A landslide identification method, system, computer device and medium

Abstract

The application provides a landslide identification method, system, computer equipment and medium, and belongs to the technical field of disaster identification. The method obtains two-dimensional RGB images and corresponding three-dimensional laser scanning measured point cloud images of known landslides and non-landslide slopes, and combines NYU Depth V2 data sets to construct a training data set and a verification data set. An improved encoder-decoder convolutional neural network model is used to reconstruct the point cloud images of the verification data set and verify the reconstruction effect, three-dimensional information such as the number of cracks in the slope, the length of the cracks in the slope, the relative height difference of the slope body, the length of the longest crack and the maximum crack width is extracted, an improved nonlinear NT hyperplane algorithm is used to screen landslide identification indexes, a landslide identification model is constructed, and the optimal model is selected through the verification data set, so that the landslide identification based on the two-dimensional RGB images is realized. The application reduces the complexity of landslide identification modeling and avoids falling into the 'dimension disaster'.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of disaster identification, in particular to a landslide identification method and system, computer equipment and medium. BACKGROUND

[0002] Landslide is a natural phenomenon that the rock-soil mass on the slope slides downward along a certain weak surface (weak zone) under the action of gravity. From the development history, landslide is divided into three stages: creep, sudden slip and outbreak, among which, the creep stage accounts for more than 95% of the whole life cycle of landslide. Therefore, landslide identification in the creep stage is the premise of accurate monitoring and early warning and engineering treatment, which helps to reduce economic losses caused by landslide.

[0003] The existing landslide identification technology has the following problems:

[0004] (1) The existing landslide identification based on remote sensing interpretation generally adopts object-oriented classification method and SBAS-InSAR technology, which is limited by the precision of satellite and aerial images and image processing methods, and is only suitable for regional landslide identification, and cannot realize the identification of single slope;

[0005] (2) The existing identification of single landslide only relies on traditional disaster-causing factors, such as topography, hydrogeology, land use, vegetation cover, etc., but the crack index which plays an important role in landslide development is not included in the identification system, resulting in large errors in landslide identification results;

[0006] (3) The traditional three-dimensional information acquisition means of slope mostly adopts three-dimensional laser scanner, which is complicated in scanning work and expensive in equipment, and the number of slopes to be scanned is large, so this kind of research is still in the initial stage, and the application range is greatly limited. SUMMARY

[0007] The purpose of the present application is to provide a landslide identification method, system, computer equipment and medium, which improves the accuracy of landslide identification, reduces the complexity of modeling, and avoids falling into "dimension disaster".

[0008] To achieve the above purpose, the present application provides a landslide identification method, comprising the following steps:

[0009] Step S1, obtaining the two-dimensional RGB images of known landslides and the same number of non-landslide slopes in the study area and the corresponding three-dimensional laser scanning measured point cloud images, combining the NYU Depth V2 dataset to construct a three-dimensional information extraction training dataset and a verification dataset;

[0010] Step S2, establishing an improved encoder-decoder convolutional neural network model, using the trained model to reconstruct the point cloud image of the verification dataset and verifying the reconstruction effect;

[0011] Step S3, based on Polyworks software, extracting the number of intra-slope cracks, the length of intra-slope cracks, the relative height difference of the slope body, the length of the longest crack, and the width of the largest crack of the part of the training data set and the part of the validation data set;

[0012] Step S4, using an improved nonlinear NT hyperplane algorithm to screen the three-dimensional information of the training data set and the landslide disaster-causing factors, and establishing a landslide recognition index system;

[0013] Step S5, constructing a landslide recognition training data set and a validation data set, establishing a landslide recognition model based on multiple mathematical models, selecting the optimal model through the validation data set, and realizing landslide recognition based on two-dimensional RGB images.

[0014] Preferably, step S1 comprises:

[0015] The number of landslides is known , and non-landslide slopes are randomly selected;

[0016] Obtain two-dimensional RGB images and measured point cloud images;

[0017] Crop or extend the two-dimensional RGB images of slopes and the NYU Depth V2 data set, to obtain two-dimensional RGB images;

[0018] Crop or extend the measured point cloud images to , and perform inclination correction, missing point filling, and denoising processing;

[0019] Perform augmentation operation on the two-dimensional RGB images and the point cloud images, construct groups of image pairs, and divide training samples and validation samples.

[0020] Preferably, in step S2, the improved encoder-decoder convolutional neural network model comprises:

[0021] Encoder: using an improved ResNet101 as a feature extractor, replacing the 3x3 convolution layer of the last Residual Block 4 with a dilated convolution layer with a dilation rate of 2, the output feature map resolution is increased to 1 / 16 of the input image, and the end pooling layer, the fully connected layer and the SoftMax layer are deleted;

[0022] Decoder: composed of 1 convolution layer and 4 up-sampling layers, outputting a feature map of 208x144x64 after 4 times of up-sampling;

[0023] Feature fusion module: upsample the feature maps of 4 different scales of the encoder to 208x144x16 size and fuse the channels through convolution layers;

[0024] Refinement module: extract the edge information of the input image through the Sobel operator, and connect the convolution layer with the decoder feature map to output the depth map;

[0025] The model is trained using transfer learning, and the reconstruction effect is verified by the average relative error ≤0.02 and the average absolute error ≤50mm.

[0026] Preferably, in step S3, part of the training data set and part of the validation data set refer to image pairs from known landslides and non-landslide slopes in the training data set and the validation data set;

[0027] If there are multiple image pairs of the slope, the number of cracks in the slope, the length of the slope, and the relative height difference of the slope are taken as the overall value of the slope, and the longest crack length and the maximum crack width are taken as the actual value of each image pair.

[0028] Preferably, in step S4, the improved nonlinear NT hyperplane algorithm converts the landslide identification index screening into a quadratic programming problem:

[0029] ;

[0030] ;

[0031] wherein, represents a positive semi-definite matrix, represents a column vector, represents a transpose, represents an adjusted column vector, represents a penalty parameter, and the value is between 0-1, represents a column vector with all elements being 1 with the dimension of the number of indexes to be screened.

[0032] Let , , the quadratic programming problem is converted to:

[0033] ;

[0034] ;

[0035] wherein, represents the augmented matrix of , , represents the augmented matrix of , , a unit matrix representing the number of rows, columns, etc. is equal to the number of indicators to be screened, 、 respectively represent known landslide, non-landslide slope samples.

[0036] Preferably, in step S4, the three-dimensional information includes: the number of cracks in the slope, the length of the cracks in the slope, the relative height difference of the slope body, the length of the longest crack, and the maximum crack width; and the landslide disaster factors include: elevation, slope, slope direction, land use, rock-soil type, river distance, road distance, fault distance, plan curvature, profile curvature, normalized vegetation cover index, sediment transport index, flow power index, and terrain humidity index.

[0037] Preferably, in step S5, the landslide identification data set contains landslides and non-landslide slopes, the input is the value of the landslide identification indicator of each slope, and the output is a landslide label, wherein the model output of the known landslide is 1, and the model output of the non-landslide slope is 0;

[0038] A plurality of mathematical models are trained and verified, including but not limited to analytic hierarchy process, Logistic regression model, weighted information amount model, K-means model, non-dominated sorting algorithm, BP neural network, support vector machine, random forest model, crayfish algorithm, multiverse optimization model, long short-term memory neural network, sparrow optimization model, marine predator optimization model, XGBoost model, information amount-BP neural network, Bayesian-BP neural network, gorilla optimization-BP neural network, probability ratio-support vector machine, random forest-genetic algorithm, Logistic regression-information amount model, Logistic regression-evidential weight model, information amount-random forest model, support vector machine-random forest model, geographic weighted regression-random forest model, entropy weight-TOPSIS model, Blending-XGBoost-CNN model, evidence weight-CatBoost model, maximum entropy-iterative self-organizing clustering algorithm;

[0039] The optimal model is screened based on true positive rate TPR, false positive rate FPR, AUC, precision, and modeling time consumption, and the priority from high to low is in turn TPR maximum, FPR minimum, AUC maximum, Precision maximum, and modeling time consumption shortest;

[0040] The two-dimensional RGB image of the to-be-identified slope is reconstructed into a point cloud, and the three-dimensional information is extracted and input into the optimal model to output the identification result.

[0041] The application also provides a landslide identification system, comprising:

[0042] A data acquisition module is configured to acquire two-dimensional RGB images of known landslides and non-landslide slopes in a study area and synchronously acquire corresponding measured point cloud images through a three-dimensional laser scanner;

[0043] A data set construction module is configured to combine the NYU Depth V2 data set, crop or extend the two-dimensional RGB images and the measured point cloud images, correct the inclination, fill in missing points, and remove noise, and construct a training data set and a verification data set for three-dimensional information extraction through an augmentation operation;

[0044] A point cloud reconstruction module is configured to run an improved encoder-decoder convolutional neural network model, input the two-dimensional RGB images, and output reconstructed point cloud images, the encoder adopts a ResNet101 improved by a hollow convolution, the decoder includes an upsampling layer and a feature fusion module, and edge information is embedded through a refinement module;

[0045] A three-dimensional information extraction module is configured to extract the number of intraslope cracks, the length of intraslope cracks, the relative height difference of the slope body, the length of the longest crack, and the width of the largest crack from the measured point cloud images and the reconstructed point cloud images based on the Polyworks software;

[0046] An index screening module is configured to use an improved nonlinear NT hyperplane algorithm to solve the extracted three-dimensional information and landslide disaster factors by quadratic programming, and screen to form a landslide identification index system;

[0047] A landslide identification module is configured to construct a training set and a verification set based on the screened index system, train a landslide identification model through a plurality of mathematical models, and select an optimal model according to TPR, FPR, AUC, Precision, and modeling time consumption;

[0048] An output module is configured to perform point cloud reconstruction and three-dimensional information extraction on the two-dimensional RGB images of the to-be-identified slope, input the optimal model, and output a landslide identification result.

[0049] The application also provides a computer device comprising a memory and a processor, the memory is used to store instructions, and the processor is used to execute the instructions to realize the landslide identification method as described above.

[0050] The application also provides a computer-readable storage medium having a computer program stored thereon, the computer program is executed by a processor to realize the landslide identification method as described above.

[0051] Therefore, the landslide identification method, system, computer device and medium have the following beneficial technical effects:

[0052] (1) The landslide identification method of the present application not only relies on traditional landslide disaster-causing factors, but also includes crack indicators (number of cracks in the slope, length of cracks in the slope, relative height difference of the slope body, length of the longest crack, and width of the largest crack) that play an important role in the development of landslides, comprehensively and accurately reflecting the landslide disaster-causing mechanism;

[0053] (2) When obtaining the slope crack indicators, the present application obtains a two-dimensional RGB image of the slope and reconstructs a three-dimensional point cloud image of the two-dimensional RGB image based on an improved encoder-decoder convolutional neural network, avoiding the use of a three-dimensional laser scanner which is complex in process and expensive in equipment;

[0054] (3) The present application uses an improved nonlinear NT hyperplane algorithm to screen landslide identification indicators, which reduces the modeling complexity while preserving the original landslide information, and avoids falling into the "curse of dimensionality". BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 The flowchart of the landslide identification method of the present application;

[0056] Figure 2 The structure diagram of the improved encoder-decoder convolutional neural network. DETAILED DESCRIPTION

[0057] The technical solutions of the present application are further described below through the drawings and examples.

[0058] Unless otherwise defined, the technical terms or scientific terms used in the present application shall have the usual meanings understood by those skilled in the art to which the present application belongs.

[0059] Example 1

[0060] As shown in Figure 1 , a landslide identification method, the specific steps are as follows:

[0061] Step S1, obtaining two-dimensional RGB images of known landslides and the same number of non-landslide slopes in the study area and corresponding measured point cloud images based on three-dimensional laser scanning, combining NYU Depth V2 dataset to construct three-dimensional information extraction training dataset and verification dataset.

[0062] Specifically, the number of known landslides in the study area is , and non-landslide slopes are randomly selected, wherein the known landslide information is obtained by visiting local geological disaster prevention departments, consulting relevant statistical data, combining remote sensing interpretation, unmanned aerial vehicle scanning, and field reconnaissance, etc., and the non-landslide slope information is obtained by field reconnaissance.

[0063] Two-dimensional RGB images are acquired using a high-precision camera with no less than 24 million pixels. They are taken during the winter when vegetation is sparse, the weather is clear and well-lit, and there is no interference such as fog or haze.

[0064] When the slope space is small and a single shot can capture the entire slope, the camera should be placed at an appropriate position directly in front of the slope. During the shot, vegetation, loose soil, garbage, etc. on the slope surface can be easily removed to avoid obstruction of the view by vegetation, protruding rocks, and light.

[0065] When the slope has a large spatial scale and multiple shots are needed to capture the panoramic view of the slope, the camera can be set up at multiple appropriate positions in front of the slope. During shooting, vegetation, loose soil, garbage, etc. on the slope surface can be simply processed to avoid obstruction of the view by vegetation, protruding rocks, and light. The two-dimensional RGB images taken multiple times can be stitched together into a whole.

[0066] The measured point cloud images were acquired by fine scanning with a 3D laser scanner, with a point pitch of 2mm. The scanning time was synchronized with the capture time of the 2D RGB images.

[0067] The position of the 3D laser scanner should be consistent with the position of the high-precision camera. When the slope space is large, the point cloud images scanned from multiple stations should be registered as a whole using the software of the 3D laser scanner. To improve the registration accuracy, a target sphere can be set as a feature point during scanning.

[0068] Download 2D RGB images and corresponding measured point cloud images of 1449 scenes from the NYU Depth V2 dataset, and use LabelImg software to analyze them. Expand or crop 2D RGB images of slopes and 1449 scenes from the NYU Depth V2 dataset;

[0069] If the spatial scale of a slope or scene is no greater than 25m × 25m, then its 2D RGB image is expanded to 25m × 25m. If the spatial scale of a slope or scene is greater than 25m × 25m, then its 2D RGB image is cropped into multiple 25m × 25m images. The cropped images should contain all slope or scene information and minimize duplicate information. A total of [number missing] images are obtained. A two-dimensional RGB image;

[0070] The measured point cloud images of the aforementioned slope and scene were expanded or cropped using Polyworks software. The image is divided into sections, so that they correspond one-to-one with a two-dimensional RGB image;

[0071] The eigenvalue method is used to correct the tilt of expanded or cropped point cloud images, that is, by solving the normal vector. and parameters Constructing plane equations To make 3D point clouds , The distance to that plane is the minimum. These represent the components of the normal vector. Represents the third in three-dimensional space The point is at The coordinate values ​​along the axis. Represents the third in three-dimensional space The point is at The coordinate values ​​along the axis. Represents the third in three-dimensional space The point is at The coordinate values ​​along the axis. This represents the total number of discrete points in a 3D point cloud;

[0072] Due to the normal vector Constraints must be met The problem is transformed into solving the function using the Lagrange multiplier method. Minimum value:

[0073] ;

[0074] in, Represents the Lagrange multipliers;

[0075] Linear interpolation is used to fill in missing points in the point cloud:

[0076] ;

[0077] in, , These represent the missing point number and the reconstructed depth value of the missing point after filling, respectively. , They represent The closest valid point number and depth value previously. , They represent Next, the nearest valid point number and depth value;

[0078] Denoising is achieved using a bilateral filtering algorithm:

[0079] ;

[0080] in, Represents the three-dimensional coordinates of the filtered point. Represents the three-dimensional coordinates of the point before filtering. Indicates the filter factor. express The normal vector;

[0081] Each point cloud image after inclination correction, missing point filling and denoising contains N = 156275001 points;

[0082] The above two-dimensional RGB image and the measured point cloud image are respectively subjected to mirror flipping, horizontal flipping, vertical flipping, random brightness transformation and other augmentation operations to obtain a group of image pairs, and a three-dimensional information extraction dataset is constructed, wherein 70% is a training sample, and the remaining 30% is a verification sample, the two-dimensional RGB image is a model input, and the corresponding measured point cloud image is a model output.

[0083] Step S2, an improved encoder-decoder convolutional neural network model is established, and the point cloud image of the verification dataset is reconstructed by using the trained model and the reconstruction effect is verified.

[0084] As shown in Figure 2 , the improved encoder-decoder convolutional neural network model comprises an encoder, a decoder, a feature fusion module and a refinement module.

[0085] The encoder: the improved ResNet101 is used as a feature extractor to down-sample the two-dimensional RGB image, extract details and multi-scale features of different resolutions, and pass the feature map output by the last down-sampling module to the decoder; 33 residual structures are stacked, including 1x1, 3x3 and 1x1 convolution layers, and the input feature signal is associated with the output feature after convolution by using a jump connection mode, so that the original features are retained while the deep complex features are increased.

[0086] Unlike the traditional ResNet101, the 3x3 convolution layer in the last Residual Block 4 is replaced by a dilated convolution layer with a dilated rate of 2, so that the resolution of the output feature map is increased from 1 / 32 of the input image to 1 / 16, and at the same time, the last average pooling layer, the full connection layer and the SoftMax layer are deleted, and the Residual Block 4 is directly connected with the decoder.

[0087] The decoder: composed of 1 convolution layer and 4 up-sampling layers, respectively Conv 2, Up 1, Up 2, Up 3, Up 4, the feature map is increased by 1 times after each up-sampling, and after 4 times of up-sampling, the output feature map size is 208x144x64.

[0088] Fine module: the edge information of the input image is extracted by Sobel operator, the edge information is embedded by 3x3 convolution layer (Conv4), the edge information is fused with the feature map channel output by the decoder through Conv 5, Conv 6 and Conv 7 convolution layers, and the final depth map is output;

[0089] Feature fusion module: through Up 5, Up 6, Up 7 and Up 8 upsampling modules, the four different scale feature maps generated by the encoder downsampling are enlarged by 2, 4, 8 and 16 times respectively, the output feature map size is 208x144x16, and the output of the feature fusion layer is obtained through channel transformation of the convolution layer.

[0090] The input / output channel number and output feature size of each layer are shown in Table 1.

[0091] Table 1 Input / output channel number and output feature size of each layer

[0092] ;

[0093] The encoder-decoder convolutional neural network model is trained by using the transfer learning method, and the original data is down-sampled to 108x144 for accelerating the training. First, the model training hyperparameters such as Epochs, Batch Size, initial learning rate, momentum coefficient and decay rate are set, as shown in Table 2, then the model is pre-trained using the image pairs in the training set derived from the NYU Depth V2 dataset, and finally the model parameter correction is performed using the image pairs in the training set derived from the measured point cloud image.

[0094] Table 2 Standard training environment parameters

[0095] ;

[0096] The trained model is used to analyze and verify the sample, and the reconstructed point cloud image is output. Based on the average relative error (MRE) ), the average absolute error (MAE) ) index, the reconstructed point cloud image and the measured point cloud image are compared to verify the reconstruction effect.

[0097] 、 The calculation method is as follows:

[0098] ;

[0099] ;

[0100] wherein, represents the depth reconstruction value of point , represents the depth value of point The measured depth value;

[0101] If both conditions are met ≤0.02、 If the thickness is ≤50mm, the trained model is considered to meet the requirements; otherwise, it is considered not to meet the requirements. ≤0.02、 If the result is ≤50mm, adjust the model training hyperparameters and retrain the constructed encoder-decoder convolutional neural network model using transfer learning until the verification results meet the requirements.

[0102] Step S3: Extract three-dimensional information of the number of cracks within the slope, the length of cracks within the slope, the relative height difference of the slope, the longest crack length, and the maximum crack width from the measured point cloud images of part of the training dataset and the reconstructed point cloud images of part of the validation dataset using Polyworks software.

[0103] Among them, the partial training dataset and the partial validation dataset refer to the image pairs in the training dataset and the validation dataset that originate from known landslide and non-landslide slopes;

[0104] If a slope has multiple image pairs, the three-dimensional information of the entire slope and each image pair is extracted using Polyworks software. The number of cracks, the length of cracks, and the relative height difference of the slope are uniformly taken as the number of cracks, the length of cracks, and the relative height difference of the slope as a whole. The longest crack length and the maximum crack width of each image pair are taken as the actual values ​​of that image pair.

[0105] Step S4: Use the improved nonlinear NT hyperplane algorithm to filter the three-dimensional information and commonly used landslide-causing factors of the training dataset, and establish a landslide identification index system.

[0106] The improved nonlinear NT hyperplane algorithm transforms the selection of landslide identification indicators into a quadratic programming problem:

[0107] ;

[0108] ;

[0109] in, Represents a positive semidefinite matrix. Represents a column vector. Indicates transpose. This indicates adjusting the column vector. This represents the penalty parameter, with a value between 0 and 1. This represents a column vector whose dimension is the number of indicators to be screened, and whose elements are all 1s.

[0110] set up , Then the quadratic programming problem is transformed into:

[0111] ;

[0112] ;

[0113] wherein, denotes the augmented matrix of , denotes the augmented matrix of , denotes the identity matrix with the number of rows and columns equal to the number of indicators to be screened, , denote the known landslide and non-landslide slope samples, respectively.

[0114] To solve the above problem, first, the parameter is selected, and the vector is initialized, and = 0 is set.

[0115] Secondly, calculate , denotes the two-dimensional projection into the feasible region, that is:

[0116] ;

[0117] Definition: wherein, , are lower triangular matrix and diagonal matrix, respectively.

[0118] Thirdly, if , the algorithm stops, wherein ξ denotes the allowable error; otherwise, update to , to , and go to the previous step.

[0119] The three-dimensional information (the number of internal cracks, the length of internal cracks, the relative height difference of the slope body, the length of the longest crack, and the width of the largest crack) of the training data set and the commonly used landslide disaster-causing factors (elevation, slope, slope direction, land use, rock-soil type, river distance, road distance, fault distance, plan curvature, profile curvature, normalized vegetation cover index , sediment transport index , water power index , terrain humidity index ) are screened by using the improved nonlinear NT hyperplane algorithm, the dimension of the landslide identification model is reduced, and a landslide identification index system is established, wherein the number of screened landslide indicators is .

[0120] ;

[0121] ;

[0122] ;

[0123] ;

[0124] wherein, represents the near-infrared band reflectance value, represents the red band reflectance value, represents the catchment area flowing to a specific location, represents the slope.

[0125] Step S5, constructing a landslide identification training dataset and a verification dataset, establishing a landslide identification model based on a plurality of mathematical models, selecting a model with the best identification effect based on the verification dataset, and realizing landslide identification based on two-dimensional RGB image extraction of three-dimensional information.

[0126] The constructed landslide identification dataset includes samples, i.e. known landslides and non-landslide slopes, each sample is a vector containing elements, wherein the first elements are model inputs, i.e., the values of the landslide identification indicators of each slope, and the last element is the model output, the known landslide value is 1, and the non-landslide slope value is 0, 70% is randomly selected as a training sample, and the remaining 30% is a verification sample;

[0127] The mathematical models used include but are not limited to the analytic hierarchy process, the Logistic regression model, the weighted information amount model, the K-means model, the non-dominated sorting algorithm, the BP neural network, the support vector machine, the random forest model, the crayfish algorithm, the multiverse optimization model, the long short-term memory neural network, the sparrow optimization model, the marine predator optimization model, the XGBoost model, the information amount-BP neural network, the Bayesian-BP neural network, the gorilla optimization-BP neural network, the probability ratio-support vector machine, the random forest-genetic algorithm, the Logistic regression-information amount model, the Logistic regression-evidential weight model, the information amount-random forest model, the support vector machine-random forest model, the geographic weighted regression-random forest model, the entropy weight-TOPSIS model, the Blending-XGBoost-CNN model, the evidence weight-CatBoost model, the maximum entropy-iterative self-organizing clustering algorithm, and the hardware and software environments for modeling are the same as Table 2;

[0128] The above models were trained using training samples, and the trained models were used to analyze the verification samples. The output value was between 0 and 1, where 0 represents that it is definitely not a landslide and 1 represents that it is definitely a landslide. The model was set to identify a landslide when the output value was ≥0.5 and to identify a non-landslide when the output value was <0.5.

[0129] Statistical analysis of the above models on the validation samples TP (The number of landslides correctly identified by the model is known.) FN (The number of landslides that were misidentified by the model is known.) FP (Number of non-landslide slopes misidentified by the model) TN (The number of non-landslide slopes correctly identified by the model);

[0130] The landslide identification model with the best modeling performance was selected using TPR, FPR, AUC, Precision, and modeling time metrics.

[0131] The calculation methods for TPR and FPR are as follows:

[0132] ;

[0133] ;

[0134] The closer TPR is to 1 and the closer FPR is to 0, the better the model recognition performance.

[0135] ROC curves were plotted with TPR and FPR as the ordinate and x-axis respectively. AUC was the area under the curve, with a value between 0 and 1. 0 represents that the model is completely ineffective and 1 represents that the model is completely correct. The closer the AUC is to 1, the better the model recognition effect.

[0136] Precision is calculated as follows:

[0137] ;

[0138] The closer the Precision is to 1, the better the model's recognition performance.

[0139] The model with the maximum TPR is selected as the model with the best landslide identification effect; if there are not less than two models with the maximum TPR and the same, the model with the minimum FPR in the above not less than two models is selected as the model with the best identification effect; if there are still not less than two models with the minimum FPR and the same, the model with the maximum AUC in the above not less than two models is selected as the model with the best identification effect; if there are still not less than two models with the maximum AUC and the same, the model with the maximum Precision in the above not less than two models is selected as the model with the best identification effect; if there are still not less than two models with the maximum Precision and the same, the model with the shortest modeling time in the above not less than two models is selected as the model with the best identification effect;

[0140] A two-dimensional RGB image of a to-be-identified slope is acquired, a point cloud image thereof is reconstructed based on the constructed encoder-decoder convolutional neural network model, three-dimensional information is extracted by using Polyworks software, landslide identification indexes of the slope are analyzed based on the model with the best landslide identification effect, and a recognition result is determined according to the model output, so as to realize landslide identification based on extraction of three-dimensional information from a two-dimensional RGB image.

[0141] Embodiment two

[0142] A landslide identification system comprises:

[0143] A data acquisition module is configured to acquire two-dimensional RGB images of known landslide and non-landslide slopes in a study area, and to synchronously acquire corresponding measured point cloud images by using a three-dimensional laser scanner;

[0144] A data set construction module is configured to combine the NYU Depth V2 data set, to perform cropping or extension, inclination correction, missing point filling and denoising processing on the two-dimensional RGB images and the measured point cloud images, and to construct a training data set and a verification data set for three-dimensional information extraction by using an augmentation operation;

[0145] A point cloud reconstruction module is configured to run an improved encoder-decoder convolutional neural network model, to input two-dimensional RGB images and to output reconstructed point cloud images, the encoder adopts a ResNet101 improved by using a hollow convolution, the decoder comprises an upsampling layer and a feature fusion module, and edge information is embedded by using a fine module;

[0146] A three-dimensional information extraction module is configured to extract the number of intra-slope cracks, the length of intra-slope cracks, the relative height difference of a slope body, the length of the longest crack and the width of the largest crack from the measured point cloud images and the reconstructed point cloud images based on Polyworks software;

[0147] An index screening module is configured to use an improved nonlinear NT hyperplane algorithm to perform quadratic programming solving on the extracted three-dimensional information and landslide disaster factors, and to screen to form a landslide identification index system;

[0148] The landslide identification module is configured to construct a training set and a verification set based on the screened index system, train a landslide identification model through a plurality of mathematical models, and select an optimal model according to TPR, FPR, AUC, Precision and modeling time consumption;

[0149] The output module is configured to perform point cloud reconstruction and three-dimensional information extraction on a two-dimensional RGB image of a to-be-identified slope, and input the optimal model to output a landslide identification result.

[0150] The above functions, if realized in the form of software function units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0151] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered a list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by or in connection with an instruction execution system, apparatus or device, such as a computer-based system, a system including a processor or other system that can fetch the instructions from an instruction execution system, apparatus or device and execute the instructions, or in conjunction with these instructions execution systems, apparatus or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate or transport a program for use by or in connection with an instruction execution system, apparatus or device, or in conjunction with these instruction execution systems, apparatus or devices.

[0152] More specific examples (non-exhaustive list) of the computer-readable medium include an electrical connection (electronic device) having one or more wires, a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CD ROM). In addition, the computer-readable medium can even be paper or other suitable medium on which the program can be printed, because the program can be electronically obtained, for example, by optical scanning of the paper or other medium, followed by editing, interpretation or processing, if necessary, in other suitable ways, and then stored in a computer memory.

[0153] It is worth noting that the content not elaborated in the present application is all the prior art, which is well known to those skilled in the art.

[0154] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application rather than limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: it can still modify or equivalently replace the technical solutions of the present application, and these modifications or equivalent replacements also cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A landslide identification method, characterized in that, Includes the following steps: Step S1: Obtain two-dimensional RGB images of known landslides and an equal number of non-landslide slopes within the study area, along with their corresponding three-dimensional laser scanning measured point cloud images. Combine these with the NYU Depth V2 dataset to construct a three-dimensional information extraction training dataset and a validation dataset. Step S2: Establish an improved encoder-decoder convolutional neural network model, use the trained model to reconstruct the point cloud images of the verification dataset, and verify the reconstruction effect. Encoder: The improved ResNet101 is used as the feature extractor. The 3×3 convolutional layer of the last Residual Block 4 is replaced with a dilated convolutional layer with a dilation rate of 2. The resolution of the output feature map is increased to 1 / 16 of the input image. The terminal pooling layer, fully connected layer and SoftMax layer are removed. Step S3: Extract three-dimensional information of the number of cracks in the slope, the length of cracks in the slope, the relative height difference of the slope, the longest crack length, and the maximum crack width from the measured point cloud images of part of the training dataset and the reconstructed point cloud images of part of the validation dataset based on Polyworks software. Step S4: Use the improved nonlinear NT hyperplane algorithm to screen the three-dimensional information and landslide-causing factors of the training dataset, and establish a landslide identification index system. The improved nonlinear NT hyperplane algorithm transforms the selection of landslide identification indicators into a quadratic programming problem: ； ； in, Represents a positive semidefinite matrix. Represents a column vector. Indicates transpose. This indicates adjusting the column vector. This represents the penalty parameter, with a value between 0 and 1. This represents a column vector whose dimension is the number of indicators to be screened, and whose elements are all 1s. set up , Then the quadratic programming problem is transformed into: ； ； in, express augmented matrix, , express augmented matrix, , An identity matrix representing the number of rows and columns equal to the number of indicators to be filtered. , These represent known landslide and non-landslide slope samples, respectively. The three-dimensional information includes: the number of cracks within the slope, the length of cracks within the slope, the relative elevation difference of the slope, the longest crack length, and the maximum crack width; the landslide-causing factors include: elevation, slope, aspect, land use, soil and rock type, distance from rivers, distance from roads, distance from faults, plane curvature, profile curvature, normalized vegetation cover index, sediment transport index, water flow power index, and topographic humidity index. Step S5: Construct a landslide recognition training dataset and a validation dataset. Establish a landslide recognition model based on multiple mathematical models. Select the optimal model through the validation dataset to achieve landslide recognition based on two-dimensional RGB images.

2. The landslide identification method according to claim 1, characterized in that, Step S1 includes: The number of landslides is known to be Random selection One non-landslide slope; Acquire 2D RGB images and measured point cloud images; right Cropping or expanding 2D RGB images of slopes and the NYU Depth V2 dataset to obtain A two-dimensional RGB image; The measured point cloud image is cropped or expanded accordingly. The image is then processed for tilt correction, missing point filling, and noise reduction. Augmentation operations are performed on 2D RGB images and point cloud images to construct... The image pairs are divided into training samples and validation samples.

3. The landslide identification method according to claim 1, characterized in that, In step S2, the improved encoder-decoder convolutional neural network model includes: Encoder: The improved ResNet101 is used as the feature extractor. The 3×3 convolutional layer of the last Residual Block 4 is replaced with a dilated convolutional layer with a dilation rate of 2. The resolution of the output feature map is increased to 1 / 16 of the input image. The terminal pooling layer, fully connected layer and SoftMax layer are removed. Decoder: Consists of 1 convolutional layer and 4 upsampling layers, outputting a 208×144×64 feature map after 4 upsampling steps; Feature fusion module: Upsamples the feature maps of the encoder at four different scales to a size of 208×144×16 and performs channel fusion through a convolutional layer; Refinement module: Extracts edge information from the input image using the Sobel operator, and outputs a depth map by convolutional layer and connecting it with the decoder feature map; The model is trained using transfer learning, and the average relative error is used. ≤0.02 and mean absolute error ≤50mm to verify the reconstruction effect.

4. The landslide identification method according to claim 1, characterized in that, In step S3, the partial training dataset and the partial validation dataset refer to the image pairs in the training dataset and the validation dataset that originate from known landslide and non-landslide slopes; If there are multiple image pairs of a slope, the number and length of cracks within the slope and the relative height difference of the slope are taken as the overall values ​​of the slope, while the length of the longest crack and the width of the maximum crack are taken as the actual values ​​of each image pair.

5. The landslide identification method according to claim 1, characterized in that, In step S5, the landslide identification dataset contains landslide and For non-landslide slopes, the input is the landslide identification index value of each slope, and the output is the landslide label. The model output is 1 for known landslides and 0 for non-landslide slopes. Multiple mathematical models were used for training and validation, including: Analytic Hierarchy Process (AHP), Logistic Regression Model, Weighted Information Content Model, K-means Model, Non-dominated Ranking Algorithm, Backpropagation Neural Network (BP Neural Network), Support Vector Machine (SVM), Random Forest Model, Crayfish Algorithm, Multiverse Optimization Model, Long Short-Term Memory Neural Network (LSTM), Sparrow Optimization Model, Marine Predator Optimization Model, XGBoost Model, Information Content-BP Neural Network, Bayesian-BP Neural Network, Gorilla Optimization-BP Neural Network, Probability Ratio-SVM, Random Forest-Genetic Algorithm, Logistic Regression-Information Content Model, Logistic Regression-Weighted Evidence Model, Information Content-Random Forest Model, Support Vector Machine-Random Forest Model, Geographically Weighted Regression-Random Forest Model, Entropy Weight-TOPSIS Model, Blending-XGBoost-CNN Model, Weighted Evidence-CatBoost Model, and Maximum Entropy-Iterative Self-Organizing Clustering Algorithm. The optimal model was selected based on the true positive rate (TPR), false positive rate (FPR), AUC, precision, and modeling time, with the priority from high to low as follows: maximum TPR, minimum FPR, maximum AUC, maximum precision, and shortest modeling time. The point cloud is reconstructed from the 2D RGB image of the slope to be identified, and the 3D information is extracted and then input into the optimal model to output the recognition result.

6. A landslide identification system, characterized in that, For performing the landslide identification method as described in any one of claims 1-5, comprising: The data acquisition module is used to acquire two-dimensional RGB images of known landslides and non-landslide slopes within the study area, and simultaneously acquires the corresponding measured point cloud images through a three-dimensional laser scanner; The dataset construction module is used to combine the NYU Depth V2 dataset to perform cropping or expansion, tilt correction, missing point filling and denoising on 2D RGB images and measured point cloud images, and to construct training and validation datasets for 3D information extraction through augmentation operations. The point cloud reconstruction module is used to run an improved encoder-decoder convolutional neural network model. It takes a 2D RGB image as input and outputs a reconstructed point cloud image. The encoder uses a dilated convolution improved ResNet101, and the decoder includes an upsampling layer and a feature fusion module, and embeds edge information through a refinement module. The 3D information extraction module is used to extract the number of cracks in the slope, the length of cracks in the slope, the relative height difference of the slope, the longest crack length, and the maximum crack width from the measured point cloud image and the reconstructed point cloud image based on Polyworks software. The index screening module is used to perform a quadratic programming solution on the extracted three-dimensional information and landslide disaster-causing factors using an improved nonlinear NT hyperplane algorithm, and to screen and form a landslide identification index system. The landslide identification module is used to construct training and validation sets based on the selected indicator system, train the landslide identification model through various mathematical models, and select the optimal model based on TPR, FPR, AUC, Precision and modeling time. The output module is used to reconstruct point clouds and extract three-dimensional information from the two-dimensional RGB image of the slope to be identified, and to output the landslide identification result by taking the optimal model as input.

7. A computer device, characterized in that, It includes a memory and a processor, the memory being used to store instructions, and the processor being used to execute the instructions to implement the landslide identification method as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the landslide identification method as described in any one of claims 1-5.

Images