[0053] Sub-step 12: Based on the precision weights, the at least two initial deformation data are weighted and fused to generate fused deformation data.

[0054] In sub-step 12, the deformation rate and time-series cumulative displacement retrieved by the two InSAR technologies can be fused according to the accuracy weight using an adaptive weighted fusion method to generate fused deformation data. This method can automatically allocate weights according to grid-by-grid accuracy to ensure high accuracy and full coverage of the fused data.

[0055] 1. Fusion deformation rate The calculation formula is as follows:

[0056] 2. Fusion of time-series cumulative displacement (Unit: mm), i.e., the first Imaging time, coordinates The formula for calculating the temporal cumulative displacement of the raster is as follows:

[0057] Furthermore, after fusion is completed, the accuracy of the fusion results is verified. The fused data must meet the two GNSS verification standards: overall goodness of fit R0. 2 (R) 2The goodness of linear fit between the fused data and GNSS monitoring data is measured (the closer to 1, the better the fit and the higher the fusion accuracy). The root mean square error per grid is less than or equal to 8 mm. If any of the above criteria are not met, the nugget value and range of the Kriging interpolation model need to be readjusted. The nugget value adjustment range is 0.005 to 0.02, and the range adjustment range is 400 meters to 600 meters. The interpolation and fusion operations are then repeated until the criteria are met.

[0058] The final output includes fused deformation rate raster data with a resolution of 10m×10m and fused temporal cumulative displacement data for 158 time periods in the reservoir bank study area, which serve as the unified data basis for subsequent steps.

[0059] Step 103: Generate a multi-dimensional feature vector based on the fused deformation data and the multi-source basic data.

[0060] In step 103, based on the fused deformation data generated in step 102, and combined with the reservoir water movement patterns and rainfall characteristics in the reservoir bank study area, a multi-dimensional fused feature index system is constructed to generate a multi-dimensional feature vector. This can overcome the limitations of traditional single displacement features and more accurately capture the spatiotemporal deformation patterns of landslides. Among them, the feature dimensions were optimized to 96 dimensions through experiments.

[0061] In some embodiments of this application, the multi-source basic data further includes reservoir water data and rainfall data, and the step of generating a multi-dimensional feature vector based on the fused deformation data and the multi-source fused data includes: Sub-step 21: Based on the reservoir water data, determine multiple reservoir water movement stages of the reservoir water data.

[0062] In sub-step 21, the reservoir water movement can be divided into four stages based on the hydrological year (from June to May of the following year, consistent with the hydrological change cycle of the reservoir area). Quantitative judgment indicators are used for the division to avoid subjective judgment errors and to provide a clear stage division basis for subsequent coupled feature calculations.

[0063] The specific classification criteria are as follows: (1) Reservoir water rise period: The reservoir water level rises from 145 meters to 175 meters, with a rise rate of ≥0.2 meters / day and a duration of ≥30 days; (2) High water level operation period: The reservoir water level is greater than or equal to 170 meters, the fluctuation range is less than or equal to 5 meters, and the duration is greater than or equal to 60 days; (3) Reservoir water level drop period: The reservoir water level drops from 175 meters to 145 meters at a rate of ≥0.2 meters / day and lasts for ≥30 days; (4) Low water level operation period: The reservoir water level is less than or equal to 150 meters, the fluctuation range is less than or equal to 5 meters, and the duration is greater than or equal to 60 days.

[0064] Sub-step 22: Based on the rainfall data, determine the heavy rainfall event.

[0065] In sub-step 22, the determination of heavy rainfall events can adopt a dual threshold standard of rainfall intensity and duration to avoid the determination error caused by a single threshold.

[0066] The specific criteria are as follows: daily rainfall greater than or equal to 100 mm (daily rainfall refers to the cumulative rainfall in a single day over 24 hours, in millimeters), or cumulative rainfall greater than or equal to 150 mm over three consecutive days (three consecutive days refer to three consecutive natural days, and the cumulative rainfall is the sum of the three days), or monthly rainfall greater than or equal to 200 mm (monthly rainfall refers to the cumulative rainfall in a natural month).

[0067] A heavy rainfall event is defined as one that meets any of the above conditions, and the start and end times of the event should be recorded. ,in The moment when heavy rainfall begins. The exact time when the heavy rainfall ends, accurate to the day.

[0068] Sub-step 23: Based on the fused deformation data, the multiple reservoir water movement stages, and the heavy rainfall event, generate a multi-dimensional feature vector; wherein, the multi-dimensional feature vector includes any one or more of the following: basic temporal displacement features, deformation gradient features, spatiotemporal correlation features, and reservoir water-rainfall coupling features.

[0069] Multidimensional feature vectors can include any one or more of the following: basic temporal displacement features, deformation gradient features, spatiotemporal correlation features, and reservoir water-rainfall coupling features.

[0070] In sub-step 23, based on the fused time-series cumulative displacement data output in step 102, combined with the reservoir water movement stages defined in sub-step 21 and the heavy rainfall events determined in sub-step 22, four-dimensional feature indices can be calculated. All features are normalized to the [-1,1] interval using a normalization formula to eliminate the influence of dimensions and ensure the stability of subsequent model training.

[0071] First, regarding the basic time-series displacement characteristics (36 dimensions) Basic temporal displacement characteristics (36-dimensional), based on fused temporal cumulative displacement The calculation can employ a differentiated adaptive thinning method, adjusting the thinning interval based on the deformation activity of the reservoir in four stages. This ensures that the features retain key deformation information while avoiding redundancy. The specific operation is as follows: 1. Calculate the deformation activity K (unit: mm / d) for each stage of water movement in the reservoir. The formula is:

[0072] Among them, the maximum cumulative displacement during a stage refers to the displacement of water within the reservoir during that stage. The difference between the maximum and minimum values ​​(unit: mm), and the stage duration refers to the total number of days (unit: d) of the water movement stage in the reservoir; definition High activity level (the landslide deformation is severe during this stage, requiring more frequent and denser dewatering intervals). (The landslide deformation is gentle at this stage, so the thinning interval can be increased).

[0073] 2. Dilution Rule: During periods of high water level operation and reservoir water decline, which are considered high-activity phases, a 12-day dilution interval is adopted (i.e., an imaging moment is selected every 12 days). As eigenvalues), each stage is thinned to 9 dimensions, for a total of 18 dimensions across the two stages; the periods of rising reservoir water and low water levels are considered low-activity periods, and a 36-day thinning interval is used (i.e., an imaging moment is selected every 36 days). As eigenvalues), each stage is thinned to 9 dimensions, for a total of 18 dimensions across the two stages; after thinning, the four stages total 36 dimensions. The eigenvalues ​​are the fused cumulative displacement values ​​corresponding to each thinning moment, and the dimension order is arranged according to the reservoir water movement stages (rising period → high water level period → falling period → low water level period) and the thinning moment.

[0074] II. Targeting Deformation Gradient Characteristics (24 dimensions) Deformation gradient characteristics (24 dimensions) are used to capture the spatial variation trend of landslide deformation, divided into two categories: temporal gradient and spatial gradient, each with 12 dimensions. The specific calculation method is as follows: 1. Time gradient (Unit: mm / d), reflecting the rate of change of landslide displacement over time, calculated using the following formula:

[0075] in, The first step is to thin out the basic time-series displacement features. At any moment ( ) The first after dilution At that moment, The time interval between two thinning moments (unit: days); the temporal gradient values ​​of the first 24 thinning moments are selected as temporal gradient features, totaling 12 dimensions (one gradient value is calculated for every two adjacent moments, resulting in 23 gradient values ​​for 24 moments, and the first 12 are selected as features).

[0076] 2. Spatial gradient (Unit: mm / m) Reflecting the rate of change of landslide displacement in space, the Sobel operator (3×3 window, the pixel weight matrix in the window is: [[-1,0,1],[-2,0,2],[-1,0,1]]) is used to calculate the spatial gradient of the raster in the x-direction (east-west direction) and y-direction (north-south direction), and the gradient magnitude is taken as the spatial gradient feature value; the reservoir water is thinned in four stages, each stage is thinned to 3 dimensions, and the four stages are 12 dimensions in total. The thinning rule is consistent with that for the basic time series displacement feature.

[0077] The formula for calculating the gradient magnitude is: ,in, The gradient is in the x-direction. The gradient is in the y-direction.

[0078] Deformation gradient characteristics Time gradient (12-dimensional) and spatial gradient The combination of (12 dimensions) results in a total of 24 dimensions, with the temporal gradient first and the spatial gradient second.

[0079] III. Targeting Spatiotemporal Correlation Features (16-dimensional) Spatiotemporal correlation features (16-dimensional), used to reflect the spatiotemporal correlation of landslide deformation, is divided into two categories: temporal autocorrelation coefficient and spatial neighborhood correlation coefficient, each with 8 dimensions. The specific calculation method is as follows: 1. Time autocorrelation coefficient (Unitless, value range [-1, 1]), reflecting the correlation of landslide displacement in the time series. The Pearson autocorrelation coefficient calculation method is used to calculate the first-order autocorrelation coefficient of the time series displacement after thinning the basic time series displacement characteristics (i.e., the correlation coefficient of displacement between two adjacent thinning times). It is selected according to the four stages of reservoir water movement, and one average value is selected for each stage, for a total of 4 values. At the same time, one autocorrelation coefficient value is selected before and after each heavy rainfall event, for a total of 4 values ​​for all heavy rainfall events (if there are less than 4, the actual number is selected; if there are more than 4, the 4 with the strongest correlation are selected). There are a total of 8 dimensions, with the dimension order being the value corresponding to the reservoir water movement stage first, and the value corresponding to the heavy rainfall before and after.

[0080] 2. Spatial neighborhood correlation coefficient (Unitless, value range [-1, 1]), reflecting the correlation between the target grid and the displacement of the surrounding grids. The displacement correlation coefficient between the target grid and its 8 neighboring grids (i.e., the 8 grids adjacent to the target grid vertically, horizontally, and diagonally) is calculated (using the Pearson correlation coefficient calculation method). The average of the 8 correlation coefficients is taken as the spatial neighborhood correlation coefficient of the target grid. The reservoir water is thinned in four stages, with each stage being 2-dimensional, for a total of 8 dimensions across the four stages. The thinning rules are consistent with those for the basic time-series displacement characteristics.

[0081] Spatiotemporal correlation features Time autocorrelation coefficient (8-dimensional) and spatial neighborhood correlation coefficient The combination of (8 dimensions) results in a total of 16 dimensions, with the order of dimensions being time autocorrelation coefficient first and spatial neighborhood correlation coefficient second.

[0082] IV. Targeting the coupling characteristics of reservoir water and rainfall (20 dimensions) Reservoir water-rainfall coupling characteristics (20-dimensional) This method couples external inducing factors such as reservoir water and rainfall with landslide deformation data to quantify the impact of inducing factors on landslide deformation, avoiding the limitations of analyzing inducing factors or deformation data alone. The specific calculation method is as follows: 1. Reservoir-water coupling characteristics (10-dimensional): Quantifying the coupling relationship between reservoir water changes and landslide deformation, including two types of indicators: ① Pearson correlation coefficient between reservoir water level and cumulative displacement (unitless, range [-1, 1]), reflecting the degree of linear correlation between reservoir water level changes and landslide displacement; ② Coupling coefficient between reservoir water level change rate and displacement time gradient (unitless, range [-1, 1]), calculated using the following formula:

[0083] Among them, △ For the first Reservoir water level change rate over a period of time (unit: m / d). For the first Displacement time gradient over a period of time (unit: mm / d). This represents the number of time periods in the hydrological year; the above two types of indicators each take one value for 5 hydrological years (2017-2022), for a total of 10 dimensions (5 values ​​for each type of indicator).

[0084] 2. Rainfall Coupling Characteristics (10-dimensional): Quantifying the coupling relationship between rainfall and landslide deformation, including two types of indicators: ① The ratio of cumulative rainfall to displacement increment during periods of heavy rainfall (unit: mm / mm), calculated using the following formula:

[0085] Among them, △R is the cumulative rainfall during the heavy rainfall period (unit: mm), and △D is the landslide displacement increment during the same period (unit: mm); ② Pearson correlation coefficient between rainfall and displacement during non-heavy rainfall periods (unitless, value range [-1,1]); The above two types of indicators each take 1 value for 5 hydrological years (2017-2022), for a total of 10 dimensions (5 values ​​for each type of indicator).

[0086] Reservoir water-rainfall coupling characteristics The combination of reservoir water coupling features (10-dimensional) and rainfall coupling features (10-dimensional) results in a total of 20 dimensions, with the reservoir water coupling features preceding the rainfall coupling features.

[0087] V. Final Feature Index System Output (i.e., Generation of Multi-Dimensional Feature Vectors) The above four-dimensional features are concatenated in the order of basic temporal displacement features, deformation gradient features, spatiotemporal correlation features, and reservoir water-rainfall coupling features to construct a 96-dimensional fusion feature index system. This involves concatenating the above four-dimensional features in the following order: The sequential concatenation of elements constructs a 96-dimensional fusion feature index system. Each grid cell corresponds to a set of 96-dimensional feature vectors, which serve as input features for the subsequent recognition network model. The feature vector format is as follows: ,in ( ) is the first Normalized values ​​of dimensional features, This represents the transpose of the feature vector, ensuring that the feature format of the input model is consistent.

[0088] Step 104: Based on the multi-dimensional feature vector, a pre-trained recognition network model is used to obtain the landslide recognition result.

[0089] In step 104, the constructed 96-dimensional fused feature vector can be input into the trained dual-branch attention long short-term memory (SA-LSTM) network model to identify landslides in all monitoring units in the reservoir bank study area.

[0090] In some embodiments of this application, the pre-trained recognition network model is a dual-branch attention long short-term memory network model; The bi-branch attention long short-term memory network model includes a spatial attention branch and a temporal attention branch. The spatial attention branch and the temporal attention branch are used to assign attention weights to spatially relevant features and temporally relevant features in the input features, respectively.

[0091] Specifically, considering the technical characteristics of landslides in reservoir areas involving spatiotemporal dual-feature coupling, a landslide-specific dual-branch attention LSTM (Spatial-TemporalAttentionLSTM, SA-LSTM) model is designed. By adding spatial attention branches and temporal attention branches, it can adaptively focus on key spatiotemporal features for landslide identification, solving the technical problems of insufficient attention to key features and low identification accuracy of traditional LSTM models. The structure, parameters, and formulas of the model are all fully defined and can be directly reproduced.

[0092] I. Overall Structure of the SA-LSTM Model The described dual-branch attention long short-term memory network model adopts a dual-branch input-attention fusion-fully connected output structure. The input layer is the 96-dimensional feature vector constructed in step 103, and the output layer is the binary classification result (slippery / non-slippery). The hidden layers include spatial attention branches, temporal attention branches, and attention fusion layers. The connections and functions of each layer are clearly defined. The specific structure of the model is as follows: Input layer → Spatial attention layer → First long short-term memory network layer → Attention fusion layer → Fully connected layer → Output layer; Input layer → Temporal attention layer → Second long short-term memory network layer → Attention fusion layer.

[0093] The core function of the spatial attention layer is to focus on the key spatial features of landslide deformation, assign attention weights to spatially related features (i.e., deformation gradient features and spatial neighborhood correlation coefficients in spatiotemporal correlation features) to enhance the influence of key spatial features.

[0094] II. Core Layer Parameters and Calculation Formulas in the Model 1. Spatial Attention Layer (SA) The core function of the Spatial Attention Layer (SA) is to focus on key spatial features of landslide deformation, specifically spatially relevant features (i.e., deformation gradient features). Spatiotemporal correlation characteristics Spatial neighborhood correlation coefficient Attention weights are assigned to enhance the influence of key spatial features. The specific formula and parameters are as follows:

[0095]

[0096] in, The spatial attention weight matrix has a fixed dimension of 32×32 (32 rows and 32 columns, consistent with the dimension of the input spatial features), and the initial values ​​adopt a random normal distribution (mean 0, standard deviation 0.01). The bias term has a fixed dimension of 32×1 (32 rows, 1 column) and an initial value of 0; σ(x) is the Sigmoid activation function, with the following formula:

[0097] Used to normalize the attention weights to [0,1]; For feature splicing operations, that is... (24-dimensional) and (8-dimensional) Concatenate column by column to obtain a 32-dimensional spatial feature vector; This is an element-wise multiplication operation, which means that the attention weights are multiplied by the corresponding elements of the spatial feature vectors. For spatial attention weights, with a dimension of 32×1, the larger the weight value, the higher the importance of the corresponding spatial feature; The spatial attention-weighted features, with a dimension of 32×1, are used as the input to the LSTM1 layer.

[0098] 2. Temporal Attention Layer (TA) The core function of the Time Attention (TA) layer is to focus on the key temporal characteristics of landslide deformation, specifically time-related features (i.e., basic temporal displacement characteristics). Spatiotemporal correlation characteristics Time autocorrelation coefficient Reservoir water-rainfall coupling characteristics Attention weights are assigned to enhance the influence of key temporal features. The specific formula and parameters are as follows:

[0099]

[0100] in, The time attention weight matrix has a fixed dimension of 64×64 (64 rows and 64 columns, consistent with the dimension of the input time features), and the initial values ​​adopt a random normal distribution (mean 0, standard deviation 0.01). The bias term has a fixed dimension of 64×1 (64 rows and 1 column) and an initial value of 0; σ(x) is the Sigmoid activation function. For feature splicing operations, that is... (36 dimensions) (8-dimensional) and (20-dimensional) Concatenate columns to obtain a 64-dimensional time feature vector; This is an element-wise multiplication operation; For time attention weights, the dimension is 64×1. The larger the weight value, the higher the importance of the corresponding time feature. The features, weighted by time attention and with a dimension of 64×1, are used as input to the LSTM2 layer.

[0101] 3. LSTM branch layer (LSTM1 / LSTM2) The LSTM branch layer consists of LSTM1 and LSTM2 layers, both employing an improved LSTM structure. It introduces a forget gate threshold adaptive adjustment mechanism to adapt to the non-stationarity of landslide deformation (landslide deformation exhibits non-linear and non-stationary characteristics as time and environment change). Specific parameters are as follows: LSTM1 layer: Input is spatial attention-weighted features (32-dimensional), the number of neurons is fixed at 64 (for extracting deep information of spatial features); the initial threshold of the forget gate is fixed at 0.7 (threshold range [0,1], 0 represents complete forgetting of historical information, 1 represents complete retention of historical information, 0.7 can balance the weight of historical information and current information); the activation function is the Tanh function; the recursive activation function is the Sigmoid function, used to control the input and forgetting of information; the output is... (64-dimensional feature vector), which serves as one of the inputs to the attention fusion layer.

[0102] The Tanh function is used for the nonlinear transformation of the output layer, and its formula is as follows:

[0103] LSTM2 layer: Input is the temporally attention-weighted features. (64-dimensional), with a fixed number of 128 neurons (for extracting deep information from temporal features); the initial threshold for the forget gate is fixed at 0.7; the activation function is the Tanh function (same as LSTM1 layer); the recursive activation function is the Sigmoid function (same as LSTM1 layer); the output is... (128-dimensional feature vector), which serves as the second input to the attention fusion layer.

[0104] The outputs are respectively (64-dimensional) (128 dimensions).

[0105] 4. Attention Fusion Layer The core function of the attention fusion layer is to perform weighted fusion of the output features of LSTM1 and LSTM2 layers, integrating the deep features of the spatiotemporal dual branches to obtain a unified fused feature vector. The specific formula and parameters are as follows:

[0106] Wherein: This is the fusion weight matrix output by the LSTM1 layer, with a fixed dimension of 128×64 (128 rows, 64 columns, and...). (Dimensional matching), the initial values ​​are random normal distribution (mean 0, standard deviation 0.01). This is the fusion weight matrix output by the LSTM2 layer, with a fixed dimension of 128×128 (128 rows, 128 columns, and 128 columns). (Dimensional matching), the initial values ​​are random normal distribution (mean 0, standard deviation 0.01). The fusion bias term has a fixed dimension of 128×1 (128 rows and 1 column) and an initial value of 0. This is a weighted fusion function; The fused feature vector, with a dimension of 128×1, serves as the input to the fully connected layer.

[0107] 5. Fully Connected Layer and Output Layer The fully connected layer is used to perform non-linear transformations on the fused features, and the output layer is used to output the classification probabilities of landslide / non-landslide. The specific parameters and functions are as follows: Output layer: The number of neurons is fixed at 2, corresponding to the landslide and non-landslide labels respectively; the activation function is the softmax function, which is used to normalize the output value to [0,1], obtaining the probability value P = [P_landslide / non-landslide]. 滑坡 P 非滑坡 ]; Judgment rule: P 滑坡 At that time, the grid was determined to be a landslide grid; P 滑坡 When the time is right, the grid is determined to be a non-slope grid.

[0108] The formula for the Softmax function is as follows:

[0109] Among them, For the output layer The output value of each neuron.

[0110] III. Model Training with Fixed Parameters 1. Loss Function: The cross-entropy loss function is used to measure the deviation between the model's predictions and the true labels. The formula is as follows:

[0111] in, The labels are real (landslide = 1, non-landslide = 0). Predict probabilities for the model.

[0112] 2. Optimizer: AdamW optimizer, learning rate 0.001, weight decay 0.01, β1=0.9, β2=0.999; 3. Training parameters: Batch size 32, number of iterations 200, training set / validation set / test set ratio 7:2:1, early stopping method is used, training is stopped if the validation set loss does not decrease for 10 consecutive iterations to avoid overfitting.

[0113] In some embodiments of this application, the training steps of the recognition network model include: S1, Select landslide samples and non-landslide samples, and define landslide type label and deformation stage label for each sample; wherein, the deformation stage label is based on the quantitative division of landslide deformation trend in each sample.

[0114] Specifically, the positive samples (landslide samples) consist of 400 raster samples selected from the historical landslide list (124 sites) and pre-identified deformation zones in the study area. These include 200 landslides with strong deformation (annual deformation rate greater than 10 mm / year) and 200 landslides with small deformation (annual deformation rate greater than 0 and less than or equal to 10 mm / year). The samples cover all landslide types, including those with floating overburden reduction and hydrodynamic pressure. The samples are evenly distributed throughout the study area to avoid clustering. The negative samples (non-landslide samples) consist of 600 raster samples from non-deformable areas such as ridges, bedrock, and stable terraces in the study area. These samples have an annual deformation rate less than or equal to 1 mm / year, and their temporal displacement shows no significant trend. The spatial separation between these negative and positive samples is greater than or equal to 50 meters.

[0115] Define two labels for all samples: landslide type label. (Float-off type = 0, Hydrodynamic pressure type = 1) and Deformation stage labels The deformation stage labels are based on the quantitative division of deformation trends using fused temporal displacements, avoiding subjective judgment. The specific calculation formula is as follows:

[0116] Among them, V 前期 V represents the average deformation rate over the previous two years. 后期 This represents the average deformation rate over the following three years. For the time span.

[0117] The rules for dividing the deformation stages are as follows: ① Initial deformation stage ( ): and That is, the deformation stage division factor is less than or equal to 0.1 and the average deformation rate in the first two years is less than or equal to 5 mm / year; ②Accelerated deformation stage ( ): That is, the factor for dividing the deformation stage is greater than 0.1; ③ Stable deformation stage ( ): and That is, the deformation stage division factor is less than or equal to 0.1 and the average deformation rate in the first two years is greater than 5 mm / year.

[0118] The final sample labels are This enables accurate classification using dual labels.

[0119] S2. Using a stratified sampling method, the samples are divided into training set, validation set and test set according to the landslide type label and the deformation stage label.

[0120] Specifically, the landslides are divided into six strata (2 types × 3 stages) based on “landslide type + deformation stage”. Within each stratum, the training set, validation set, and test set are divided in a 7:2:1 ratio to ensure that the sample proportions of each stratum are consistent and to avoid an excessively high proportion of samples of any one type.

[0121] After experimentation and optimization, the final ratio of positive to negative samples in the training set is 2:3 (280 landslide samples and 420 non-landslide samples), while the ratio of positive to negative samples in the validation and test sets is maintained at 1:1. This ratio ensures the representativeness of landslide samples while avoiding excessive suppression of non-landslide samples.

[0122] S3. The training set is optimized using cross-validation to obtain a target training set, and the recognition network model is trained using the target training set.

[0123] Specifically, higher training weights are assigned to hard examples in the training set to improve the model's ability to identify them. The specific method for hard example mining is as follows: for samples that are incorrectly predicted in the training set (hard examples, mainly small-deformation landslides), hard example weights are calculated. (Ordinary sample) In subsequent iterations, the training weights of difficult examples are increased to improve the model's ability to identify landslides with small deformations.

[0124] A five-fold cross-validation method is used to optimize the training set, adjusting the weights and proportions of difficult examples. Optimization is considered successful if the average precision on the validation set is greater than or equal to 98%, recall is greater than or equal to 45%, and F1-Score is greater than or equal to 0.62. If these criteria are not met, the weights of difficult examples are readjusted until they are satisfied. The final output is a standardized model training dataset, i.e., the target training set.

[0125] The recognition network model is trained using the target training set until it converges, and the optimal model is saved. The loss function used for model training is the cross-entropy loss function, as shown in the following formula:

[0126] in, The labels are real (landslide = 1, non-landslide = 0). Predict probabilities for the model.

[0127] The optimizer used is the AdamW optimizer with a learning rate of 0.001, weight decay of 0.01, β1=0.9, and β2=0.999. Training parameters are set as follows: batch size of 32, number of iterations of 200, training / validation / test set split ratio of 7:2:1, and early stopping is employed: training stops if the validation set loss does not decrease for 10 consecutive iterations to avoid overfitting.

[0128] It should be added that step 104 can be specifically described as inputting the 96-dimensional fused feature vectors of all grids (monitoring units) in the reservoir bank study area into a trained dual-branch attention long short-term memory network model, and outputting the landslide probability of each grid. At that time, if a grid was identified as a landslide, adjacent landslide grids were merged into a landslide area, and isolated grids (with a number less than or equal to 5) were removed to obtain the preliminary results of the landslide areas in the study area.

[0129] Landslides were classified into high-deformation landslides (annual deformation rate greater than 10 mm / year) and low-deformation landslides (annual deformation rate greater than 0 and less than or equal to 10 mm / year) based on their deformation rate. Combining the geological background of the study area (strata lithology, topographic slope, and drainage system distribution) and a historical landslide inventory, low-deformation landslides were re-verified. Misidentified areas due to topographic shadows and vegetation disturbance were eliminated, ultimately determining the list of potential landslide sites in the study area.

[0130] Step 105: For the deformed landslides in the landslide identification results, determine the dynamic early warning threshold, and determine the corresponding early warning level based on the dynamic early warning threshold.

[0131] To address the challenge of identifying landslides with small deformations (i.e., those difficult to identify), a dynamic early warning threshold model can be constructed based on deformation stages and reservoir water-rainfall coupling, with the threshold being the time-series cumulative displacement increment threshold. It changes dynamically with the deformation stage and reservoir water-rainfall conditions.

[0132] In step 105, a dynamic early warning threshold is determined for landslides with small deformation in the landslide identification results. Then, the early warning level corresponding to the landslide with small deformation can be determined based on the dynamic early warning threshold.

[0133] In some embodiments of this application, for deformed landslides in the landslide identification results, a dynamic early warning threshold is determined, including: Sub-step 31: Extract deformation stage information, reservoir water change information, and rainfall information of the deformed landslide.

[0134] In sub-step 31, deformation stage information, reservoir water change information, and rainfall information can be used as three core influencing factors, namely deformation stage factor, reservoir water factor, and rainfall factor, and then each factor is normalized to the [0,1] interval. The details are as follows: 1. Deformation stage factor Initial deformation stage Accelerated deformation stage Stable deformation stage ; 2. Reservoir water factors : Assign values ​​according to the stages of reservoir water movement, during the rising phase of reservoir water. High water level operation period Reservoir water level drop period Low water level operation period ; 3. Rainfall factors : Assign values ​​based on rainfall intensity; no rainfall ordinary rainfall heavy rainfall .

[0135] Sub-step 32: Determine the coupling influence coefficient based on the deformation stage information, reservoir water change information, and rainfall information.

[0136] In sub-step 32, the deformation stage factor, reservoir water factor and rainfall factor can be coupled and calculated to obtain the coupled influence factor as the coupled influence coefficient.

[0137] The formula for calculating the coupling effect factor is as follows: Coupling Influence Factors Normalized to [0,1].

[0138] Sub-step 33: Determine the dynamic early warning threshold of the deformed landslide based on the preset benchmark displacement increment threshold and the coupling influence coefficient.

[0139] In sub-step 33, the baseline displacement increment threshold for small deformation landslides is used. Based on this, the preset benchmark displacement increment threshold was determined to be 5 mm through experiments. The dynamic early warning threshold was then calculated using the coupling influence factor, and the calculation formula is as follows:

[0140] in, for Dynamic warning threshold (mm) at any given time. for Coupling influence factors at different times .

[0141] In some embodiments of this application, determining the corresponding warning level based on the dynamic warning threshold includes... Sub-step 41: Obtain the actual displacement increment of the deformed landslide.

[0142] In sub-step 41, the actual displacement increment of the small deformation landslide within the current monitoring period is obtained, in millimeters. The actual displacement increment comes from the fused time-series cumulative displacement data output in step 102, specifically the difference between the cumulative displacement at the current time and the previous time.

[0143] Sub-step 42: Determine the warning ratio between the actual displacement increment and the dynamic warning threshold, and determine the corresponding warning level based on the warning ratio.

[0144] In sub-step 42, the ratio of the actual displacement increment to the dynamic early warning threshold is calculated. Based on the range of this ratio, small deformation landslides are divided into three early warning levels, and corresponding disposal measures are specified.

[0145] The formula for calculating the early warning ratio is as follows: Early warning ratio = Actual displacement increment / Dynamic early warning threshold The tiered early warning rules are as follows: (1) Blue Alert (Attention Level): Alert ratio is less than 0.5. The response measures are: strengthen monitoring and update data once a week.

[0146] (2) Yellow Alert (Warning Level): The alert ratio is greater than or equal to 0.5 and less than 0.8. The response measures are: intensified monitoring, data updates once a day, and on-site investigations.

[0147] (3) Red Alert (Danger Level): The alert ratio is greater than or equal to 0.8. The response measures are: immediately activate emergency monitoring, issue early warning information, and organize the evacuation of personnel from the danger zone.

[0148] In some embodiments of this application, it also includes: Step 106: Extract the spatial attention weights and temporal attention weights from the recognition network model.

[0149] In step 106, spatial attention weights and temporal attention weights can be extracted from the trained bi-branch attention long short-term memory network model.

[0150] Spatial attention weights are derived from the output of the spatial attention layer, denoted as... A larger weight value indicates a higher importance of the corresponding spatial feature, which can be used to identify key deformation areas of landslides. The temporal attention weights are derived from the output of the temporal attention layer and are denoted as... The larger the weight value, the higher the importance of the corresponding time feature, which can be used to identify the key inducing time nodes of landslide deformation.

[0151] Step 107: Determine the key deformation area of ​​the landslide based on the spatial attention weight.

[0152] In step 107, spatial attention weights can be based on the bi-branch attention long short-term memory model. Identify the key deformation areas of landslide deformation.

[0153] Specifically, the spatial attention weights of all grids within a typical landslide are extracted, and the grids with high spatial attention weights (such as...) are assigned priority. The grid is identified as a high-attention area, i.e., the critical deformation area of ​​the landslide.

[0154] In addition, the deformation rate and deformation gradient characteristics of key deformation areas can be statistically analyzed to clarify the spatial distribution pattern of key deformation areas (such as the leading edge, trailing edge, or middle).

[0155] Step 108: Determine the key triggering time nodes of the landslide based on the time attention weight.

[0156] In step 108, the key inducing time nodes of landslide deformation are identified based on the temporal attention weights of the bi-branch attention long short-term memory model.

[0157] Specifically, the temporal attention weights of all grids within a typical landslide are extracted, and those with high temporal attention weights (such as...) are assigned priority. The moment of thinning was identified as the high-attention moment, which is the key triggering time node for landslides.

[0158] At the same time, by combining reservoir water data and rainfall data, the dominant inducing factors of landslide deformation can be determined by the reservoir water movement stage and rainfall intensity corresponding to the key inducing time points.

[0159] Step 109: Based on the key deformation area of ​​the landslide and the key induction time node of the landslide, determine the dominant inducing factor and critical parameters of the landslide deformation.

[0160] In step 109, based on the key deformation area and key induction time point of the landslide, combined with the reservoir water data and the rainfall data, the deformation mechanism of a typical landslide can be quantitatively analyzed to determine the dominant inducing factors and their corresponding contribution rates, as well as the critical parameters of landslide deformation.

[0161] As an example, for a landslide with reduced buoyancy, the contribution rate of buoyancy force is calculated. and contribution rate of rainfall The calculation formula is as follows:

[0162]

[0163] when The study determined that buoyancy-induced weight reduction was the dominant triggering factor, while rainfall was a contributing factor (with a contribution rate of less than 40%). The critical reservoir water level for buoyancy-induced weight reduction (168 meters) and the critical rainfall intensity for contributing to landslide deformation (cumulative rainfall of ≥100 mm over 3 consecutive days) were also determined. The critical reservoir water level for buoyancy-induced weight reduction and the critical rainfall intensity for contributing to landslide deformation can be considered as critical parameters for landslide deformation.

[0164] For hydrodynamic pressure-type landslides, calculate the contribution rate of hydrodynamic pressure. and contribution rate of rainfall , This is the ratio of the displacement increment during the reservoir water drop period to the total displacement increment. The study determined that hydrodynamic pressure was the dominant inducing factor, and also identified the critical rate of hydrodynamic pressure decline (reservoir water level decline rate greater than or equal to 0.3 m / day) and the hydrodynamic pressure lag time (calculated to be 36 days). The critical rate of hydrodynamic pressure decline and the hydrodynamic pressure lag time can be considered as critical parameters for landslide deformation.

[0165] Step 110: Generate landslide prevention and control recommendations based on the dominant inducing factors and the critical parameters of landslide deformation.

[0166] In step 110, based on the quantitative analysis of the deformation mechanism of typical landslides, the dominant inducing factors and critical parameters of landslide deformation are determined. Thus, targeted landslide prevention and control suggestions linked to the quantitative parameters can be generated based on the dominant inducing factors and the critical parameters of landslide deformation.

[0167] For example, for floating weight reduction landslides, the landslide prevention recommendation is to complete the leading edge anti-slide project before the water level of the floating weight reduction reservoir rises to 165 meters.

[0168] For landslides caused by dynamic water pressure, the recommended prevention and control measure is to complete the drainage project before the reservoir water level drops (the rate of decrease in dynamic water pressure is greater than or equal to 0.3 meters per day).

[0169] In this embodiment, multi-source basic data of the reservoir bank research area is acquired. This multi-source basic data includes synthetic aperture radar (SAR) image data. Deformation inversion is performed on the SAR image data to obtain at least two initial deformation data sets. Based on these initial deformation data sets, fused deformation data is generated. A multi-dimensional feature vector is generated based on the fused deformation data and the multi-source basic data. A pre-trained recognition network model is used based on the multi-dimensional feature vector to obtain landslide recognition results. For deformed landslides identified in the landslide recognition results, a dynamic early warning threshold is determined. Based on this dynamic early warning threshold, the corresponding early warning level is determined. This achieves the goal of generating fused deformation data by weighted fusion of at least two initial deformation data sets. This approach balances the accuracy and spatial coverage of deformation monitoring, enabling a more comprehensive capture of the spatiotemporal deformation patterns of landslides. Furthermore, it improves the accuracy of landslide recognition and enables dynamic adjustment of landslide early warning, providing reliable technical support for disaster prevention and control of reservoir bank landslides.

[0170] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0171] Reference Figure 2 The diagram shows a structural block diagram of a deformation identification and analysis device for reservoir bank landslides provided in some embodiments of the present invention, which may specifically include the following modules: The data acquisition module 201 is used to acquire multi-source basic data of the reservoir shore research area; wherein, the multi-source basic data includes synthetic aperture radar image data; The data fusion module 202 is used to perform deformation inversion on the synthetic aperture radar image data to obtain at least two initial deformation data, and generate fused deformation data based on the at least two initial deformation data; The feature generation module 203 is used to generate a multi-dimensional feature vector based on the fused deformation data and the multi-source basic data; The identification module 204 is used to obtain landslide identification results based on the multi-dimensional feature vector and a pre-trained identification network model. The early warning module 205 is used to determine a dynamic early warning threshold for deformed landslides in the landslide identification results, and to determine the corresponding early warning level based on the dynamic early warning threshold.

[0172] In one embodiment of this application, the multi-source basic data includes deformation verification data, and the data fusion module 202 includes: Based on the deformation verification data, the accuracy of the at least two initial deformation data is evaluated, and the accuracy weights corresponding to the at least two initial deformation data are determined. Based on the precision weights, the at least two initial deformation data are weighted and fused to generate fused deformation data.

[0173] In one embodiment of this application, the multi-source basic data further includes reservoir water data and rainfall data, and the feature generation module 203 includes: Based on the reservoir water data, multiple reservoir water movement stages are determined. Based on the rainfall data, a heavy rainfall event is identified; Based on the fused deformation data, the multiple reservoir water movement stages, and the heavy rainfall event, a multi-dimensional feature vector is generated; wherein, the multi-dimensional feature vector includes any one or more of the following: basic temporal displacement features, deformation gradient features, spatiotemporal correlation features, and reservoir water-rainfall coupling features.

[0174] In one embodiment of this application, the early warning module 205 includes: Extract deformation stage information, reservoir water change information, and rainfall information of the deformed landslide; Based on the deformation stage information, reservoir water change information, and rainfall information, the coupling influence coefficient is determined; The dynamic early warning threshold for the deformed landslide is determined based on the preset benchmark displacement increment threshold and the coupling influence coefficient.

[0175] In one embodiment of this application, the early warning module 205 includes Obtain the actual displacement increment of the deformed landslide; Determine the warning ratio between the actual displacement increment and the dynamic warning threshold, and determine the corresponding warning level based on the warning ratio.

[0176] In one embodiment of this application, it further includes: The weight extraction module is used to extract the spatial attention weights and temporal attention weights in the recognition network model. The key area determination module is used to determine the key deformation area of ​​the landslide based on the spatial attention weight; The time node determination module is used to determine the key triggering time nodes of the landslide based on the time attention weight. The analysis module is used to determine the dominant inducing factors and critical parameters of the deformed landslide based on the key deformation area and the key inducing time node of the landslide. The suggestion generation module is used to generate landslide prevention suggestions based on the dominant inducing factors and the critical parameters of landslide deformation.

[0177] In one embodiment of this application, the pre-trained recognition network model is a dual-branch attention long short-term memory network model; The bi-branch attention long short-term memory network model includes a spatial attention branch and a temporal attention branch. The spatial attention branch and the temporal attention branch are used to assign attention weights to spatially relevant features and temporally relevant features in the input features, respectively.

[0178] In one embodiment of this application, the training step of the recognition network model includes: The sample selection module is used to select landslide samples and non-landslide samples, and to define a landslide type label and a deformation stage label for each sample; wherein, the deformation stage label is based on the quantitative classification of landslide deformation trend in each sample; The sample generation module is used to divide the samples into training set, validation set and test set according to the landslide type label and the deformation stage label using a hierarchical sampling method. The training module is used to optimize the training set using cross-validation to obtain a target training set, and to train the recognition network model using the target training set.

[0179] The deformation identification and analysis device for reservoir bank landslides 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), televisions (TVs), ATMs, or self-service machines, etc. This application embodiment does not impose specific limitations.

[0180] The deformation identification and analysis device for reservoir bank landslides in this application embodiment can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application embodiment does not specifically limit it.

[0181] The deformation identification and analysis device for reservoir bank landslides provided in this application embodiment can achieve... Figure 1 The various processes implemented by the deformation identification and analysis device for reservoir bank landslides in the method embodiment are not described in detail here to avoid repetition.

[0182] In this embodiment, multi-source basic data of the reservoir bank research area is acquired. This multi-source basic data includes synthetic aperture radar (SAR) image data. Deformation inversion is performed on the SAR image data to obtain at least two initial deformation data sets. Based on these initial deformation data sets, fused deformation data is generated. A multi-dimensional feature vector is generated based on the fused deformation data and the multi-source basic data. A pre-trained recognition network model is used based on the multi-dimensional feature vector to obtain landslide recognition results. For deformed landslides identified in the landslide recognition results, a dynamic early warning threshold is determined. Based on this dynamic early warning threshold, the corresponding early warning level is determined. This achieves the goal of generating fused deformation data by weighted fusion of at least two initial deformation data sets. This approach balances the accuracy and spatial coverage of deformation monitoring, enabling a more comprehensive capture of the spatiotemporal deformation patterns of landslides. Furthermore, it improves the accuracy of landslide recognition and enables dynamic adjustment of landslide early warning, providing reliable technical support for disaster prevention and control of reservoir bank landslides.

[0183] Optionally, this application embodiment also provides an electronic device, including a processor 300, a memory 309, and a program or instructions stored in the memory 309 and executable on the processor 310. When the program or instructions are executed by the processor 310, they implement the various processes of the above-described deformation identification and analysis method embodiment for reservoir bank landslides and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0184] It should be noted that the electronic devices in the embodiments of this application include the mobile electronic devices and non-mobile electronic devices described above.

[0185] Figure 3 A schematic diagram of the hardware structure of an electronic device to implement an embodiment of this application.

[0186] The electronic device 300 includes, but is not limited to, components such as: a radio frequency unit 301, a network module 302, an audio output unit 303, an input unit 304, a sensor 305, a display unit 306, a user input unit 307, an interface unit 308, a memory 309, and a processor 310. The user input unit 307 includes a touch panel 3071 and other input devices 3072; the display unit 306 includes a display panel 3061; and the input unit includes a graphics processor 3041 and a microphone 3042.

[0187] Those skilled in the art will understand that the electronic device 300 may also include a power supply (such as a battery) for supplying power to various components. The power supply may be logically connected to the processor 310 through a power management system, thereby enabling functions such as managing charging, discharging, and power consumption through the power management system. Figure 3 The electronic device structure shown does not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown, or combine certain components, or have different component arrangements, which will not be elaborated here.

[0188] 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 deformation identification and analysis method for reservoir bank landslides and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0189] 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.

[0190] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface and the processor are coupled. The processor is used to run programs or instructions to implement the various processes of the above-described deformation identification and analysis method embodiment for reservoir bank landslides, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0191] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0192] 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.

[0193] 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.

[0194] 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 deformation identification and analysis method for reservoir bank landslides, characterized in that, The method includes: Acquire multi-source basic data of the reservoir shore research area; wherein, the multi-source basic data includes synthetic aperture radar image data; Deformation inversion is performed on the synthetic aperture radar image data to obtain at least two initial deformation data, and fused deformation data is generated based on the at least two initial deformation data; Based on the fused deformation data and the multi-source basic data, a multi-dimensional feature vector is generated; Based on the aforementioned multi-dimensional feature vectors, a pre-trained recognition network model is used to obtain landslide recognition results; For the deformed landslides in the landslide identification results, a dynamic early warning threshold is determined, and the corresponding early warning level is determined based on the dynamic early warning threshold.

2. The method according to claim 1, characterized in that, The multi-source basic data includes deformation verification data, and the generation of fused deformation data based on the at least two initial deformation data includes: Based on the deformation verification data, the accuracy of the at least two initial deformation data is evaluated, and the accuracy weights corresponding to the at least two initial deformation data are determined. Based on the precision weights, the at least two initial deformation data are weighted and fused to generate fused deformation data.

3. The method according to claim 1, characterized in that, The multi-source basic data also includes reservoir water data and rainfall data. The generation of a multi-dimensional feature vector based on the fused deformation data and the multi-source fused data includes: Based on the reservoir water data, multiple reservoir water movement stages are determined. Based on the rainfall data, a heavy rainfall event is identified; Based on the fused deformation data, the multiple reservoir water movement stages, and the heavy rainfall event, a multi-dimensional feature vector is generated; wherein, the multi-dimensional feature vector includes any one or more of the following: basic temporal displacement features, deformation gradient features, spatiotemporal correlation features, and reservoir water-rainfall coupling features.

4. The method according to claim 1, characterized in that, For deformed landslides identified in the landslide identification results, a dynamic early warning threshold is determined, including: Extract deformation stage information, reservoir water change information, and rainfall information of the deformed landslide; Based on the deformation stage information, reservoir water change information, and rainfall information, the coupling influence coefficient is determined; The dynamic early warning threshold for the deformed landslide is determined based on the preset benchmark displacement increment threshold and the coupling influence coefficient.

5. The method according to claim 1, characterized in that, The step of determining the corresponding warning level based on the dynamic warning threshold includes... Obtain the actual displacement increment of the deformed landslide; Determine the warning ratio between the actual displacement increment and the dynamic warning threshold, and determine the corresponding warning level based on the warning ratio.

6. The method according to claim 1, characterized in that, Also includes: Extract the spatial attention weights and temporal attention weights from the recognition network model; Based on the spatial attention weights, the key deformation areas of the landslide are determined; Based on the aforementioned time attention weights, the key triggering time points for landslides are determined; Based on the key deformation area of ​​the landslide and the key induction time node of the landslide, the dominant inducing factors and critical parameters of the landslide deformation are determined. Based on the dominant inducing factors and the critical parameters of landslide deformation, landslide prevention and control recommendations are generated.

7. The method according to any one of claims 1-6, characterized in that, The pre-trained recognition network model is a dual-branch attention long short-term memory network model; The bi-branch attention long short-term memory network model includes a spatial attention branch and a temporal attention branch. The spatial attention branch and the temporal attention branch are used to assign attention weights to spatially relevant features and temporally relevant features in the input features, respectively.

8. The method according to any one of claims 1-6, characterized in that, The training steps of the recognition network model include: Select landslide samples and non-landslide samples, and define landslide type label and deformation stage label for each sample; wherein, the deformation stage label is based on the quantitative classification of landslide deformation trend in each sample; A stratified sampling method was adopted, and the samples were divided into training set, validation set and test set according to the landslide type label and the deformation stage label; The training set is optimized using cross-validation to obtain a target training set, which is then used to train the recognition network model.

9. A deformation identification and analysis device for reservoir bank landslides, characterized in that, The method includes: The data acquisition module is used to acquire multi-source basic data of the reservoir shore research area; wherein, the multi-source basic data includes synthetic aperture radar image data; The data fusion module is used to perform deformation inversion on the synthetic aperture radar image data to obtain at least two initial deformation data, and generate fused deformation data based on the at least two initial deformation data; The feature generation module is used to generate multi-dimensional feature vectors based on the fused deformation data and the multi-source basic data; The identification module is used to obtain landslide identification results based on the multi-dimensional feature vector and a pre-trained identification network model. The early warning module is used to determine a dynamic early warning threshold for deformed landslides in the landslide identification results, and to determine the corresponding early warning level based on the dynamic early warning threshold.

10. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and capable of running on the processor, wherein the computer program, when executed by the processor, implements the method as described in any one of claims 1 to 8.

11. A readable storage medium, characterized in that, A computer program is stored on the readable storage medium, which, when executed by a processor, implements the method as described in any one of claims 1 to 8.

Images