A flood-type landslide disaster monitoring and early warning method

Abstract

The application discloses a flood-type landslide disaster monitoring and early warning method, and belongs to the technical field of geological disaster prediction. The method comprises the following steps: step 1, obtaining a flood-type landslide disaster case in which a rainstorm event and a landslide event are spatiotemporally overlapped, collecting landslide factor data, environmental data, monitoring data and historical landslide data, and constructing a historical database; step 2, constructing and training a Bayesian network model based on the historical database; step 3, obtaining landslide monitoring data and performing preprocessing; step 4, training an LSTM time sequence prediction model; step 5, inputting real-time monitoring data into the LSTM model to obtain future landslide monitoring data through prediction; and step 6, inputting the future monitoring data into the Bayesian network to obtain a slope body instability probability based on a future trend. Through the organic fusion of the Bayesian network and the LSTM, dynamic prediction and real-time response of the landslide instability probability are realized, and the timeliness, accuracy and robustness of early warning are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of natural disaster early warning technology, and in particular to a method for monitoring and early warning of flood-induced landslide disasters. Background Technology

[0002] Landslides are a common geological hazard in cities, especially flood-induced landslides triggered by extreme rainfall events, which pose a serious threat to people's lives, property, and infrastructure safety. Accurate and timely prediction and early warning of landslide instability probability are of great significance for disaster prevention and mitigation.

[0003] Currently, landslide prediction and early warning methods can be mainly classified into the following categories:

[0004] The first category is experience-based judgment methods based on geological surveys and engineering geology analogies. These methods analyze the geological conditions and historical landslide occurrences of the target area, and refer to the stability of nearby similar terrain features to provide a qualitative stability assessment or a semi-quantitative probability of instability. However, these methods heavily rely on expert experience, are highly subjective, lack objective quantitative standards, and struggle to reflect the dynamic evolution of landslides under external influences, thus failing to provide real-time dynamic early warning.

[0005] The second category is early warning methods based on a single monitoring indicator threshold. This type of method involves deploying sensors within the landslide body to monitor indicators such as deformation and rainfall. An early warning is triggered when the monitored value exceeds a preset threshold. While the method is simple and intuitive, a single indicator cannot fully reflect the complex instability mechanism of a landslide. It is prone to false alarms due to localized abnormal fluctuations, and often by the time the indicator exceeds the limit, the landslide has already entered an accelerated deformation phase, leaving very little time for disaster prevention and response, thus failing to achieve truly meaningful early warning.

[0006] The third category is prediction methods based on statistical models or machine learning. In recent years, researchers have attempted to establish statistical relationships or machine learning models between landslide occurrence and various influencing factors. However, existing methods have the following significant shortcomings:

[0007] Most models are static models, such as traditional Bayesian networks and logistic regression. Although they have strong real-time reasoning capabilities and interpretability, and can quickly calculate the probability of instability based on multi-source monitoring data at the current moment, they are essentially assessments based on the current state and do not have the ability to model temporal evolution. They cannot predict the trend of slope state changes in the future, resulting in delayed prediction results and making it difficult to meet the needs of early warning.

[0008] The fourth category is early warning methods based on multi-source data fusion. With the development of technologies such as satellite remote sensing, drones, and IoT sensors, it has become possible to obtain multi-source monitoring data on landslides. However, existing methods often simply overlay or process various types of data separately, lacking a unified data fusion framework and scientific quantitative modeling methods. They fail to fully utilize the temporal evolution patterns and causal relationships between variables inherent in historical landslide cases, resulting in limited prediction accuracy and reliability.

[0009] Therefore, there is an urgent need for a dynamic prediction method for landslide instability probability that can integrate geological background, historical cases and real-time monitoring data, take into account both immediate response and future trend prediction, and possess high accuracy, strong robustness and good interpretability. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide a flood-type landslide disaster monitoring and early warning method that can integrate geological background, historical cases and real-time monitoring data, take into account both immediate response and future trend prediction, and has high precision, strong robustness and good interpretability.

[0011] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0012] This invention provides a method for monitoring and early warning of flood-induced landslide disasters, comprising the following steps:

[0013] Historical data on rainstorm events and landslide events in the target area were obtained. Cases of flood-type landslide disasters where the rainstorm events and landslide events overlapped in time and space were selected. Landslide factor data, landslide environmental data, landslide monitoring data, and historical landslide data of each case were obtained and preprocessed to obtain a historical database of flood-type landslide disasters.

[0014] A Bayesian network model was trained using a historical database of flood-type landslide disasters to output the slope instability probability corresponding to the landslide monitoring data in the region. The Bayesian network model is divided into a "target node-parent node layer" with one target node and four types of parent nodes. It adopts a two-level tree structure of target node-parent node layer. The target node represents the landslide instability risk and directly represents the overall stability of the slope. The four types of parent nodes include landslide factor nodes, landslide environment nodes, landslide monitoring nodes, and historical landslide nodes.

[0015] Real-time landslide monitoring data is acquired, preprocessed, and then real-time landslide monitoring data is obtained.

[0016] A time-series prediction model was trained using landslide monitoring data from a historical database to predict future landslide monitoring data based on real-time landslide monitoring data.

[0017] The real-time landslide monitoring data is input into a trained time-series prediction model to predict future landslide monitoring data.

[0018] By inputting future landslide monitoring data into a trained Bayesian network model, the predicted probability of future slope instability is obtained.

[0019] The above setup achieves the following effect: Addressing the technical shortcomings of existing landslide prediction methods where static models lack temporal prediction capabilities and cannot provide early warnings, this invention organically links a Bayesian network model with a temporal prediction model, forming an innovative technical path of "LSTM predicting future data + Bayesian network real-time inference." While Bayesian networks possess powerful real-time inference capabilities and interpretability, enabling rapid calculation of the current instability probability based on multi-source monitoring data, they are essentially static models and cannot predict future trends. This invention utilizes the temporal evolution patterns inherent in historical landslide cases to transform the time series of real-time monitoring data into predicted values ​​for future monitoring data. Subsequently, the Bayesian network infers based on these predicted values, outputting the instability probability at future moments. This process upgrades the Bayesian network from "assessment based on the current state" to "prediction based on future trends," achieving a leap from "post-event interpretation" to "pre-event warning." It can identify rising risk trends days or even weeks before a landslide occurs, buying valuable time for evacuation and emergency response.

[0020] Meanwhile, this method takes into account both the time sensitivity of LSTM and the interpretability of Bayesian networks. When the probability of future instability increases, it can be backtracked to analyze which predictor variables caused this result, making it easier for engineers to take targeted measures.

[0021] Furthermore, this method makes full use of the temporal and correlation information in historical landslide cases, maximizing the value of historical data and adapting to the nonlinear characteristics of landslide evolution, significantly improving the accuracy, timeliness, and practicality of prediction.

[0022] Furthermore, the landslide factor data includes climate change data, population distribution data, vegetation cover data, and urban development data; the climate change data includes the return period of extreme precipitation; the population distribution data includes population density; and the urban development data includes the proportion of impervious surfaces.

[0023] The landslide environmental data includes slope, aspect, and elevation;

[0024] The landslide monitoring data includes rainfall intensity, pore water pressure, soil saturation, and displacement rate in the radar line-of-sight direction.

[0025] The historical landslide data includes historical instability frequency, historical instability intensity, and historical disaster scale.

[0026] The above settings achieve the following effects: By comprehensively collecting multi-dimensional data on landslide disasters, this invention constructs a complete indicator system covering triggering factors, environmental conditions, real-time dynamics, and historical evolution. This enables the Bayesian network model to comprehensively assess the risk of landslide instability from multiple perspectives, avoiding misjudgments caused by single indicators and improving the comprehensiveness and prediction accuracy of the model.

[0027] Furthermore, the method for obtaining the radar line-of-sight displacement rate includes:

[0028] Acquire SAR image data, perform image registration, select one image as the master image, and register the remaining images to the imaging space of the master image; SAR stands for Synthetic Aperture Radar.

[0029] Set spatial baseline thresholds and temporal baseline thresholds, and perform stacking interferometry on SAR images that meet the baseline threshold conditions to generate landslide differential interferometric phase maps;

[0030] Based on the differential interferometric phase map of the landslide, the surface deformation rate was obtained using Stacking-InSAR and SBAS-InSAR techniques respectively, and then fused using a weighted average method to obtain the displacement rate in the radar line-of-sight direction. Stacking-InSAR refers to stacked interferometric synthetic aperture radar measurement, and SBAS-InSAR refers to small baseline aggregating aperture radar interferometry.

[0031] The weighted average fusion formula is:

[0032] ;

[0033] in, The deformation rate after fusion. The average deformation rate obtained from Stacking-InSAR The deformation rate obtained from SBAS-InSAR, and the Stacking-InSAR weights. and SBAS-InSAR weights The deformation rate variances obtained from Stacking-InSAR and SBAS-InSAR are calculated using the following formulas:

[0034] ;

[0035] ;

[0036] in, For Stacking-InSAR weights, For SBAS-InSAR weights, The average deformation rate variance obtained from Stacking-InSAR The variance of the deformation rate obtained from SBAS-InSAR.

[0037] The above setup achieves the following effect: This invention integrates deformation rates obtained from Stacking-InSAR and SBAS-InSAR technologies and employs a variance-based weighted averaging method, fully leveraging the advantages of both technologies: Stacking-InSAR is suitable for obtaining average deformation rates over long time series, while SBAS-InSAR can better capture the temporal evolution details of deformation. The weighted fusion determines the weights based on the variances of each technology; the smaller the variance (higher the accuracy), the greater the technological contribution, resulting in more accurate and reliable deformation rate data, providing high-quality input for subsequent Bayesian networks and temporal prediction models.

[0038] Furthermore, the rainfall intensity is obtained in real time by rain gauges deployed in an open and unobstructed area at the top of the slope.

[0039] The pore water pressure is obtained in real time by pore water pressure gauges deployed near the potential sliding surface.

[0040] The soil saturation was obtained in real time by soil water potential sensors deployed at different locations along the upper, middle, and lower slopes.

[0041] Furthermore, the method for acquiring the landslide environmental data includes:

[0042] Drones were used to conduct aerial photography of the landslide area to obtain drone images. At the same time, the lidar equipment on the drone was used to scan the landslide area and collect three-dimensional point cloud data.

[0043] Geometric correction and image matching and stitching are performed on the collected UAV images to generate a complete image map of the landslide area, which is then processed and output as a digital orthophoto map using digital photogrammetry software.

[0044] The lidar point cloud data is filtered to remove noise points and non-ground points, resulting in a clean ground point cloud. A digital ground model is then generated based on this point cloud, and an interpolation process is used to generate a digital land surface model.

[0045] Slope, aspect, and elevation information of various features are extracted from digital orthophoto maps and digital surface models.

[0046] The above setup achieves the following effects: This invention integrates UAV aerial photography and LiDAR technology to achieve high-precision, all-around acquisition of the topography and landforms of landslide areas. UAV imagery provides rich texture information, while LiDAR point cloud data has strong vegetation penetration capabilities, enabling the acquisition of high-precision three-dimensional surface information. Through a series of refined processing steps, including geometric correction, filtering, and model generation, accurate key environmental factors such as slope, aspect, and elevation are ultimately extracted, providing refined spatial data support for the Bayesian network model.

[0047] Furthermore, the training method for the Bayesian network model includes:

[0048] The Bayesian network model is subjected to structural learning. A structure learning method based on scoring search is adopted, combined with domain knowledge constraints. Prior constraints are set for unidirectional connection of parent nodes and conditional independence of four types of parent nodes. The Bayesian information criterion scoring is used as the optimization objective. The hill climbing method is used to iteratively perform edge addition, deletion and reversal to determine the dependency relationship between nodes and construct a directed acyclic graph that conforms to the landslide disaster mechanism.

[0049] The formula for calculating the Bayesian information criterion score is as follows:

[0050] ;

[0051] in Let G represent the BIC score, G be the Bayesian network structure to be learned, and D be the multi-source fusion training dataset. This indicates that, given the model structure G and its maximum likelihood estimate... Under the given conditions, the likelihood probability of observing dataset D. This is the maximum likelihood estimate of the network parameters; n is the sample size of the dataset, and dim(G) is the parameter dimension of the network structure.

[0052] After completing the structure learning, parameter learning is performed. The data in the historical database is divided into training and test sets according to the proportion. The frequency of child nodes is counted by grouping the parent node value combination. The node probability under each combination is calculated by the conditional probability formula. Laplace smoothing is used to optimize the bias caused by insufficient sample size and construct a structured conditional probability table.

[0053] The conditional probability formula is:

[0054] ;

[0055] Where Y is a child node variable, y is the specific value of the child node variable, X1, X2...X k Let x1, x2...x be the set of variables of the 1st, 2nd...kth parent nodes. kThe values ​​of the variables for the 1st, 2nd...kth parent nodes are given, where k is the number of parent nodes; M represents the frequency of Y=y under the given combination of parent node values; and D represents the total number of samples under the given combination of parent node values.

[0056] The above setup achieves the following effects: This invention employs a combination of structure learning and parameter learning in the construction of Bayesian networks. In the structure learning phase, domain knowledge constraints (unidirectional parent node connections, four types of conditional parent node independence) are combined with data-driven BIC scoring. This ensures consistency between the network structure and the landslide disaster mechanism while avoiding the potential for illogical physical interpretations that might result from pure algorithmic learning. The hill-climbing method iteratively optimizes the network topology by adding, deleting, and reversing edges. In the parameter learning phase, conditional probability formulas are used to calculate the probability of each node, and Laplace smoothing is introduced to address the problem of insufficient sample size, improving the model's robustness. This rigorous modeling method ensures that the Bayesian network accurately expresses the causal relationships between variables and possesses good generalization ability.

[0057] Furthermore, in the Bayesian network model, the values ​​of all nodes are discretized into three states {0, 1, 2} based on the data distribution and the landslide disaster-causing pattern, corresponding to safe, critical, and dangerous states, respectively.

[0058] Establish a full-node design table for the landslide Bayesian network, including:

[0059] Target node: Landslide instability risk node;

[0060] The landslide factor node has the following sub-nodes: climate change, population distribution, vegetation coverage, and urban development;

[0061] The landslide environment node has sub-nodes: slope, aspect, and elevation;

[0062] The landslide monitoring node has the following sub-nodes: rainfall intensity, pore water pressure, soil saturation, and displacement rate in the radar line of sight direction;

[0063] The historical landslide nodes are further divided into sub-nodes: historical instability frequency, historical instability intensity, and historical disaster scale.

[0064] The output of the Bayesian network model includes qualitative output and quantitative output. The qualitative output is the state code of the target node {0,1,2}, and the quantitative output is the landslide instability probability P1, which is the probability that the slope is in a critical state. P1=1-P(R=0) is taken as the landslide instability probability, and P(R=0) represents the probability when the target node takes the value 0.

[0065] The above settings achieve the following effects: This invention discretizes continuous data into three states: safe, critical, and dangerous. This simplifies the model's complexity and aligns with the intuitive understanding of risk levels in engineering practice. By establishing a full-node design table, the quantitative thresholds for each node are clearly defined, making the model operable and reproducible. The qualitative output directly provides the risk level, facilitating rapid decision-making; the quantitative output P1 provides continuous probability values, providing a basis for refined risk assessment and early warning threshold setting. Taking P1 = 1 - P (R = 0) as the landslide instability probability, and defining the complement of the target node with a value of 0 (safe) as the instability probability, conforms to conventional risk expression practices.

[0066] Furthermore, the time-series prediction model employs a long short-term memory network to predict future landslide monitoring data based on the time series of real-time landslide monitoring data.

[0067] The architecture of the Long Short-Term Memory network model includes:

[0068] Input layer: The input is a sequence of monitoring data from the past T time steps;

[0069] Long Short-Term Memory Network Layer: Stack two layers of Long Short-Term Memory network. The first layer sets several hidden units and returns the complete sequence. The second layer sets several hidden units and only returns the output of the last time step. A dropout layer is added after each Long Short-Term Memory network layer to prevent overfitting.

[0070] Fully connected layer: Maps the output of the Long Short-Term Memory network layer to the predicted values ​​at H future time steps.

[0071] Furthermore, the method also includes:

[0072] The warning level is determined based on the predicted future slope instability probability P1:

[0073] If P1 < 0.3, then the risk of landslide instability is low, and the warning level is no warning.

[0074] If 0.3≤P1<0.6, then it is considered a medium-sized landslide instability risk, and the warning level is alert.

[0075] If 0.6 ≤ P1 < 1, then it is a high risk of landslide instability, and the warning level is an immediate action warning.

[0076] The above setup achieves the following effect: Based on the future slope instability probability P1, this invention establishes a clear standard for classifying early warning levels. Probability thresholds of 0.3 and 0.6 are used as the grading boundaries, corresponding to low, medium, and high risk levels and corresponding early warning actions (no warning, warning alert, and immediate action warning). This tiered early warning mechanism allows even non-professionals to quickly understand the risk level and take appropriate measures, improving the operability and practicality of the early warning system.

[0077] Furthermore, the method also includes:

[0078] The acquired real-time landslide monitoring data is directly input into the trained Bayesian network model to calculate the slope instability probability P2=1-P(R=0) based on the current real-time state.

[0079] Based on the slope instability probability P2=1-P(R=0) based on the current real-time state, the second warning level is determined;

[0080] The criteria for classifying the second warning level are as follows:

[0081] If P2 < 0.3, the risk of landslide instability is low, and the warning level is no warning.

[0082] If 0.3≤P2<0.6, then it is considered a medium-sized landslide instability risk, and the warning level is alert.

[0083] If 0.6 ≤ P2 < 1, then it is a high risk of landslide instability, and the warning level is an immediate action warning.

[0084] The above setup achieves the following effect: This invention adds a parallel channel for direct input of real-time monitoring data into the Bayesian network, constructing a two-layer early warning architecture of "instant response + trend prediction". This design has multiple beneficial effects:

[0085] First, it compensates for the response delay of time series forecasting models in sudden situations:

[0086] When landslides rapidly enter the accelerated deformation stage due to sudden factors such as extreme rainfall or earthquakes, the predictions made by LSTM based on previous trends may be distorted or delayed. However, the real-time channel can immediately reflect the current abnormal state and ensure that an early warning is triggered the instant a landslide is about to occur.

[0087] Secondly, a dual-channel redundant prediction mechanism was constructed. When sensor failure or abnormal data transmission causes a decrease in the quality of time-series prediction input, the real-time channel can still operate independently. Conversely, if a momentary jump in individual sensors causes an artificially high real-time probability, the look-ahead channel can be verified through trend smoothing, effectively reducing the risk of single model failure.

[0088] Third, it has achieved multi-timescale early warning. Real-time probability is used to issue immediate alerts to guide emergency evacuation, while future probability is used to deploy preventive measures such as patrols and reinforcement in advance, realizing the transformation from "post-event emergency response" to "pre-event early warning + in-event response".

[0089] Fourth, by comparing the dual-channel output results, the accuracy of the time series prediction model can be dynamically evaluated. If the long-term trends of the two are consistent, the model is reliable. If there is a significant deviation, the model will be retrained or the parameters will be adjusted to achieve online adaptive optimization.

[0090] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0091] (1) The flood-type landslide early warning of the present invention integrates satellite remote sensing, UAV, Internet of Things sensors and landslide disaster data, which can comprehensively and accurately warn of flood-type landslides, and improve the accuracy and reliability of the early warning.

[0092] (2) Extending static reasoning to dynamic prediction. The LSTM model utilizes the temporal evolution patterns inherent in historical landslide cases to transform the time series of real-time monitoring data into predicted values ​​for future monitoring data. Subsequently, the Bayesian network infers based on these predicted values ​​and outputs the probability of instability at future moments. This process upgrades the Bayesian network from "assessment based on the current state" to "prediction based on future trends," achieving a leap from "post-event explanation" to "pre-event warning."

[0093] (3) This invention solves the problem of “insufficient precision” in traditional prevention and control through scientific quantitative methods, and solves the problem of “poor implementation” through simplified adaptation and visualization. Ultimately, it realizes the transformation of flood-type landslide disasters “from passive response to active prevention and control, from fragmented response to whole-chain governance”, and provides an efficient and sustainable technical solution for urban flood landslide prevention and control.

[0094] (4) This invention constructs a dual-channel redundant prediction mechanism to improve system robustness: This method sets up two parallel probability output channels: one is the "instant probability" based on real-time data, and the other is the "look-ahead probability" based on future prediction data. The two channels are independent of each other and complementary to each other: when sensor failure or abnormal data transmission causes a decrease in the quality of time-series prediction input, the real-time channel can still operate independently; conversely, if a momentary jump in individual sensors causes an artificially high real-time probability, the look-ahead channel can be verified through trend smoothing. The dual-channel mechanism effectively reduces the risk of single model failure and improves the reliability and fault tolerance of the early warning system. Attached Figure Description

[0095] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0096] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0097] Example 1:

[0098] This embodiment provides a method for monitoring and early warning of flood-induced landslide disasters, such as... Figure 1 As shown, it includes the following steps:

[0099] Step 1: Obtain historical data of rainstorm events and landslide events in the target area, filter out flood-type landslide disaster cases where rainstorm events and landslide events overlap in time and space, and obtain landslide factor data, landslide environmental data, landslide monitoring data, and historical landslide data for each case, and perform preprocessing to obtain a historical database of flood-type landslide disasters;

[0100] Step 1.1: Obtain nearly 100 years of data on Chinese cities from the National Meteorological Science Data Center. This data includes annual precipitation, maximum daily precipitation, and number of days with heavy rainfall per year. Extract heavy rainfall events as triggering factors for floods and landslides. Obtain the distribution of landslide points from the "National Geological Hazard Dataset" of the National Earth System Science Data Center. Select cases that overlap with heavy rainfall events in time and space from the landslide point data as flood-type landslide disaster cases.

[0101] Step 1.2: Obtain landslide factor data, landslide environmental data, landslide monitoring data, and historical landslide data for each case, and perform preprocessing to obtain a historical database of flood-type landslide disasters:

[0102] (1) The landslide monitoring data includes rainfall intensity, pore water pressure, soil saturation, and displacement rate in the radar line-of-sight direction;

[0103] Rainfall intensity: The rain gauge is placed on an open, unobstructed area at the top of the slope in the target area to obtain the rainfall intensity. The rainfall intensity is monitored and recorded in real time through the rain gauge.

[0104] Pore ​​water pressure: A pore water pressure gauge is placed near the potential sliding surface to obtain the pore water pressure. The pore water pressure is then monitored and recorded in real time using the pore water pressure gauge.

[0105] Soil saturation: Soil water potential sensors are deployed at different locations on the upper, middle and lower slopes of the slope to obtain soil saturation data. The soil saturation is then monitored and recorded in real time using the soil water potential sensors.

[0106] Radar line-of-sight displacement rate:

[0107] InSAR technology was used to collect multiple Sentinel-1A satellite SAR image data of the monitored area.

[0108] During the image registration stage, one image is selected as the master image. A feature-matching image registration method is used to register the remaining images to the master image's imaging space, ensuring that the offset of the same ground feature pixel in both the azimuth and range directions is controlled within 0.25 pixels. After registration, the image area is cropped using an image cropping tool based on the boundary vector file of the monitoring area. Subsequent preprocessing steps include: orbit correction based on the orbit parameter file, radiometric correction using radiometric calibration coefficients, and Goldstein filtering to eliminate image noise.

[0109] A spatial baseline threshold of 100 meters and a temporal baseline threshold of 50 days were set. Using InSAR data processing software, stacking interferometry was performed on SAR images that met the baseline threshold conditions to generate landslide differential interferometric phase maps.

[0110] Weighted averaging calculations are performed on the differential interferometric phase map to obtain the average surface deformation rate of the monitored area, and the variance of the average deformation rate is also calculated. Based on the orbital information of the satellite imaging and the geometric parameters of the reference ellipsoid, a lookup table is generated in the data processing software. The coordinates of each pixel are then calculated and transformed using the lookup table to obtain the surface deformation data in the geographic coordinate system.

[0111] Average deformation rate obtained from Stacking-InSAR The average deformation rate variance obtained from Stacking-InSAR It can be calculated using the following formula,

[0112]

[0113] in, is the average deformation rate obtained from Stacking-InSAR; n is the number of interferograms involved in the weighted average calculation; and These are the time baseline and corresponding phase value of the i-th interferogram, respectively;

[0114]

[0115] in, Here, represents the variance of the average deformation rate obtained from Stacking-InSAR, and n is the number of interferograms involved in the weighted average calculation. Rate estimates for each interferogram;

[0116] When using SBAS-InSAR technology, a connection map is first constructed using data processing software to select suitable image combinations for interferometric processing.

[0117] Next, interferometric operations are performed on the selected image combinations to generate interferograms;

[0118] Subsequently, through track optimization and re-leveling, interference components such as track deviation and terrain phase are eliminated;

[0119] Then, the inversion method is used to solve for the surface deformation rate in the time series, the time series variation curve is plotted, and the variance of the average deformation rate is calculated.

[0120] Finally, following the aforementioned orbit optimization and interference removal processing logic, the results are converted to a geographic coordinate system to obtain surface deformation data in that coordinate system.

[0121] Deformation rate obtained by SBAS-InSAR The calculation formula is:

[0122]

[0123] in, The deformation rate is obtained from SBAS-InSAR. For the total deformation, For the deformation at this moment, For the initial shape variables, For the total time interval, At this moment, This is the initial time.

[0124] Deformation rate variance obtained from SBAS-InSAR The calculation method refers to the calculation formula given by the average deformation rate variance obtained from Stacking-InSAR;

[0125] The deformation rates obtained from Stacking-InSAR and SBAS-InSAR were fused using a weighted average method. Based on the variance of the average deformation rates obtained from both methods, a weighted calculation was performed according to the formula to obtain the fused deformation rate. ;

[0126] The formula for calculating the weighted average is:

[0127]

[0128] in, The deformation rate after fusion, Stacking-InSAR weights and SBAS-InSAR weights The deformation rate variances obtained from Stacking-InSAR and SBAS-InSAR are calculated using the following formulas:

[0129]

[0130]

[0131] in, For Stacking-InSAR weights, For SBAS-InSAR weights, The average deformation rate variance obtained from Stacking-InSAR The variance of the deformation rate obtained from SBAS-InSAR.

[0132] (2) The landslide environmental data includes slope, aspect, and elevation;

[0133] Unmanned aerial vehicles (UAVs) were used to conduct aerial photography of the landslide area. During flight, the entire landslide area needed to be covered, ensuring sufficient forward and lateral overlap. The flight altitude needed to be dynamically adjusted based on the size of the landslide area and the complexity of the terrain to acquire high-resolution image data. Simultaneously, the landslide area was scanned using LiDAR equipment mounted on the UAVs. During the scanning process, key parameters such as flight altitude, scanning angle, and point cloud density needed to be set appropriately to ensure that the acquired 3D point cloud data accurately reflects the regional terrain features.

[0134] Subsequent geometric corrections were performed on the acquired UAV images to eliminate image distortion caused by camera lens distortion and flight attitude fluctuations. Adjacent images were stitched together using an image matching algorithm to generate a complete image map of the landslide area. Digital photogrammetry software was used to process the images and output a digital orthophoto map (DOM). For LiDAR point cloud data, filtering was first performed to remove noise points and non-ground points, resulting in a clean ground point cloud. A digital terrain model (DEM) was then generated based on this point cloud, and a digital surface model (DSM) was generated through interpolation.

[0135] By combining detailed features of landslide surface fracturing (such as ground fissures, mounds, troughs, faults, and landslide walls), automatic identification is performed on digital orthophoto maps (DOM) and digital surface models (DSM) using either manual identification or image recognition algorithms. Ground fissures are identified by analyzing image texture features and color differences; mounds and faults are identified based on topographic elevation changes; and landslide walls are identified through image edge feature analysis. After identification, data extraction techniques are used to obtain key information such as slope, aspect, and elevation for each feature.

[0136] (3) The landslide factor data includes climate change data, population distribution data, vegetation coverage, and urban development data; the climate change data includes the return period of extreme precipitation; the population distribution data includes population density; and the urban development data includes the proportion of impervious surfaces.

[0137] The historical landslide data includes historical instability frequency, historical instability intensity, and historical disaster scale. Historical instability intensity includes the historical maximum displacement rate and landslide volume; historical disaster scale includes the economic loss from a single incident.

[0138] Specific methods for obtaining it include:

[0139] Daily precipitation data and annual average temperature data of target cities in China are obtained from the National Meteorological Science Data Center to calculate indicators such as the number of annual rainstorms (daily precipitation ≥ 50 mm), the maximum annual 24-hour precipitation, and the frequency of extreme precipitation.

[0140] Data sets of landslide disaster points in target cities across the country over the past century were obtained from the National Earth System Science Data Center. For each city, the annual maximum daily precipitation sequence was extracted for subsequent extreme value analysis.

[0141] Based on the meteorological records, geological records, and urban statistical data of the target city, obtain climate change data, population distribution data, vegetation coverage, urban development data, historical instability frequency, historical instability intensity, and historical disaster scale;

[0142] Z-score standardization was applied to all continuous independent variables to eliminate the influence of dimensions.

[0143] Calculate the variance inflation factor (VIF), remove variables with a VIF greater than 10, and solve the multicollinearity problem.

[0144] Plot time series graphs of landslide frequency and intensity for each city (or region) on an annual basis to visually observe growth trends, periodicity, and abrupt change points. Use the Mann-Kendall trend test (a non-parametric method) to determine the statistical significance of the trend.

[0145] The year of mutation in a sequence can be identified using methods such as the Pettt test.

[0146] For the annual maximum daily precipitation of each city, a generalized extreme value distribution is used for fitting.

[0147] Return Period Calculation: Extreme precipitation amounts with different return periods (e.g., 50-year return period, 100-year return period) are calculated. The temporal evolution trend of extreme precipitation intensity is analyzed. Cases of spatiotemporal overlap between landslides and torrential flood events are screened. Finally, a database of "city-year-landslide frequency-landslide intensity" is established. With time (year) as the sole independent variable, linear regression models for landslide frequency and intensity are constructed separately. "Annual landslide frequency (Y1)" and "Annual average landslide intensity (Y2)" are used as dependent variables, and "time (T)" is used as the independent variable (year code: e.g., 1925=1, 1926=2, ..., 2024=100). Two univariate linear regression models are constructed:

[0148] Frequency trend model:

[0149] ;

[0150] Intensity trend model:

[0151] ;

[0152] in , For the intercept term, , The trend slope (core indicator). , This is the random error term.

[0153] Use Python to solve for the parameters using the least squares method and output the regression results.

[0154] If b > 0: the frequency / intensity of flood-type landslides shows a significant upward trend over time;

[0155] If b < 0: the frequency / intensity of flood-type landslides shows a significant decreasing trend;

[0156] The return period T of extreme precipitation can be calculated using the cumulative distribution function F(x) of extreme precipitation frequency, as follows:

[0157] ;

[0158] F(x) is the probability of the cumulative distribution function, and T is the return period of extreme precipitation.

[0159] Step 1.3: Integrate four types of data: landslide factors, landslide environment, landslide monitoring, and historical landslides. Unify all data to the same coordinate system, spatial resolution, and temporal resolution.

[0160] Step 2: Based on the historical database of flood-type landslide disasters, construct a Bayesian network flood-type landslide model to quantify the probability of landslide instability. The Bayesian network model is constructed by dividing the model into "target node-parent node layers," with one target node and four types of parent nodes. Node values ​​are based on data distribution. The model adopts a two-level tree structure of target node-parent node layers. All nodes are discretized into values ​​{0, 1, 2} according to data distribution and landslide disaster patterns, corresponding to safe, critical, and dangerous states, respectively.

[0161] The core target node is the landslide instability risk (R), which directly represents the overall stability of the slope. The four types of parent nodes cover the core dimensions of landslide disaster, including landslide factor nodes, landslide environment nodes, landslide monitoring nodes, and historical landslide nodes. Each parent node has sub-nodes.

[0162] The node values ​​of landslide factor nodes, landslide environment nodes, landslide monitoring nodes, and historical landslide nodes correspond to the landslide factor data, landslide environment data, landslide monitoring data, and historical landslide data in the historical database of flood-type landslide disasters.

[0163] The landslide factor node has sub-nodes: climate change, population distribution, vegetation coverage, and urban development. The node values ​​correspond to climate change data, population distribution data, vegetation coverage data, and urban development data, respectively.

[0164] In the Bayesian network model, the node weights for climate change, population distribution, vegetation cover, and urban development are determined using the following methods:

[0165] Calculating climate change using weighted normalization Population distribution Vegetation coverage Urban development The degree of influence of each factor on the landslide was calculated using data analysis software. - Rank correlation coefficient with "annual landslide frequency (Y1)" and "annual average landslide intensity (Y2)" The weights are calculated using the following formula:

[0166] ;

[0167] in, For the i-th factor x i The weight, For factor x i The rank correlation coefficient is given by k, where k is the total number of rank correlation coefficients.

[0168] The landslide environment node has sub-nodes: slope, aspect, and elevation;

[0169] The landslide monitoring node has the following sub-nodes: rainfall intensity, pore water pressure, soil saturation, and displacement rate in the radar line of sight direction;

[0170] The historical landslide nodes are further divided into sub-nodes: historical instability frequency, historical instability intensity, and historical disaster scale.

[0171] The training input for the Bayesian network model is spatiotemporally coordinated historical data of flood-type landslide disasters, including four types of quantitative data: landslide factors, environment, real-time monitoring, and historical landslides. Continuous data is standardized and converted into {0,1,2} discrete codes to form a dataset that the model can identify. A full-node design table for the landslide Bayesian network is established, as shown in Table 1 and Table 1-continued.

[0172] Table 1. Full Node Design Table for Landslide Bayesian Network

[0173]

[0174] Table 1 - Continued Landslide Bayesian Network Full Node Design Table - Continued

[0175]

[0176] The output of a Bayesian network is divided into two categories: qualitative and quantitative. The qualitative output is the state code of each node {0,1,2}, and the quantitative output is the landslide instability probability P1=1-P(R=0).

[0177] Node dependencies are determined through network structure learning. A score-based search-based structure learning method combined with domain knowledge constraints is used to determine the dependencies (parent-child node connections) between nodes in the Bayesian network. This avoids the irrationality of pure algorithm learning and ensures that the network structure is consistent with the landslide disaster mechanism. Specifically, a combination of BIC scoring and hill climbing is adopted. First, prior constraints such as unidirectional parent node connections and four types of parent node conditional independence are set. Then, the BIC score is used as the optimization objective. The BIC score is a classic scoring function for Bayesian network structure learning. It comprehensively considers the network's fit and complexity to avoid overfitting. The formula is:

[0178] ;

[0179] in Let G represent the BIC score, G be the Bayesian network structure to be learned, and D be the training dataset from the historical database. This indicates that, given the model structure G and its maximum likelihood estimate... Under the given conditions, the likelihood probability of observing dataset D; This is the maximum likelihood estimation of the network parameters; n is the sample size of the dataset, and dim(G) is the parameter dimension of the network structure (the number of parameters in the conditional probability table). The goal is to find the network structure that maximizes BIC(G,D). This is the optimal structure.

[0180] Hill climbing is a local greedy search algorithm that combines BIC scoring to iteratively optimize the network structure. By iteratively adding, deleting, and reversing edges through hill climbing, the algorithm selects the optimal directed acyclic graph that conforms to the landslide disaster mechanism and clarifies the dependencies between nodes.

[0181] Based on this, the flood-type landslide disaster data in the historical database were divided into training and testing sets in a 7:3 ratio. The frequency of child nodes was statistically analyzed by grouping them according to the combination of parent node values. The node probability under each combination was calculated using the conditional probability formula. Laplace smoothing was used to optimize the bias caused by insufficient sample size. Finally, a structured conditional probability table (CPT) was constructed as the core parameter for model inference.

[0182] For each child node, group them according to "all combinations of values ​​of the parent node", and count the frequency of each value of the child node (W=0 / 1 / 2). For each combination of parent nodes, calculate the conditional probability P of each value of the child node according to the following formula:

[0183] ;

[0184] Where Y is a child node variable, y is the specific value of the child node variable, X1, X2...X k Let x1, x2...x be the set of variables of the 1st, 2nd...kth parent nodes. k The values ​​of the variables for the 1st, 2nd...kth parent nodes are given, where k is the number of parent nodes; M represents the frequency of Y=y under the given combination of parent node values; and D represents the total number of samples under the given combination of parent node values.

[0185] The frequency of child nodes is statistically analyzed by grouping the values ​​of parent nodes. The probability is calculated using the conditional probability formula. The probability results of all parent node combinations are compiled into a structured conditional probability table (CPT). P1=1-P(R=0) is taken as the probability of landslide instability. P(R=0) represents the probability when the target node (landslide instability risk) takes a value of 0 (safe).

[0186] In a well-trained Bayesian network, the landslide factor nodes, landslide environment nodes, and historical landslide nodes remain largely unchanged. Only the values ​​of the landslide monitoring nodes need to be updated to obtain the corresponding landslide instability probability.

[0187] Step 3: Acquire landslide monitoring data in real time, perform preprocessing, and obtain real-time landslide monitoring data;

[0188] Rainfall intensity: Rain gauges are placed on the top of the slope in an open, unobstructed area of ​​the target region, and the rainfall intensity is monitored and recorded in real time.

[0189] Pore ​​water pressure: A pore water pressure gauge is placed near the potential sliding surface to monitor and record the pore water pressure in real time.

[0190] Soil saturation: Soil water potential sensors are deployed along different slope positions (upper, middle, and lower) to monitor and record soil saturation in real time.

[0191] Radar line-of-sight displacement rate:

[0192] Stacking-InSAR and SBAS-InSAR technologies were employed, and the deformation rates obtained from both technologies were fused using a weighted averaging method. Based on the variance of the average deformation rates obtained from both technologies, a weighted calculation was performed according to a formula to obtain the fused deformation rate as the displacement rate in the radar line-of-sight direction.

[0193] Step 4: Train a time series prediction model using landslide monitoring data from the historical database to predict future landslide monitoring data based on real-time landslide monitoring data;

[0194] (1) The raw monitoring data obtained from field sensors (such as InSAR, pore water pressure gauge, rain gauge, etc.) usually contains noise, missing values ​​or outliers, and needs to be preprocessed to form a high-quality time series.

[0195] Data cleaning: Use interpolation methods (such as linear interpolation and spline interpolation) to fill in missing data caused by sensor failure or transmission problems; use the 3σ principle or the Local Outlier Factor (LOF) algorithm to identify and remove outliers, or replace them with reasonable estimates from adjacent time points.

[0196] Time alignment: Since different sensors may have different sampling frequencies (e.g., InSAR data may be sampled every half hour, while rain gauge data may be sampled every 15 minutes), all data needs to be resampled to a uniform time resolution (e.g., every 15 minutes). Alignment can be achieved using mean aggregation or interpolation methods.

[0197] Normalization: To avoid the influence of different units on model training, each monitored variable is normalized. Common methods include Min-Max scaling to the [0,1] interval or Z-score standardization.

[0198] After preprocessing, the time series of landslide monitoring data is obtained, and the historical database of landslide monitoring data is compiled.

[0199] (2) Selection and architecture design of time series forecasting models:

[0200] Landslide monitoring data is characterized by nonlinearity, non-stationarity, and long-term time-series dependencies. This invention selects a Long Short-Term Memory (LSTM) network as the core architecture of the time-series prediction model. LSTM effectively captures long-term dependencies in the time series through a gating mechanism, making it particularly suitable for predicting landslide deformation evolution trends. Gated Recurrent Units (GRUs) or Transformers can also be selected depending on the specific circumstances; this embodiment uses LSTM as an example for detailed explanation.

[0201] LSTM model architecture:

[0202] Input layer: The input is a sequence of monitoring data from the past T time steps, where T is the length of the historical window (e.g., 60 days) and m is the number of variables (including original variables and derived features).

[0203] LSTM Layers: Two LSTM layers are stacked to enhance the model's expressive power. The first LSTM layer has 128 hidden units and returns the complete sequence as input to the second layer; the second LSTM layer has 64 hidden units and returns only the output of the last time step. To prevent overfitting, a Dropout layer is added after each LSTM layer with a dropout rate of 0.2.

[0204] Fully connected layer: The output of the LSTM layer passes through a fully connected layer and is mapped to the predicted values ​​for the next H time steps.

[0205] The activation function uses linear activation because the predicted values ​​are continuous.

[0206] Model hyperparameters:

[0207] Historical window length T: determined by grid search or experience, usually 30 to 90 days, and 60 days in this embodiment.

[0208] Prediction step size H: Set according to the early warning requirements, such as the next 7 days, 15 days or 30 days. In this embodiment, 7 days (one week) is used.

[0209] Batch size: 64.

[0210] Learning rate: Initial value 0.001, adaptively adjusted using the Adam optimizer.

[0211] Loss function: Mean Squared Error (MSE).

[0212] This invention addresses the nonlinear, non-stationary, and long-term time-series dependencies inherent in landslide monitoring data by employing LSTM as the time-series prediction model. LSTM effectively captures long-term dependencies in the time series through a gating mechanism, enabling it to learn typical patterns of landslide evolution. The model adopts a two-layer stacked architecture: the first layer extracts time-series features and transmits the complete sequence, while the second layer further abstracts and focuses on features at the final time step, enhancing the model's expressive power. The addition of a dropout layer effectively prevents overfitting, and the fully connected layer maps the extracted features to predicted values ​​for the next H time steps. This architectural design allows LSTM to accurately predict future trends in monitoring data, providing input for the Bayesian network at future time steps.

[0213] Model training process:

[0214] Training data preparation:

[0215] Extract landslide monitoring data sequences prior to instability from the historical database constructed in step 2. Each case includes landslide monitoring data for L consecutive days prior to instability, ensuring sufficient length for sample generation.

[0216] Training samples are generated using a sliding window approach: For each case, starting from day 1, data from T consecutive days is taken as input, and data from the subsequent H consecutive days is taken as output labels. The sliding step size is set to 1 day to fully expand the sample size.

[0217] The samples generated from all cases are merged, randomly shuffled, and then divided into a training set (80%), a validation set (10%), and a test set (10%). The validation set is used for parameter tuning and early stopping, while the test set is used for final evaluation.

[0218] Loss function:

[0219] The mean squared error (MSE) is used as the loss function to calculate the difference between the predicted and actual values:

[0220] Training strategy:

[0221] Gradient descent is performed using the Adam optimizer with an initial learning rate of 0.001. A learning rate decay strategy is set: if the validation set loss no longer decreases after every 10 training iterations, the learning rate is multiplied by 0.5.

[0222] The maximum number of training iterations is set to 200, and an early stopping mechanism is adopted: monitor the validation set loss, and if it does not decrease after 20 consecutive iterations, stop training and restore to the optimal model weights.

[0223] To prevent gradient explosion, gradient clipping is applied with a threshold set to 1.0.

[0224] During training, the loss curves of the training and validation sets can be recorded to observe overfitting.

[0225] Step 5: Input the real-time landslide monitoring data into the trained time series prediction model to predict future landslide monitoring data;

[0226] Raw monitoring data obtained from field sensors (InSAR, pore water pressure gauges, rain gauges, etc.) usually contains noise, missing values, or outliers, and needs to be preprocessed to form high-quality time series.

[0227] Data cleaning: Use interpolation methods (such as linear interpolation and spline interpolation) to fill in missing data caused by sensor failure or transmission problems; use the 3σ principle or local outlier algorithm to identify and remove outliers, or replace them with reasonable estimates from adjacent time points.

[0228] Time alignment: Since different sensors may have different sampling frequencies (e.g., InSAR data may be sampled every half hour, while rain gauge data may be sampled every hour), all data needs to be resampled to a uniform time resolution (e.g., every half hour). Alignment can be achieved using mean aggregation or interpolation methods.

[0229] Normalization: To avoid the influence of different units on model training, each monitored variable is normalized. Common methods include Min-Max scaling to the [0,1] interval or Z-score standardization.

[0230] After preprocessing, a time series of real-time landslide monitoring data is obtained.

[0231] By inputting the time series of real-time landslide monitoring data into the time series prediction model constructed above, future landslide monitoring data can be obtained.

[0232] After obtaining the future landslide monitoring data, it is necessary to perform an inverse transformation based on the normalization parameters saved during normalization in step 3 to restore the predicted values ​​to their original physical dimensions before inputting them into the Bayesian network model.

[0233] Step 6: Input the future landslide monitoring data into the Bayesian network model trained in Step 2 to obtain the predicted probability of future slope instability, and determine the warning level based on the probability of future slope instability.

[0234] If P1 < 0.3, then the risk of landslide instability is low, and the warning level is no warning.

[0235] If 0.3≤P1<0.6, then it is considered a medium-sized landslide instability risk, and the warning level is alert.

[0236] If 0.6≤P1≤1, then it is a high risk of landslide instability, and the warning level is an immediate action warning.

[0237] This invention innovatively integrates a Bayesian network model with a Long Short-Term Memory (LSTM) time-series prediction model, overcoming the limitations of a single model in predicting landslide instability probability. This achieves complementary advantages and significantly improves the accuracy, timeliness, and practicality of the prediction.

[0238] 1. Advantages and limitations of Bayesian network models:

[0239] Bayesian networks, as a probabilistic graphical model, have unique advantages in predicting the probability of landslide instability:

[0240] Advantage 1: Powerful real-time inference capabilities. Based on conditional probability tables and Bayes' theorem, Bayesian networks can quickly calculate the corresponding posterior instability probability from multi-source monitoring data (such as deformation rate, rainfall, soil saturation, pore water pressure, etc.) input at the current moment. This inference process has low computational cost and rapid response, making it very suitable for real-time monitoring scenarios. When new monitoring data arrives, the instability probability can be updated immediately, enabling dynamic early warning.

[0241] Advantage 2: High interpretability. The graph structure of Bayesian networks intuitively expresses the causal relationship between each monitored variable and the probability of instability (for example, rainfall leads to an increase in soil saturation, which in turn increases pore water pressure, ultimately affecting slope stability). Its reasoning process is transparent and traceable, making it easy for engineers to understand and verify.

[0242] Advantage 3: Capable of handling uncertainty. Bayesian networks express the relationships between variables in probabilistic form, which can naturally integrate heterogeneous data from multiple sources and has a certain degree of robustness to data noise and missing values.

[0243] However, Bayesian networks also have inherent limitations in landslide prediction:

[0244] Bayesian networks have insufficient long-term predictive capabilities. Essentially, a Bayesian network is a static model; its inference is based on current observation data and lacks the ability to model temporal evolution. When it's necessary to predict the probability of slope instability in the future (e.g., 7 or 15 days), a Bayesian network cannot provide a direct answer because it lacks the foresight to anticipate changes in future monitoring data. Simply repeating the current monitoring data ignores the nonlinear trends and abrupt changes in the landslide evolution process, leading to delayed predictions and preventing early warning.

[0245] To compensate for the limitations of Bayesian networks in long-term prediction capabilities, this invention introduces Long Short-Term Memory (LSTM) networks as a time-series prediction model. LSTM offers the following core advantages:

[0246] Advantage 1: It excels at capturing long-term dependencies in time series. Through gating mechanisms (forget gate, input gate, output gate), LSTM can effectively learn and remember the evolutionary patterns in historical monitoring data, such as the trend of deformation rate accelerating over time, and the delayed response relationship between rainfall and soil saturation.

[0247] Advantage 2: It can predict future trends. After being trained based on historical landslide case data, the LSTM model can predict monitoring data at multiple future time steps (such as deformation rate, rainfall, soil saturation, etc. in the next 7 days) based on the time series of current real-time monitoring data, thus extending the prediction perspective from "present" to "future".

[0248] This invention organically combines LSTM and Bayesian networks, forming an innovative technical path of "LSTM predicting future data + Bayesian network real-time inference," resulting in the following significant synergistic effects:

[0249] First, static inference is extended to dynamic prediction. The LSTM model utilizes the temporal evolution patterns inherent in historical landslide cases to transform the time series of real-time monitoring data into predicted values ​​for future monitoring data. Subsequently, the Bayesian network infers based on these predicted values, outputting the probability of instability at future moments. This process upgrades the Bayesian network from "assessment based on the current state" to "prediction based on future trends," achieving a leap from "post-event explanation" to "pre-event warning."

[0250] Secondly, it enables true early warning. Traditional landslide early warning systems often rely on threshold triggers (such as deformation rates exceeding a certain limit), but by this time, the landslide may have already entered an accelerated deformation phase, leaving very little time for disaster prevention and response. This invention uses LSTM to predict monitoring data for the next few days, and then combines this with a Bayesian network to calculate the probability of future instability. This allows for the identification of rising risk trends days or even weeks before a landslide occurs, buying valuable time for evacuation and emergency response.

[0251] Third, it balances accuracy and interpretability. While LSTM models have strong predictive power, their internal mechanisms are complex, making it difficult to directly explain "why a high probability of instability is predicted"; Bayesian networks, on the other hand, provide a transparent reasoning path. The combination of the two allows the prediction results to have both the time-series sensitivity of LSTM and the interpretability of Bayesian networks—when the probability of future instability increases, it is possible to backtrack and analyze which predictor variables (such as accelerated deformation rate, sudden increase in pore water pressure) led to this result, facilitating engineers to take targeted measures.

[0252] Fourth, fully utilize historical case information. The training of the LSTM model relies on the historical landslide case database constructed in step 2, which contains the complete evolutionary process before landslide instability. Through training, the LSTM learns the typical evolutionary pattern of landslides from stability to instability; while the Bayesian network learns the static correlation between each monitoring variable and the probability of instability. The combination of the two maximizes the use of empirical knowledge from historical cases, both for understanding variable relationships and for predicting future trends.

[0253] Finally, it adapts to the nonlinear characteristics of landslide evolution. Landslide instability is often not a linear process, but rather goes through stages such as initial deformation, uniform deformation, accelerated deformation, and finally instability. LSTM models can capture this nonlinear evolution pattern and predict that the deformation rate may suddenly accelerate at some point in the future; Bayesian networks can then accurately calculate the jump in the instability probability based on this accelerated deformation rate and other predictive variables. The combination of these two approaches allows the prediction results to truly reflect the nonlinear risks of landslide evolution.

[0254] In summary, this invention achieves the following comprehensive advantages by combining the LSTM time-series prediction model with the Bayesian network model:

[0255] Improved forecast timeliness: Upgraded from "current status assessment" to "future trend prediction" to achieve early warning.

[0256] Enhanced prediction accuracy: LSTM captures the evolutionary patterns of time series, while Bayesian networks integrate multi-factor probabilistic reasoning, and the two complement each other.

[0257] Model interpretability is preserved: The causal structure of Bayesian networks makes the prediction results transparent and reliable, which is convenient for engineering applications.

[0258] Maximizing the value of historical data: Fully utilize the temporal and relational information in historical landslide cases, resulting in a high degree of knowledge reuse.

[0259] Dynamic update capability: With the continuous input of real-time monitoring data, LSTM can continuously update future predictions, and the Bayesian network can synchronously update the instability probability, realizing dynamic adaptive early warning.

[0260] This technological approach overcomes the inherent limitations of single models and provides a scientific, reliable, and practical solution for urban landslide disaster early warning, with significant social and economic benefits.

[0261] The flood-type landslide early warning system of this invention integrates technologies such as satellite remote sensing, drones, and IoT sensors, as well as landslide disaster data, enabling comprehensive and accurate early warning of flood-type landslides, thus improving the accuracy and reliability of the early warning.

[0262] This invention achieves seamless integration of "risk zoning → trend prediction → real-time monitoring → early warning push → material allocation", solving the core pain points of "data fragmentation and link breakage" in flood-type landslide prevention and control.

[0263] This invention solves the problem of "insufficient precision" in traditional prevention and control through scientific quantitative methods, and addresses the pain point of "poor implementation" through simplified adaptation and visualization. Ultimately, it realizes the transformation of flood-induced landslide disasters "from passive response to proactive prevention and control, and from fragmented response to whole-chain governance", providing an efficient and sustainable technical solution for urban flood and landslide prevention and control.

[0264] Example 2:

[0265] This embodiment provides a method for monitoring and early warning of flood-induced landslide disasters, which differs from Embodiment 1 in that:

[0266] To prevent distortion or response delays in time-series prediction models, which could hinder timely responses to sudden landslide disasters, this method also sets:

[0267] The real-time landslide monitoring data obtained in step 3 is directly input into the Bayesian network model trained in step 2 to calculate the slope instability probability P2=1-P(R=0) based on the current real-time state.

[0268] The second warning level is determined based on the slope instability probability P2 based on the current real-time state.

[0269] The criteria for classifying the second warning level are as follows:

[0270] If P2 < 0.3, the risk of landslide instability is low, and the warning level is no warning.

[0271] If 0.3≤P2<0.6, then it is considered a medium-sized landslide instability risk, and the warning level is alert.

[0272] If 0.6≤P2≤1, then it is a high risk of landslide instability, and the warning level is an immediate action warning.

[0273] The main purpose of this step is:

[0274] 1. To compensate for the response delay of time series forecasting models in sudden events:

[0275] Time-series prediction models (such as LSTM) rely on historical data patterns and future trend extrapolation, and their predictions are effective when landslides are progressing smoothly. However, in the event of sudden events such as extreme rainfall or earthquakes, landslides may rapidly enter an accelerated deformation phase. In such cases, the predictions based on previous trends by time-series prediction models may become distorted or outdated. Directly inputting real-time monitoring data into a Bayesian network can immediately reflect the current abnormal state, ensuring that early warnings are triggered the instant a landslide is about to occur, and avoiding missed detections due to model prediction delays.

[0276] 2. Construct a dual-channel redundant prediction mechanism to improve system robustness:

[0277] This method employs two parallel probability output channels: an "instant probability" based on real-time data and a "look-ahead probability" based on future prediction data. These two channels are independent yet complementary: when sensor malfunctions or abnormal data transmission degrade the quality of the time-series prediction input, the real-time channel can still operate independently; conversely, if a momentary jump in individual sensor readings causes an artificially high real-time probability, the look-ahead channel can verify this through trend smoothing. This dual-channel mechanism effectively reduces the risk of single-model failure and improves the reliability and fault tolerance of the early warning system.

[0278] 3. Enable multi-timescale early warning to meet diverse emergency response needs:

[0279] Real-time instability probability reflects the current dangerous state of the slope and can be used to issue immediate warnings (such as red alerts) to guide personnel to take emergency evacuation. Instability probability based on future forecast data provides risk trends for the next few days to weeks and can be used to proactively deploy preventative measures such as inspections and reinforcement. The combination of these two methods ensures both immediate response to sudden disasters and provides a scientific basis for medium- and long-term disaster prevention decisions, achieving a shift from "post-disaster emergency response" to "pre-disaster warning + in-disaster response."

[0280] 4. Enhance the intuitiveness and operability of early warning systems:

[0281] The calculated real-time probability can be directly correlated with preset thresholds (e.g., low risk <0.3, medium risk 0.3-0.6, high risk >0.6), generating clear warning levels that are easy for non-professionals to understand and implement. Simultaneously, this real-time probability can be combined with the inference path of a Bayesian network to trace key monitoring indicators that lead to high risk (e.g., a sudden increase in pore water pressure), guiding on-site personnel to take targeted measures.

[0282] 5. Verify the reliability of the time series prediction model:

[0283] By comparing real-time probabilities with probabilities derived from future predictions, the accuracy of time-series forecasting models can be dynamically evaluated. If the long-term trends of the two are consistent, the model's predictions are reliable; if there is a significant deviation, it can trigger model retraining or parameter adjustment, achieving online adaptive optimization of the model.

[0284] In summary, this method constructs a two-layer early warning architecture of "instant response + trend prediction" by adding a parallel path for direct input of real-time monitoring data into the Bayesian network. This significantly improves the timeliness, accuracy, and robustness of landslide disaster early warning, providing more comprehensive technical support for urban disaster prevention and mitigation.

[0285] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0286] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0287] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0288] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0289] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for monitoring and early warning of flood-induced landslide disasters, characterized in that, include: Historical data on rainstorm events and landslide events in the target area were obtained. Cases of flood-type landslide disasters where the rainstorm events and landslide events overlapped in time and space were selected. Landslide factor data, landslide environmental data, landslide monitoring data, and historical landslide data of each case were obtained and preprocessed to obtain a historical database of flood-type landslide disasters. A Bayesian network model was trained using a historical database of flood-type landslide disasters to output the slope instability probability corresponding to the landslide monitoring data in the region. The Bayesian network model is divided into "target node-parent node layers", with one target node and four types of parent nodes. It adopts a two-level tree structure of target node-parent node layers. The target node represents the landslide instability risk and directly represents the overall stability of the slope. The four types of parent nodes include landslide factor nodes, landslide environment nodes, landslide monitoring nodes, and historical landslide nodes. Real-time landslide monitoring data is acquired, preprocessed, and then real-time landslide monitoring data is obtained. A time-series prediction model was trained using landslide monitoring data from a historical database to predict future landslide monitoring data based on real-time landslide monitoring data. The real-time landslide monitoring data is input into a trained time-series prediction model to predict future landslide monitoring data. By inputting future landslide monitoring data into a trained Bayesian network model, the predicted probability of future slope instability can be obtained. The landslide factor data includes climate change data, population distribution data, vegetation cover data, and urban development data; the climate change data includes the return period of extreme precipitation; the population distribution data includes population density; and the urban development data includes the proportion of impervious surfaces. The landslide environmental data includes slope, aspect, and elevation; The landslide monitoring data includes rainfall intensity, pore water pressure, soil saturation, and displacement rate in the radar line-of-sight direction. The historical landslide data includes historical instability frequency, historical instability intensity, and historical disaster scale. The training methods for the Bayesian network model include: The Bayesian network model is subjected to structural learning. A structure learning method based on scoring search is adopted, combined with domain knowledge constraints. Prior constraints are set for unidirectional connection of parent nodes and conditional independence of four types of parent nodes. The Bayesian information criterion scoring is used as the optimization objective. The hill climbing method is used to iteratively perform edge addition, deletion and reversal to determine the dependency relationship between nodes and construct a directed acyclic graph that conforms to the landslide disaster mechanism. The formula for calculating the Bayesian information criterion score is as follows: ； in Let G represent the BIC score, G be the Bayesian network structure to be learned, and D be the multi-source fusion training dataset. This indicates that, given the model structure G and its optimal parameter estimates... Under the given conditions, the likelihood probability of observing dataset D; This is the maximum likelihood estimate of the network parameters; n is the sample size of the dataset, and dim(G) is the parameter dimension of the network structure. After completing the structure learning, parameter learning is performed. The data in the historical database is divided into training and test sets according to the proportion. The frequency of child nodes is counted by grouping the parent node value combination. The node probability under each combination is calculated by the conditional probability formula. Laplace smoothing is used to optimize the bias caused by insufficient sample size and construct a structured conditional probability table. The conditional probability formula is: ； Where Y is the parent node variable, y is the specific value of the parent node variable, and X1, X2...X k Let x1, x2...x be the set of variables for the 1st, 2nd...kth child nodes. k This refers to the specific values ​​of the variables for the 1st, 2nd...kth child nodes, where k is the number of child nodes; In the Bayesian network model, the values ​​of all nodes are discretized into three states {0, 1, 2} based on the data distribution and the landslide disaster-causing pattern, corresponding to safe, critical, and dangerous states, respectively. Establish a full-node design table for the landslide Bayesian network, including: Target node: Landslide instability risk node; The landslide factor node has the following sub-nodes: climate change, population distribution, vegetation coverage, and urban development; The landslide environment node has sub-nodes: slope, aspect, and elevation; The landslide monitoring node has the following sub-nodes: rainfall intensity, pore water pressure, soil saturation, and displacement rate in the radar line of sight direction; The historical landslide nodes are further divided into sub-nodes: historical instability frequency, historical instability intensity, and historical disaster scale. The output of the Bayesian network model includes qualitative output and quantitative output. The qualitative output is the state code of the target node {0, 1, 2}, and the quantitative output is the landslide instability probability P1. P1 = 1 - P(R = 0) is taken as the landslide instability probability, and P(R = 0) represents the probability when the target node takes the value 0. In a trained Bayesian network, the landslide factor nodes, landslide environment nodes, and historical landslide nodes remain unchanged. Only the values ​​of the landslide monitoring nodes need to be updated to obtain the corresponding landslide instability probability.

2. The method for monitoring and early warning of flood-induced landslide disasters according to claim 1, characterized in that, The method for obtaining the displacement rate in the radar line-of-sight direction includes: Acquire SAR image data, perform image registration, select one image as the master image, and register the remaining images to the imaging space of the master image. Set spatial baseline thresholds and temporal baseline thresholds, and perform stacking interferometry on SAR images that meet the baseline threshold conditions to generate landslide differential interferometric phase maps; Based on the differential interferometric phase map of the landslide, the surface deformation rate was obtained by Stacking-InSAR and SBAS-InSAR technologies respectively, and then fused by a weighted average method to obtain the displacement rate in the radar line-of-sight direction. The weighted average fusion formula is: ； in, The deformation rate after fusion. The average deformation rate obtained from Stacking-InSAR The deformation rate obtained from SBAS-InSAR, and the Stacking-InSAR weights. and SBAS-InSAR weights The deformation rate variances obtained from Stacking-InSAR and SBAS-InSAR are calculated using the following formulas: ； ； in, The average deformation rate variance obtained from Stacking-InSAR The variance of the deformation rate obtained from SBAS-InSAR.

3. The method for monitoring and early warning of flood-induced landslide disasters according to claim 1, characterized in that, The rainfall intensity is obtained in real time by rain gauges deployed in an open, unobstructed area at the top of the slope. The pore water pressure is obtained in real time by pore water pressure gauges deployed near the potential sliding surface. The soil saturation was obtained in real time by soil water potential sensors deployed at different locations along the upper, middle, and lower slopes.

4. The method for monitoring and early warning of flood-induced landslide disasters according to claim 1, characterized in that, The method for acquiring the landslide environmental data includes: Drones were used to conduct aerial photography of the landslide area to obtain drone images. At the same time, the lidar equipment on the drone was used to scan the landslide area and collect three-dimensional point cloud data. Geometric correction and image matching and stitching are performed on the collected UAV images to generate a complete image map of the landslide area, which is then processed and output as a digital orthophoto map using digital photogrammetry software. The lidar point cloud data is filtered to remove noise points and non-ground points, resulting in a clean ground point cloud. A digital ground model is then generated based on this point cloud, and an interpolation process is used to generate a digital land surface model. Slope, aspect, and elevation information of various features are extracted from digital orthophoto maps and digital surface models.

5. The method for monitoring and early warning of flood-induced landslide disasters according to claim 1, characterized in that, The time-series prediction model employs a long short-term memory network to predict future landslide monitoring data based on the time series of real-time landslide monitoring data. The architecture of the Long Short-Term Memory (LSTM) network model includes: Input layer: The input is a sequence of monitoring data from the past T time steps; Long Short-Term Memory Network Layer: Stack two layers of Long Short-Term Memory network. The first layer sets some hidden units and returns the complete sequence. The second layer sets some hidden units and only returns the output of the last time step. Add a dropout layer after each Long Short-Term Memory network layer to prevent overfitting. Fully connected layer: Maps the output of the Long Short-Term Memory network layer to the predicted values ​​at the next H time steps.

6. The method for monitoring and early warning of flood-induced landslide disasters according to claim 1, characterized in that, The method further includes: The warning level is determined based on the predicted probability of future slope instability. If P1 < 0.3, then the risk of landslide instability is low, and the warning level is no warning. If 0.3≤P1<0.6, then it is a medium-sized landslide instability risk, and the warning level is alert. If 0.6 ≤ P1 < 1, then it is a high risk of landslide instability, and the warning level is an immediate action warning.

7. The method for monitoring and early warning of flood-induced landslide disasters according to claim 6, characterized in that, The method further includes: The acquired real-time landslide monitoring data is directly input into the trained Bayesian network model to calculate the slope instability probability P2 based on the current real-time state. The second warning level is determined based on the slope instability probability P2 based on the current real-time state. The criteria for classifying the second warning level are as follows: If P2 < 0.3, the risk of landslide instability is low, and the warning level is no warning. If 0.3≤P2<0.6, then it is considered a medium-sized landslide instability risk, and the warning level is alert. If 0.6 ≤ P2 < 1, then it is a high risk of landslide instability, and the warning level is an immediate action warning.

Images