Landslide monitoring and early warning method and system for complex environment

Abstract

The invention provides a landslide monitoring and early warning method and system for a complex environment, and the method comprises the steps: carrying out the field scanning of a target region, collecting original point cloud data, carrying out the data preprocessing, obtaining the point cloud data of a mountain, carrying out the semantic division of the point cloud data of the mountain and corresponding feature vectors, and obtaining the feature vectors of the mountain; extracting high-level semantic features of the point cloud, obtaining a ground feature classification result, analyzing mountain point cloud data in different periods, obtaining ground surface displacement data of the target area, obtaining a three-dimensional displacement vector field of the surface of the mountain in the target area, calculating a ground surface deformation rate, and judging the stability of the mountain in the target area. And carrying out landslide prediction on the mountain point cloud data, and providing early warning information. According to the method, local and global features are fused through a multi-scale feature extraction structure, the surface deformation data are extracted based on feature points of normal vectors and curvatures, dynamic monitoring of mountain surface changes is realized, and occurrence of landslide events is prevented.

Description

Technical Field

[0001] The invention belongs to the technical field of landslide monitoring, and in particular relates to a landslide monitoring and early warning method and system for complex environments. Background Art

[0002] Landslide monitoring is crucial in disaster prevention and mitigation and ensuring people's safety. With the development of remote sensing technology, especially the application of technologies such as LiDAR, satellite imaging and drones, it is now possible to more accurately monitor small changes in the mountains, helping us better understand the stability of the mountains and provide strong data support for landslide warning. In addition, the combination of big data and artificial intelligence makes the analysis of landslide monitoring data more efficient and can detect potential dangers and changing trends more quickly. However, the current landslide monitoring and early warning solutions in complex environments have the following shortcomings:

[0003] (1) When classifying point clouds, most traditional feature extraction methods are used, which are difficult to effectively process multi-scale features in complex scenes and easily lead to low classification accuracy. In particular, in complex or changing environments, the accuracy of distinguishing landslide-related objects is low;

[0004] (2) They mostly rely on simple differential methods or local area analysis, which makes it difficult to fully describe the three-dimensional deformation of the entire mountain surface. Especially in complex scenes, it is difficult to accurately capture the tiny deformation changes during the landslide process;

[0005] (3) They often rely solely on data from a certain field, such as displacement monitoring or geological surveys, and lack multi-dimensional data fusion and analysis, resulting in insufficient accuracy in identifying potential landslide risks, slow response speed, and difficulty in achieving timely warnings and effective responses.

[0006] Therefore, we need to develop a landslide monitoring and early warning method and system for complex environments, which has high classification accuracy for land objects and accurate and reliable deformation monitoring results, and combines multi-dimensional data to identify and warn landslide risks, providing a reference for landslide monitoring in complex scenarios. Summary of the invention

[0007] The purpose of the present invention is to provide a landslide monitoring and early warning method and system for complex environments, so as to solve the problems of low classification accuracy, relatively simple data and incomplete deformation monitoring of existing complex environment landslide monitoring and early warning schemes mentioned in the above background technology.

[0008] To achieve the above objectives, on the one hand, the present invention provides a landslide monitoring and early warning method for complex environments, the method is specifically as follows:

[0009] Use professional measurement equipment to conduct field scanning of the target area that needs landslide monitoring, collect original point cloud data reflecting the surface information of the target area, and perform data preprocessing to obtain mountain point cloud data;

[0010] Based on different surface coverage types, the mountain point cloud data and corresponding feature vectors are semantically divided, and high-level semantic features of the point cloud are automatically extracted to complete the classification of objects in the mountain point cloud data and obtain the object classification results;

[0011] The feature vector represents the geometric properties and radiation properties of the point cloud in the mountain point cloud data, and is composed of two parts: geometric features and radiation features;

[0012] The geometric features describe the local spatial shape of the point cloud, including normal vectors, curvature, and roughness;

[0013] By analyzing the mountain point cloud data at different periods, the surface displacement data of the target area is obtained, and the three-dimensional displacement vector field of the surface of the mountain in the target area is obtained by combining the surface deformation detection algorithm. Based on the three-dimensional displacement vector field, the surface deformation rate is calculated, and the stability of the mountain in the target area is judged based on the surface deformation rate to complete the surface deformation detection;

[0014] Based on the surface deformation rate, the ground feature classification results and the geological structure data, a landslide warning model is constructed to predict landslides based on the mountain point cloud data and provide warning information;

[0015] The surface deformation rate includes an average displacement rate and an average displacement direction. The acquisition of the average displacement rate includes:

[0016] Land feature area The mountain point cloud data includes points, Points in arrive The horizontal displacement during the time period is , the vertical displacement is , then the feature area The average horizontal displacement rate and the average vertical displacement rate Calculated as: ,in For the time interval.

[0017] Based on the above scheme, the surface deformation rate is calculated by decomposing the three-dimensional displacement vector field into horizontal displacement components and vertical displacement components. For each point in the point cloud, , its horizontal displacement and vertical displacement It can be expressed as: ;

[0018] After calculating the horizontal displacement component and the vertical displacement component of each point, the surface deformation rates of different ground object areas are counted based on the ground object classification results;

[0019] A principal component analysis is performed on the three-dimensional displacement vectors of all points, and the direction of the first principal component is taken as the average displacement direction.

[0020] Based on the above scheme, the mountain point cloud data is The form of coordinates represents the spatial position of each point;

[0021] The radiation characteristics are derived from the echo signal of the laser radar, including reflection intensity and echo number;

[0022] The geometric features are as follows:

[0023] The normal vector reflects the orientation of the local surface and is estimated by principal component analysis or least squares plane fitting method;

[0024] The curvature reflects the curvature of the local surface, and the calculation formula is as follows: ,in is the coefficient of the quadratic term of the fitted parabola;

[0025] The roughness reflects the fluctuation of the local surface, and the calculation formula is as follows: ,in represents the number of points in the neighborhood, that is, the total number of local points used for calculation, Indicate point The attribute value of point The height relative to a reference plane or reference point, Represents the average of all point attribute values ​​in the neighborhood.

[0026] Based on the above scheme, the automatic extraction of high-level semantic features of point clouds uses the point cloud deep learning framework PointNet++ as the backbone network, and simultaneously extracts local features and global features of the point cloud geometric features in the mountain point cloud data, specifically including:

[0027] Local feature extraction: Using multi-scale group convolution, convolution operations are performed on point clouds at different neighborhood scales to obtain multi-scale local descriptors;

[0028] Global feature extraction: Through the maximum pooling layer, the multi-scale local descriptors of different areas in the point cloud are aggregated into an overall global descriptor as the high-level semantic feature.

[0029] Based on the above solution, the local feature extraction specifically includes:

[0030] Select a set of key points in the point cloud by sampling the farthest point , with each As the center point, select the radius Neighborhood ball , each field ball The points in the group form a group, and point-by-point convolution is performed in each group. Geometric features and the characteristics of other points in the group Cascade and send to a multi-layer perceptron (MLP) to learn local features. The expression is: ,in Indicate point The local descriptor of Indicated by point is the center and the radius is The set of points in the neighborhood ball of It is the maximum pooling operation.

[0031] Based on the above scheme, the acquisition of the surface displacement data of the target area is specifically as follows:

[0032] Selecting a series of benchmark feature points in the point cloud of the mountain point cloud data as benchmarks for surface deformation analysis, wherein the benchmark feature points are selected in a stable surface area;

[0033] Find the same-name points of the reference feature points in the point clouds of different periods, establish the corresponding relationship between the point clouds of different periods through the geometric features and radiation features of the points in the point clouds, and find the closest reference feature point pairs in different periods as matching point pairs by calculating the Euclidean distance between the reference feature points of different periods;

[0034] Based on the matching point pairs, the three-dimensional displacement vector of each reference feature point is calculated. , the calculation formula is as follows: ,in and The reference feature points exist and The three-dimensional coordinates of the moment.

[0035] Based on the above scheme, the professional measuring equipment is deployed at multiple measuring sites, which are distributed in a ring around the target area. The straight-line distance between two adjacent measuring sites is 1 / 3 to 1 / 2 of the effective ranging of the professional measuring equipment. The professional measuring equipment also includes a satellite navigation system and a measuring robot to assist in achieving high-precision three-dimensional positioning of the measuring sites.

[0036] Based on the above scheme, the construction of the landslide early warning model includes:

[0037] Prepare positive samples and negative samples, each sample corresponds to a prediction unit, extract the horizontal displacement rate and the average vertical displacement rate and the ground feature classification results of each period in the sample, and combine them with the geological structure data to form a complete training sample set as the training set of the landslide early warning model;

[0038] Using a feature selection method, the key features in the training sample set that are highly correlated with the occurrence of landslides are screened out, the landslide early warning model is trained based on the key features, the key feature importance ranking is output, and the top-ranked key feature subset is selected for the final landslide early warning model training;

[0039] The landslide early warning model is constructed based on the machine learning algorithm of support vector machine, which is expressed as:

[0040]

[0041] in The input feature vector contains the features of the prediction unit: x = [ v ¯ h,k , v ¯ v,k , r 1 , r 2 , ..., r m , D] in and are the average horizontal displacement rate and the average vertical displacement rate, For this prediction unit The proportion of landform types, is the distance from the prediction unit to the nearest geological structure, , is the category label of the sample, represents the stable region, Indicates the landslide area. is the Lagrange multiplier, is the kernel function, is the bias term;

[0042] The performance of the landslide early warning model is evaluated.

[0043] Based on the above scheme, the landslide prediction and early warning information provision include:

[0044] The target area is divided into a number of prediction units according to a preset size, and the average horizontal displacement rate is and the average vertical displacement rate As the displacement feature of the prediction unit, the proportion of each type of ground objects in the prediction unit is calculated, and the distance from the prediction unit to the nearest geological structure is extracted;

[0045] The above displacement characteristics, the proportion of various types of landforms and the distance are input into the landslide early warning model to obtain the probability of landslide occurrence in each prediction unit; the landslide occurrence probability results of all prediction units are summarized to generate a landslide hazard zoning map of the target area to intuitively express the spatial distribution of landslides;

[0046] Set warning thresholds for landslide probability, set quantitative landslide risk levels, and monitor the prediction units;

[0047] When the landslide risk level occurs in the target area, early warning information is issued in a timely manner.

[0048] On the other hand, the present invention provides a landslide monitoring and early warning system for complex environments, which is used to implement a landslide monitoring and early warning method for complex environments described in the above scheme, comprising:

[0049] Data collection layer: responsible for scanning the target area for landslide monitoring using professional measurement equipment, and collecting original point cloud data reflecting the surface information of the target area;

[0050] Data processing layer: responsible for data preprocessing of the original point cloud data, including denoising, splicing, and registration, to obtain mountain point cloud data; semantically dividing the mountain point cloud data and the corresponding feature vectors based on different surface coverage types, and automatically extracting high-level semantic features of the point cloud, completing the classification of objects in the mountain point cloud data, and obtaining the object classification results; by analyzing the mountain point cloud data of different periods, obtaining the surface displacement data of the target area, combining with the surface deformation detection algorithm, obtaining the three-dimensional displacement vector field of the surface of the mountain in the target area, calculating the surface deformation rate based on the three-dimensional displacement vector field, and judging the stability of the mountain in the target area based on the surface deformation rate, completing the surface deformation detection;

[0051] Data storage layer: responsible for storing and managing all point cloud data and processing result data;

[0052] Business application layer: As a user-oriented function and service integration layer, it is responsible for the following businesses: building a landslide warning model based on long-term monitoring data and machine learning algorithms, predicting landslides based on the mountain point cloud data through the landslide warning model, comprehensively analyzing the displacement rate, landform distribution and geological structure characteristics of the mountain, predicting areas where landslides may occur, and providing warning information.

[0053] Compared with the prior art, the present invention has the following advantages and effects:

[0054] (1) Based on the PointNet++ deep learning framework, a multi-scale feature extraction structure was designed, and local and global features were integrated, which effectively improved the accuracy and efficiency of object classification in complex scenes, especially in mountainous areas or landslide areas, and was able to more accurately and efficiently identify and classify different types of objects;

[0055] (2) By extracting feature points through normal vector and curvature analysis and combining it with spatial interpolation technology, the complete three-dimensional displacement field of the mountain surface can be depicted. This can comprehensively and accurately reflect the spatial distribution of mountain surface deformation, provide high-precision monitoring data for landslides, and provide accurate reference for landslide prevention.

[0056] (3) In terms of early warning mechanism, a support vector machine (SVM) early warning model integrating displacement field, terrain attributes and geological structure was constructed, and a scientific graded early warning indicator system was established to achieve reliable identification and rapid response to potential landslide risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0058] Figure 1 is a flow chart of a landslide monitoring and early warning method for a complex environment provided by an embodiment of the present invention;

[0059] Figure 2 It is a structural diagram of a landslide monitoring and early warning system for complex environments provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to more clearly explain the purpose, technical solutions and advantages of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. The example implementation methods can be implemented in various forms and should not be understood as being limited to the examples described herein. On the contrary, these implementation methods are provided to make the present invention more comprehensive and complete, and to fully convey the concepts of the example implementation methods to those skilled in the art.

[0061] In addition, the described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present invention. However, those skilled in the art will appreciate that the technical solution of the present invention can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known methods, devices, implementations or operations are not shown or described in detail to avoid blurring various aspects of the present invention.

[0062] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0063] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0064] The present invention is described in detail below in conjunction with specific embodiments:

[0065] Example 1

[0066] As attached Figure 1 As shown, Embodiment 1 of the present invention provides a landslide monitoring and early warning method for a complex environment, and the specific steps of the method are as follows:

[0067] Step S1: Performing a field scan of the target area for landslide monitoring by means of professional measuring equipment to collect original point cloud data reflecting the surface information of the target area;

[0068] Step S2: performing data preprocessing on the original point cloud data, including but not limited to denoising, splicing, and registration, to obtain mountain point cloud data;

[0069] Step S3: Based on different surface cover types, the mountain point cloud data and the corresponding feature vectors are semantically divided, and high-level semantic features of the point cloud are automatically extracted to complete the classification of objects in the mountain point cloud data and obtain the object classification result;

[0070] Step S4: Analyze the mountain point cloud data at different periods to obtain the surface displacement data of the target area, and combine the surface deformation detection algorithm to obtain the three-dimensional displacement vector field of the surface of the mountain in the target area. Based on the three-dimensional displacement vector field, calculate the surface deformation rate, and judge the stability of the mountain in the target area based on the surface deformation rate to complete the surface deformation detection;

[0071] Step S5: Based on the long-term monitoring data and machine learning algorithm, a landslide warning model is constructed, and the landslide is predicted on the mountain point cloud data through the landslide warning model. The displacement rate, landform distribution and geological structure characteristics of the mountain are comprehensively analyzed to predict the area where landslides may occur and provide warning information.

[0072] Preferably, in this embodiment, the professional measurement equipment in step S1 includes but is not limited to ground three-dimensional laser radar equipment;

[0073] It should be noted that the ground-based three-dimensional laser radar is an active remote sensing measurement system that determines the slant distance between the laser emission point and the target by emitting laser pulses to the target and receiving the laser echo signal reflected by the target. Combined with the directional parameters (azimuth and zenith angle) of the laser rangefinder, one laser pulse can realize the rapid determination of the three-dimensional coordinates of a point on the target, and the collection of multiple points forms three-dimensional point cloud data;

[0074] Specifically, the professional measuring equipment is arranged at a plurality of measuring sites, and each measuring site is located at a position with a wide field of view and a gentle slope. The measuring sites are distributed in a ring around the target area, and the straight-line distance between two adjacent measuring sites is 1 / 3 to 1 / 2 of the effective ranging of the professional measuring equipment. The professional measuring equipment also includes a satellite navigation system and a measuring robot to assist in achieving high-precision three-dimensional positioning of the measuring sites.

[0075] Preferably, the data preprocessing in step S2 includes:

[0076] Denoising: using statistical filtering or radius filtering to remove noise points in the original point cloud data to obtain denoised point cloud data;

[0077] Stitching: Stitching the denoised point cloud data obtained from multiple measurement sites into a unified three-dimensional mountain model through the Iterative Closest Point (ICP) algorithm, and optimizing the registration efficiency of the point cloud registration algorithm through feature point selection and downsampling;

[0078] Registration: Select stable and reliable feature points in the original point cloud data as control points. Based on the control points, use the least squares method to solve the transformation relationship between point clouds in different periods to achieve temporal registration, control the registration residual within 3 cm, and obtain the mountain point cloud data; divide and index the massive spatiotemporal point clouds in the pre-processed data to improve data retrieval and analysis speed.

[0079] Preferably, the surface cover types in step S3 include but are not limited to bare soil, low vegetation, tall vegetation, rocks, roads, and buildings, and can be appropriately adjusted according to the specific mountain scene of the target area.

[0080] Preferably, in step S3, the mountain point cloud data is The form of coordinates represents the spatial position of each point;

[0081] The feature vector represents the geometric properties and radiation properties of the point cloud in the mountain point cloud data, and is composed of two parts: geometric features and radiation features;

[0082] The geometric features describe the local spatial shape of the point cloud, including but not limited to normal vector, curvature, and roughness; the normal vector Reflects the orientation of the local surface; curvature Reflects the curvature of the local surface. The calculation formula is as follows: ,in To fit the parabola The quadratic coefficients are: a, b, c, d, e, and f are fitting coefficients, x and y are the plane coordinates of the point, z is the elevation value, a and b are the quadratic coefficients, c is the cross coefficient, d and e are the linear coefficients, and f is the constant term; roughness Reflects the fluctuation of the local surface, and the calculation formula is: ,in represents the number of points in the neighborhood, that is, the total number of local points used for calculation, Indicates The attribute value of a point is the height of the point relative to a reference plane or reference point, such as Coordinate values, Represents the average value of all point attribute values ​​in the neighborhood. The larger the value, the more obvious the surface fluctuations are, that is, the higher the roughness;

[0083] The radiation characteristics are derived from the echo signal of the lidar, including but not limited to reflection intensity and echo number. The reflection intensity reflects the surface characteristics of the target object, such as material and roughness, and the reflection intensity is normalized to a grayscale between 0 and 255; the echo number reflects the degree of scattering of the laser when penetrating complex objects such as vegetation, and can be used to distinguish between bare land and vegetation; after experimental comparison, in this embodiment, 128 dimensions are preferably used as the dimension of the feature vector to achieve a good balance between performance and efficiency.

[0084] Preferably, the automatic extraction of high-level semantic features of point clouds in step S3 uses the point cloud deep learning framework PointNet++ as a backbone network, and simultaneously extracts local features and global features of point cloud geometric features in the mountain point cloud data, specifically including:

[0085] Step S301, local feature extraction: using a multi-scale group convolution (MSG) method, convolution operations are performed on the point cloud at different neighborhood scales to obtain a multi-scale local descriptor;

[0086] Specifically, a set of key points are selected in the point cloud through farthest point sampling. , with each As the center point, select the radius Neighborhood ball , each field ball The points in the group form a group, and point-by-point convolution is performed in each group. The feature vector and the feature vectors of other points in the group Cascade and send to a multi-layer perceptron (MLP) to learn local features. The expression is: ,in Indicate point The local descriptor of Indicated by point is the center and the radius is The set of points in the neighborhood ball of It is the maximum pooling operation;

[0087] Specifically, in order to extract local features of different scales, the multi-scale group convolution uses a plurality of neighborhood balls of different radii to respectively extract the local descriptors and cascade them as the multi-scale local descriptors;

[0088] Exemplarily, the extraction of the geometric features requires the selection of an appropriate neighborhood scale. In this embodiment, a multi-scale geometric feature fusion strategy is adopted to calculate the geometric features of the point cloud at different neighborhood radii such as 0.2m, 0.5m, and 1.0m, and then cascade them into a high-dimensional feature vector, namely the global descriptor.

[0089] Step S302, global feature extraction: through the maximum pooling layer, the multi-scale local descriptors of different regions in the point cloud are aggregated into an overall global descriptor as the high-level semantic feature;

[0090] Step S303, classification head setting: three fully connected layers and one Softmax layer are used to form a classification head, and the global descriptor is mapped into a ground object category probability vector for each point in the point cloud through the classification head;

[0091] It should be noted that in order to extract local features of different scales, the multi-scale group convolution MSG uses multiple neighborhood balls of different radii to extract local descriptors respectively and cascade them; global feature extraction is mainly achieved through the maximum pooling layer, which aggregates multi-scale local descriptors in different regions into an overall global descriptor; finally, the classification head consists of three fully connected layers and a Softmax layer, which maps the global descriptor to the probability vector of the object category for each point.

[0092] Preferably, the classification of objects in the mountain point cloud data in step S3 is achieved through a ground object classification prediction model, including:

[0093] Step S304, constructing a labeled point cloud training set: for the target area, manually select a point cloud with representative ground object types and annotate point by point to obtain high-quality training samples; in this embodiment, to ensure the training effect, the effective sample size of each ground object category is not less than 10,000 points; before training, the training samples are data enhanced to improve the robustness and generalization of the model to obtain a point cloud training set, and the data enhancement methods used include but are not limited to: random rotation, translation, scaling, and Gaussian noise perturbation. These operations can expand the variability of samples without changing the semantics of the point cloud; specifically, in this embodiment, a combination strategy of random horizontal rotation (0~360°), random translation (±1m), and Gaussian noise perturbation (mean is 0, standard deviation is 0.01) is used to perform data enhancement to obtain the point cloud training set;

[0094] Step S305, training and optimizing the object classification prediction model: inputting the point cloud training set into the object classification prediction model, calculating the predicted category distribution of the point cloud and the predicted probability of the object classification; using the cross entropy loss function to measure the difference between the predicted category distribution and the true distribution: ,in is the number of point cloud samples, is the number of land feature categories, for The true category label (one-hot form), For point Belongs to category The prediction probability is ; the Adam optimizer is used to adaptively adjust the learning rate of each parameter in the land object classification prediction model, and the parameter update formula is: ,in For the The parameters of the step, is the initial learning rate, and are the bias-corrected estimates of the first and second moments of the gradient, is a smoothing term (take a smaller value, such as );

[0095] Specifically, in this embodiment, the initial learning rate of the Adam optimizer is set to 0.001, and a step-by-step attenuation strategy is used for dynamic adjustment, that is, the learning rate is reduced by 10 times every certain number of steps until the model training is completed; the number of samples in each training iteration is set to 8, and the number of training steps, that is, the number of times all samples are traversed for training, is set to 200 to ensure that the model fully converges;

[0096] Step S306, performing object classification prediction on the point cloud: inputting the high-level semantic features into the object classification prediction model, performing classification prediction, and obtaining the object classification results of the point cloud in the mountain point cloud data, including but not limited to the predicted category distribution of the point cloud and the predicted probability of the object classification.

[0097] Preferably, the acquisition of the surface displacement data of the target area in step S4 is specifically as follows:

[0098] Step S401: Select a series of reference feature points from the point cloud of the mountain point cloud data as references for surface deformation analysis. The reference feature points are selected in a stable surface area, including but not limited to hardened roads and buildings, so as to prevent the mountain deformation from affecting the stable surface area. Meanwhile, in this embodiment, in order to ensure the reliability of the surface deformation analysis, the number of the reference feature points is not less than 1000 and is evenly distributed in the entire point cloud area.

[0099] Specifically, points with small normal vector changes and low curvature are selected as reference feature points, and then a uniformly distributed subset of reference feature points is obtained through the farthest point sampling method; The normal vector It can be obtained by fitting a plane to its neighborhood points, and the curvature It can be estimated by calculating the eigenvalues ​​of the neighborhood covariance matrix: ,in are the eigenvalues ​​of the covariance matrix.

[0100] Step S402: finding the same-name points of the reference feature points in the point clouds of different periods, i.e., matching the reference feature points; establishing the corresponding relationship between the point clouds of different periods through the geometric features and radiation features of the points in the point clouds; the three-dimensional point feature descriptors used include but are not limited to the fast point feature histogram; by calculating the Euclidean distance between the reference feature points of different periods, finding the closest reference feature point pairs of different periods as matching point pairs;

[0101] Step S403: Calculate the three-dimensional displacement vector of each reference feature point based on the matching point pair. , the calculation formula is as follows: ,in and The reference feature points exist and The three-dimensional coordinates at a certain moment; the surface deformation data includes but is not limited to the three-dimensional displacement vector.

[0102] Preferably, the three-dimensional displacement vector field of the mountain surface in the target area is obtained by combining the ground deformation detection algorithm in step S4, specifically:

[0103] Step S404: the surface deformation detection algorithm includes but is not limited to the spatial interpolation method. The three-dimensional displacement vector of the known points in the neighborhood is used to treat the interpolation point Establish a linear relationship between displacement components and spatial coordinates and estimate The three-dimensional displacement vector of the point; the spatial interpolation method includes but is not limited to Kriging interpolation and radial basis function (RBF) interpolation;

[0104] Specifically, let the interpolation point In the neighborhood of Known points , and the displacement components are , then the interpolation point exist Displacement component in the direction It can be estimated as: ,in, is the interpolation coefficient, and The displacement components in the direction can be estimated in the same way;

[0105] It should be noted that, since the benchmark feature points are discretely distributed, in order to obtain a complete three-dimensional displacement vector field, it is necessary to perform three-dimensional scattered point spatial interpolation on all points in the mountain point cloud data; through spatial interpolation, the three-dimensional displacement vector field of the dense point cloud on the mountain surface can be obtained.

[0106] Preferably, the calculating of the surface deformation rate in step S4 includes:

[0107] Step S405: decompose the three-dimensional displacement vector field into horizontal displacement components and vertical displacement components. For each point in the point cloud , its horizontal displacement and vertical displacement It can be expressed as: ;

[0108] Step S406: After calculating the horizontal displacement component and the vertical displacement component of each point, based on the ground object classification result, the surface deformation rates of different ground object areas are counted, including but not limited to the average displacement rate and the average displacement direction, specifically:

[0109] Step S4061: Set K feature area The mountain point cloud data includes points, Points in arrive The horizontal displacement during the time period is , the vertical displacement is , then the K feature area The average horizontal displacement rate and the average vertical displacement rate It can be calculated as: ,in is the time interval;

[0110] Step S4062: the K feature area The average displacement direction is obtained by calculating the main direction of the displacement vector, that is, performing principal component analysis on the three-dimensional displacement vectors of all points, and taking the direction of the first principal component as the average displacement direction;

[0111] Preferably, the step S4 of determining the stability of the mountain in the target area is specifically as follows:

[0112] Step S407, by setting a reasonable displacement rate threshold, determine the stability of different areas; illustratively, the displacement rate threshold is set to 5 cm / month. If the average displacement rate of the K area is greater than 5 cm / month, the K area can be determined to be an unstable deformation area, and its deformation trend needs to be paid special attention to; the selection of the displacement rate threshold needs to comprehensively consider factors such as geological conditions, monitoring accuracy, and deformation level. The specific threshold provided in this embodiment is only for reference and can be adjusted according to actual conditions in specific applications.

[0113] Preferably, the monitoring data in step S5 includes but is not limited to: the surface deformation rate calculated based on step S406, the ground feature classification result and geological structure data;

[0114] Specifically, the object classification result gives the object category of each point in the form of point cloud attributes, including but not limited to bare soil, vegetation, rock, and road; by calculating the proportion of each object category point cloud, the distribution of different objects in the landslide area is reflected; for example, the total number of point clouds in the mountain point cloud data is , category is The number of point clouds is The proportion of this category for: ;

[0115] Specifically, the geological structure data is obtained by collecting regional geological survey data, including but not limited to the spatial distribution information of faults, weak interlayers and other unfavorable geological structures; since the unfavorable geological structures are represented in vector form (such as lines and surfaces), in order to be included in the model analysis and prediction, the geological structure data is rasterized, and the target area is divided into a number of prediction units according to a preset size, and the distance from each prediction unit (such as a 10m×10m grid) to the nearest geological structure is calculated; for example, the prediction unit The center coordinates of , to The shortest distance between the geological structure lines is ,but Geological structure distance characteristics It can be expressed as: ,in is the total number of geological structure lines in the area; The smaller it is, the closer the predicted unit is to the unfavorable geological structure and the greater the impact on the landslide.

[0116] Preferably, the construction of the landslide early warning model in step S5 is specifically as follows:

[0117] Step S501, prepare the model training set: prepare positive samples and negative samples, each sample corresponds to a prediction unit, each sample contains the monitoring data of each period in the landslide development process, extract the average horizontal displacement rate of each period in the sample and the average vertical displacement rate , the ground feature classification results, combined with the geological structure data, constitute a complete training sample set as the training set of the landslide early warning model; the positive sample is a typical landslide body in a high-incidence area of ​​landslides, and the negative sample is a slope with good overall stability; specifically, to ensure the training effect, the number of positive samples and negative samples should be roughly balanced, and each should be no less than 5;

[0118] Step S502, feature selection and optimization: adopting a feature selection method to screen out key features in the training sample set that are highly correlated with the occurrence of landslides, train the landslide early warning model based on the key features, and calculate the importance score of each key feature, remove the feature with the lowest score, repeat the above steps based on a new training sample set until the key feature number requirement is met, output the key feature importance ranking, and select the top-ranked key feature subset for the final landslide early warning model training;

[0119] Step S503, constructing a model: Considering the limited number of training samples and the high feature dimension, this embodiment uses a support vector machine (SVM) machine learning algorithm to construct the landslide warning model, which is expressed as:

[0120]

[0121] in The input feature vector contains the features of the prediction unit: x = [ v ¯ h,k , v ¯ v,k , r 1 , r 2 , ..., r m , D] in and are the average horizontal displacement rate and the average vertical displacement rate, For this prediction unit The proportion of landform types, is the distance from the prediction unit to the nearest geological structure, , is the category label of the sample, represents the stable region, Indicates the landslide area. is the Lagrange multiplier, is the kernel function, is the bias term; model training is to solve and , making the classification accuracy of training samples the highest.

[0122] Step S504, evaluate model performance: adopt the cross-validation method to divide, train and test the training sample set for multiple times; specifically, adopt 5 times of 10-fold cross-validation, randomly divide the training sample set into 10 parts, select 9 parts as training sets each time, and the remaining 1 part as test set, repeat 10 times, and take the average index of 5 times of 10-fold validation as the model performance evaluation result; model evaluation is based on the accuracy evaluation index (Accuracy), and when calculating the accuracy, the result of the test set obtained in the cross-validation is used: ,in is the true number of cases, is the number of true negative examples, is the number of false positive cases, is the number of false negative cases.

[0123] Preferably, the step S5 predicts the area where landslide may occur and provides early warning information, specifically:

[0124] Step S505: Use a gridding strategy to implement regionalized early warning, that is, divide the target area into a number of prediction units according to a preset size, the preset size is not limited to 10 meters × 10 meters, for each prediction unit: based on the average horizontal displacement rate and the average vertical displacement rate As the displacement characteristics of the prediction unit; calculate the proportion of various types of landforms in the prediction unit; extract the distance from the prediction unit to the nearest geological structure. Input the above characteristics into the landslide warning model to obtain the landslide occurrence probability of the prediction unit; summarize the landslide occurrence probability results of all prediction units to generate a landslide hazard zoning map of the target area, which intuitively expresses the spatial distribution of landslides;

[0125] Specifically, a time range may be set for the probability of landslide occurrence, and the set probability of landslide occurrence includes but is not limited to: the probability of landslide occurrence within the next 24 hours, the probability of landslide occurrence within the next 3 days;

[0126] Step S506: set an early warning threshold for the probability of landslide occurrence, set a quantitative landslide risk level, and monitor the prediction unit; specifically, this embodiment refers to the statistical analysis results of existing landslide cases and preliminarily sets two levels of early warning thresholds: when the probability of landslide occurrence of the prediction unit is greater than 50%, an orange early warning is issued; when it is greater than 75%, a red early warning is issued;

[0127] Specifically, an orange warning indicates that a landslide is more likely to occur in the area, and landslide monitoring efforts need to be strengthened; a red warning indicates that a landslide is very likely to occur, and the emergency plan should be activated;

[0128] Step S507: When the target area has the landslide risk level, timely release warning information to prompt relevant departments and the public to take preventive measures; the ways of releasing the warning information include but are not limited to: visually marking the spatial location and scope of the warning area on the landslide hazard zoning map; automatically generating a warning briefing, in which the system describes the basic characteristics, risk level, possible impact range, etc. of the warning area, and reports it to the relevant management and decision-making departments; issuing warning information to residents near the warning area through various channels such as text messages, television, radio, and the Internet, prompting them to evacuate the danger zone and take other response measures.

[0129] In this embodiment, the target area is scanned on the spot, the original point cloud data is collected, and the data is preprocessed to obtain the mountain point cloud data, the mountain point cloud data and the corresponding feature vector are semantically divided, the high-level semantic features of the point cloud are extracted, and the ground object classification result is obtained. By analyzing the mountain point cloud data at different periods, the surface displacement data of the target area is obtained, and the three-dimensional displacement vector field of the surface of the mountain in the target area is obtained, the surface deformation rate is calculated, the stability of the mountain in the target area is judged, the mountain point cloud data is predicted for landslides, and early warning information is provided. Through the multi-scale feature extraction structure, local and global features are integrated, and surface deformation data is extracted based on the feature points of normal vectors and curvature, so as to realize dynamic monitoring of mountain surface changes and prevent the occurrence of landslides. In terms of monitoring accuracy, the point cloud registration accuracy has been improved to the centimeter level, and the deformation monitoring accuracy is better than 1cm, which is an order of magnitude higher than the existing technology; in terms of processing efficiency, the speed of ground object classification has been increased by 2-3 times, and the early warning analysis time has been shortened by 50%; in terms of early warning capability, reliable identification and rapid response of potential landslide risks have been achieved; in terms of system performance, it supports efficient management and elastic computing of massive data, providing a new technical path for landslide monitoring in complex scenarios.

[0130] Example 2

[0131] As attached Figure 2 As shown, Embodiment 2 of the present invention provides a landslide monitoring and early warning system for complex environments, which adopts a layered architecture design and consists of four layers: data collection, data processing, data storage, and business application. The layers are decoupled and interoperable through standard interfaces, as follows:

[0132] Data collection layer: responsible for scanning the target area for landslide monitoring using professional measurement equipment, and collecting original point cloud data reflecting the surface information of the target area;

[0133] Specifically, the data acquisition layer is responsible for the acquisition and aggregation of field point cloud data, which is achieved by using professional measurement equipment in conjunction with a local area network. The professional measurement equipment includes a laser radar scanner, the scanning frequency of which is not less than 50 Hz, and the ranging accuracy is better than 1 cm. The original point cloud data is uploaded to the data storage layer in real time via Wi-Fi or 4G network, and the data transmission adopts a multi-path, multi-mode network redundancy design;

[0134] Data processing layer: responsible for data preprocessing of the original point cloud data, including but not limited to denoising, splicing, and registration, to obtain mountain point cloud data; semantically dividing the mountain point cloud data and the corresponding feature vectors based on different surface coverage types, and automatically extracting high-level semantic features of the point cloud, completing the classification of objects in the mountain point cloud data, and obtaining the object classification results; by analyzing the mountain point cloud data of different periods, obtaining the surface displacement data of the target area, combining with the surface deformation detection algorithm, obtaining the three-dimensional displacement vector field of the surface of the mountain in the target area, calculating the surface deformation rate based on the three-dimensional displacement vector field, and judging the stability of the mountain in the target area based on the surface deformation rate, completing the surface deformation detection;

[0135] Exemplarily, the data processing layer is responsible for tasks such as decoding, preprocessing, object classification and deformation detection of raw point cloud data. The data processing layer includes at least one workstation to form a workstation cluster for preprocessing the raw point cloud data. The workstation subscribes to new point cloud data from the data center through a message queue, performs parallel processing, and generates structured data products, including but not limited to point cloud classification results and deformation quantification indicators, which are then stored in the master control server.

[0136] Data storage layer: responsible for storing and managing all point cloud data and processing result data;

[0137] Specifically, the data storage layer is built based on the master control server, and the master control server adopts a RAID 5 mode hard disk array to ensure the security of data storage. The master control server is configured with at least: Intel Xeon 16-core CPU, 128GB memory and 20TB hard disk array. After unified standardization, the data storage layer stores data in the form of files and databases, and adopts a hot backup mechanism to regularly synchronize to an off-site backup center; a GIS database is used to uniformly store and manage all original data and processing results; in order to improve the scalability and service quality of the system, the database in this embodiment adopts a Netflix-like microservice architecture, and different data services are split into multiple loosely coupled microservices according to business needs, using RESTful API communication, and the deployment of microservices adopts the current mainstream Docker containerization technology to achieve rapid deployment, elastic expansion and fault isolation.

[0138] Business application layer: As a user-oriented function and service integration layer, its business includes but is not limited to building a landslide warning model based on long-term monitoring data and machine learning algorithms, predicting landslides based on the mountain point cloud data through the landslide warning model, comprehensively analyzing the displacement rate, ground object distribution and geological structure characteristics of the mountain, predicting areas where landslides may occur, and providing warning information;

[0139] Specifically, in the business application layer, the landslide warning model is deployed on the master control server as a core business and implemented through a background scheduled task; the scheduled task is responsible for extracting the latest monitoring data and processing results, inputting them into the warning model for analysis, evaluating regional landslide risks, generating landslide hazard zoning maps and warning briefings and other warning information, and publishing them through the WebGIS platform for timely access by managers and the public.

[0140] Specifically, to ensure the security of the system, this embodiment performs necessary isolation and reinforcement in terms of network configuration; independent firewalls are set up for the data acquisition network, data processing network, and business access network, and network isolation is performed using measures such as host whitelists and application access to prevent cross-segment access; servers and workstations are deployed with security protection software, system patches are updated in a timely manner, and complex password policies are adopted; database access adopts strict user authority control, and sensitive data must be stored in encrypted form; system logs are collected and monitored in a unified manner in real time, and alarms are issued in a timely manner when abnormalities are found.

[0141] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. This application is intended to cover any modification, use or adaptation of the present invention, which follows the general principles of the present invention and includes common knowledge or customary techniques in the art that are not disclosed by the present invention. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present invention are indicated by the claims. It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. A landslide monitoring and early warning method for complex environments, characterized in that: include: Use professional measurement equipment to conduct field scanning of the target area that needs landslide monitoring, collect original point cloud data reflecting the surface information of the target area, and perform data preprocessing to obtain mountain point cloud data; Based on different surface coverage types, the mountain point cloud data and corresponding feature vectors are semantically divided, and high-level semantic features of the point cloud are automatically extracted to complete the classification of objects in the mountain point cloud data and obtain the object classification results; The feature vector represents the geometric properties and radiation properties of the point cloud in the mountain point cloud data, and is composed of two parts: geometric features and radiation features; The geometric features describe the local spatial shape of the point cloud, including normal vectors, curvature, and roughness; By analyzing the mountain point cloud data at different periods, the surface displacement data of the target area is obtained, and the three-dimensional displacement vector field of the surface of the mountain in the target area is obtained by combining the surface deformation detection algorithm. Based on the three-dimensional displacement vector field, the surface deformation rate is calculated, and the stability of the mountain in the target area is judged based on the surface deformation rate to complete the surface deformation detection; Based on the surface deformation rate, the ground feature classification results and the geological structure data, a landslide warning model is constructed to predict landslides based on the mountain point cloud data and provide warning information; The surface deformation rate includes an average displacement rate and an average displacement direction. The acquisition of the average displacement rate includes: Land feature area The mountain point cloud data includes points, Points in arrive The horizontal displacement during the time period is , the vertical displacement is , then the feature area The average horizontal displacement rate and the average vertical displacement rate Calculated as: ,in and For each point The horizontal displacement in the xy plane and the vertical displacement in the z-axis direction are: is the time interval.

2. A landslide monitoring and early warning method for complex environments according to claim 1, characterized in that: The calculation of the surface deformation rate is to decompose the three-dimensional displacement vector field into horizontal displacement components and vertical displacement components. For each point in the point cloud , its horizontal displacement and vertical displacement It can be expressed as: ,in , and For the separation point Displacement on the x, y, and z axes; After calculating the horizontal displacement component and the vertical displacement component of each point, the surface deformation rates of different ground object areas are counted based on the ground object classification results; A principal component analysis is performed on the three-dimensional displacement vectors of all points, and the direction of the first principal component is taken as the average displacement direction.

3. A landslide monitoring and early warning method for complex environments according to claim 1, characterized in that: The mountain point cloud data is The form of coordinates represents the spatial position of each point; The radiation characteristics are derived from the echo signal of the laser radar, including reflection intensity and echo number; The geometric features are as follows: The normal vector reflects the orientation of the local surface and is estimated by principal component analysis or least squares plane fitting method; The curvature reflects the curvature of the local surface, and the calculation formula is as follows: ,in is the coefficient of the quadratic term of the fitted parabola; The roughness reflects the fluctuation of the local surface, and the calculation formula is as follows: ,in represents the number of points in the neighborhood, that is, the total number of local points used for calculation, Indicate point The attribute value of point The height relative to a reference plane or reference point, Represents the average of all point attribute values ​​in the neighborhood.

4. The landslide monitoring and early warning method for complex environments according to claim 1, characterized in that: The automatic extraction of high-level semantic features of point clouds uses the point cloud deep learning framework PointNet++ as the backbone network, and simultaneously extracts local features and global features of the point cloud geometric features in the mountain point cloud data, specifically including: Local feature extraction: Using multi-scale group convolution, convolution operations are performed on point clouds at different neighborhood scales to obtain multi-scale local descriptors; Global feature extraction: Through the maximum pooling layer, the multi-scale local descriptors of different areas in the point cloud are aggregated into an overall global descriptor as the high-level semantic feature.

5. A landslide monitoring and early warning method for complex environments according to claim 4, characterized in that: The local feature extraction specifically includes: Select a set of key points in the point cloud by sampling the farthest point , with each As the center point, select the radius Neighborhood ball , each field ball The points in the group form a group, and point-by-point convolution is performed in each group. Geometric features and the characteristics of other points in the group Cascade and send to a multi-layer perceptron (MLP) to learn local features. The expression is: ,in Indicate point The local descriptor of Indicated by point is the center and the radius is The set of points in the neighborhood ball of It is the maximum pooling operation.

6. The landslide monitoring and early warning method for complex environments according to claim 1, characterized in that: The acquisition of the surface displacement data of the target area is specifically as follows: Selecting a series of benchmark feature points in the point cloud of the mountain point cloud data as benchmarks for surface deformation analysis, wherein the benchmark feature points are selected in a stable surface area; Find the same-name points of the reference feature points in the point clouds of different periods, establish the corresponding relationship between the point clouds of different periods through the geometric features and radiation features of the points in the point clouds, and find the closest reference feature point pairs in different periods as matching point pairs by calculating the Euclidean distance between the reference feature points of different periods; Based on the matching point pairs, the three-dimensional displacement vector of each reference feature point is calculated. , the calculation formula is as follows: ,in and The reference feature points exist and The three-dimensional coordinates of the moment.

7. The landslide monitoring and early warning method for complex environments according to claim 1, characterized in that: The professional measuring equipment is arranged at a plurality of measuring sites, which are distributed in a ring around the target area, and the straight-line distance between two adjacent measuring sites is 1 / 3 to 1 / 2 of the effective ranging of the professional measuring equipment; the professional measuring equipment also includes a satellite navigation system and a measuring robot to assist in achieving high-precision three-dimensional positioning of the measuring sites.

8. The landslide monitoring and early warning method for complex environments according to claim 1, characterized in that: The construction of the landslide early warning model comprises: Prepare positive samples and negative samples, each sample corresponds to a prediction unit, extract the horizontal displacement rate and the average vertical displacement rate and the ground feature classification results of each period in the sample, and combine them with the geological structure data to form a complete training sample set as the training set of the landslide early warning model; Using a feature selection method, the key features in the training sample set that are highly correlated with the occurrence of landslides are screened out, the landslide early warning model is trained based on the key features, the key feature importance ranking is output, and the top-ranked key feature subset is selected for the final landslide early warning model training; The landslide early warning model is constructed based on the machine learning algorithm of support vector machine, which is expressed as: ,in The input feature vector contains the features of the prediction unit: in and are the average horizontal displacement rate and the average vertical displacement rate, For this prediction unit The proportion of landform types, is the distance from the prediction unit to the nearest geological structure, , is the category label of the sample, represents the stable region, Indicates the landslide area. is the Lagrange multiplier, is the kernel function, is the bias term; The performance of the landslide early warning model was evaluated.

9. The landslide monitoring and early warning method for complex environments according to claim 1, characterized in that: The landslide prediction and early warning information provided include: The target area is divided into a number of prediction units according to a preset size, and the average horizontal displacement rate is and the average vertical displacement rate As the displacement feature of the prediction unit, the proportion of each type of ground objects in the prediction unit is calculated, and the distance from the prediction unit to the nearest geological structure is extracted; The above displacement characteristics, the proportion of various types of landforms and the distance are input into the landslide early warning model to obtain the probability of landslide occurrence in each prediction unit; the landslide probability results of all prediction units are summarized to generate a landslide hazard zoning map of the target area to intuitively express the spatial distribution of landslides; Set warning thresholds for landslide probability, set quantitative landslide risk levels, and monitor the prediction units; When the landslide risk level occurs in the target area, early warning information is issued in a timely manner.

10. A landslide monitoring and early warning system for complex environments, used to execute a landslide monitoring and early warning method for complex environments as claimed in any one of claims 1 to 9, characterized in that: include: Data collection layer: responsible for scanning the target area for landslide monitoring using professional measurement equipment, and collecting original point cloud data reflecting the surface information of the target area; Data processing layer: responsible for data preprocessing of the original point cloud data, including denoising, splicing, and registration, to obtain mountain point cloud data; semantically dividing the mountain point cloud data and the corresponding feature vectors based on different surface coverage types, and automatically extracting high-level semantic features of the point cloud, completing the classification of objects in the mountain point cloud data, and obtaining the object classification results; by analyzing the mountain point cloud data of different periods, obtaining the surface displacement data of the target area, combining with the surface deformation detection algorithm, obtaining the three-dimensional displacement vector field of the surface of the mountain in the target area, calculating the surface deformation rate based on the three-dimensional displacement vector field, and judging the stability of the mountain in the target area based on the surface deformation rate, completing the surface deformation detection; Data storage layer: responsible for storing and managing all point cloud data and processing result data; Business application layer: As a user-oriented function and service integration layer, it is responsible for the following businesses: building a landslide warning model based on long-term monitoring data and machine learning algorithms, predicting landslides based on the mountain point cloud data through the landslide warning model, comprehensively analyzing the displacement rate, landform distribution and geological structure characteristics of the mountain, predicting areas where landslides may occur, and providing warning information.

Images