An intelligent mountain landslide remote monitoring method and system

Abstract

This invention relates to the field of geological disaster monitoring and discloses an intelligent remote monitoring method and system for landslides. The method first constructs a basic dataset by acquiring unique IDs and physical coordinates of marker points in the target monitoring area. Based on a unified time reference, mountain images are acquired, and the coordinates of light spots are extracted through differential processing and converted into a point cloud dataset. Next, displacement analysis is performed to identify anomalous marker points, and the monitoring frequency is adjusted. Spatial clustering analysis is conducted to determine clustered areas and calculate their geometric characteristics. Weighted time series fitting of the displacement rates of anomalous marker points within the clustered areas yields a weighted average displacement acceleration. Multimodal environmental data is then acquired, and correlation analysis is performed to identify environmental factors and time delay characteristics. This data is input into a machine learning prediction model to output the probability of landslide occurrence, determine whether an early warning is triggered, and send it to a remote terminal. This method can accurately monitor mountain displacement, promptly identify potential landslide risk areas, and provide a scientific basis for disaster prevention and mitigation to reduce disaster losses.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring technology, and more specifically to an intelligent remote monitoring method and system for landslides. Background Technology

[0002] Landslides are highly destructive geological hazards, making remote monitoring and early warning technologies crucial for protecting people's lives and property. Traditional technologies primarily rely on periodic manual inspections or the deployment of discrete sensors (such as GPS, crack gauges, and extensometers) for single-point monitoring. These methods suffer from limitations such as limited monitoring range, high deployment costs, and limited data dimensionality, and they also struggle to capture the systematic evolution of landslides from localized deformation to overall instability. To overcome these limitations, current technologies have incorporated non-contact remote sensing techniques such as 3D laser scanning and synthetic aperture radar interferometry, enabling high-precision, large-scale deformation monitoring of mountain surfaces.

[0003] However, existing technologies still have the following drawbacks:

[0004] Firstly, existing technologies lack in-depth mining of massive point cloud data, making it difficult to effectively identify anomalous regions with spatial clustering, thus making it difficult to assess the scale and morphology of potential slip bodies.

[0005] Secondly, existing technologies neglect the fact that landslide evolution is a complex process involving the combined effects of internal and external factors (such as deformation, rainfall, and groundwater), and fail to establish dynamic correlations and causal inferences between multi-source heterogeneous data, resulting in insufficient accuracy and timeliness of early warning information. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides an intelligent remote monitoring method and system for landslides to solve the problems existing in the background art.

[0007] This invention provides the following technical solution: an intelligent remote monitoring method for landslides, comprising:

[0008] S1: Obtain the unique IDs of all marker points within the target monitoring area and their corresponding physical coordinate information to construct a basic dataset for displacement calculation;

[0009] S2: Based on a unified time reference, the background image of the mountain and the image of the light-bearing point in the target monitoring area are acquired sequentially. The background image and the image of the light-bearing point are differentially processed to extract the coordinates of the light-bearing point image. Combined with the basic dataset, the coordinates of the light-bearing point image are converted into a point cloud dataset.

[0010] S3: Based on the point cloud datasets of the current and historical times, perform displacement analysis, identify abnormal marker points, and adaptively adjust the monitoring frequency according to the displacement analysis results;

[0011] S4: Perform spatial clustering analysis based on the generated point cloud dataset, identify the clustered regions formed by all outlier markers in space, and calculate the spatial geometric features of the clustered regions;

[0012] S5: Based on the identified clustered regions and their spatial geometric features, perform weighted time series fitting on the displacement rates of all anomalous marker points contained within the clustered regions, and calculate the weighted average displacement acceleration.

[0013] S6: Acquire multimodal environmental data synchronized with the target monitoring area, and perform correlation analysis between the calculated weighted average displacement acceleration and the multimodal environmental data to identify environmental factors and their time delay characteristics.

[0014] S7: Input the current value of the weighted average displacement acceleration, the identified environmental factors and the time delay characteristics of the environmental factors into the preset machine learning prediction model, and output the probability of landslide occurrence.

[0015] S8: Based on the probability of landslide occurrence, determine whether to trigger a landslide early warning and send the warning information to a remote terminal.

[0016] Preferably, S1 obtaining the unique ID of the marker point and its corresponding physical coordinate information includes:

[0017] The LoRa gateway module integrated into the remote monitoring camera unit broadcasts a low-power wake-up command to the monitoring network.

[0018] Receive the unique ID identifier and status information returned by multiple active light-emitting marker units activated by the wake-up command;

[0019] The data processing center constructs a basic dataset for displacement calculation based on the unique ID identifier and status information returned. This dataset contains marker point IDs, status information, and preset physical coordinates.

[0020] Preferably, step S2 sequentially acquires the background image and the image containing illuminated points based on a unified time reference, and performs differential processing, including the following steps:

[0021] S211: In At any given moment, capture a background image without any marker light spots. ;

[0022] S212: Broadcast a synchronization lighting command through the LoRa gateway module so that all marker point units emit pulsed lasers synchronously after a fixed delay;

[0023] S213: In At any given moment, capture an image of the light-bearing points that includes all the marked light spots. ,in and Based on a unified time reference, the aforementioned Later than the stated And the This is associated with the fixed delay to ensure that the synchronously emitted pulsed laser spot is captured;

[0024] S214: Calculate the image containing light points With background image The differential image is used to eliminate background interference, and the image coordinates of each light point are extracted from the differential image;

[0025] The extracted light spot image coordinates are converted into a point cloud dataset, specifically including:

[0026] The coordinates of the light spot image are matched and associated with the constructed basic dataset. Based on the matching and association results, a point cloud dataset containing a unique identifier ID, image coordinates, and three-dimensional physical coordinates is generated.

[0027] Preferably, in step S3, displacement analysis is performed and the monitoring frequency is adaptively adjusted based on the analysis results, including:

[0028] The point cloud dataset at the current moment is compared point by point with the point cloud dataset at the previous historical moment, and the displacement vector of each marked point is calculated within a preset time interval.

[0029] Based on the displacement vector, the displacement rate of each marker point is calculated, and combined with the preset deformation evaluation model, a quantified single-point deformation state evaluation index is generated for each marker point.

[0030] Based on the single-point deformation state evaluation index, all abnormal marker points are identified, wherein marker points whose single-point deformation state evaluation index exceeds a preset abnormal threshold are designated as abnormal marker points.

[0031] Based on the distribution of the single-point deformation state assessment index of all marked points within the region, the monitoring frequency for the next monitoring cycle is dynamically adjusted, specifically including the following steps:

[0032] A set of tiered thresholds is defined, including: the upper limit of the stable range, a warning threshold, and a high-risk threshold, and the three satisfy the following relationship: the upper limit of the stable range... Warning threshold High-risk threshold;

[0033] When the single-point deformation state assessment index of at least one marked point in the area exceeds the high-risk threshold, the monitoring frequency will be increased to high-frequency monitoring mode.

[0034] When the single-point deformation state assessment index of all marked points in the region is within the stable range defined by the upper limit of the stable range, the monitoring frequency will be reduced to low-frequency monitoring mode.

[0035] When the single-point deformation state assessment index of at least one marked point in the area exceeds the warning threshold but does not exceed the high-risk threshold, the current monitoring frequency is maintained or the system is operated in medium-frequency monitoring mode.

[0036] Preferably, in step S4, the three-dimensional physical coordinates of the abnormal markers identified in step S3 are used as input, and a preset clustering algorithm is used to perform spatial clustering analysis to identify at least one clustered region composed of multiple abnormal markers. Any abnormal marker in the clustered region can be connected to at least one other abnormal marker in the region through a path whose three-dimensional spatial distance between them is less than or equal to the preset neighborhood radius.

[0037] For each identified cluster region, the spatial geometric features of the cluster region are calculated. The spatial geometric features include at least: the volume and surface area of ​​the minimum bounding box or convex hull used to characterize the region's extent, the point density or average nearest neighbor distance used to characterize the region's density, and the principal direction vector or elongation obtained from principal component analysis used to characterize the region's morphology.

[0038] Preferably, in step S5, for each clustering region identified in step S4, the geometric centroid of that region is calculated;

[0039] Each anomaly marker point contained within the region is assigned a spatial weight, which is inversely proportional to the spatial distance from the marker point to the geometric centroid.

[0040] The historical displacement rate data of each abnormal marker point in the region are used to form an independent time series. The corresponding spatial weights are used as weighting coefficients to perform weighted fitting on each time series, thereby generating a weighted average displacement rate curve that characterizes the overall dynamic evolution trend of the region.

[0041] Calculate the rate of change of the slope of the weighted average displacement rate curve, and use the rate of change of the slope as the weighted average displacement acceleration of the region.

[0042] Preferably, step S6 deploys multimodal environmental sensors within the target monitoring area to collect multimodal environmental data synchronized with displacement monitoring in real time or access it through an external data interface. The environmental data includes at least one or more of rainfall, groundwater level, soil moisture content, and seismic activity data.

[0043] For each cluster region, a joint data sequence with time as a common reference is constructed. This sequence includes the calculated weighted average displacement acceleration time series and the synchronously acquired multimodal environmental data time series.

[0044] The sliding window cross-correlation algorithm is used to calculate the cross-correlation function between the weighted average displacement acceleration sequence and each environmental factor sequence at different time offsets;

[0045] By analyzing the peak position and magnitude of the cross-correlation function, the target environmental factor that has a significant statistical correlation with the acceleration change is identified, and the time offset corresponding to the peak is determined as the time delay characteristic of the acceleration response induced by the target environmental factor.

[0046] Preferably, in step S7, the current value of the weighted average displacement acceleration in each cluster area, the identified environmental factors, and the time delay features of the environmental factors are fused to form a multi-dimensional dynamic feature vector, and the feature vector is input into a machine learning prediction model pre-trained based on a historical landslide disaster case dataset, and finally outputs a probability of landslide occurrence within a future time window.

[0047] The machine learning prediction model is a gradient boosting tree model or a neural network model. The model achieves probabilistic prediction by learning the nonlinear mapping relationship between feature vectors in historical landslide disaster case datasets and landslide results.

[0048] Preferably, step S8 compares the landslide occurrence probability with a preset landslide occurrence probability threshold to determine the corresponding warning level;

[0049] When the probability of a landslide is greater than or equal to the preset landslide probability threshold, a structured early warning message containing the location of the target monitoring area, key triggering environmental factors, and the predicted instability time window is automatically generated, and the structured early warning message is sent to one or more designated remote terminal devices through a preset communication protocol.

[0050] To achieve the above objectives, the present invention provides the following technical solution: an intelligent landslide remote monitoring system, comprising the implementation of the aforementioned intelligent landslide remote monitoring method, including:

[0051] Marker point basic information acquisition module: Obtain the unique ID of all marker points within the target monitoring area and their corresponding physical coordinate information to construct a basic dataset for displacement calculation;

[0052] Image Differentiation and Point Cloud Generation Module: Based on a unified time base, the module sequentially acquires the mountain background image and the image containing light spots in the target monitoring area, performs differential processing on the background image and the image containing light spots, extracts the coordinates of the light spot images, and combines them with the basic dataset to convert the coordinates of the light spot images into a point cloud dataset.

[0053] Displacement analysis module: Based on the point cloud datasets of the current and historical times, it performs displacement analysis, identifies abnormal marker points, and adaptively adjusts the monitoring frequency according to the displacement analysis results;

[0054] Spatial clustering analysis module: Performs spatial clustering analysis based on the generated point cloud dataset, identifies the clustered regions formed by all outlier markers in space, and calculates the spatial geometric features of the clustered regions;

[0055] Displacement acceleration calculation module: Based on the identified clustered region and its spatial geometric features, the displacement rate of all anomaly marker points contained within the clustered region is weighted and time-series fitted to calculate the weighted average displacement acceleration.

[0056] Environmental correlation analysis module: Acquire multimodal environmental data synchronized with the target monitoring area, and perform correlation analysis between the calculated weighted average displacement acceleration and the multimodal environmental data to identify environmental factors and their time delay characteristics;

[0057] Landslide probability prediction module: Input the current value of the weighted average displacement acceleration, the identified environmental factors and the time delay characteristics of the environmental factors into the preset machine learning prediction model, and output the probability of landslide occurrence.

[0058] Early warning module: Based on the probability of landslide occurrence, determine whether to trigger a landslide early warning and send the warning information to a remote terminal.

[0059] The technical effects and advantages of this invention are as follows:

[0060] (1) By intelligently filtering out marker points with abnormal displacement behavior from the global point cloud, the data was initially focused. Then, spatial clustering analysis was performed on these discrete anomaly points to determine whether they constitute statistically significant clusters in space. Furthermore, by calculating their spatial geometric features, the disordered set of points was upgraded to the concept of a "potential slip body" with a clear location, range, and shape. Finally, the points within the region were further weighted and fitted based on these geometric features to calculate the weighted average displacement acceleration that characterizes the overall motion state. This series of processes successfully transformed massive, discrete point cloud data into structured risk body information with clear location, quantifiable scale, identifiable shape, and dynamic characterization, completely solving the fundamental defect of existing technologies that cannot assess the scale and shape of potential slip bodies.

[0061] (2) By performing correlation analysis between the calculated weighted average displacement acceleration and multimodal environmental data such as rainfall and groundwater level, not only were key inducing factors identified, but the "time delay characteristics" of these factors affecting deformation response were also quantified, thus revealing the causal chain of landslide evolution. The internal deformation state, external inducing factors, and their time delay characteristics were input into a machine learning prediction model, outputting a comprehensive landslide occurrence probability, achieving a leap from simple threshold alarms to intelligent prediction based on multi-source information fusion. Finally, an early warning was issued based on this probability, and decision-making information containing risk causes was pushed. This entire mechanism makes the early warning no longer based solely on deformation results, but deeply integrates the dynamic correlation and causal inference of internal and external factors, greatly improving the accuracy, timeliness, and decision support value of the early warning. Attached Figure Description

[0062] Figure 1 This is a diagram illustrating the method steps of the present invention.

[0063] Figure 2 This is a system structure block diagram of the present invention. Detailed Implementation

[0064] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. The intelligent landslide remote monitoring method and system involved in the present invention are not limited to the structures described in the following embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] like Figure 1 The embodiment shown provides an intelligent remote monitoring method for landslides, including:

[0066] S1: Obtain the unique IDs of all marker points within the target monitoring area and their corresponding physical coordinate information to construct a basic dataset for displacement calculation.

[0067] In this embodiment, S1 obtaining the unique ID of the marker point and its corresponding physical coordinate information includes:

[0068] The LoRa gateway module integrated into the remote monitoring camera unit broadcasts a low-power wake-up command to the monitoring network.

[0069] Receive the unique ID identifier and status information returned by multiple active light-emitting marker units activated by the wake-up command;

[0070] The data processing center constructs a basic dataset for displacement calculation based on the unique ID identifier and status information returned. This dataset contains marker point IDs, status information, and preset physical coordinates.

[0071] It should be specifically noted that the process of constructing the basic dataset is implemented through a Low Power Wide Area Network (LPWAN) communication system. The remote monitoring camera unit not only handles image acquisition but also integrates a LoRa gateway module, serving as the communication center for the entire monitoring network. During system initialization or periodic maintenance, the LoRa gateway module broadcasts a low-power wake-up command, which activates multiple dormant active luminous marker units deployed within the target monitoring area. Each awakened marker unit then transmits its pre-stored, unique ID and current status information such as battery level and signal strength via LoRa wireless communication technology. After receiving all the transmitted information, the data processing center matches and associates it with the pre-surveyed and recorded physical coordinates (such as latitude and longitude or 3D engineering coordinates) corresponding to each ID, ultimately constructing a structured basic dataset. This dataset is the foundation for all subsequent displacement calculations, with each record precisely corresponding to a marker point in the physical world.

[0072] S2: Based on a unified time reference, the background image of the mountain and the image containing light points in the target monitoring area are acquired sequentially. The background image and the image containing light points are differentially processed to extract the coordinates of the light point images. Combined with the basic dataset, the coordinates of the light point images are converted into a point cloud dataset.

[0073] In this embodiment, step S2 sequentially acquires the background image and the image containing light points based on a unified time reference, and performs differential processing, including the following steps:

[0074] S211: In At any given moment, capture a background image without any marker light spots. ;

[0075] S212: Broadcast a synchronization lighting command through the LoRa gateway module so that all marker point units emit pulsed lasers synchronously after a fixed delay;

[0076] S213: In At any given moment, capture an image of the light-bearing points that includes all the marked light spots. ,in and Based on a unified time reference, the aforementioned Later than the stated And the This is associated with the fixed delay to ensure that the synchronously emitted pulsed laser spot is captured;

[0077] S214: Calculate the image containing light points With background image The differential image is used to eliminate background interference, and the image coordinates of each light point are extracted from the differential image;

[0078] The extracted light spot image coordinates are converted into a point cloud dataset, specifically including:

[0079] The coordinates of the light spot image are matched and associated with the constructed basic dataset. Based on the matching and association results, a point cloud dataset containing a unique identifier ID, image coordinates, and three-dimensional physical coordinates is generated.

[0080] It should be specifically explained that a technique based on temporal synchronization and image differencing is used to accurately convert two-dimensional image information into a three-dimensional point cloud dataset. Firstly, in... At a specific moment, the remote monitoring camera unit captures an image of the mountain background without any marker light spots, which is used for subsequent background removal. Next, the LoRa gateway module integrated into the camera unit broadcasts a synchronization lighting command. Upon receiving this command, all actively emitting marker units deployed on-site synchronously emit high-brightness pulsed lasers after a very short, fixed delay (e.g., 50 milliseconds). The system precisely calculates this delay and... time( = (With a delay and preparation time) A second image is captured, clearly showing all the synchronously emitting marker points. The system then calculates the difference between the image containing the light points and the background image. Through pixel-level subtraction, it efficiently eliminates interference from static background elements such as vegetation, shadows, and lighting variations, leaving only the pure light point information. Finally, image processing algorithms (such as the centroid method) are used to extract the precise image coordinates (pixel positions) of each light point from the difference image and match them with the unique D in the base dataset constructed in S1, thereby generating a structured point cloud dataset. Each record in this dataset contains complete information including {unique ID, image coordinates, and 3D physical coordinates}.

[0081] S3: Based on the point cloud datasets of the current and historical times, perform displacement analysis, identify abnormal marker points, and adaptively adjust the monitoring frequency according to the displacement analysis results.

[0082] In this embodiment, step S3 involves performing displacement analysis and adaptively adjusting the monitoring frequency based on the analysis results, including:

[0083] The point cloud dataset at the current moment is compared point by point with the point cloud dataset at the previous historical moment, and the displacement vector of each marked point is calculated within a preset time interval.

[0084] Based on the displacement vector, the displacement rate of each marker point is calculated, and combined with the preset deformation evaluation model, a quantified single-point deformation state evaluation index is generated for each marker point.

[0085] Based on the single-point deformation state evaluation index, all abnormal marker points are identified, wherein marker points whose single-point deformation state evaluation index exceeds a preset abnormal threshold are designated as abnormal marker points.

[0086] Based on the distribution of the single-point deformation state assessment index of all marked points within the region, the monitoring frequency for the next monitoring cycle is dynamically adjusted, specifically including the following steps:

[0087] A set of tiered thresholds is defined, including: the upper limit of the stable range, a warning threshold, and a high-risk threshold, and the three satisfy the following relationship: the upper limit of the stable range... Warning threshold High-risk threshold;

[0088] When the single-point deformation state assessment index of at least one marked point in the area exceeds the high-risk threshold, the monitoring frequency will be increased to high-frequency monitoring mode.

[0089] When the single-point deformation state assessment index of all marked points in the region is within the stable range defined by the upper limit of the stable range, the monitoring frequency will be reduced to low-frequency monitoring mode.

[0090] When the single-point deformation state assessment index of at least one marked point in the area exceeds the warning threshold but does not exceed the high-risk threshold, the current monitoring frequency is maintained or the system is operated in medium-frequency monitoring mode.

[0091] It should be specifically explained that an adaptive frequency adjustment mechanism based on single-point deformation state assessment is used to achieve efficient and intelligent allocation of monitoring resources. Specifically, the system first compares the current point cloud dataset with the dataset from the previous historical moment point by point, accurately calculating the displacement vector of each marker point per unit time, and deriving the displacement rate from this. Subsequently, the system inputs these rates into a preset deformation assessment model (which comprehensively considers factors such as displacement rate and rate change) to generate a quantified "single-point deformation state assessment index" for each marker point. This index objectively reflects the degree of deformation at that point. To achieve adaptive adjustment, the system sets a set of graded thresholds, for example: an upper limit of 0.5 mm / d for the stable range, a warning threshold of 2.0 mm / d, and a high-risk threshold of 5.0 mm / d. When the system detects that the evaluation index of any marker point in the area exceeds 5.0 mm / d, it determines that the mountain has entered a high-risk state and immediately increases the monitoring frequency from the usual once per hour to a high-frequency mode (such as once every 5 minutes) to capture subtle changes before instability. Conversely, when the evaluation index of all marker points is below 0.5 mm / d, the system determines that the mountain is in a stable state and automatically reduces to a low-frequency mode (such as once every 12 hours) to significantly save power consumption and communication resources. When the index is between the two, medium-frequency monitoring is maintained.

[0092] S4: Perform spatial clustering analysis based on the generated point cloud dataset to identify the clustered regions formed by all outlier marker points in space, and calculate the spatial geometric features of the clustered regions.

[0093] In this embodiment, S4 takes the three-dimensional physical coordinates of the abnormal markers identified in S3 as input, and uses a preset clustering algorithm to perform spatial clustering analysis to identify at least one clustered region composed of multiple abnormal markers. Any abnormal marker in the clustered region can be connected to at least one other abnormal marker in the region through a path whose three-dimensional spatial distance between them is less than or equal to the preset neighborhood radius.

[0094] For each identified cluster region, the spatial geometric features of the cluster region are calculated. The spatial geometric features include at least: the volume and surface area of ​​the minimum bounding box or convex hull used to characterize the region's extent, the point density or average nearest neighbor distance used to characterize the region's density, and the principal direction vector or elongation obtained from principal component analysis used to characterize the region's morphology.

[0095] It should be specifically explained that spatial clustering analysis elevates discrete outliers into potential slip bodies with clear physical meaning. Specifically, the system uses the 3D physical coordinates of all outlier markers identified by S3 as the input dataset and employs a density-based clustering algorithm, such as DBSCAN, for spatial clustering analysis. The advantage of this algorithm is its ability to automatically identify clusters of arbitrary shapes composed of spatially adjacent points and effectively eliminate isolated noise points. In the DBSCAN algorithm, a preset neighborhood radius (Eps) is used to quantify the spatial proximity between outlier markers. When the 3D spatial distance between two outlier markers is less than or equal to this preset neighborhood radius, they are determined to be spatially reachable and belong to the same cluster. The 3D spatial distance between two outlier markers is calculated from their 3D physical coordinates. Furthermore, the anomalous markers within a cluster must meet the condition of "path connectivity." Specifically, this means that when identifying clusters, not only must the distance between any two adjacent points be less than or equal to a preset neighborhood radius, but any anomalous marker within the cluster must also be able to establish a connected transmission path by finding a series of intermediate markers whose pairwise distances are all less than or equal to the preset neighborhood radius. Based on this connectivity logic, if the distance between an anomalous marker and any marker within the current cluster is greater than the preset neighborhood radius, and the aforementioned connected path cannot be established through other intermediate points, then that point cannot be assigned to that cluster. This effectively avoids erroneously merging spatially distant anomalous points belonging to different independent slip bodies into the same cluster (i.e., preventing cross-regional merging caused by the chain effect). Correspondingly, based on the above spatial clustering analysis constrained by both density and connectivity, at least one mutually independent cluster may ultimately be identified, while isolated anomalous markers that cannot establish a connected path with any point are excluded as noise, ensuring that each identified cluster consists of multiple physically closely related anomalous markers. The specific value of the preset neighborhood radius can be reasonably set according to the geological structure characteristics of the target monitoring area and the spacing of the marker points. For example, in monitoring areas with densely distributed marker points, a smaller neighborhood radius (such as 5 meters) can be set to finely characterize local deformation; in areas with sparsely distributed marker points, a larger neighborhood radius (such as 20 meters) can be set to capture the overall slip trend. For each identified cluster area, the system will further calculate its multi-dimensional spatial geometric features to comprehensively characterize its physical properties. For example, the system will calculate the volume and surface area of ​​the minimum bounding box or convex hull of the area to quantify its scale; calculate the point density or average nearest neighbor distance to assess the concentration of deformation within it; and simultaneously, extract the principal direction vector and elongation of the point cloud through principal component analysis (PCA) to determine the main movement direction and geometric shape of the potential slip body.

[0096] S5: Based on the identified clustered regions and their spatial geometric features, perform weighted time series fitting on the displacement rates of all anomalous marker points contained within the clustered regions to calculate the weighted average displacement acceleration.

[0097] In this embodiment, step S5 calculates the geometric centroid of each cluster region identified in step S4;

[0098] Each anomaly marker point contained within the region is assigned a spatial weight, which is inversely proportional to the spatial distance from the marker point to the geometric centroid.

[0099] The historical displacement rate data of each abnormal marker point in the region are used to form an independent time series. The corresponding spatial weights are used as weighting coefficients to perform weighted fitting on each time series, thereby generating a weighted average displacement rate curve that characterizes the overall dynamic evolution trend of the region.

[0100] Calculate the rate of change of the slope of the weighted average displacement rate curve, and use the rate of change of the slope as the weighted average displacement acceleration of the region.

[0101] It should be specifically explained that a spatially weighted fitting algorithm is used to calculate a unified and representative dynamic evolution index for each identified cluster region. Specifically, for any cluster region identified by S4, the system first calculates the geometric centroid of the three-dimensional coordinates of all its anomaly marker points, using this as the reference center for the region. Next, each anomaly marker point within the region is assigned a spatial weight inversely proportional to its spatial distance from the geometric centroid; that is, the closer a point is to the centroid and the more representative it is of the overall movement of the region, the higher its weight. For example, a point 1 meter from the centroid might receive a weight of 0.8, while an edge point 5 meters from the centroid might only receive a weight of 0.2. Subsequently, the system constructs an independent time series using the historical displacement rate data of each marker point, and uses its corresponding spatial weight as coefficients to perform a weighted least-squares fitting on all time series, thereby generating a weighted average displacement rate curve that smooths out noise and highlights the core dynamics. The slope of this curve represents the weighted average displacement rate of the region, and the rate of change of this slope with time, i.e., the second derivative, is defined as the "weighted average displacement acceleration" of the region.

[0102] S6: Acquire multimodal environmental data synchronized with the target monitoring area, and perform correlation analysis between the calculated weighted average displacement acceleration and the multimodal environmental data to identify environmental factors and their time delay characteristics.

[0103] In this embodiment, S6 deploys multimodal environmental sensors within the target monitoring area to collect multimodal environmental data synchronized with displacement monitoring in real time or access it through an external data interface. The environmental data includes at least one or more of rainfall, groundwater level, soil moisture content, and seismic activity data.

[0104] For each cluster region, a joint data sequence with time as a common reference is constructed. This sequence includes the calculated weighted average displacement acceleration time series and the synchronously acquired multimodal environmental data time series.

[0105] The sliding window cross-correlation algorithm is used to calculate the cross-correlation function between the weighted average displacement acceleration sequence and each environmental factor sequence at different time offsets;

[0106] By analyzing the peak position and magnitude of the cross-correlation function, the target environmental factor that has a significant statistical correlation with the acceleration change is identified, and the time offset corresponding to the peak is determined as the time delay characteristic of the acceleration response induced by the target environmental factor.

[0107] It should be specifically noted that a causal inference method based on time-series analysis reveals the deep dynamic relationship between landslide deformation and environmental inducing factors. First, the system acquires multimodal environmental data that is strictly synchronized with displacement monitoring by deploying multimodal environmental sensors (such as rain gauges, groundwater level gauges, and soil moisture meters) in the monitoring area or by connecting to external data interfaces such as meteorological and seismic data. For each clustered area, the system constructs a joint data sequence with time as a common reference. This sequence simultaneously includes the weighted average displacement acceleration time series calculated by S5 and the time series of each environmental factor. To quantify the causal relationship between them, the system uses a sliding window cross-correlation algorithm to calculate the cross-correlation function between the acceleration sequence and each environmental factor sequence at different time offsets. For example, calculations show that the acceleration sequence and the rainfall sequence exhibit the largest cross-correlation peak at a time offset of 24 hours. The magnitude of this peak indicates the degree to which rainfall explains the acceleration change (i.e., statistical correlation), and the time offset (24 hours) corresponding to the peak is precisely determined as the time delay characteristic of the "rainfall-induced deformation response" in that area.

[0108] S7: Input the current value of the weighted average displacement acceleration, the identified environmental factors and their time delay characteristics into the preset machine learning prediction model, and output the probability of landslide occurrence.

[0109] In this embodiment, S7 fuses the current value of the weighted average displacement acceleration in each cluster area, the identified environmental factors, and the time delay features of the environmental factors to form a multi-dimensional dynamic feature vector, and inputs the feature vector into a machine learning prediction model pre-trained based on a historical landslide disaster case dataset, and finally outputs a probability of landslide occurrence within a future time window.

[0110] The machine learning prediction model is a gradient boosting tree model or a neural network model. The model achieves probabilistic prediction by learning the nonlinear mapping relationship between feature vectors in historical landslide disaster case datasets and landslide results.

[0111] It should be specifically explained that an intelligent prediction of landslide probability from multi-source features is achieved through a pre-trained machine learning model. First, for each cluster area, the system fuses key information such as the current value of the weighted average displacement acceleration calculated by S5, key environmental factors identified by S6 (e.g., 24-hour cumulative rainfall), and their time delay characteristics (e.g., 24 hours) to form a multi-dimensional dynamic feature vector. For example, a feature vector might be represented as [acceleration = 0.15 m / s²]. 2 [Environmental factor 1 = 50mm (rainfall), delay 1 = 24h; Environmental factor 2 = 2.5m (water level), delay 2 = 12h]. This feature vector is then input into a machine learning prediction model pre-trained on a large dataset of historical landslide disaster cases. This embodiment preferentially uses a gradient boosting tree model (such as XGBoost or LightGBM), which builds a powerful predictor by learning the complex nonlinear mapping relationship between thousands of similar feature vectors in historical data and the final landslide outcome (occurred / not occurred). When a new real-time feature vector is input, the model can output a probability value (e.g., 85%) of a landslide occurring within a specific future time window (e.g., the next 48 hours).

[0112] S8: Based on the probability of landslide occurrence, determine whether to trigger a landslide early warning and send the warning information to a remote terminal.

[0113] In this embodiment, step S8 compares the landslide occurrence probability with a preset landslide occurrence probability threshold to determine the corresponding warning level;

[0114] When the probability of a landslide is greater than or equal to the preset landslide probability threshold, a structured early warning message containing the location of the target monitoring area, key triggering environmental factors, and the predicted instability time window is automatically generated, and the structured early warning message is sent to one or more designated remote terminal devices through a preset communication protocol.

[0115] It should be specifically noted that the generated structured early warning information not only includes the precise geographical location of the target monitoring area, but also integrates key triggering environmental factors analyzed by S6 (such as "24-hour cumulative rainfall has reached 80mm") and the instability time window predicted by S7 (such as "the probability of landslide occurrence within the next 48 hours is 85%). Finally, this structured information is pushed instantly and reliably to one or more designated remote terminal devices, such as the monitoring center's large screen, the emergency management personnel's mobile APP, or the relevant personnel's email address, through a preset communication protocol (such as MQTT, SMS, or API interface).

[0116] like Figure 2 This embodiment provides an implementation system for an intelligent remote monitoring method for landslides, including a marker point basic information acquisition module, an image difference and point cloud generation module, a displacement analysis module, a spatial clustering analysis module, a displacement acceleration calculation module, an environmental correlation analysis module, a landslide probability prediction module, and an early warning module. The marker point basic information acquisition module is connected to the image difference and point cloud generation module, the image difference and point cloud generation module is connected to the displacement analysis module, the displacement analysis module is connected to the spatial clustering analysis module, the spatial clustering analysis module is connected to the displacement acceleration calculation module, the displacement acceleration calculation module is connected to the environmental correlation analysis module, the environmental correlation analysis module is connected to the landslide probability prediction module, and the landslide probability prediction module is connected to the early warning module.

[0117] The marker point basic information acquisition module obtains the unique IDs of all marker points within the target monitoring area and their corresponding physical coordinate information to construct a basic dataset for displacement calculation;

[0118] The image difference and point cloud generation module acquires the mountain background image and the light-containing point image of the target monitoring area sequentially based on a unified time base. It performs difference processing on the background image and the light-containing point image to extract the coordinates of the light-containing point image and converts the coordinates of the light-containing point image into a point cloud dataset by combining it with the basic dataset.

[0119] The displacement analysis module performs displacement analysis based on the point cloud datasets of the current and historical times, identifies abnormal marker points, and adaptively adjusts the monitoring frequency based on the displacement analysis results.

[0120] The spatial clustering analysis module performs spatial clustering analysis based on the generated point cloud dataset, identifies the clustered regions formed by all abnormal marker points in space, and calculates the spatial geometric features of the clustered regions.

[0121] The displacement acceleration calculation module performs weighted time series fitting on the displacement rates of all abnormal marker points contained within the identified clustered region and its spatial geometric features, and calculates the weighted average displacement acceleration.

[0122] The environmental correlation analysis module acquires multimodal environmental data synchronized with the target monitoring area, and performs correlation analysis between the calculated weighted average displacement acceleration and the multimodal environmental data to identify environmental factors and their time delay characteristics.

[0123] The landslide probability prediction module inputs the current value of the weighted average displacement acceleration, the identified environmental factors, and the time delay characteristics of the environmental factors into a preset machine learning prediction model, and outputs the probability of landslide occurrence.

[0124] The early warning module determines whether to trigger a landslide early warning based on the probability of landslide occurrence, and sends the early warning information to a remote terminal.

[0125] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0126] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent remote monitoring method for landslides, characterized in that, include: S1: Obtain the unique IDs of all marker points within the target monitoring area and their corresponding physical coordinate information to construct a basic dataset for displacement calculation; S2: Based on a unified time reference, the background image of the mountain and the image of the light-bearing point in the target monitoring area are acquired sequentially. The background image and the image of the light-bearing point are differentially processed to extract the coordinates of the light-bearing point image. Combined with the basic dataset, the coordinates of the light-bearing point image are converted into a point cloud dataset. S3: Based on the point cloud datasets of the current and historical times, perform displacement analysis, identify abnormal marker points, and adaptively adjust the monitoring frequency according to the displacement analysis results; S4: Perform spatial clustering analysis based on the generated point cloud dataset, identify the clustered regions formed by all outlier markers in space, and calculate the spatial geometric features of the clustered regions; S5: Based on the identified clustered regions and their spatial geometric features, perform weighted time series fitting on the displacement rates of all anomalous marker points contained within the clustered regions, and calculate the weighted average displacement acceleration. S6: Acquire multimodal environmental data synchronized with the target monitoring area, and perform correlation analysis between the calculated weighted average displacement acceleration and the multimodal environmental data to identify environmental factors and their time delay characteristics. S7: Input the current value of the weighted average displacement acceleration, the identified environmental factors and the time delay characteristics of the environmental factors into the preset machine learning prediction model, and output the probability of landslide occurrence. S8: Based on the probability of landslide occurrence, determine whether to trigger a landslide early warning and send the warning information to a remote terminal.

2. The intelligent remote monitoring method for landslides according to claim 1, characterized in that, The step of S1 to obtain the unique ID of the marker point and its corresponding physical coordinate information includes: The LoRa gateway module integrated into the remote monitoring camera unit broadcasts a low-power wake-up command to the monitoring network. Receive the unique ID identifier and status information returned by multiple active light-emitting marker units activated by the wake-up command; The data processing center constructs a basic dataset for displacement calculation based on the unique ID identifier and status information returned. This dataset contains marker point IDs, status information, and preset physical coordinates.

3. The intelligent remote monitoring method for landslides according to claim 2, characterized in that, S2 sequentially acquires the background image and the image containing illuminated points based on a unified time reference, and performs differential processing, including the following steps: S211: In At any given moment, capture a background image without any marker light spots. ; S212: Broadcast a synchronization lighting command through the LoRa gateway module so that all marker point units emit pulsed lasers synchronously after a fixed delay; S213: In At any given moment, capture an image of the light-bearing points that includes all the marked light spots. ,in and Based on a unified time reference, the aforementioned Later than the stated And the This is associated with the fixed delay to ensure that the synchronously emitted pulsed laser spot is captured; S214: Calculate the image of the light-containing points With background image The differential image is used to eliminate background interference, and the image coordinates of each light point are extracted from the differential image; The extracted light spot image coordinates are converted into a point cloud dataset, specifically including: The coordinates of the light spot image are matched and associated with the constructed basic dataset. Based on the matching and association results, a point cloud dataset containing a unique identifier ID, image coordinates, and three-dimensional physical coordinates is generated.

4. The intelligent remote monitoring method for landslides according to claim 3, characterized in that, The displacement analysis performed in step S3 and the monitoring frequency adaptively adjusted based on the analysis results include: The point cloud dataset at the current moment is compared point by point with the point cloud dataset at the previous historical moment, and the displacement vector of each marked point is calculated within a preset time interval. The displacement rate of each marker point is calculated based on the displacement vector, and a quantified single-point deformation state evaluation index is generated for each marker point in combination with the preset deformation evaluation model. Based on the single-point deformation state evaluation index, all abnormal marker points are identified, wherein marker points whose single-point deformation state evaluation index exceeds a preset abnormal threshold are designated as abnormal marker points. Based on the distribution of the single-point deformation state assessment index of all marked points within the region, the monitoring frequency for the next monitoring cycle is dynamically adjusted, specifically including the following steps: A set of tiered thresholds is defined, including: the upper limit of the stable range, a warning threshold, and a high-risk threshold, and the three satisfy the following relationship: the upper limit of the stable range... Warning threshold High-risk threshold; When the single-point deformation state assessment index of at least one marked point in the area exceeds the high-risk threshold, the monitoring frequency will be increased to high-frequency monitoring mode. When the single-point deformation state assessment index of all marked points in the region is within the stable range defined by the upper limit of the stable range, the monitoring frequency will be reduced to low-frequency monitoring mode. When the single-point deformation state assessment index of at least one marked point in the area exceeds the warning threshold but does not exceed the high-risk threshold, the current monitoring frequency is maintained or the system is operated in medium-frequency monitoring mode.

5. The intelligent remote monitoring method for landslides according to claim 4, characterized in that, S4 takes the three-dimensional physical coordinates of the abnormal markers identified in S3 as input, and uses a preset clustering algorithm to perform spatial clustering analysis to identify at least one clustered region composed of multiple abnormal markers. Any abnormal marker in the clustered region can be connected to at least one other abnormal marker in the region through a path whose three-dimensional spatial distance between them is less than or equal to the preset neighborhood radius. For each identified cluster region, the spatial geometric features of the cluster region are calculated. The spatial geometric features include at least: the volume and surface area of ​​the minimum bounding box or convex hull used to characterize the region's extent, the point density or average nearest neighbor distance used to characterize the region's density, and the principal direction vector or elongation obtained from principal component analysis used to characterize the region's morphology.

6. The intelligent remote monitoring method for landslides according to claim 5, characterized in that, For each clustering region identified in S4, S5 calculates the geometric centroid of that region; Each anomaly marker point contained within the region is assigned a spatial weight, which is inversely proportional to the spatial distance from the marker point to the geometric centroid. The historical displacement rate data of each abnormal marker point in the region are used to form an independent time series. The corresponding spatial weights are used as weighting coefficients to perform weighted fitting on each time series, thereby generating a weighted average displacement rate curve that characterizes the overall dynamic evolution trend of the region. Calculate the rate of change of the slope of the weighted average displacement rate curve, and use the rate of change of the slope as the weighted average displacement acceleration of the region.

7. The intelligent remote monitoring method for landslides according to claim 6, characterized in that, The S6 deploys multimodal environmental sensors within the target monitoring area to collect multimodal environmental data synchronized with displacement monitoring in real time or access it through an external data interface. The environmental data includes at least one or more of rainfall, groundwater level, soil moisture content, and seismic activity data. For each cluster region, a joint data sequence with time as a common reference is constructed. This sequence includes the calculated weighted average displacement acceleration time series and the synchronously acquired multimodal environmental data time series. The sliding window cross-correlation algorithm is used to calculate the cross-correlation function between the weighted average displacement acceleration sequence and each environmental factor sequence at different time offsets; By analyzing the peak position and magnitude of the cross-correlation function, the target environmental factor that has a significant statistical correlation with the acceleration change is identified, and the time offset corresponding to the peak is determined as the time delay characteristic of the acceleration response induced by the target environmental factor.

8. The intelligent remote monitoring method for landslides according to claim 7, characterized in that, S7 integrates the current value of the weighted average displacement acceleration in each cluster area, the identified environmental factors, and the time delay features of the environmental factors to form a multi-dimensional dynamic feature vector. This feature vector is then input into a machine learning prediction model pre-trained based on a historical landslide disaster case dataset, and finally outputs a probability of landslide occurrence within a future time window. The machine learning prediction model is a gradient boosting tree model or a neural network model. The model achieves probabilistic prediction by learning the nonlinear mapping relationship between feature vectors in historical landslide disaster case datasets and landslide results.

9. The intelligent remote monitoring method for landslides according to claim 8, characterized in that, S8 compares the landslide occurrence probability with a preset landslide occurrence probability threshold to determine the corresponding warning level; When the probability of a landslide is greater than or equal to the preset landslide probability threshold, a structured early warning message is automatically generated, which includes the location of the target monitoring area, key triggering environmental factors, and the predicted instability time window. This structured early warning message is then sent to one or more designated remote terminal devices via a preset communication protocol.

10. An intelligent landslide remote monitoring system, implementing the intelligent landslide remote monitoring method as described in any one of claims 1-9, characterized in that, include: Marker point basic information acquisition module: Obtain the unique ID of all marker points within the target monitoring area and their corresponding physical coordinate information to construct a basic dataset for displacement calculation; Image Differentiation and Point Cloud Generation Module: Based on a unified time base, the module sequentially acquires the background image of the mountain and the image containing light spots in the target monitoring area. It performs differential processing on the background image and the image containing light spots to extract the coordinates of the light spots. Combined with the basic dataset, the coordinates of the light spots are converted into a point cloud dataset. Displacement analysis module: Based on the point cloud datasets of the current and historical times, it performs displacement analysis, identifies abnormal marker points, and adaptively adjusts the monitoring frequency according to the displacement analysis results; Spatial clustering analysis module: Performs spatial clustering analysis based on the generated point cloud dataset, identifies the clustered regions formed by all outlier markers in space, and calculates the spatial geometric features of the clustered regions; Displacement acceleration calculation module: Based on the identified clustered region and its spatial geometric features, the displacement rate of all anomaly marker points contained within the clustered region is weighted and time-series fitted to calculate the weighted average displacement acceleration. Environmental correlation analysis module: Acquire multimodal environmental data synchronized with the target monitoring area, and perform correlation analysis between the calculated weighted average displacement acceleration and the multimodal environmental data to identify environmental factors and their time delay characteristics; Landslide probability prediction module: Input the current value of the weighted average displacement acceleration, the identified environmental factors and the time delay characteristics of the environmental factors into the preset machine learning prediction model, and output the probability of landslide occurrence. Early warning module: Based on the probability of landslide occurrence, determine whether to trigger a landslide early warning and send the warning information to a remote terminal.

Images