A method and system for monitoring dynamic displacement of riverbank slopes

Abstract

This application provides a method and system for dynamic displacement monitoring of riverbank slopes, belonging to the field of computer models and systems for dynamic monitoring of riverbank slope displacement. It addresses the problems of poor overall coverage, low data fusion accuracy, and lack of dynamic early warning in related technologies. The method employs an unmanned aerial vehicle (UAV) equipped with an ultra-wideband microwave dielectric sensor and a hyperspectral camera. Based on multi-source data, a dynamic risk heat map is constructed, an adaptive cruise path is planned, and microwave dielectric and hyperspectral data are collected simultaneously. After denoising, a dielectric distribution matrix is ​​constructed. An improved DeepLabv3+ model is used to identify displacement regions, and an Attention-BiGRU network is used to predict displacement trends. A stability index is constructed by combining soil parameters, and graded early warnings are issued. This application achieves integrated closed-loop monitoring of riverbank slope displacement, improving detection accuracy and intelligence.

Description

Technical Field

[0001] This application relates to the field of computer models and systems for dynamically monitoring riverbank slope displacement, and more particularly to a method and system for monitoring dynamic displacement of riverbank slopes. The core technology is a specific computer model for dynamically monitoring riverbank slope displacement and a computer system based on the specific computer model. Background Technology

[0002] With the rapid development of water conservancy projects, the stability monitoring of riverbanks has received increasing attention. Slope displacement is a key indicator for assessing levee safety and providing early warning of collapse risks. Traditional slope displacement monitoring mainly relies on ground measurement methods such as GNSS displacement monitoring stations and total stations, obtaining local displacement data by setting up a small number of monitoring points.

[0003] In recent years, the rapid development of drone technology has provided new technical means for bank slope monitoring. Existing technologies include schemes that use drones equipped with visible light cameras or lidar for bank slope inspection, acquiring images or point cloud data of the bank slope surface and comparing images or point clouds from different periods to identify displacement areas. Some technologies also attempt to apply deep learning models to displacement identification or use time-series prediction models for displacement trend prediction.

[0004] However, existing monitoring technologies still have the following shortcomings: First, single-sensor monitoring methods are insufficient to comprehensively perceive the condition of the bank slope, and surface images cannot reflect changes in the physical properties of the soil inside the bank slope, resulting in insufficient accuracy and timeliness of displacement identification; Second, UAV inspections typically use fixed flight paths or equal-precision scanning modes, lacking targeted monitoring of high-risk areas, leading to unreasonable monitoring efficiency and resource allocation; Third, displacement identification and trend prediction are mostly independent modules, failing to form an integrated closed loop from risk prediction, data collection, displacement identification, trend prediction to graded early warning; Fourth, early warning mechanisms are mostly based on static threshold judgments, without integrating real-time hydrological conditions and soil physical and mechanical parameters, limiting the accuracy and engineering guidance value of early warning results. Summary of the Invention

[0005] This application provides a method and system for monitoring dynamic displacement of riverbank slopes. It can achieve intelligent monitoring of the entire process of riverbank slope displacement by integrating risk heat map-driven adaptive acquisition, multi-source data fusion identification, hybrid-driven trend prediction, and dynamic threshold-based hierarchical early warning, thereby improving the accuracy of displacement identification and the reliability of early warning.

[0006] Firstly, this application provides a method for monitoring the dynamic displacement of riverbank slopes. The method involves acquiring multi-source basic data of the target riverbank slope, including static basic data and dynamic time-series data, wherein the static basic data includes soil physical and mechanical parameters; based on the multi-source basic data, constructing a dynamic risk heat map characterizing the spatiotemporal distribution of slope risk; based on the dynamic risk heat map, planning an adaptive cruise path for a drone, and controlling the drone equipped with multiple sensors to fly along the adaptive cruise path to simultaneously collect microwave dielectric data and hyperspectral image data of the slope; denoising the microwave dielectric data and combining it with its spatial location information to construct a dielectric distribution characterizing the spatial distribution of dielectric properties within the slope. The dielectric distribution matrix is ​​input into a trained deep learning model for pixel-level anomaly detection to identify potential displacement regions. The hyperspectral image data is then used to verify the authenticity of these potential displacement regions, determining the true displacement regions. The multi-source time-series data corresponding to the true displacement regions are input into a trained time-series prediction model to predict slope displacement trend parameters within a preset timeframe. Based on the displacement trend parameters and the soil physical and mechanical parameters, a dynamic stability assessment model is constructed to calculate the slope stability index. According to the relationship between the stability index and a preset dynamic threshold, a corresponding early warning level is generated and output.

[0007] By adopting the above technical solutions, the computer model provided in this application for dynamically monitoring riverbank slope displacement integrates risk prediction, dynamic data acquisition, precise identification, trend prediction, and tiered early warning into a unified monitoring closed loop. A dynamic risk heat map is constructed based on multi-source basic data, allowing monitoring resources to be tilted towards high-risk areas, thus improving monitoring efficiency. The fusion of microwave dielectric data and hyperspectral image data enables the coordinated perception of the internal dielectric properties and surface deformation of the bank slope, enhancing the accuracy of displacement identification. Furthermore, by combining deep learning time-series prediction results with soil physical and mechanical parameters, a dynamic stability assessment model is constructed, making the early warning results more valuable for engineering guidance.

[0008] Furthermore, the step of constructing the dynamic risk heat map includes: spatially dividing the target riverbank slope into multiple regular grids; dividing each parameter in the multi-source basic data into intervals and assigning a label value to different intervals of each parameter; introducing time decay weights to historical data at different time points and introducing feature weights to different parameters, wherein the time decay weight is a positive number less than 1, and the weight of recent data is greater than the weight of distant data; calculating the comprehensive risk weighted frequency of each grid according to the label value, the time decay weight, and the feature weight; classifying the risk level of each grid according to the comprehensive risk weighted frequency, and generating the dynamic risk heat map.

[0009] By adopting the above technical solution, this application introduces time decay weight and feature weight, so that the risk heat map can not only reflect the degree of influence of different parameters on slope stability, but also reflect the higher reference value of recent data, thereby improving the accuracy of risk prediction.

[0010] Furthermore, the step of planning the adaptive cruise path of the UAV includes: assigning corresponding priority weights to grids of different risk levels; constructing a multi-objective optimization function, wherein the optimization objectives of the multi-objective optimization function are to minimize the total length of the flight path, maximize the sum of the priority weights of the covered grids, and minimize flight energy consumption; solving the multi-objective optimization function to obtain the adaptive cruise path; wherein, when controlling the UAV to fly according to the adaptive cruise path, the frequency of coverage of high-risk level grids is higher than that of low-risk level grids.

[0011] By adopting the above technical solution, this application achieves optimized allocation of monitoring resources, prioritizes coverage of high-risk areas and increases monitoring frequency, and maximizes the ability to perceive potential risk areas under limited energy and time constraints.

[0012] Further, the step of constructing the dielectric distribution matrix includes: processing the microwave dielectric data using a combined denoising method, the combined denoising method including notch filtering, wavelet thresholding denoising and adaptive Wiener filtering performed sequentially; mapping the denoised microwave dielectric data onto a pre-divided spatial grid; calculating the mean of all microwave dielectric data in each spatial grid, and using the mean as a matrix element, filling the grid according to the spatial arrangement order to construct the dielectric distribution matrix.

[0013] By adopting the above technical solutions, this application effectively eliminates environmental electromagnetic noise and Gaussian white noise through a combined denoising method, thereby improving the signal-to-noise ratio of microwave dielectric data; the dielectric distribution matrix realizes the spatial structured expression of the dielectric properties inside the bank slope, providing a high-quality data foundation for subsequent deep learning recognition.

[0014] Further, the step of identifying potential displacement regions and verifying their authenticity includes: inputting the dielectric distribution matrix into an improved DeepLabv3+ model, outputting the anomaly probability of each spatial grid, and marking grids with anomaly probabilities greater than a preset threshold as potential displacement regions, wherein the improved DeepLabv3+ model replaces its backbone network with a MobileNetV3 network; aligning the grid coordinates of the marked potential displacement regions with the pixel coordinates of the hyperspectral image data; calculating the weighted structural similarity index W-SSIM between the current hyperspectral image and historical hyperspectral images in the potential displacement regions, wherein pixels in key areas of the bank slope are assigned higher weights than those in other areas; and determining potential displacement regions with W-SSIM indices less than a preset similarity threshold as real displacement regions.

[0015] By adopting the above technical solution, this application applies the semantic segmentation model to the anomaly detection of the dielectric distribution matrix and verifies it by combining multiple hyperspectral images. This not only leverages the sensitivity of microwave dielectric data to internal anomalies, but also eliminates false detection areas through image verification. The introduction of regional weights enhances the focus on key parts such as the water-facing side and slope toe of the dike, thereby improving the accuracy and relevance of displacement identification.

[0016] Furthermore, the step of predicting the slope displacement trend parameters within a preset future time period includes: inputting multi-dimensional time-series data corresponding to the actual displacement area into an attention-enhanced bidirectional gated recurrent unit network; the attention-enhanced bidirectional gated recurrent unit network extracts the positive and negative features of the time-series data through forward gated recurrent units and backward gated recurrent units, respectively, and weights and fuses the hidden states output by the bidirectional gated recurrent units through an attention mechanism layer to obtain a comprehensive hidden state; mapping the comprehensive hidden state to a displacement distance sequence within a preset future time period through a fully connected layer; calculating the displacement rate based on the displacement distance sequence, and extracting the maximum displacement distance from the displacement distance sequence.

[0017] By adopting the above technical solution, this application captures the forward and reverse evolution patterns of displacement time series through a bidirectional GRU, and enhances the feature extraction of key time nodes through an attention mechanism, thereby improving the accuracy of displacement trend prediction.

[0018] Furthermore, the step of predicting the slope displacement trend parameters within a preset time period further includes: calculating the radius of the displacement influence range based on a simplified formula of geotechnical elasticity theory, combined with the maximum displacement distance, the thickness of the main compressible layer in the soil physical and mechanical parameters, and the empirical coefficient of soil compressibility. The radius of the displacement influence range is proportional to the square root of the maximum displacement distance and the square root of the thickness of the main compressible layer, and the proportionality coefficient is the empirical coefficient of soil compressibility.

[0019] By adopting the above technical solution, this application combines data-driven prediction results with geotechnical mechanics physical models to calculate the radius of displacement influence range with clear engineering significance. This solves the problems of poor interpretability and difficulty in engineering application of pure data-driven models, and provides a direct basis for delineating warning zones and guiding reinforcement scope.

[0020] Further, the step of constructing a dynamic stability assessment model and calculating the stability index of the bank slope includes: constructing the calculation formula for the stability index SI as follows: the stability index SI equals the product of the first weighting coefficient and the reciprocal of the displacement rate, plus the product of the second weighting coefficient and the reciprocal of the maximum displacement distance, plus the product of the third weighting coefficient and the reciprocal of the radius of the displacement influence range, plus the product of the fourth weighting coefficient and the soil cohesion, plus the product of the fifth weighting coefficient and the soil internal friction angle, minus the product of the sixth weighting coefficient and the soil moisture content, minus the product of the seventh weighting coefficient and the groundwater level, wherein the first to seventh weighting coefficients are preset weighting coefficients; obtaining the displacement distance sequence, the displacement rate and the maximum displacement distance, and the soil physical and mechanical parameters, and substituting them into the formula to calculate the stability index SI; correcting the stability index SI based on real-time acquired hydrological data to obtain the corrected stability index SI'.

[0021] By adopting the above technical solution, this application constructs a dynamic stability assessment model that integrates predicted parameters and soil physical and mechanical parameters, and introduces a hydrological dynamic correction factor, so that the assessment results can respond to changes in hydrological conditions in real time, thereby improving the dynamism and accuracy of stability assessment.

[0022] Further, the step of generating and outputting the corresponding early warning level includes: dynamically setting a first early warning threshold and a second early warning threshold according to the engineering grade and revetment structure type of the target riverbank; comparing the corrected stability index SI' with the first early warning threshold and the second early warning threshold; if the corrected stability index SI' is less than the first early warning threshold, it is determined to be a level one early warning; if the corrected stability index SI' is greater than or equal to the first early warning threshold and less than the second early warning threshold, it is determined to be a level two early warning; if the corrected stability index SI' is greater than or equal to the second early warning threshold, it is determined to be a level three early warning; retrieving a preset reinforcement and maintenance scheme library, obtaining a preset reinforcement and maintenance scheme associated with the early warning level, and outputting the corresponding early warning level and the preset reinforcement and maintenance scheme.

[0023] By adopting the above technical solution, this application achieves dynamic adaptation of early warning thresholds to engineering level and revetment structure type, and associates early warning level with specific reinforcement and maintenance plan, so that the early warning results have clear engineering guidance value and facilitate rapid response by operation and maintenance personnel.

[0024] Secondly, this application provides a system for monitoring dynamic displacement of riverbank slopes, used to execute the monitoring method described in any of the first aspects, comprising: a data acquisition module, including an ultra-wideband microwave dielectric radiation sensor, a hyperspectral imaging spectrometer, an inertial measurement unit, and a real-time dynamic positioning module mounted on an unmanned aerial vehicle (UAV) platform, and a ground data acquisition terminal, for acquiring the multi-source basic data, the microwave dielectric data, and the hyperspectral image data; a dynamic risk heat map construction module, for constructing and updating the dynamic risk heat map based on the multi-source basic data; a UAV path planning module, for planning the adaptive cruise path based on the dynamic risk heat map and controlling the flight of the UAV; a data processing module, for denoising the microwave dielectric data and constructing the dielectric distribution matrix; a displacement region identification module, equipped with the trained deep learning model, for identifying and confirming the real displacement region; a displacement trend prediction module, equipped with the trained time-series prediction model, for predicting the displacement trend parameters; and a stability assessment and early warning module, for constructing the dynamic stability assessment model, calculating the stability index, generating and outputting the early warning level and the corresponding reinforcement and maintenance plan.

[0025] By adopting the above technical solution, this application provides a complete system implementation corresponding to the method. The modules are linked by data flow to form an integrated monitoring closed loop, which can realize unattended, fully automatic monitoring of dynamic displacement of riverbank slopes.

[0026] In summary, this application has at least the following beneficial effects:

[0027] A method and system for monitoring dynamic displacement of riverbank slopes are provided, realizing integrated closed-loop monitoring of risk prediction, dynamic data acquisition, accurate identification, trend prediction, and hierarchical early warning;

[0028] By using dynamic risk heatmaps to drive adaptive path planning, monitoring resources are tilted towards high-risk areas, significantly improving monitoring efficiency and resource utilization.

[0029] By integrating microwave dielectric data and hyperspectral image data, and combining them with an improved deep learning model, we can achieve collaborative identification of internal anomalies and surface deformations of bank slopes, thereby improving the accuracy of displacement detection.

[0030] By combining Attention-BiGRU time series prediction with a geotechnical mechanics physical model, the radius of displacement influence with engineering significance is calculated, thereby enhancing the interpretability of the prediction results.

[0031] A dynamic stability index is constructed that integrates prediction parameters, soil parameters, and hydrological correction factors to achieve dynamic adaptation of early warning thresholds with engineering levels and revetment structures. Furthermore, the early warning level is linked to reinforcement schemes to enhance the engineering guidance value of the early warning.

[0032] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0033] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0034] Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented is shown.

[0035] Figure 2 A flowchart of a method for monitoring dynamic displacement of riverbank slopes according to an embodiment of this application is shown. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, 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.

[0037] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0038] This application provides a method and system for monitoring dynamic displacement of riverbank slopes. It achieves optimized allocation of monitoring resources, accurate displacement identification and interpretable trend prediction through an integrated approach of risk heat map-driven adaptive acquisition, multi-source data fusion identification, hybrid-driven trend prediction, and dynamic threshold-based hierarchical early warning, thereby improving the level of monitoring intelligence and engineering application value.

[0039] Figure 1 A schematic diagram of an exemplary operating environment in which embodiments of this application can be implemented is shown.

[0040] Reference Figure 1 The operating environment consists of four core components: an aerial data acquisition platform layer, a ground infrastructure layer, a data processing and computing layer, and a communication and interaction layer. These four components work together to support the complete closed-loop implementation of the method for monitoring the dynamic displacement of riverbank slopes.

[0041] The aerial data acquisition platform utilizes industrial-grade drones as its flight platform. These drones are preferably hexacopter or quadcopter models, possessing a flight endurance of over 45 minutes and an IP43 or higher protection rating to meet the requirements of continuous operation under complex weather conditions in river channels. The drones integrate an ultra-wideband microwave dielectric radiation sensor, a hyperspectral imaging spectrometer, an inertial measurement unit, and a real-time dynamic positioning module. The ultra-wideband microwave dielectric radiation sensor is equipped with a dual-band antenna with a center frequency of 400MHz to 6GHz, preferably operating in a 1GHz and 5GHz dual-band mode. The 1GHz band is used to collect dielectric property data penetrating deep soil at depths of 8 to 10 meters, while the 5GHz band is used to collect high-resolution dielectric data at depths of 0.5 to 1.5 meters in the surface layer. The hyperspectral imaging spectrometer has a spectral range of 400 to 1000 nanometers in the visible and near-infrared bands, with a spectral resolution better than 5 nanometers and a spatial resolution of 0.1 meters. It is used to acquire high-resolution spectral images of the riverbank surface to identify material composition and surface deformation. The inertial measurement unit (IMU) employs fiber optic gyroscopes or microelectromechanical systems (MEMS) to record the UAV's three-axis attitude angles and acceleration information at sampling frequencies above 200Hz, aiding in the spatiotemporal synchronization and geometric correction of multi-source data. The real-time dynamic positioning module supports joint positioning using GPS, BeiDou, and GLONASS systems, enabling real-time acquisition of spatial coordinates with centimeter-level accuracy during flight, with horizontal positioning accuracy better than 2 cm and vertical positioning accuracy better than 3 cm. The platform autonomously flies according to an adaptive cruise path planned by the data processing and computation layer, achieving spatiotemporal synchronous acquisition of microwave dielectric data and hyperspectral imagery data. During acquisition, precise timestamps and spatial coordinate labels are added to each image frame and each dielectric data point using data from the real-time dynamic positioning module and the IMU.

[0042] The ground infrastructure layer includes UAV take-off and landing sites deployed along the monitored river section, ground data acquisition terminals, and reference stations providing differential signals for real-time dynamic positioning. The UAV take-off and landing sites adopt a modular design, with each site covering a monitoring range of 8 to 10 kilometers. Four to five sites are deployed along the 32-kilometer river channel. Each site is equipped with automatic charging or battery replacement devices, supporting multiple take-off and landing missions for UAVs without human intervention. The ground data acquisition terminals include portable vane shear testers, heavy-duty dynamic penetrometers, soil moisture meters, and water level gauges. These are used to supplement the collection of in-situ soil cohesion, internal friction angle, compressibility layer thickness, moisture content, and groundwater level parameters. The data collection cycle is set weekly or monthly depending on monitoring needs, increasing to daily during the flood season. Reference stations are deployed at intervals of 10 to 15 kilometers along both banks of the river. Working in conjunction with the airborne real-time dynamic positioning module, they broadcast differential correction signals to the UAVs via data communication links, ensuring consistent positioning accuracy across the entire system.

[0043] The data processing and computing layer includes edge computing units, data processing servers, and a data storage system. Edge computing units are deployed near the UAV takeoff and landing site, equipped with high-performance graphics processors and at least 32GB of RAM, for real-time or near-real-time preprocessing of the acquired raw data. Preprocessing includes using a combined denoising method to sequentially perform notch filtering to remove 50Hz power frequency noise, wavelet thresholding to remove Gaussian white noise, and adaptive Wiener filtering for smoothing the microwave dielectric data. The processed dielectric data is then mapped using a spatial grid to construct a dielectric distribution matrix. The data processing server is deployed at the monitoring center, equipped with multiple high-performance graphics processors and at least 128GB of RAM, for running an improved DeepLabv3+ model for displacement region identification, running an attention-enhanced bidirectional gated recurrent unit network for displacement trend prediction, and constructing dynamic risk heat maps and dynamic stability assessment models. The data storage system adopts a distributed storage architecture, including relational databases and object storage, to store multi-source basic data, historical monitoring data, model parameters, and generated dynamic risk heat maps, dielectric distribution matrices, early warning results, etc. The data storage capacity is designed to be more than 50TB according to the monitoring cycle, and supports online query of at least three years of historical data.

[0044] The communication and interaction layer includes a reliable data transmission network and monitoring and early warning terminals. The data transmission network combines 4G or 5G mobile communication networks with microwave communication to establish a real-time bidirectional data link between the UAV and the ground station. The downlink bandwidth is no less than 50Mbps for transmitting collected data, and the uplink bandwidth is no less than 10Mbps for transmitting control commands and path planning parameters. The monitoring and early warning terminals include a large visualization screen deployed at the monitoring center, a PC application for technical personnel, and a mobile application for maintenance personnel. The large visualization screen uses a 55-inch or larger splicing screen to display real-time slope risk heat maps, displacement area distribution, prediction curves, and the current warning level. The PC application is based on a browser or client architecture and provides interactive functions such as data query, model parameter adjustment, and historical trend analysis. The mobile application supports Android and iOS systems and pushes early warning information and corresponding reinforcement and maintenance plans to maintenance personnel via SMS and WeChat interfaces. The push content includes the warning level, displacement area location, suggested handling measures, and response time limit.

[0045] The four layers mentioned above are interconnected through data flow and control flow: microwave dielectric data and hyperspectral image data collected by the aerial acquisition platform are transmitted in real-time or near real-time to the edge computing unit for preprocessing via the data transmission network. The preprocessed data is then synchronized to the data processing server for in-depth analysis and model calculations. The dynamic risk heat map generated by the data processing server is fed back to the UAV path planning module to optimize the adaptive cruise path for the next cycle. The displacement area identification results and displacement trend prediction results are pushed to the monitoring and early warning terminal for visualization. Soil parameters, groundwater levels, and other data collected by ground infrastructure are accessed to the data processing and computing layer through manual input or automatic upload, and are fused with the UAV-collected data to participate in the stability assessment model calculation. In the entire operating environment, data acquisition, processing, analysis, and feedback form a closed loop. The aerial platform and ground facilities collaborate, and edge computing and central computing cooperate to jointly support the complete implementation of the method for monitoring dynamic displacement of riverbank slopes.

[0046] Based on the above operating environment, this application embodiment also provides a method for monitoring the dynamic displacement of riverbank slopes. Figure 2 A flowchart of a method for monitoring dynamic displacement of riverbank slopes according to an embodiment of this application is shown.

[0047] Reference Figure 2 The method specifically includes the following steps:

[0048] S1: Obtain multi-source basic data of the target riverbank slope. The multi-source basic data includes static basic data and dynamic time-series data. The static basic data includes soil physical and mechanical parameters.

[0049] In this step, multi-source foundation data is acquired through various methods. Static foundation data includes soil parameters, topographic parameters, and revetment structure parameters, among which the soil physical and mechanical parameters specifically include cohesion. (Unit: kPa), internal friction angle (Unit: °), Soil compressibility coefficient (unit MPa) Main compression layer thickness (unit: m) and moisture content (Unit: %) These parameters were obtained through in-situ tests such as vane shear tests and heavy dynamic penetration tests, and recorded as spatial discrete point data according to the location of the monitoring section. Here, cohesion c is defined as a constant in the soil shear strength that is independent of normal stress; internal friction angle ϕ is defined as the angle corresponding to the slope of the soil shear strength as a function of normal stress; soil compressibility coefficient a is defined as the ratio of the change in void ratio to the change in pressure under lateral confinement conditions; main compressibility layer thickness H is defined as the vertical distance from the ground surface to the bottom boundary of the compressibility layer; and water content w is defined as the ratio of the mass of water to the mass of solid particles in the soil. Topographic parameters were obtained through GIS and total station measurements, and revetment structure parameters included embankment height, slope, and structure type. These data together constitute the basic properties of the bank slope.

[0050] Dynamic time-series data includes historical displacement monitoring data, multi-period hyperspectral imagery data, hydrological data, and meteorological data. Historical displacement monitoring data is continuously collected by GNSS monitoring stations along the river channel at fixed sampling intervals (e.g., 1 day) to form a time-series displacement data. (unit: m), where Indicates a point in time. Displacement. Defined as the monitoring point at time The change in position relative to a reference point. Multi-period hyperspectral imagery data is acquired periodically by drones or satellites; each pixel contains hundreds of consecutive bands of spectral reflectance for subsequent change detection. Hydrological data is obtained from hydrological stations, showing real-time flow rates. (Unit: m³ / s) Flow velocity (unit: m / s) and water level (Unit: m) Meteorological data is obtained from weather stations, including rainfall. (Unit: mm) and temperature (Unit: °C). These dynamic data are also recorded in chronological order, forming a multidimensional time series.

[0051] After all basic data is collected, preprocessing is required to eliminate dimensional differences and outliers. First, the min-max normalization method is used to map data of different dimensions to... Interval, for any parameter Its normalized value The calculation is as follows:

[0052]

[0053] in and These are the minimum and maximum values ​​of the parameter in historical data, respectively. For missing data, linear interpolation is used to complete it, assuming that at time point... and If there are missing values, then at any point in between... The estimated value for:

[0054]

[0055] For abnormal data, use 3 Criterion elimination: Calculate the mean of the data series and standard deviation If a certain data satisfy If a value is found to be outlier, it is identified and removed. The replacement value is then filled in using interpolation. Here, the mean is... Defined as the arithmetic mean and standard deviation of a data series. Defined as the square root of the arithmetic mean of the squared deviations of each value in the data sequence from the mean. After preprocessing, all data are integrated into a standardized base dataset and stored in a local database. This dataset contains the static parameter vector for each spatial grid (or monitoring point). and the dynamic observation vector at each time point This provides a data foundation for the subsequent construction of dynamic risk heat maps.

[0056] S2: Based on the aforementioned multi-source basic data, construct a dynamic risk heat map characterizing the spatiotemporal distribution of bank slope risk.

[0057] The specific method in this step includes: spatially dividing the target riverbank slope into multiple regular grids. The grid size is adaptively adjusted according to the complexity of the slope topography and the monitoring accuracy requirements. The grid size for high-risk potential areas is set to 2m × 2m, and the grid size for regular areas is set to 8m × 8m. Let the entire monitoring area be divided into... There are 1 grid, each grid is indexed by its spatial index. Unique identifier, among which , Here Defined as the number of rows in the grid. Defined as the number of columns in the grid, all of which are positive integers.

[0058] The parameters in the multi-source basic data are divided into intervals. Key parameters that significantly affect slope stability are selected, including soil moisture content. Historical displacement and water level The value range of each parameter is divided into 5 consecutive numerical intervals. For example, for soil moisture content... It can be divided into , , , , ,in to The threshold is determined based on historical statistics. A label value is assigned to each interval of each parameter. ,in Indicates the parameter type. This represents the interval number, with a value ranging from 1 to 5. A higher value indicates a higher risk for that interval. Defined as parameters In the The risk quantification score corresponding to the value taken within each interval.

[0059] The characteristic weights of each parameter are determined using the analytic hierarchy process (AHP). A judgment matrix is ​​constructed to compare the importance of each parameter pairwise. The normalized eigenvector corresponding to the largest eigenvalue of the matrix is ​​calculated to obtain the weight coefficients of each parameter. For the selected key parameter, the characteristic weight of soil moisture content is set. Historical displacement feature weights Water level feature weight ,satisfy Feature weights Defined as parameters The relative importance coefficient of the contribution to slope risk. , , The weights correspond to water content, displacement, and water level, respectively.

[0060] A time decay weight is introduced for historical data at different time points to reflect the physical fact that recent data contributes more to current risk. The time decay weight coefficient is defined. It is a positive number less than 1, preferably ranging from 0.9 to 0.98. In this embodiment, it is taken as... Time decay weighting coefficient Defined as a decay factor where the weight of historical data decreases with age. For data collection times... Historical data, relative to the current moment Time decay weight The calculation is as follows:

[0061]

[0062] in This refers to a time interval, typically one year. For historical data collection time, For the current moment, The unit time interval is denoted as . Therefore, the th Historical data weighted before the year This means that recent data has a higher weight, while older data has a lower weight.

[0063] For any grid Statistical analysis of its comprehensive risk-weighted frequency Traverse all historical time points. and all key parameters If at time Grid parameters The value of falls within the interval If it is inside, then its corresponding tag value is The contribution value of this record is the label value multiplied by the time decay weight, and then multiplied by the parameter feature weight. The contribution values ​​of all records are summed to obtain the overall risk-weighted frequency:

[0064]

[0065] in Indicates at time Grid parameters The interval number to which it belongs For parameters Feature weights, For a moment Time decay weight. Overall risk-weighted frequency. Defined as a grid The cumulative risk value is calculated by adjusting the weights of all historical data records. This accumulation process incorporates the current state of the parameters, their historical evolution patterns, and differences in their respective importance.

[0066] Based on the frequency of comprehensive risk weighting The risk level is determined by the scope of the risk assessment. Two thresholds are set. and In this embodiment, we take , .like Then determine the grid. It is classified as low-risk level one; if If so, it is determined to be a medium-risk level 2; if If so, it is classified as a high-risk level three. Defined as the threshold for dividing low risk into medium risk. Defined as the threshold for distinguishing between medium and high risk. Risk level for each grid. Defined as:

[0067]

[0068] Risk level Defined as a grid The risk level is quantified as follows: 1 indicates low risk, 2 indicates medium risk, and 3 indicates high risk.

[0069] Finally, based on the ArcGIS geographic information system, the risk level of each grid was determined. Its geographical coordinates By combining different colors to map different risk levels, a dynamic risk heatmap with spatiotemporal labels is generated. This heatmap is stored in matrix form, with each element corresponding to the risk level of a grid, and is updated in real time as new monitoring data is added to the local database. When new water level, displacement, or moisture content data arrives, the risk levels of the affected grids are recalculated. The risk level is updated to ensure that the heat map always reflects the latest risk distribution, providing a dynamic basis for subsequent drone path planning.

[0070] S3: Based on the dynamic risk heat map, plan the adaptive cruise path of the UAV and control the UAV equipped with multi-source sensors to fly along the adaptive cruise path to simultaneously collect microwave dielectric data and hyperspectral image data of the bank slope.

[0071] The specific method in this step includes: assigning corresponding priority weights to grids with different risk levels. Among them, the low-risk first-level grid Medium-risk level 2 grid High-risk level 3 grid The monitoring frequency for high-risk grids is set three times that of low-risk grids. Priority weight. Defined as a grid In path planning, coverage is prioritized, with higher values ​​indicating higher priority. A multi-objective optimization function is constructed, whose objectives are to minimize the total flight path length, maximize the sum of priority weights for coverage grids, and minimize flight energy consumption. The optimization function expression is:

[0072]

[0073] in These are weighting coefficients, corresponding to the relative importance of path length, coverage priority, and flight energy consumption in the optimization objective, respectively. In the example, they are respectively taken as... , , ,satisfy . For the first The segment path length is defined as the Euclidean distance between two adjacent waypoints and is calculated from the Euclidean distance between two adjacent waypoints. This represents the total number of path segments; This represents the sum of priority weights for all grids covered by the path. This term is normalized and then optimized in the negative direction (i.e., maximizing the coverage weight is equivalent to minimizing its normalized missing values). The flight energy consumption of the UAV is defined as the energy consumed by the UAV to complete the entire flight path. It is related to factors such as path length, flight attitude, and wind speed. In this embodiment, it is simplified to a linear function that is directly proportional to the path length. ,in The energy consumption coefficient per unit distance is defined as the energy consumed per unit flight distance and is determined by the performance parameters of the UAV.

[0074] To solve this multi-objective optimization function, a genetic algorithm is used for path optimization. First, the monitoring area is discretized into a candidate set of waypoints, with each waypoint corresponding to the coordinates of a grid center. Chromosomes encode a waypoint sequence, representing the UAV's access order. The initial population randomly generates several paths covering all high-risk grids. The fitness function is defined as the optimization objective... The reciprocal of the fitness, i.e., the higher the fitness, the better. The smaller the value, the better. Iterative evolution is achieved through selection, crossover, and mutation operations until the maximum number of iterations or fitness convergence is reached. The final optimal chromosome is decoded as an adaptive cruise path, which prioritizes covering high-risk areas and achieves multiple coverages of high-risk grids (three times the frequency of low-risk areas) within the constraints of total flight distance and energy consumption. The genetic algorithm parameters are set as follows: population size 100, crossover probability 0.8, mutation probability 0.1, and number of iterations 200.

[0075] When the drone flies along the adaptive cruise path, high-risk grids are covered more frequently than low-risk grids. The drone used is a DJI M300 RTK or an equivalent industrial-grade model, equipped with an ultra-wideband microwave dielectric radiation sensor (center frequency 1GHz / 5GHz), a hyperspectral imaging spectrometer (spectral range 400-1000nm), an inertial measurement unit (IMU), and a real-time dynamic positioning module. The flight altitude is set to 50m, with a flight path overlap of 70%. A zigzag grid scanning mode is used to simultaneously collect microwave dielectric data and hyperspectral image data from the surface and deep layers of the shoreline. Precise spatiotemporal coordinate labels are added to all collected data through the IMU and the real-time dynamic positioning module, achieving spatiotemporal synchronization of multi-source data. The data is then correlated according to the spatial grid, and invalid data caused by drone jitter is removed to obtain valid data. During the data acquisition process, the real-time dynamic positioning module provides centimeter-level latitude, longitude, and elevation information, which is fused with the attitude data from the IMU to ensure that each data point can be accurately mapped to the corresponding grid (i,j), providing a spatial registration basis for the subsequent construction of the dielectric distribution matrix.

[0076] S4: Denoise the microwave dielectric data and, in conjunction with its spatial location information, construct a dielectric distribution matrix that characterizes the spatial distribution of dielectric properties within the bank slope.

[0077] The specific steps of this method include: processing the microwave dielectric data using a combined denoising method, which includes sequential notch filtering, wavelet thresholding denoising, and adaptive Wiener filtering. First, a notch filter is used to filter out 50Hz power frequency noise from the microwave dielectric data. The notch filter attenuates specific frequencies and their neighborhoods in the frequency domain, and its transfer function can be expressed as:

[0078]

[0079] in For frequency variables, it is defined as the frequency component of the signal; The center frequency of the notch filter is the power frequency noise frequency that needs to be filtered out. The notch bandwidth is defined as the width of the attenuation frequency band, and is usually taken as... ; The filter order determines the steepness of the filter attenuation curve; in this embodiment, we take... After notch filtering, power frequency interference in the data is effectively suppressed, and the signal-to-noise ratio is improved to over 28dB.

[0080] Then, wavelet thresholding denoising is used, selecting the db4 wavelet as the basis function to decompose the signal into 6 layers. Let the original signal be... The wavelet coefficients of each layer are obtained after wavelet decomposition. ,in Indicates the number of decomposition layers. For coefficient indexing. Wavelet coefficients Defined as the signal at the 1st In the layer decomposition Wavelet transform values ​​at each position. For high-frequency detail coefficients ( to Soft thresholding is used, and the soft thresholding function is defined as follows:

[0081]

[0082] Where the threshold Calculated using a general threshold formula:

[0083]

[0084] In the formula The noise standard deviation, defined as the statistical dispersion of the noise component in the signal, is estimated from the median of the first-level wavelet detail coefficients.

[0085]

[0086] The median() function takes the median value, where 0.6745 is the proportionality constant between the standard deviation of the Gaussian noise and the median. The length is denoted as . Soft thresholding can suppress Gaussian white noise while preserving the abrupt changes in the signal.

[0087] Finally, an adaptive Wiener filter is used to smooth the wavelet threshold-denoised data. The adaptive Wiener filter estimates based on local statistical properties, for data in pixels... The center's neighborhood window (window size set to) ), its output The calculation is as follows:

[0088]

[0089] in The local mean of the pixels within the window is defined as the arithmetic mean of all pixel values ​​within the window. The local variance of pixels within the window is defined as the mean of the squared deviations of each pixel value from the local mean within the window. The global noise variance is defined as the variance of noise across the entire image, estimated from the residual noise after wavelet thresholding. This filter is effective in flat regions (…). Smaller areas tend to be smoother, especially at the edges. If the value is relatively large, it remains unchanged, thus preserving signal details while further eliminating residual noise. After combined denoising, the denoised dielectric constant distribution data is obtained. ,in For spatial coordinates, The dielectric constant is defined as a characteristic parameter of a material storing electrical energy in an electric field, and its range is between 15 and 30.

[0090] The denoised microwave dielectric data is mapped onto a pre-divided spatial grid. For each grid... Its spatial extent is defined by latitude and longitude or planar coordinates, and it collects the set of all dielectric data points falling within this grid. ,in This represents the number of data points within the grid. Defined as a grid The number of dielectric data points contained within. Calculate the mean of all microwave dielectric data within each spatial grid:

[0091]

[0092] Using this mean as matrix elements, fill the matrix according to the row and column order of the grid to construct a matrix of size [size missing]. Dielectric distribution matrix :

[0093]

[0094] Dielectric distribution matrix Defined as a two-dimensional matrix composed of grid average dielectric constants as elements, it is used to characterize the spatial distribution of dielectric properties within the bank slope. In this embodiment... , ,matrix Each element represents the average dielectric constant of the corresponding grid, realizing a spatially structured expression of the dielectric properties inside the bank slope from discrete measurement points to a regular grid, providing a regular input data format for subsequent deep learning-based displacement region identification. This matrix also preserves spatial topological relationships, enabling convolution operations to effectively extract local dielectric anomaly features.

[0095] S5: Input the dielectric distribution matrix into the trained deep learning model to perform pixel-level anomaly detection and identify potential displacement regions; combine the hyperspectral image data to verify the authenticity of the potential displacement regions and determine the real displacement regions.

[0096] The specific methods in this step include: [determining the dielectric distribution matrix...] Convert to a four-dimensional tensor format suitable for improving DeepLabv3+ model input, i.e., add batch and channel dimensions, resulting in a shape of... The tensor, in which For the number of grid rows, Let be the number of grid columns. Min-max normalization is used to map the tensor elements to . The normalization formula for the interval is:

[0097]

[0098] in and These are the minimum and maximum values ​​of the entire matrix, respectively. Defined as the normalized dielectric distribution matrix. The normalized tensor is input into the improved DeepLabv3+ model, which replaces the original backbone network with MobileNetV3 to reduce computation and improve real-time detection capabilities. The pre-trained model can output the anomaly probability of each spatial grid belonging to a potential displacement region. Anomaly probability Defined as a grid The probability value of belonging to the potential displacement region, with a range of values. The pre-training process uses a historically collected dielectric distribution matrix sample set, which includes positive samples (confirmed displacement regions) and negative samples (stable regions). The loss function is cross-entropy, the optimizer is Adam, and the learning rate is set to 0.001. Training continues until the loss converges.

[0099] Set an anomaly probability threshold In the example, take , will satisfy The mesh is marked as the potential displacement region, and a binary mask matrix is ​​obtained. The element 1 represents a potential displacement mesh, and 0 represents a normal mesh. Anomaly probability threshold. Defined as the critical value for determining whether a mesh is a potential displacement region; Defined as a binary mask matrix for the potential displacement region. The grid coordinates of the marked potential displacement region are precisely aligned with the pixel coordinates of the hyperspectral image data. Since precise latitude, longitude, and elevation information has been added to each data point during UAV acquisition via a real-time dynamic positioning module and inertial measurement unit, the grid coordinates can be transformed. Pixel coordinate system mapped to hyperspectral image Nearest neighbor interpolation is used to determine the image pixel region corresponding to each grid.

[0100] Calculate the weighted structural similarity index W-SSIM between the current hyperspectral image and historical hyperspectral images in the potential displacement region. First, define the standard structural similarity index SSIM for two image patches. and SSIM is defined as:

[0101]

[0102] in and The mean of the image patch. and For variance, For covariance, Defined as an image block The average pixel value, Defined as an image block The average pixel value, Defined as an image block pixel variance Defined as an image block pixel variance Defined as an image block and The pixel covariance; , It is the stability constant. For pixel value dynamic range (for normalized images, take...) To highlight the importance of key areas on the bank slope (such as the water-facing side of the dike and the toe of the slope), a weighting factor is introduced. Each pixel is assigned a different weight, with a weight of 1.2 for key regions and 1.0 for other regions. The weights of all pixels within the window are then normalized. Weighting factors Defined as pixel The importance coefficient in structural similarity calculation. The weighted mean, variance, and covariance are then calculated as follows:

[0103]

[0104]

[0105]

[0106] Substituting the weighted statistics above into the SSIM formula yields the weighted structural similarity index W-SSIM. For each potential displacement region, a neighborhood window centered on it (the size of which can be set to...) is taken. (pixels), calculate the W-SSIM value of the current image and historical images within this window. Preset similarity threshold. , if W-SSIM < If the similarity threshold is met, it indicates a significant change in the image and is identified as a true displacement region; otherwise, it is considered a false detection and removed. Defined as the critical W-SSIM value for determining whether a significant change has occurred in the image. After verification, a binary mask of the actual displacement region is obtained. . Defined as a binary mask matrix for the actual displacement region.

[0107] right Morphological operations are performed on the identified actual displacement regions to optimize region boundaries and merge adjacent micro-regions. First, a closing operation (dilation followed by erosion) is used to fill the small voids within the regions. The structural element selected is... A rectangular kernel. The closing operation is defined as:

[0108]

[0109] in Indicates expansion. Indicates corrosion. It is a structural element. Defined as the mask matrix after the closing operation. Then, an expansion operation is performed, using... The rectangular kernel further expands the region boundary, connecting adjacent separated regions into a complete region. After morphological processing, the final true displacement region mask is obtained. . Defined as the final true displacement region mask matrix after morphological processing.

[0110] Statistically analyze the characteristic information of each real displacement region, including area, deformation, center coordinates, and boundary coordinates. Area By calculating the number of grids in the region multiplied by the area of ​​a single grid (grids in high-risk areas). Regular area ) is obtained. Deformation variables By comparing the positional offset of corresponding regions in the current hyperspectral image with that in historical images, the displacement vector of each pixel can be obtained using optical flow or feature matching methods, and then the average displacement within the region can be calculated. Area Defined as the surface area covered by the actual displacement region; deformation Defined as the average displacement magnitude of the actual displacement region. Center coordinates The geometric center of all grid coordinates in the region:

[0111]

[0112] Here Defined as the number of grid cells within the region. Boundary coordinates are obtained by extracting the mask contour, forming a polygonal point set. This feature information provides the basic data for subsequent displacement trend prediction.

[0113] S6: Input the multi-source time series data corresponding to the actual displacement area into the trained time series prediction model to predict the slope displacement trend parameters within a preset time period in the future.

[0114] In this step, a multidimensional time-series dataset is first constructed. For each identified real displacement region, its historical time-series data is extracted. The time window length is set to... (In this embodiment, we take) (days), taking the current time as the reference, taking the previous time. The input sequence consists of data from consecutive time points. Time window length. Defined as the length of historical data used for prediction. Each time point... eigenvectors Includes: the average dielectric constant of the region at that time. (Obtained from the values ​​of the corresponding grid in the dielectric distribution matrix), hyperspectral image deformation. (Displacement obtained through image registration or optical flow methods), and synchronously acquired hydrological data, including flow rate. Flow rate Water level and meteorological data such as rainfall and temperature In addition, regional soil parameters extracted from static foundation data (such as cohesion, internal friction angle, etc.) can be included. However, these parameters are considered constant in the short term and can be repeatedly added to each time point to enhance the model's perception of regional characteristics. Therefore, It is dimensional vector, This represents the total number of selected features. Feature vector. Defined as time A vector consisting of all the features used for prediction; Defined as the dimension of the feature. For each real displacement region, construct its own time-series dataset and organize it in chronological order. After merging the data from all regions, organize them according to... The training set, validation set, and test set are randomly divided into sets in proportion to ensure the model's generalization ability.

[0115] The multidimensional time-series data corresponding to the actual displacement region is input into an attention-enhanced bidirectional gated recurrent unit (GRU) network. This network consists of a forward GRU, a backward GRU, and an attention mechanism layer. The basic computation of the gated recurrent unit is introduced first. For each time step... GRU's hidden state Update using the following formula:

[0116]

[0117]

[0118]

[0119]

[0120] in and These are the update door and the reset door, respectively. Defined as the update gate vector, it controls the degree to which the hidden state from the previous time step is preserved; Defined as the reset gate vector, it controls the degree to which the hidden state of the previous time step is reset; For the sigmoid function, For element-wise multiplication, , , These are trainable parameters. The forward GRU is arranged in ascending time order. Process the sequence to obtain the forward hidden state sequence. Backward GRU in reverse chronological order Process to obtain the backward hidden state sequence . Defined as a forward GRU at time... The hidden state, Defined as a backward GRU at time... The hidden states. In this embodiment, the number of hidden layer neurons in both the forward and backward GRUs is set to 128, so each hidden state is a 128-dimensional vector.

[0121] To integrate bidirectional temporal information and enhance the role of key time steps, an attention mechanism is introduced. For each time step... By concatenating the forward and backward hidden states, we obtain The dimension is 256. Defined as time The concatenated bidirectional hidden state. Additive attention is used to calculate the weights at each time step. :

[0122]

[0123]

[0124] in , , These are trainable parameters. Defined as time Attention score Defined as time The attention weights are then calculated. Finally, the hidden states at all time steps are summed using weighted averages to obtain the comprehensive hidden state. :

[0125]

[0126] Overall Hidden Status Defined as a fusion vector of all hidden states at all time steps after attention weighting, it integrates bidirectional information from the entire sequence and highlights time points that contribute significantly to the prediction task through attention weights.

[0127] Comprehensive hidden state The input fully connected layer is mapped to a sequence of displacement distances over a predetermined future time interval. The number of prediction steps is set to... (In this embodiment) (corresponding to the next 30 days), the number of neurons in the fully connected layer is set to 64, and the number of neurons in the output layer is... The activation function is a linear function. Number of prediction steps. Defined as the number of future time points that need to be predicted. The mapping relationship is as follows:

[0128]

[0129] in This is the weight matrix. This is a bias term. Indicates the number of times from the current moment. The cumulative displacement distance at each time step, in meters. Defined as the future The cumulative displacement distance over each time step.

[0130] Calculate displacement rate based on displacement distance sequence That is, the displacement increment between adjacent time steps divided by the time step interval. (In this embodiment) sky):

[0131]

[0132] Where the definition Displacement rate Defined as the first To the The average displacement rate between time steps. The maximum value is extracted from the displacement distance sequence as the future maximum displacement distance. Future maximum displacement distance Defined as the maximum cumulative displacement distance within the prediction period.

[0133] During model training, mean squared error (MSE) is used as the loss function to measure the difference between the predicted displacement distance and the actual observed value.

[0134]

[0135] in The number of training samples, For the first The first sample Step-by-step predicted value The values ​​represent the corresponding true values. The Adam optimizer is used with a learning rate of 0.0005, 200 training epochs, and an early stopping strategy (training stops if the validation set loss does not decrease for 10 consecutive epochs). The model is considered trained successfully when the MSE on the test set is less than 0.01.

[0136] Furthermore, the step of predicting the slope displacement trend parameters within a preset time period also includes: calculating the radius of the displacement influence range based on a simplified formula of geotechnical elasticity theory, combined with the maximum displacement distance, the thickness of the main compressible layer in the soil physical and mechanical parameters, and the empirical coefficient of soil compressibility. This formula originates from the approximate relationship between surface displacement and the range of influence in the theory of elastic half-space, and its expression is:

[0137]

[0138] in The maximum future displacement distance (in meters) predicted by the model. The thickness of the main compressible layer of soil (in meters) is derived from the physical and mechanical parameters of the soil in the static foundation data. The soil compressibility empirical coefficient is dimensionless and defined as an empirical constant reflecting the influence of soil compressibility on the displacement range. Its value is determined according to the soil type: for silty clay, it is [value missing]. For clay extraction For sandy soil The specific value can be determined through indoor experiments or regional experience. This radius... This indicates the area that the displacement may affect, and is used to quantify the potential impact of the displacement in subsequent stability assessments.

[0139] S7: Based on the displacement trend parameters and the soil physical and mechanical parameters, construct a dynamic stability assessment model and calculate the stability index of the bank slope.

[0140] This step first clarifies the parameters involved in the stability assessment and their sources. Displacement rate. The average displacement rate for the next critical period (e.g., the next 30 days) predicted in step S6 is taken, and its calculation formula is as follows: ,in For the first The instantaneous displacement rate at each time step. Number of prediction steps. Average displacement rate. Defined as the arithmetic mean of displacement rates at each time step within the prediction period. Maximum displacement distance. The maximum value of the displacement distance sequence output from step S6 is taken directly, in meters. Displacement influence radius. According to the formula in step S6 The calculation is as follows, in meters. This is the empirical coefficient for soil compressibility. The thickness of the main compressible layer. Soil cohesion. (Unit: kPa), internal friction angle (Unit: °), Moisture content (Unit: %), Groundwater Level (Unit: m) All parameters are derived from the soil physical and mechanical parameters in the static foundation data collected in step S1, and are updated regularly with the latest field test results.

[0141] Constructing a stability index The calculation formula integrates displacement trend parameters and soil physical and mechanical parameters to quantify the overall stability of the slope. The formula expression is:

[0142]

[0143] in to The pre-defined weighting coefficients are defined as the relative importance weights of each parameter to the stability index, reflecting the relative importance of each parameter to slope stability. The weighting coefficients are determined using the Analytic Hierarchy Process (AHP), a multi-criteria decision-making method suitable for weight allocation problems combining qualitative and quantitative factors. The specific steps are as follows: First, establish a hierarchical structure, using the stability index as the target layer and the seven parameters as the criterion layer. Then, construct the judgment matrix. Its elements Indicates parameters Relative to parameters The importance of each factor is determined using a 1-9 scale (1 represents equal importance, and 9 represents extreme importance). The judgment matrix must meet consistency requirements, which are determined by calculating the consistency ratio. To test, among them , To determine the largest eigenvalue of a matrix, For the random consistency index (for , ).when At this point, the judgment matrix is ​​considered to have satisfactory consistency. Finally, the normalized eigenvector corresponding to the largest eigenvalue of the judgment matrix is ​​calculated, thus obtaining the weight coefficients of each parameter. In this embodiment, based on expert experience and engineering practice, the determined weight coefficients are as follows: , , , , , , And satisfy These coefficients reflect displacement trend parameters ( The soil moisture content and groundwater level have a greater impact on stability, while soil moisture content and groundwater level have a negative impact.

[0144] The stability index is calculated by substituting the displacement distance sequence, displacement rate, and maximum displacement distance, along with the soil physical and mechanical parameters, into the above formula. Stability Index Defined as a comprehensive index for quantifying the overall stability of a bank slope. A higher value indicates a more stable bank slope, and vice versa. Since bank slope stability is significantly affected by real-time hydrological conditions—for example, rising water levels reduce effective soil stress and exacerbate seepage damage—a hydrological dynamic correction factor is introduced. right After making corrections, the corrected stability index is obtained. :

[0145]

[0146] Hydrological dynamic correction factor Defined as a correction factor reflecting the impact of real-time hydrological conditions on bank slope stability, it is determined based on real-time hydrological data, primarily considering water level. and traffic The impact. In this embodiment, an empirical function based on water level is used to calculate. :

[0147]

[0148] in Reference water level (e.g., normal water level or historical average water level) is defined as a benchmark water level value used for comparison. The adjustment coefficient, defined as the sensitivity coefficient to the impact of water level changes on stability, is determined based on regional experience, for example... When the real-time water level is higher than the reference water level, A decrease indicates a reduction in stability; when the water level is lower than the reference water level... Increase, but not exceeding the upper limit of 1.0. In this embodiment, based on the real-time water level... and preset reference water level ,Pick ,but , compared with the example While the approximation is close, in practical applications, the function form can be determined through more precise regression analysis. Alternatively, a flow-based correction formula can also be used. Alternatively, a multivariate function of water level and flow rate can be considered to ensure that the correction factor dynamically reflects the real-time impact of changes in hydrological conditions on bank slope stability. The corrected stability index... This will be used directly for subsequent warning level determination.

[0149] S8: Generate and output the corresponding warning level based on the relationship between the stability index and the preset dynamic threshold.

[0150] In this step, the first warning threshold needs to be dynamically set. Second warning threshold These two thresholds are used to classify the three warning levels. The setting of these thresholds needs to comprehensively consider the engineering grade and revetment structure type of the target riverbank, because banks of different engineering grades and structural types have different tolerances for displacement and risk acceptance levels. Therefore, an engineering grade coefficient is introduced. and revetment structure type coefficient The base threshold is adjusted to a dynamic threshold applicable to the current bank slope.

[0151] Set a basic low-risk threshold and basic high-risk threshold These are reference values ​​predetermined based on a large amount of historical data and engineering experience, for example... , . Defined as a basic low-risk threshold, Defined as a basic high-risk threshold. Engineering grade coefficient. Reflecting the different safety requirements for different engineering levels: For Class I critical projects (such as flood control dikes in large cities), the requirements are higher, taking... Class II projects Projects of Grade III and below shall be subject to the following approval procedures: Revetment structure type coefficient Reflecting the deformation resistance of different structural forms: concrete structures are taken as... Masonry structure Earthen or dry-laid stone structures Engineering grade coefficient Defined as a coefficient for adjusting the early warning threshold based on the engineering level; revetment structure type coefficient. These are defined as coefficients used to adjust the early warning threshold based on the type of revetment structure. These coefficients can be adjusted according to local water conservancy engineering specifications or expert experience and are pre-stored in the system.

[0152] The dynamic threshold is calculated as follows:

[0153]

[0154]

[0155] Defined as a dynamically set first warning threshold. Defined as a dynamically set second early warning threshold. In this embodiment, the target river is a Class II engineering project, and the revetment is a masonry structure; therefore, the threshold is set to... , ,thereby , .

[0156] The corrected stability index calculated in step S7 The warning level is determined by comparing the results with the two thresholds mentioned above.

[0157]

[0158] Level 1 warning indicates that the bank slope is extremely unstable and poses a significant safety hazard of collapse and landslide, requiring immediate reinforcement measures; Level 2 warning indicates that the bank slope has a certain risk of displacement, with the displacement rate showing an upward trend, requiring increased monitoring frequency; Level 3 warning indicates that the bank slope is in a stable state, with displacement and deformation within the allowable range, and routine inspections can be maintained.

[0159] After determining the warning level, the system retrieves a pre-set reinforcement and maintenance plan database. This database is a relational database table, with each record containing fields such as warning level, plan name, specific measures description, required resources, and suggested response time. The corresponding reinforcement and maintenance plan is retrieved based on the warning level as the query key. For example, a Level 1 warning corresponds to a specialized reinforcement plan, which may include grouting reinforcement (injecting cement-water glass grout into the soil to fill pores), riprap protection (filling the slope toe with riprap to enhance erosion resistance), increased monitoring (increasing the monitoring frequency to once every 2 hours), and traffic control (prohibiting heavy vehicles from passing through), etc.; a Level 2 warning corresponds to an enhanced monitoring plan, such as increasing the monitoring frequency from once a day to three times a week and strengthening on-site inspections; a Level 3 warning corresponds to a routine inspection plan, maintaining daily monitoring and regular inspections.

[0160] Finally, the warning level and corresponding reinforcement and maintenance plan are output. Output methods include: real-time display of the warning level and key indicators on the monitoring center's large visualization screen; push of detailed reports via PC applications; and sending warning notifications and handling suggestions to designated maintenance personnel via mobile applications (such as WeChat and SMS). The notification content should include the bank slope location, warning level, stability index value, recommended reinforcement measures, and response timeframe. The entire warning process achieves a closed loop from data collection, analysis, and evaluation to decision support, providing a scientific basis for the safety management of riverbank slopes.

[0161] In another embodiment, this application also provides an optimized method for monitoring dynamic displacement of riverbank slopes. Based on the aforementioned embodiments, this method introduces edge intelligence, federated causal discovery, and a generative decision engine to construct a closed-loop intelligent monitoring system that is self-sensing, self-cognizing, and self-determining, thereby further improving the real-time performance, interpretability, and scientific nature of the monitoring and decision-making.

[0162] First, trusted edge computing nodes are deployed on the UAV platform, ground monitoring station, and data processing center. Each node integrates a lightweight neural network inference engine, a federated learning client, and a blockchain light node. While the UAV executes adaptive cruise path flight, it simultaneously collects microwave dielectric data and hyperspectral image data. The onboard edge nodes process the raw data in real time. Specifically, a lightweight semantic segmentation network, obtained through knowledge distillation, is used to process the dielectric data stream. This network, based on ESPNetv2 and trained using a distillation loss function, has one-fifth the number of parameters of the original DeepLabv3+ network, resulting in an inference speed improvement of more than five times. The network input is a denoised dielectric data slice, and the output is a depth feature vector for each grid cell. ,in This feature vector encodes the dielectric anomaly probability and local spatial structure information. Simultaneously, a lightweight optical flow network (such as a miniature version of FlowNetS) is used to extract deformation features from continuous hyperspectral images. The aforementioned feature vectors and real-time water levels Rainfall Soil physical and mechanical parameters (cohesion) internal friction angle Moisture content Main compression layer thickness The basic data, such as , are concatenated to form a joint feature vector. The local node calculates the data fingerprint hash value for each joint feature vector. The hash value is broadcast to the consortium blockchain network for notarization, and the original feature vector is uploaded to the central federated server after differential privacy perturbation to ensure data privacy and immutability.

[0163] The central server runs a global federated causal discovery model based on a hybrid architecture of variational autoencoders and structural causal models. Let the set of encrypted features collected from each edge node be... The model is composed of an encoder. decoder It consists of a causal structure learner, in which As a latent variable, in this embodiment we take Causal structure learners aim to discover the directed acyclic graph (DAG) structure between the dimensions of latent variables, and the causal relationships between latent and observed variables. An adjacency matrix is ​​defined. Indicates causal connections between latent variables. Indicates the first The nth latent variable is the th The reasons for the potential variables are as follows. The objective function is optimized using a differentiable DAG learning algorithm:

[0164]

[0165] The first term represents the negative value of the variational lower bound, i.e., the reconstruction loss, which encourages latent variables to retain information from the original data; the second term represents the sparsity constraint. As a weighting factor, this embodiment takes The constraints guarantee the acyclicity of the cause-effect graph. This represents element-wise multiplication. Let be the matrix exponent. Solving this optimization problem yields the global causal graph. The structural equation parameters (such as linear coefficients or neural network weights) for each node are also included. This causal graph reveals the causal chain between water level, rainfall, soil properties, and displacement evolution, and is transferable across different scenarios.

[0166] Based on this, a deep generative counterfactual reasoning engine is constructed. For the current moment... Collected joint feature vector The latent representation is obtained through the encoder. According to the cause-and-effect diagram With the corresponding structural equations, intervention operations can be performed on any target variable. For example, to assess the effect of water level changes on maximum displacement... If there is a causal effect, then intervention is necessary. That is, the water level variable is a specified value. By propagating forward along the causal graph, the values ​​of the latent and observed variables are recalculated to obtain the maximum displacement under the counterfactual condition. In this embodiment, the intervention is achieved by modifying the input variables in the structural equations, specifically by calculating the counterfactual displacement using the following formula:

[0167]

[0168] in These are the latent variables re-inferred through a causal graph after intervention. All counterfactual inference results are output in the form of probability distributions, such as "If the water level drops to the historical average of 3.0m, the expected decrease in displacement over the next 30 days is 0.03m, with a probability of 85%."

[0169] When based on the corrected stability index calculated in step S7 When the warning threshold is triggered, the system automatically activates the generative decision engine. This engine, based on a conditional variational autoencoder (CVAE), learns the joint distribution of state, intervention measures, and effects in historical successful reinforcement cases. The historical case library contains joint feature vectors from past incidents. The reinforcement measures taken (Represented as discrete codes, such as grouting as 1, riprap as 2, drainage as 3, etc.) and the effect indicators after the implementation of the measures. (e.g., reduction in displacement, improvement in stability index). The CVAE model consists of an encoder. and decoder The composition and training objective are to maximize the lower bound of the conditional log-likelihood. For the current state of danger... The engine selects from the candidate measure set Multiple intervention measures in the middle sampling For each measure, from the prior distribution Sampling is performed, and the decoder generates the distribution of the effects of the measure after its implementation. And calculate the expected effect indicators (such as the expected reduction in displacement). All candidate measures are based on their expected effects and estimated costs. and risk index A multi-objective ranking is performed, and the Pareto front is used to select the optimal solution. Specifically, the comprehensive utility function is defined as follows:

[0170]

[0171] in The weighting coefficient is determined based on engineering experience in this embodiment. Choose the measure with the greatest utility as the recommended option. .

[0172] Finally, the optimal solution The corresponding effect predictions are input into a large language model (based on the Transformer architecture) fine-tuned from a water conservancy engineering corpus. This model takes structured data (including hazard descriptions, recommended measures, expected effect curves, and resource requirement templates) as input and generates a natural language emergency report. The report includes the plan name, detailed implementation steps, a list of required personnel and equipment, expected displacement evolution curves (in text description and numerical table form), potential risk warnings, and a reference index. The generated report is automatically hashed and stored on the blockchain via a smart contract, and then pushed to operations and maintenance personnel through visual terminals and mobile applications.

[0173] Through the aforementioned optimizations, this method achieves a seamless closed loop from real-time perception, causal discovery, counterfactual inference, to intelligent decision-making. The combination of edge computing and federated causal models ensures data privacy and cross-scenario knowledge transfer capabilities, while the generative decision engine endows the system with proactive planning and self-evolution capabilities, significantly improving the intelligence level and engineering practical value of riverbank dynamic displacement monitoring.

[0174] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to the embodiments of this application, 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 all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0175] Corresponding to the above method, this application also provides a system for monitoring the dynamic displacement of riverbank slopes, which can operate in... Figure 1 The operating environment shown.

[0176] The system includes a data acquisition module, a dynamic risk heat map construction module, a UAV path planning module, a data processing module, a displacement area identification module, a displacement trend prediction module, and a stability assessment and early warning module. The modules are interconnected through data flow and control flow, forming an integrated monitoring closed loop, which is used to execute the method for dynamic displacement monitoring of riverbank slopes described in the first aspect and its embodiments.

[0177] The data acquisition module comprises an aerial acquisition unit and a ground acquisition unit. The aerial acquisition unit, mounted on an unmanned aerial vehicle (UAV) platform, includes an ultra-wideband microwave dielectric radiation sensor, a hyperspectral imaging spectrometer, an inertial measurement unit (IMU), and a real-time dynamic positioning module. The ultra-wideband microwave dielectric radiation sensor penetrates vegetation and shallow soil to collect dielectric property data at different depths within the bank slope. The hyperspectral imaging spectrometer acquires high-resolution spectral images of the bank slope surface, identifying material composition and surface deformation information. The IMU records the three-axis attitude angles and acceleration data during UAV flight, providing attitude information for spatiotemporal synchronization and geometric correction of multi-source data. The real-time dynamic positioning module adds centimeter-level precision spatial coordinate labels to all acquired data, ensuring that each data point can be accurately mapped to its corresponding geographical location. The ground acquisition unit includes portable in-situ testing equipment and an automated monitoring station to supplement the acquisition of data that is difficult for UAVs to obtain, including soil physical and mechanical parameters such as soil cohesion, internal friction angle, and moisture content, as well as environmental data such as groundwater level, rainfall, and temperature.

[0178] The dynamic risk heatmap construction module is deployed on the data processing server to receive multi-source basic data output by the data acquisition module, including static basic data and dynamic time-series data. This module performs a series of processing steps on the received data, including gridding, parameter interval labeling, time decay weighting and feature weighting allocation, and comprehensive risk weighted frequency statistics, to generate a dynamic risk heatmap characterizing the spatiotemporal distribution of slope risk. The heatmap is updated in real time based on new data, providing a dynamic basis for subsequent route planning.

[0179] The UAV path planning module is connected to the dynamic risk heatmap construction module, used to read the risk level information of each grid in the dynamic risk heatmap. This module assigns priority weights to different grids based on their risk levels, constructs a multi-objective optimization function with the goals of minimizing the total flight path length, maximizing the sum of coverage priority weights, and minimizing flight energy consumption, and obtains the adaptive cruise path through a genetic algorithm. After the path is generated, it is uploaded to the UAV flight control system via a communication link, controlling the UAV to fly autonomously according to the planned path, and prioritizing high-risk areas during flight to achieve high-frequency monitoring of high-risk grids.

[0180] The data processing module, deployed on an edge computing unit or data processing server, receives raw microwave dielectric data transmitted from the UAV. This module sequentially performs notch filtering, wavelet thresholding, and adaptive Wiener filtering on the microwave dielectric data to eliminate power frequency noise, Gaussian white noise, and residual noise, obtaining denoised dielectric constant data. The denoised data is then mapped to a pre-divided grid according to its spatial location. The mean dielectric constant within each grid is calculated, and the grid is filled in row and column order to construct a dielectric distribution matrix, achieving a spatially structured representation of the dielectric properties within the bank slope.

[0181] The displacement region identification module is connected to the data processing module and receives dielectric distribution matrix and hyperspectral image data. Internally, this module carries a trained improved DeepLabv3+ deep learning model. This model uses MobileNetV3 as its backbone network to perform pixel-level anomaly detection on the input dielectric distribution matrix, outputting the anomaly probability for each grid cell, and marking grid cells with anomaly probabilities greater than a preset threshold as potential displacement regions. Simultaneously, this module aligns the grid coordinates of potential displacement regions with the pixel coordinates of the hyperspectral image, calculates the weighted structural similarity index (W-SSIM) between the current image and historical images in the potential displacement regions, determines the true displacement region based on whether the W-SSIM is less than a preset similarity threshold, and performs morphological operations and feature information statistics on the confirmed regions.

[0182] The displacement trend prediction module is connected to the displacement region identification module to receive multi-source time-series data corresponding to the actual displacement region, including dielectric constant time-series data, hyperspectral image deformation time-series data, and synchronized hydrological and meteorological data. This module internally incorporates a trained attention-enhanced bidirectional gated recurrent unit network (GRU). Forward and backward GRUs extract the forward and backward features of the time-series data, respectively. After weighted fusion through an attention mechanism layer, a comprehensive hidden state is obtained. This hidden state is then mapped to a displacement distance sequence within a preset future time period through a fully connected layer. The displacement rate and maximum displacement distance are calculated based on this sequence. Furthermore, this module combines the thickness of the main compressible layer and the empirical coefficient of soil compressibility from the soil's physical and mechanical parameters, and uses a simplified formula based on geotechnical elasticity theory to calculate the radius of the displacement influence range.

[0183] The stability assessment and early warning module is connected to the displacement trend prediction module. It receives displacement trend parameters (including displacement rate, maximum displacement distance, and radius of displacement influence) and soil physical and mechanical parameters (including cohesion, internal friction angle, water content, and groundwater level) provided by the data acquisition module. Internally, this module constructs a dynamic stability assessment model. It weights and fuses the parameters according to preset weighting coefficients to calculate the stability index SI. The SI is then dynamically corrected based on real-time hydrological data to obtain the corrected stability index SI'. Next, a first and second early warning threshold are dynamically set according to the engineering grade and revetment structure type of the target riverbank. SI' is compared with the threshold to determine the early warning level. This module also has a built-in library of preset reinforcement and maintenance schemes. It can retrieve corresponding reinforcement and maintenance schemes based on the early warning level and displays the early warning level and scheme details in real time through a visual dashboard, PC application, and mobile application. Simultaneously, it pushes early warning information to maintenance personnel via SMS and WeChat.

[0184] The connections between the modules are as follows: The output of the data acquisition module connects to the dynamic risk heatmap construction module, the data processing module, and the stability assessment and early warning module, providing them with basic data support. The output of the dynamic risk heatmap construction module connects to the UAV path planning module, providing it with risk distribution information. The output of the UAV path planning module connects to the UAV flight control system, enabling the issuance of path commands. The output of the data processing module connects to the displacement region identification module, providing it with a dielectric distribution matrix. The output of the displacement region identification module connects to the displacement trend prediction module, providing it with characteristic information of the actual displacement region. The output of the displacement trend prediction module connects to the stability assessment and early warning module, providing it with displacement trend parameters. The output of the stability assessment and early warning module connects to the visualization terminal and the communication push interface, enabling the output of early warning information. The entire system forms a complete closed loop from data acquisition, risk prediction, path planning, data denoising, displacement identification, trend prediction to early warning output. All modules work collaboratively to achieve intelligent monitoring of riverbank slope dynamics.

[0185] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described herein can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0186] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the foregoing disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for monitoring dynamic displacement of riverbank slopes, characterized in that, Includes the following steps: Acquire multi-source basic data of the target riverbank slope, including static basic data and dynamic time-series data, wherein the static basic data includes soil physical and mechanical parameters; Based on the aforementioned multi-source basic data, a dynamic risk heat map characterizing the spatiotemporal distribution of bank slope risk is constructed; Based on the dynamic risk heat map, an adaptive cruise path for the UAV is planned, and the UAV equipped with multiple source sensors is controlled to fly along the adaptive cruise path to simultaneously collect microwave dielectric data and hyperspectral image data of the bank slope. The microwave dielectric data is denoised, and combined with its spatial location information, a dielectric distribution matrix characterizing the spatial distribution of dielectric properties inside the bank slope is constructed. The dielectric distribution matrix is ​​input into a trained deep learning model to perform pixel-level anomaly detection and identify potential displacement regions. By combining the hyperspectral image data, the authenticity of the potential displacement region is verified to determine the actual displacement region; The multi-source time series data corresponding to the actual displacement area is input into the trained time series prediction model to predict the slope displacement trend parameters within a preset time period in the future. Based on the displacement trend parameters and the physical and mechanical parameters of the soil, a dynamic stability assessment model is constructed to calculate the stability index of the bank slope. According to the relationship between the stability index and the preset dynamic threshold, the corresponding early warning level is generated and output.

2. The method for monitoring dynamic displacement of riverbank slopes according to claim 1, characterized in that, The steps for constructing the dynamic risk heatmap include: The target riverbank slope is spatially divided into multiple regular grids; The parameters in the multi-source basic data are divided into intervals, and different intervals of each parameter are assigned label values; A time decay weight is introduced for historical data at different time points, and a feature weight is introduced for different parameters. The time decay weight is a positive number less than 1, and the weight of recent data is greater than the weight of distant data. Based on the marker value, the time decay weight, and the feature weight, the comprehensive risk weighted frequency of each grid is calculated; Based on the comprehensive risk weighting frequency, risk levels are assigned to each grid, and the dynamic risk heat map is generated.

3. The method for monitoring dynamic displacement of riverbank slopes according to claim 2, characterized in that, The steps for planning the adaptive cruise path of the UAV include: Assign corresponding priority weights to grids with different risk levels; A multi-objective optimization function is constructed, wherein the optimization objectives of the multi-objective optimization function are to minimize the total length of the flight path, maximize the sum of the priority weights of the covering grid, and minimize the flight energy consumption; Solving the multi-objective optimization function yields the adaptive cruise path; Specifically, when controlling the drone to fly along the adaptive cruise path, the high-risk grid is covered more frequently than the low-risk grid.

4. The method for monitoring dynamic displacement of riverbank slopes according to claim 1, characterized in that, The step of constructing the dielectric distribution matrix includes: The microwave dielectric data is processed using a combined denoising method, which includes notch filtering, wavelet threshold denoising, and adaptive Wiener filtering performed sequentially. The denoised microwave dielectric data is mapped onto a pre-divided spatial grid. The mean value of all microwave dielectric data within each spatial grid is calculated, and the mean value is used as a matrix element to fill the grid according to the spatial arrangement order, thereby constructing the dielectric distribution matrix.

5. The method for monitoring dynamic displacement of riverbank slopes according to claim 1, characterized in that, The step of identifying potential displacement regions and verifying the authenticity of the potential displacement regions includes: The dielectric distribution matrix is ​​input into the improved DeepLabv3+ model, the anomaly probability of each spatial grid is output, and the grids with anomaly probabilities greater than a preset threshold are marked as potential displacement regions, wherein the improved DeepLabv3+ model replaces its backbone network with a MobileNetV3 network. Align the grid coordinates of the marked potential displacement regions with the pixel coordinates of the hyperspectral image data; Calculate the weighted structural similarity index W-SSIM between the current hyperspectral image and historical hyperspectral images in the potential displacement region, where pixels in key areas of the bank slope are assigned higher weights than those in other areas; Potential displacement regions whose W-SSIM index is less than a preset similarity threshold are identified as real displacement regions.

6. The method for monitoring dynamic displacement of riverbank slopes according to claim 1, characterized in that, The step of predicting the slope displacement trend parameters within a preset time period includes: The multidimensional time-series data corresponding to the real displacement region is input into the attention-enhanced bidirectional gated recurrent unit network. The attention-enhanced bidirectional gated recurrent unit network extracts the positive and negative features of the time-series data through the forward gated recurrent unit and the backward gated recurrent unit, respectively. The hidden state output by the bidirectional gated recurrent unit is weighted and fused through the attention mechanism layer to obtain the comprehensive hidden state. The integrated hidden state is mapped to a sequence of displacement distances within a preset future time period through a fully connected layer; The displacement rate is calculated based on the displacement distance sequence, and the maximum displacement distance is extracted from the displacement distance sequence.

7. The method for monitoring dynamic displacement of riverbank slopes according to claim 6, characterized in that, The step of predicting the slope displacement trend parameters within a preset time period further includes: Based on the simplified formula of the elastic theory of geotechnical engineering, and combined with the maximum displacement distance, the thickness of the main compressible layer in the physical and mechanical parameters of the soil, and the empirical coefficient of soil compressibility, the radius of the displacement influence range is calculated. The radius of the displacement influence range is proportional to the square root of the maximum displacement distance and the square root of the thickness of the main compressible layer, and the proportionality coefficient is the empirical coefficient of soil compressibility.

8. The method for monitoring dynamic displacement of riverbank slopes according to claim 1 or 7, characterized in that, The steps of constructing a dynamic stability assessment model and calculating the stability index of the bank slope include: The formula for calculating the stability index SI is as follows: The stability index SI is equal to the product of the first weighting coefficient and the reciprocal of the displacement rate, plus the product of the second weighting coefficient and the reciprocal of the maximum displacement distance, plus the product of the third weighting coefficient and the reciprocal of the radius of the displacement influence range, plus the product of the fourth weighting coefficient and the soil cohesion, plus the product of the fifth weighting coefficient and the soil internal friction angle, minus the product of the sixth weighting coefficient and the soil moisture content, minus the product of the seventh weighting coefficient and the groundwater level, wherein the first to seventh weighting coefficients are preset weighting coefficients; The stability index SI is calculated by substituting the displacement distance sequence, the displacement rate, the maximum displacement distance, and the soil physical and mechanical parameters into the formula. The stability index SI is corrected based on real-time acquired hydrological data to obtain the corrected stability index SI'.

9. The method for monitoring dynamic displacement of riverbank slopes according to claim 8, characterized in that, The step of generating and outputting the corresponding warning level includes: Based on the engineering grade and revetment structure type of the target riverbank slope, the first and second early warning thresholds are dynamically set. The corrected stability index SI' is compared with the first warning threshold and the second warning threshold; If the corrected stability index SI' is less than the first warning threshold, it is determined to be a Level 1 warning; If the corrected stability index SI' is greater than or equal to the first warning threshold and less than the second warning threshold, it is determined to be a level two warning. If the corrected stability index SI' is greater than or equal to the second warning threshold, it is determined to be a level three warning; Retrieve the preset reinforcement and maintenance scheme library, obtain the preset reinforcement and maintenance scheme associated with the warning level, and output the corresponding warning level and the preset reinforcement and maintenance scheme.

10. A system for monitoring dynamic displacement of riverbank slopes, used to execute the monitoring method according to any one of claims 1 to 9, characterized in that, include: The data acquisition module includes an ultra-wideband microwave dielectric radiation sensor, a hyperspectral imaging spectrometer, an inertial measurement unit, and a real-time dynamic positioning module mounted on an unmanned aerial vehicle platform, as well as a ground data acquisition terminal, for acquiring the multi-source basic data, the microwave dielectric data, and the hyperspectral image data; A dynamic risk heatmap construction module is used to construct and update the dynamic risk heatmap based on the multi-source basic data; The UAV path planning module is used to plan the adaptive cruise path based on the dynamic risk heat map and control the flight of the UAV. The data processing module is used to denoise the microwave dielectric data and construct the dielectric distribution matrix. The displacement region identification module is equipped with the trained deep learning model, which is used to identify and confirm the real displacement region. The displacement trend prediction module is equipped with the trained time-series prediction model for predicting the displacement trend parameters. The stability assessment and early warning module is used to construct the dynamic stability assessment model, calculate the stability index, generate and output the early warning level and the corresponding reinforcement and maintenance plan.

Images