A method for identifying loess sinkholes by using lidar point cloud data

Abstract

The application discloses a method for recognizing loess sinkholes by using laser radar point cloud data, and relates to the technical field of sinkhole recognition, and comprises the following steps: collecting laser radar point cloud data of a region to be detected; pre-processing the laser radar point cloud data, and marking a cave area; constructing an improved PoinNet++ model, inputting the laser radar point cloud data of the marked cave area into the improved PoinNet++ model for recognition, and extracting a rough loess sinkhole area; and finely adjusting the rough loess sinkhole by using a region growing algorithm. The application can automatically learn distinguishing features based on the technology of laser point cloud, so that the sinkholes can be more accurately recognized. Meanwhile, the application technology based on deep learning can adaptively learn features in different environments, so that the application technology can better cope with complex terrains and environmental changes, and has better robustness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of sinkhole recognition, and particularly relates to a method for recognizing loess sinkholes by using laser radar point cloud data. BACKGROUND

[0002] Soil pipe erosion is considered to be a widespread and important process, but the research on underground soil erosion is far from that on surface soil erosion in terms of width, breadth and depth. The loess cave in China is rooted in the world's most typical loess stratum and loess landform, and the Loess Plateau is one of the regions with the most frequent geological disasters and the most serious soil erosion in the world, which has many distinctive and unique regional characteristics and features. Sinkholes are widely distributed in the Loess Plateau, and these caves have the characteristics of harmfulness, concealment, unpredictability and surface response, which has brought great hidden dangers and inconvenience to transportation, agriculture, urban construction and other undertakings. Therefore, the detection and recognition of loess sinkholes have become an important issue that the departments of transportation, land, water conservancy, mineral resources and urban construction in the Loess Plateau region are very concerned about.

[0003] The development of loess cave sensitivity assessment and mapping can better serve the disaster prevention and mitigation of loess cave geological disasters. The distribution and extension of loess caves in underground space are extremely complex, and different caves are interconnected, connected with each other, and even form a complex underground cave system on slopes and in channels. The complexity, concealment and connectivity of loess cave system determine that it is difficult to understand their three-dimensional morphological structure in underground space using conventional remote sensing and unmanned aerial photogrammetry.

[0004] There are still some problems and deficiencies in the existing loess sinkhole detection methods. Specifically, the main problems are as follows:

[0005] Large amount of manpower and low efficiency. The traditional loess sinkhole detection method often relies on manual driving routine patrol and walking exploration, which consumes a large amount of manpower and material resources. Moreover, due to limited visibility and human subjective factors, the efficiency and accuracy need to be improved, and the data quality is difficult to guarantee. The traditional loess sinkhole detection method has the problems of low sampling point density, complex terrain, wide range, and discontinuous sampling, which are not comprehensive. These problems have caused great difficulties in the recognition and positioning of loess sinkholes, resulting in different analysis accuracy and effect. The traditional loess sinkhole detection method also relies on Yang's variance analysis, support vector machine and other statistical methods and machine learning methods, and the analysis accuracy and effect are affected by many factors, which cannot achieve ideal effect. Moreover, the traditional image-based loess sinkhole recognition algorithm model is large in size and complex in training. SUMMARY

[0006] The present application aims at the deficiencies of the prior art, and provides a method for identifying loess sinkholes by using lidar point cloud data, so as to solve the problems of low detection and identification efficiency of sinkholes and artificial errors and missed detection in the prior art.

[0007] The present application specifically provides the following technical scheme: a method for identifying loess sinkholes by using lidar point cloud data, comprising the following steps:

[0008] Collecting lidar point cloud data of a region to be detected;

[0009] Pretreating the lidar point cloud data and marking out a cave region;

[0010] Constructing an improved PoinNet++ model: adding a Transformer feature extraction module to the PoinNet++ model;

[0011] Inputting the lidar point cloud data of the marked-out cave region into the improved PoinNet++ model for identification, and extracting a preliminary loess sinkhole region;

[0012] Using a region growing algorithm to take all the preliminary loess sinkhole regions as seed points, and defining a neighborhood for each seed point;

[0013] Regarding each seed point as an anchor node, and regarding k seed points in the neighborhood of each seed point as neighbor nodes, judging the normal vector from each neighbor node to the anchor node, and if the normal vector is greater than a threshold value, the seed point is added to the preliminary loess sinkhole region;

[0014] Merging the neighborhood containing multiple seed points into a new region by setting rules, and obtaining a final loess sinkhole region through the new region; wherein the final loess sinkhole region has a higher degree of detail than the preliminary loess sinkhole region.

[0015] Preferably, the collecting of the lidar point cloud data of the region to be detected comprises the following steps:

[0016] Obtaining ground point cloud data by using a low-altitude airborne lidar unmanned aerial vehicle;

[0017] Obtaining internal point cloud by using a handheld lidar device to scan around the top of the cave;

[0018] For the blind area of the cave and other devices, a lidar unmanned aerial vehicle is flown into the interior of the cave to collect point cloud data;

[0019] Merging the ground point cloud data, the internal point cloud data of the cave, and the point cloud data of the blind area of measurement to obtain final lidar point cloud data.

[0020] Preferably, the ground point cloud data, the internal point cloud data of the cave and the point cloud data of the measurement blind area are merged to obtain the final lidar point cloud data, including the following steps:

[0021] The point cloud data obtained by the low-altitude airborne radar unmanned device with the highest measurement accuracy is taken as the registration reference data;

[0022] The point cloud data obtained by the handheld lidar device and the lidar unmanned device are registered by using an iterative closest point fine registration algorithm.

[0023] Preferably, the lidar point cloud data is preprocessed to mark the cave area, including the following steps:

[0024] A point cloud filtering algorithm is used to remove noise in the point cloud data;

[0025] The denoised point cloud data is segmented to mark the sinkhole area and the ground area respectively, and then the point cloud data is merged;

[0026] The point cloud data is divided into a training set, a test set and a validation set in a ratio of 7:1:2.

[0027] Preferably, the PoinNet++ model is constructed, including the following steps:

[0028] A hierarchical sampling module and a hierarchical aggregation module are added to the PointNet model to construct the PoinNet++ model;

[0029] The hierarchical sampling module uses the farthest point sampling algorithm to uniformly downsample the point cloud at each level;

[0030] The hierarchical aggregation module uses the nearest neighbor algorithm or the spherical query method to aggregate the neighborhood nodes of a single anchor node, and the sampling results below the threshold level are used as the aggregation input above the threshold level.

[0031] Preferably, the lidar point cloud data marked with the cave area is input into the PoinNet++ model for identification, including the following steps:

[0032] The lidar point cloud data is taken as the input, and a multilayer perceptron is used to increase the dimension of the input point cloud coordinate data;

[0033] The lidar point cloud data after dimension increase is input into the hierarchical sampling module, and the farthest point sampling algorithm is used to uniformly downsample the point cloud at each level to reduce the number of point clouds;

[0034] The laser radar point cloud data with reduced point cloud quantity is input into the hierarchical aggregation module, and K nearest neighbor algorithm or spherical query is used to aggregate k neighbor nodes around each point, and a Transformer feature extraction module is used to extract global features and local features of the k neighbor nodes;

[0035] The point cloud quantity after global feature extraction and local feature extraction is gradually increased to the original size by using up-sampling interpolation, and the features are updated by using the k neighbor nodes.

[0036] Preferably, the global feature extraction and local feature extraction of the k neighbor nodes by using the Transformer feature extraction module comprises the following steps:

[0037] The point cloud Ψ is sequentially sorted in three dimensions (x, y, z) to obtain a sub-point cloud X F ;

[0038] The global feature extraction module GT converts the sub-point cloud X F into Q, K, and V, all of which are in R G×MD , to generate a global attention matrix;

[0039]

[0040] wherein Q, K, and V represent query, key, and value vectors, respectively, R G×MD is a three-dimensional real space, γ is a mapping function, σ is a position encoding function, is a same or symbol, L is the distance between sub-point clouds, R G×MD is a three-dimensional real space, and d k is a constant for limiting the Softmax function within a gradient region;

[0041] The global attention matrix attn G is multiplied by V to obtain global features;

[0042] The local feature extraction module LT is used to calculate the distance d between the anchor node and its k neighbor nodes, and a local attention matrix α i is generated;

[0043] α i = 1-softmax(LN(d))

[0044] wherein LN represents layer normalization;

[0045] The local attention matrix is assigned to the neighbor nodes f i of the anchor node in the form of weight to obtain local features F;

[0046]

[0047] Among them, f i These are the eigenvalues ​​of the corresponding points.

[0048] Preferably, the point cloud Ψ is sorted sequentially along three dimensions (x, y, z) to obtain the sub-point cloud X. F It includes the following steps:

[0049] Sort by the first dimension to obtain the index. x ;

[0050] Using Index x Obtain the sorted coordinates Ψ i and corresponding features

[0051] coordinates Ψ i and corresponding features The point cloud is divided into G sub-point clouds, each containing N / G points;

[0052] The second and third dimensions are used to obtain the G group of sub-point clouds in the same way;

[0053] In obtaining the G-group sub-point cloud X with three dimensions F Then, the sub-point cloud X is calculated. F The distance L between them;

[0054]

[0055] Where r is the centroid of each sub-point cloud, defined as follows:

[0056]

[0057] Where g∈{1,...,G},k∈{1,2,3},M is the number of parameters, and i,j∈g.

[0058] Preferably, the feature update using k neighbor nodes is specifically expressed as follows:

[0059]

[0060] Where d(x,x) i f is the distance between two points. i These are the eigenvalues ​​of the corresponding points.

[0061] Preferably, before inputting the lidar point cloud data marking the cave area into the PoinNet++ model, the PoinNet++ model needs to be trained, including the following steps:

[0062] 16,000 points were randomly collected from the preprocessed point cloud for training;

[0063] The point cloud is subjected to data enhancement in a random flipping, random rotation around the Z axis and random scaling manner.

[0064] When the loss of the verification set is the smallest, the PoinNet++ model weight at this moment is saved as the optimal weight.

[0065] Compared with the prior art, the present application has the following remarkable advantages:

[0066] The present application acquires point cloud through a laser radar scanner, extracts point cloud features by using an improved PointNet++ of a deep learning algorithm, realizes preliminary identification of loess sinkholes, and realizes fine segmentation of loess sinkholes by using a region growing algorithm. The present application can automatically learn distinguishing features based on laser point cloud technology, so that the sinkholes can be more accurately identified. The present application has a high sampling point density, can continuously collect data on a complex loess land surface, and accurately improves the identification and positioning of loess sinkholes. Meanwhile, the present application can adaptively learn features in different environments, so that it can better cope with complex terrain and environmental changes, improve the fine acquisition of loess sinkholes, and has better robustness. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 A flow chart of a method for automatically identifying loess sinkholes by using laser radar point cloud data is provided in the present application.

[0068] Figure 2 An improved PointNet++ model structure diagram is provided in the present application. DETAILED DESCRIPTION

[0069] The technical solutions of the embodiments of the present application will be described clearly and completely in combination with the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the protection scope of the present application.

[0070] The present application aims to provide a sinkhole identification algorithm based on an improved PointNet++, to further improve the cave detection accuracy, and to be used for loess cave three-dimensional structure, hydrogeological structure detection and modeling.

[0071] In order to facilitate understanding and illustration, as shown in the accompanying drawings, Figure 1 The present application provides a method for identifying loess sinkholes by using laser radar point cloud data, which comprises the following steps:

[0072] Step S1: determining a loess plateau sinkhole risk area, and collecting laser radar point cloud data of a to-be-detected area by using a laser radar device.

[0073] The low-altitude airborne laser radar unmanned aerial vehicle, the handheld laser radar device and the anti-collision small laser radar unmanned aerial vehicle are respectively used for collecting terrain point cloud data in low altitude, on the ground and in the cave. Specifically:

[0074] The low-altitude airborne laser radar unmanned aerial vehicle is used for collecting large-scale ground point cloud data; the handheld laser radar device is used for collecting internal point cloud by scanning around the cave roof; the small laser radar unmanned aerial vehicle is used for collecting point cloud data in the cave for narrow caves and other blind areas.

[0075] The large-scale ground point cloud data, the internal point cloud data of the cave and the point cloud data of the blind area are merged to obtain the final laser radar point cloud data.

[0076] Due to the influence of the measurement coordinate system and the measurement error, the laser point clouds obtained by different laser radar devices lack a unified reference. If the point clouds are simply merged, problems such as point cloud layering, crossing and rotation may occur. In order to solve this problem, in the CloudCompare software, the point cloud data obtained by the low-altitude airborne radar unmanned aerial vehicle device with the highest measurement accuracy is used as the registration reference data, and the point cloud data obtained by the handheld laser radar device and the anti-collision small laser radar unmanned aerial vehicle are respectively registered with the reference data by using the built-in iterative closest point (ICP, Iterative Closest Point) fine registration algorithm.

[0077] The large-scale ground point cloud data, the internal point cloud data of the cave and the point cloud data of the blind area are merged to obtain the final laser radar point cloud data, including the following steps:

[0078] The point cloud data obtained by the low-altitude airborne radar unmanned aerial vehicle device with the highest measurement accuracy is used as the registration reference data.

[0079] The point cloud data obtained by the handheld laser radar device and the anti-collision small laser radar unmanned aerial vehicle are registered by using the iterative closest point fine registration algorithm.

[0080] Step S2: Preprocess the laser radar point cloud data (cleaning) to mark out the cave area.

[0081] A point cloud filtering algorithm is used to remove noise in the point cloud data; the original collected dense point cloud data temporarily uses a large amount of computing resources, and the point cloud is operated, thinned, filtered and the like to obtain terrain point cloud after removing noise.

[0082] The denoised point cloud data is segmented by using the CloudCompare software, the falling water hole area and the ground area are respectively marked out, and then the point cloud data is merged.

[0083] The point cloud data is divided into a training set, a test set, and a verification set in a ratio of 7:1:2.

[0084] Step S3: Construct a PoinNet++ model, and introduce a Transformer feature extraction module based on the PoinNet++ model.

[0085] The PointNet module is a basic module for feature extraction of point clouds. It uses a multilayer perceptron (MLP) network to extract features for each point in the point cloud and calculate its global feature vector. These global feature vectors are aggregated to form a representation of the point cloud, which can be used for tasks such as classification and semantic segmentation.

[0086] The PoinNet++ model is an enhanced PointNet-based point cloud hierarchical clustering model that generates more rich feature representations by progressively aggregating features of point sets with different sizes and resolutions. The PoinNet++ model is constructed by adding a hierarchical sampling module and a hierarchical aggregation module to the PointNet model. The hierarchical sampling module uses the farthest point sampling algorithm to uniformly downsample the point cloud at each level. The hierarchical aggregation module uses the nearest neighbor algorithm or spherical query to aggregate the neighborhood nodes of a single anchor node. The sampling results of lower levels (below a threshold) are used as input for higher levels (above the threshold). Points within any class of the sampling results are aggregated into a vector to reduce the number of features. Finally, the PointNet++ module provides a more rich point cloud feature representation for tasks such as semantic segmentation and object detection.

[0087] Step S4: Input the laser radar point cloud data marked with the cave area into the PoinNet++ model for recognition, and extract the preliminary (rough) loess sinkhole area.

[0088] The improved PointNet++ uses 16000 laser radar point cloud data as input and uses a multilayer perceptron to upgrade the dimension of the input point cloud coordinate data.

[0089] The upgraded laser radar point cloud data is input into the hierarchical sampling module, which uses the farthest point sampling algorithm to uniformly downsample the point cloud by a factor of N=4 at each level to reduce the number of point clouds. After one downsample, the number of point clouds becomes 4000, and after two downsamples, it becomes 1000, and so on, with a minimum of 64 points.

[0090] The laser radar point cloud data for reducing the number of point clouds is input into a hierarchical aggregation module, the aggregation module uses a K-nearest neighbor algorithm or a spherical query to aggregate k neighbor nodes around each point, and uses maximum pooling to aggregate local features after multi-layer perceptron feature transformation. A Transformer feature extraction module is used to extract global and local features from the k neighbor nodes.

[0091] The number of point clouds after global and local feature extraction is gradually increased to the original size using up-sampling interpolation, and the features are updated using the k neighbor nodes.

[0092]

[0093] where d(x,x i ) is the distance between two points, and f i is the feature value of the corresponding point.

[0094] A Transformer-based feature extraction module based on PointNet++ is adopted, which includes a global feature extraction module GT and a local feature extraction module LT. The Transformer feature extraction module is used to extract global and local features from the k neighbor nodes, including the following steps:

[0095] The point cloud Ψ is sequentially sorted in three dimensions (x, y, z) to obtain a sub-point cloud X F . The point cloud Ψ = {P (1) , P (2) , …, P (B)}, where P (j) ∈ R N×3 and represents the jth point cloud and its corresponding feature. R G×MD is a three-dimensional real space, the number of points in the original point cloud is N, and the number of sub-point clouds is G.

[0096] The global feature extraction module GT converts the sub-point cloud X F into Q, K, and V ∈ R G×MD , generating a global attention matrix.

[0097]

[0098] where Q, K, and V represent query, key, and value vectors, respectively, R G×MD is a three-dimensional real space, γ is a mapping function, σ is a position encoding function, is a XOR symbol, L is the distance between sub-point clouds, d k is a constant to limit the Softmax function in a region with larger gradient; and the global attention matrix attn GMultiply by V to obtain global features.

[0099] Although global attention also has some ability to extract local information, in order to obtain more refined local information, the local feature extraction module LT is used to calculate the distance d between the anchor node and its k neighbor nodes, generating a local attention matrix:

[0100] α i = 1 - softmax(LN(d))

[0101] Where LN represents layer normalization.

[0102] The local attention matrix is ​​assigned as weights to the neighbor nodes f of the anchor node. i To obtain local features:

[0103]

[0104] The point cloud Ψ is sorted sequentially along three dimensions (x, y, z) to obtain the sub-point cloud X. F It includes the following steps:

[0105] Sort by the first dimension to obtain the index. x .

[0106] Using Index x Obtain the sorted coordinates Ψ i and corresponding features

[0107] coordinates Ψ i and corresponding features The point cloud is divided into G sub-groups, each containing N / G points.

[0108] The second and third dimensions are used to obtain G-group sub-point clouds in the same way; G-group sub-point clouds X with three dimensions are obtained. F Then, the sub-point cloud X is calculated. F The distance L between them.

[0109]

[0110] Where r is the centroid of each sub-point cloud, defined as follows:

[0111]

[0112] Where g∈{1,...,G},k∈{1,2,3},M is the number of parameters, and i,j∈g.

[0113] Before inputting the lidar point cloud data marking the cave area into the PoinNet++ model, the PoinNet++ model needs to be trained, including the following steps:

[0114] Randomly collect 16000 points from the pretreated point cloud for training; data augmentation is performed on the point cloud in the form of random flipping, random rotation around the Z axis, and random scaling to improve the robustness and expression ability of the model.

[0115] When the loss of the validation set is the smallest, save the PoinNet++ model weight at this moment as the optimal weight.

[0116] After the data point cloud obtained by the unmanned aerial vehicle is preprocessed, it is input into the PointNet++ loaded with the optimal model weight, and the sinkhole region is preliminarily identified.

[0117] Step S5: Use the region growing algorithm to take all the preliminary loess sinkhole regions as seed points, and define a neighborhood for each seed point, i.e. all points within a certain distance from the seed point. The size of the neighborhood is determined according to factors such as point cloud density and resolution.

[0118] Step S6: Determine whether the points in the neighborhood meet the conditions: treat each seed point as an anchor node, and treat the k points in the neighborhood of each seed point as neighboring nodes. Set a threshold, and determine the normal vector from each neighboring node to the anchor node. If the normal vector is greater than the threshold, add the point to the preliminary loess sinkhole region.

[0119] Step S7: Merge regions: if there are multiple seed points in the neighborhood, merge the neighborhoods containing multiple seed points into a new region according to the set rules, and obtain the final (fine) loess sinkhole region through the new region. For example, the nearest seed point can be selected as the basis for merging, or the order of merging can be determined according to factors such as the area and density of the region.

[0120] The above is a further detailed description of the present application in combination with a specific preferred embodiment. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, which should be considered as falling within the scope of protection of the present application.

Claims

1. A method for identifying loess sinkholes using lidar point cloud data, characterized in that, Includes the following steps: Collect lidar point cloud data of the area to be detected; The lidar point cloud data is preprocessed to mark the cave area; Build a PoinNet++ model and set up the Transformer feature extraction module in the PoinNet++ model; The lidar point cloud data of the marked cave area is input into the PoinNet++ model for identification, and the preliminary loess sinkhole area is extracted. The region growing algorithm is used to take all the preliminary loess sinkhole regions as seed points, and a neighborhood is defined for each seed point. Treat each seed point as an anchor node, and consider the neighbors of each seed point as anchor nodes. k Seed points are treated as neighboring nodes, and the normal vector from each neighboring node to the anchor node is judged. If the normal vector is greater than the threshold, the seed point is added to the initial loess sinkhole area. The neighborhood containing multiple seed points is merged into a new region by setting rules, and the final loess sinkhole region is obtained through the new region; The final loess sinkhole area is more detailed than the preliminary loess sinkhole area. The lidar point cloud data marking the cave area is input into the PoinNet++ model for identification, including the following steps: Using LiDAR point cloud data as input, a multilayer perceptron is used to upgrade the dimensionality of the input point cloud coordinate data; The upgraded lidar point cloud data is input into a hierarchical sampling module, and the point cloud is downsampled uniformly by N times through the farthest point sampling algorithm to reduce the number of point clouds. The reduced point cloud data from the LiDAR system is input into the hierarchical aggregation module, where the K-nearest neighbor algorithm or spherical search is used to aggregate and find the neighboring points of each point. k The neighboring nodes were analyzed, and the Transformer feature extraction module was used to extract features from them. k Global and local features are extracted from each neighboring node; The point cloud size is gradually increased to the original size using upsampling interpolation after the extraction of global and local features. k Each neighbor node updates its characteristics; The Transformer feature extraction module is used for... k Extracting global and local features from each neighboring node, including the following steps: The point cloud Ψ is sorted sequentially along three dimensions (x, y, z) to obtain the sub-point cloud X. F ; The global feature extraction module GT transforms the sub-point cloud X using linear transformation. F Transform to Q, K, V ∈ R G×MD Generate a global attention matrix; Where Q, K, and V represent the query, key, and value vectors, respectively, and R G×MD For three-dimensional real space, For mapping functions, Here is the positional encoding function, and ⊙ is the XOR symbol. L The distance between the dots in the clouds. d k In order to put Softmax The function is a constant constrained within a region where the gradient converges. global attention matrix Multiplying by V yields global features; The anchor node and its local feature extraction module (LT) are used to calculate the local feature extraction module. k The distance d between each neighbor node is used to generate the local attention matrix. ; in, LN Representation layer normalization; The local attention matrix is ​​assigned as weights to the neighboring nodes of the anchor node. To obtain local features F ; in, For neighboring nodes.

2. The method for identifying loess sinkholes using lidar point cloud data as described in claim 1, characterized in that, The process of collecting lidar point cloud data of the area to be detected includes the following steps: Using low-altitude airborne lidar drones to acquire surface point cloud data; Use a handheld lidar device to scan around the top of the cave to obtain internal point clouds; For the blind spots in the measurement of cavities and other equipment, a lidar drone is used to fly into the cave to collect point cloud data. The surface point cloud data, the cave interior point cloud data, and the point cloud data of the measurement blind zone are merged to obtain the final lidar point cloud data.

3. The method for identifying loess sinkholes using lidar point cloud data as described in claim 2, characterized in that, The point cloud data from the surface, the interior of the cave, and the blind zone are merged to obtain the final lidar point cloud data, including the following steps: Point cloud data acquired by the low-altitude airborne radar UAV equipment with the highest measurement accuracy was used as the registration reference data; The point cloud data acquired by handheld LiDAR devices and LiDAR drones are registered using an iterative nearest-point fine registration algorithm.

4. The method for identifying loess sinkholes using lidar point cloud data as described in claim 1, characterized in that, The preprocessing of the lidar point cloud data to mark the cave area includes the following steps: Point cloud filtering algorithms are used to remove noise from point cloud data; The denoised point cloud data is segmented, and the sinkhole area and the ground area are labeled separately. Then the point cloud data are merged. The point cloud data was then divided into a training set, a test set, and a validation set in a 7:1:2 ratio.

5. The method for identifying loess sinkholes using lidar point cloud data as described in claim 4, characterized in that, The construction of the PoinNet++ model includes the following steps: The PoinNet++ model is constructed by adding a hierarchical sampling module and a hierarchical aggregation module to the PointNet model. The hierarchical sampling module uses the farthest point sampling algorithm to uniformly downsample the point cloud step by step; The hierarchical aggregation module uses the nearest neighbor algorithm or spherical query to aggregate and find the neighboring nodes of a single anchor node. The sampling results below the threshold level are used as the aggregation input above the threshold level.

6. The method for identifying loess sinkholes using lidar point cloud data as described in claim 1, characterized in that, The point cloud Ψ is sorted sequentially along three dimensions (x, y, z) to obtain the sub-point cloud X. F It includes the following steps: Sort by the first dimension to obtain the index. x ; Using Index x Obtain the sorted coordinates Ψ i and corresponding features ₣ i ; coordinates Ψ i and corresponding features ₣ i The point cloud is divided into G sub-point clouds, each containing N / G points; The second and third dimensions are used to obtain the G group of sub-point clouds in the same way; In obtaining the G-group sub-point cloud X with three dimensions F Then, the sub-point cloud X is calculated. F Distance between ; Where r is the centroid of each sub-point cloud, defined as follows: in, M is the number of parameters. i j .

7. The method for identifying loess sinkholes using lidar point cloud data as described in claim 1, characterized in that, The use of k The feature is updated for each neighboring node, and the specific expression is as follows: in The distance between the two points is denoted as .

8. The method for identifying loess sinkholes using lidar point cloud data as described in claim 1, characterized in that, Before inputting the lidar point cloud data marking the cave area into the improved PoinNet++ model, the improved PoinNet++ model needs to be trained, including the following steps: 16,000 points were randomly collected from the preprocessed point cloud for training; Point cloud data augmentation is performed using random flipping, random rotation around the Z-axis, and random scaling. Save the PoinNet++ model weights at the moment when the loss on the validation set is minimized as the optimal weights.

Images