Deep learning-based dominant-recessive landslide identification method

Abstract

The invention relates to a deep learning-based dominant-recessive landslide identification method. The method specifically comprises the steps of obtaining dominant landslide hidden danger identification data of a target area; sBAS-InSAR differential interference measurement is carried out, and an InSAR deformation rate diagram is formed; correcting, enhancing and fusing the high-resolution remote sensing image data; processing the dominant landslide hidden danger identification data; inputting the processed identification data into a convolutional neural network model to generate a convolutional feature map; making a sample of dominant landslide identification; the convolutional neural network model outputs an identification result to obtain a dominant landslide hidden danger identification result; correcting and checking a dominant landslide recognition result, and constructing a dominant landslide hidden danger database; and analyzing and processing the dominant landslide hidden danger database and the hidden landslide identification factors to obtain a Pearson correlation absolute value display result in which the characteristic factors are sequenced according to importance. According to the method, geological disasters and easily-occurring slopes can be identified to the maximum extent, and life and property losses caused by the geological disasters are reduced.

Description

Technical Field

[0001] The present invention relates to a method for identifying and predicting geological disasters, and in particular to a method for identifying explicit and implicit landslides based on deep learning. Background Art

[0002] In the high mountain valley areas of my country, there are both obvious landslides with clear morphological and deformation characteristics and hidden landslides that are difficult to effectively identify. The number of these landslides is large and the formation conditions are complex. Quantitative evaluation of the influence of the main controlling factors on landslides is an important way to predict their distribution range. In recent years, remote sensing technology has become one of the main means of hidden danger identification and disaster assessment. The common methods for identifying obvious landslides using remote sensing mainly include manual visual interpretation, pixel-based image classification, multi-source remote sensing comprehensive identification, and deep learning model identification. For obvious landslides, the use of InSAR and high-resolution remote sensing data combined with deep learning model methods can quickly and accurately obtain the scope of landslides, which has important guiding significance for emergency rescue, disposal and recovery after landslide disasters. According to the spectral and texture characteristics of the obvious landslides in high-resolution images, the disaster characteristics and scope can be effectively interpreted manually, and the maximum deformation range can be well delineated using the InSAR surface deformation extraction results. On this basis, combined with deep learning methods, the scope of landslide disasters can be obtained more quickly and accurately. Common methods for identifying hidden landslides mainly include traditional artificial ground high-density electrical method, multi-source remote sensing and disaster-pregnant conditions collaborative identification method, etc. For hidden landslides, it is not easy to identify them by visual interpretation and multi-source remote sensing. However, deep learning identification using disaster-controlling conditions such as elevation, slope, slope direction, and engineering rock formation that are favorable for landslide occurrence has better identification effect and is more suitable for wide-area landslide hazard investigation and disaster emergency analysis.

[0003] At present, the effect of using remote sensing technology to identify explicit geological disasters is particularly significant, but there are still major problems in identifying implicit geological disasters: on the one hand, the use of traditional optical remote sensing and InSAR technology to identify geological disaster bodies is not perfect, and it is not possible to systematically integrate disaster-prone slopes into the remote sensing identification system, resulting in the identification target not being comprehensive and clear, and the weight of each indicator cannot be distinguished; on the other hand, the current research cannot systematically carry out intelligent remote sensing identification that integrates "implicit disaster bodies + prone slopes". Therefore, how to accurately and quickly find explicit geological disaster bodies, implicit geological disaster bodies and their prone slopes, and establish an integrated intelligent identification system to maximize the identification of geological disasters and their prone slopes, is a problem that needs to be solved in the industry. Summary of the invention

[0004] The purpose of the present invention is to provide an explicit-implicit landslide identification method based on deep learning to solve the problems of the existing identification methods that the identification targets are not comprehensive and clear enough and the identification of implicit geological disaster bodies is unclear, so as to find out geological disasters and their prone slopes to the greatest extent.

[0005] The object of the present invention is achieved in that:

[0006] A method for identifying explicit and implicit landslides based on deep learning, comprising the following steps:

[0007] S1. Obtaining explicit landslide hazard identification data in the target area; the explicit landslide hazard identification data includes existing landslide inventory points, InSAR images, and high-resolution remote sensing image data.

[0008] S2. Based on InSAR images, SBAS-InSAR differential interferometry is performed, and image registration is performed in combination with DEM data to form a differential interferometry phase map. The flat ground effect is removed from the differential interferometry phase map so that it can intuitively reflect the terrain changes in the target area. Then, through operations including coherence estimation, filtering, phase unwrapping, orbit refinement and geocoding, the ground target height is inverted, and the surface phase related to the terrain in the differential interferometry phase map is removed to obtain the deformation extraction result. Based on the deformation extraction result, the larger deformation area is delineated to form an InSAR deformation rate map.

[0009] S3. Perform radiation correction, orthorectification, image enhancement and image fusion on high-resolution remote sensing image data to produce image base maps and perform optical remote sensing identification. Establish identification marks based on high-resolution remote sensing data according to the spectral characteristics and spatial distribution characteristics of the target area's own objects, combined with geological data and the internal structure characteristics of the disaster body.

[0010] S4. Clean, integrate, standardize and regulate the data on identification of obvious landslide hazards in the target area.

[0011] S5. Input the processed explicit landslide hazard identification data into the convolutional neural network model for convolution operation; the convolutional neural network model includes an input layer Input, a convolution layer Conv, a pooling layer Pool, a fully connected layer FC and an output layer Output; the convolution kernel in the convolution layer is used as a feature extractor to perform convolution operation on the feature map of each layer input, and then the convolution layer data is output through the convolution kernel to generate a convolution feature map.

[0012] S6. Input the InSAR deformation extraction results and high-resolution remote sensing image data into the convolutional neural network model to produce explicit landslide hazard identification samples and train the produced samples.

[0013] S7. Place the pooling layer in the convolutional neural network model between the convolutional layers, transfer the convolutional feature map to the fully connected layer, and output the recognition result; then fuse the recognition result into the original input size to obtain the recognition result of the explicit landslide hazard.

[0014] S8. Correct and verify the identification results of obvious landslides and build a database of obvious landslide hazards.

[0015] S9. Obtain the characteristic factor data of hidden landslide hazards in the target area; the characteristic factors of hidden landslide hazards include elevation factor, slope factor, aspect factor, plane curvature factor, profile curvature factor, vegetation coverage factor, lithology factor, structural factor and water system factor.

[0016] S10. Use the Pearson correlation coefficient to perform correlation analysis and absolute value processing on the data of explicit landslide hazards and the characteristic factors of implicit landslides, and obtain the absolute value display results of the Pearson correlation in which each characteristic factor is sorted according to importance; according to the absolute value display results of the Pearson correlation, select the characteristic factors of implicit landslides with high correlation as the main controlling factors of implicit landslide hazards for identification, and use the random point method to exclude cloud areas and landslide areas in the target area to generate negative samples with the same number of landslides.

[0017] Furthermore, the specific method of step S4 includes the following sub-steps:

[0018] S4-1 uses the interpolation method to eliminate missing values ​​in the target area's explicit landslide hazard identification data, uses the elimination method to remove duplicate values, and treats jump values ​​as missing values ​​or duplicate values.

[0019] S4-2 uses ArcGIS software to process topographic data, geological data and remote sensing data of the target area.

[0020] S4-3 uses the minimum-maximum normalization method to standardize the explicit landslide hazard identification data in the target area and normalize different types of data to the same value range.

[0021] S4-4 uses ArcGIS software to divide the topographic data, geological data, and remote sensing data into grid cells of the same size, and exports the topographic data, geological data, and remote sensing data as CSV files for model training.

[0022] Furthermore, in step S5, the convolution kernel of the convolution layer in the convolutional neural network model transforms and maps the input explicit landslide hazard identification data to extract local feature information and output it to the pooling layer; the weight sharing feature in the convolution layer is used to reduce the number of parameters; the pooling layer is used after the convolution layer to compress the volume of data and parameters and reduce overfitting; the pooling layer uses a filter to downsample each data body to change its spatial size; all neurons in the fully connected layer are connected through weights to complete the classification task.

[0023] Furthermore, in step S5, the calculation formula used for the convolutional layer output is:

[0024]

[0025] Among them, k is the number of convolution kernels; C j represents the output of the jth convolution kernel; f represents the nonlinear activation function; i is the spatial position of the convolution operation; x i Represents the input data corresponding to the convolution window; w j and b j denote weight and bias respectively.

[0026] Furthermore, in step S5, the calculation formula for downsampling in the pooling layer is:

[0027]

[0028] Among them, O i is the output of the pooling layer; i is the spatial position of the pooling; It is the input data corresponding to the spatial position i of this pooling process.

[0029] Furthermore, step S10 includes the following sub-steps:

[0030] S10-1 used the Pearson correlation coefficient to conduct correlation analysis and absolute value processing on the data of explicit landslide hazards and the characteristic factors of implicit landslide hazards, and analyzed the control effect of each characteristic factor on implicit landslides; the random forest model was used to screen and train the characteristic factors with high correlation in implicit landslide identification, and the Bootstrap replacement resampling method was used to construct a decision tree, and the final result was obtained by voting based on the training results of each decision tree.

[0031] S10-2 After the characteristic factors with high correlation in the identification of hidden landslides were trained using the random forest model, SHAP was used to analyze the characteristic factors used for training to quantify the control effect of each characteristic factor on the formation of hidden landslides; the predicted value of the SHAP interpretation model is the sum of the contribution values ​​of each characteristic factor, which is defined as:

[0032]

[0033] In the formula, g(z') is the explanatory model; Φ0 is the model baseline value; k is the total number of characteristic factors input into the model; Φ i Represents the SHAP value of the i-th feature factor, which is used to indicate its contribution to the model prediction results.

[0034] S10-3 draws the relationship between the change of each characteristic factor value and the change of SHAP value, and obtains the mechanism of the interaction of characteristic factors on the formation of landslide;

[0035] S10-4 uses the data of the remaining samples that did not participate in the model to test the accuracy of the model; according to the field verification situation, a confusion matrix of the comprehensive identification results of explicit landslides and implicit landslides is prepared, and the recognition accuracy is quantitatively evaluated using the accuracy rate ACC, recall rate Recall, precision rate Pre and comprehensive evaluation F1 indicators.

[0036] The deep learning adopted in the present invention can construct a characteristic explicit landslide and implicit landslide related factor identification model and conduct training to automatically learn the controlling effect of changes in different characteristic factors on landslides, thereby objectively and quantitatively evaluating the importance of factors.

[0037] Based on the acquisition of GF2 remote sensing image data and Sentinel-1A data, the present invention first uses a deep learning model to quickly and automatically identify a large range of explicit landslides in the study area, performs manual correction and verification, and establishes an explicit landslide database; then, based on the use of Pearson to analyze the correlation between landslides and various environmental factors, the main characteristic factors are screened out, and the random forest model is used to perform sample training on the main characteristic factors in the area, and finally the hidden landslide hazard area is delineated, and the accuracy ACC, recall rate Recall, precision rate Pre and comprehensive evaluation F1 index of the intelligent recognition results are analyzed through precision verification. The research results have great advantages in the investigation and evaluation of geological disasters in alpine canyon areas, and can provide important technical support for the identification of disaster risks and the investigation of hidden dangers in mountainous areas.

[0038] The present invention utilizes InSAR and high-resolution remote sensing image data, combined with historical landslide data, to perform data preprocessing and data integration processing, construct landslide identification marks, divide training sets and verification sets, and complete the construction, training parameter adjustment and optimization of convolutional neural network models in deep learning methods based on the training sets; utilizes the set model to automatically identify the dominant landslides in the study area, manually corrects and verifies the results, delineates the dominant landslide hidden danger areas, and constructs a landslide database; based on the identified dominant landslide hidden dangers, utilizes Pearson to analyze the correlation between the landslide and various environmental factors such as elevation, slope, and slope direction, screens out the main characteristic factors, uses the random forest model to perform sample training on the main characteristic factors in the landslide sample library in the area, thereby delineating the distribution area of ​​the hidden landslide hidden dangers, and analyzes the accuracy ACC, recall Recall, precision Pre and comprehensive evaluation F1 index of the intelligent recognition results through precision verification.

[0039] The present invention can embody the rapid identification of explicit landslides and the accurate prediction of implicit landslide areas, and can identify geological disasters and their prone slopes to the greatest extent, so as to reduce the loss of life and property caused by such disasters to the people. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flowchart of the identification method of the present invention. DETAILED DESCRIPTION

[0041] The present invention will be further described below in conjunction with the accompanying drawings.

[0042] like Figure 1 As shown, the explicit-implicit landslide identification method based on deep learning of the present invention comprises the following steps:

[0043] Step 1: Obtain explicit landslide hazard identification data in the target area; the explicit landslide hazard identification data includes existing landslide inventory points, InSAR images and high-resolution remote sensing image data.

[0044] Step 2: Based on the InSAR image, SBAS-InSAR differential interferometry is performed, and image registration is performed in combination with DEM data to weaken the influence of spatial decorrelation and form a differential interferometry phase map. The flat ground effect is removed from the differential interferometry phase map so that the differential interferometry phase map can intuitively reflect the terrain changes. Through coherence estimation, filtering, phase unwrapping, orbit refinement, geocoding and other measures, the ground target height is inverted, and the surface phase related to the terrain in the differential interferometry phase map is removed to obtain the deformation extraction result; atmospheric delay phase correction is used to improve the accuracy of deformation estimation. The deformation extraction result is used to complete the delineation of the larger deformation area, thereby forming an InSAR deformation rate map.

[0045] Step 3: Perform radiation correction, orthorectification, image enhancement, image fusion and other steps on the high-resolution remote sensing image data to form data that meets the research requirements for the production of image base maps and optical remote sensing identification. According to the spectral characteristics and spatial distribution characteristics of the target area's own objects, combined with geological data, analyze the internal structural characteristics of the disaster body, and establish identification marks based on high-resolution remote sensing data.

[0046] Step 4: Clean, integrate, standardize and reduce the existing landslide record points, high-resolution remote sensing image data and InSAR image data of the dominant landslides in the target area. The specific processing methods are:

[0047] Firstly, for the explicit landslide hazard identification data in the target area, the interpolation method is used to eliminate missing values, the elimination method is used to remove duplicate values, and the jump values ​​are treated as missing values ​​or duplicate values ​​depending on the situation.

[0048] Secondly, ArcGIS software is used to process the topographic data, geological data and remote sensing data of the target area, solve the inconsistencies in data format, coordinate system, scale, etc., and output them into a format that can be recognized by ArcGIS software.

[0049] Thirdly, the minimum-maximum normalization method is used to standardize the explicit landslide hazard identification data in the target area, and normalize different types of data to the same value range.

[0050] Finally, GIS software was used to divide the topographic data, geological data, and remote sensing data into grid cells of the same size, and the topographic data, geological data, and remote sensing data were exported as CSV files to facilitate model training.

[0051] Step 5: Input the existing landslide inventory points, InSAR image data and high-resolution remote sensing image data after integration and standardization into the convolutional neural network model for convolution operation. The convolutional neural network model consists of several neural network layers, including the input layer (Input), convolution layer (Conv), pooling layer (Pool), fully connected layer (FC) and output layer (Output). The convolution layer and pooling layer are usually alternating. The convolution layer is the core layer in the CNN model. The convolution kernel is the basic computing unit of the convolution layer. As a feature extractor, it performs convolution operation on the feature map of each layer input, and then outputs the convolution layer data through the convolution kernel to generate a convolution feature map.

[0052] The convolution kernel in the convolutional neural network model is used to transform and map the input explicit landslide hazard identification data to extract local feature information. The weight sharing feature in the convolution layer is used to reduce the number of parameters. The pooling layer is after the convolution layer and is used to compress the volume of data and parameters and reduce overfitting. The pooling layer uses filters to downsample each data volume to change its spatial size. All neurons in the fully connected layer are connected through weights to complete the classification task.

[0053] The calculation formula used for the convolutional layer output is:

[0054]

[0055] Among them, k is the number of convolution kernels; C j represents the output of the jth convolution kernel; f represents the nonlinear activation function; i is the spatial position of the convolution operation; x i Represents the input data corresponding to the convolution window; w j and b j denote weight and bias respectively.

[0056] The calculation formula for downsampling in the pooling layer is:

[0057]

[0058] Among them, O i is the output of the pooling layer; i is the spatial position of the pooling; It is the input data corresponding to the spatial position i of this pooling process.

[0059] Step 6: Using the above convolution model algorithm, the InSAR deformation extraction results and high-resolution remote sensing image data are input into the convolution neural network model as samples of explicit landslides. By analyzing the length and width of the circumscribed rectangles of all landslides, the landslide sample size is obtained. The two types of data are cropped in the same way and input into the model for identification. After input into the model, they are divided into two categories: landslide and no landslide. Images with landslides are positive samples, represented by 1; images without landslides are negative samples, represented by 0. Since the area without landslides is much larger than the area with landslides, negative samples with a balanced number of positive samples are selected for identification.

[0060] Step 7: Place the pooling layer in the convolutional neural network model between the convolutional layers. Each pooling layer uses a 2×2 pooling window with a step size of 2, which can reduce the size of the feature map by half. Then transfer the convolutional feature map to the fully connected layer, output the recognition result, and complete the data output. After the intelligent recognition is completed, the recognition result is fused to the original input size to obtain the final explicit landslide hazard recognition result.

[0061] Step 8: According to the results of the intelligent identification of dominant landslides, manual visual correction and verification are performed to build a database of dominant landslide hazards. The number and area of ​​landslides are statistically analyzed based on the database, and the four key parameters of learning rate (lr), batch size (bs), number of iterations (ep), and optimizer (op) are dynamically adjusted to determine the best parameter combination. The convolutional neural network model directly outputs discrete numerical variables, and the open source Gdal library is used to convert the output results into raster images in Tif format.

[0062] Step nine, obtain the characteristic factor data of hidden landslide hazards in the target area. The characteristic factors of hidden landslide hazards include elevation factor, slope factor, aspect factor, plane curvature factor, profile curvature factor, vegetation cover factor, lithology factor, structural factor and water system factor.

[0063] Step 10: Use the Pearson correlation coefficient to perform correlation analysis and absolute value processing on the data of explicit landslide hazards and the characteristic factors of implicit landslide hazards to obtain the importance ranking of each factor. Display the results according to the absolute value of the Pearson correlation, select the characteristic factors with high correlation as the main characteristic factors of implicit landslides for intelligent identification, and use the random point method to exclude cloud areas and landslide areas in the target area to generate negative samples with the same number of landslides.

[0064] Firstly, the Pearson correlation coefficient was used to conduct correlation analysis and absolute value processing on the data of explicit landslide hazards and the factors for identifying implicit landslides, and the control effect of different characteristic factors on implicit landslides was analyzed. The characteristic factors with high correlation in identifying implicit landslides were screened and trained using the random forest model, and the decision tree was constructed using the Bootstrap replacement resampling method. The final result was obtained by voting based on the training results of each decision tree.

[0065] Secondly, after the characteristic factors with high correlation in the identification of hidden landslides were trained using the random forest model, SHAP was used to analyze the characteristic factors used for training to quantify the control effect of each characteristic factor on the formation of hidden landslides. The predicted value of the SHAP interpretation model is interpreted as the sum of the contribution values ​​of each input factor, which is defined as:

[0066]

[0067] Where g(z') is the explanatory model; Φ0 is the model baseline value; k is the total number of factors input into the model; Φ i Represents the SHAP value of the i-th factor, which is used to indicate the contribution to the model prediction results, that is, the contribution value.

[0068] Thirdly, by plotting the relationship between the changes in the values ​​of each characteristic factor and its SHAP value, the mechanism of the interaction of characteristic factors on the formation of landslides was obtained.

[0069] Finally, the data of the remaining samples that did not participate were used to test the accuracy of the model. According to the field verification results, the confusion matrix of the comprehensive identification results of explicit landslides and implicit landslides was made, and the identification accuracy was quantitatively evaluated using indicators such as accuracy (ACC), recall (Recall), precision (Pre) and comprehensive evaluation (F1).

Claims

1. A method for identifying explicit and implicit landslides based on deep learning, characterized in that: The following steps are involved: S1. Obtaining explicit landslide hazard identification data in the target area; the explicit landslide hazard identification data includes existing landslide inventory points, InSAR images, and high-resolution remote sensing image data; S2. Perform SBAS-InSAR differential interferometry based on InSAR images, and perform image registration in combination with DEM data to form a differential interferometry phase map; remove the flat ground effect from the differential interferometry phase map so that it can intuitively reflect the terrain changes in the target area, and then perform operations including coherence estimation, filtering, phase unwrapping, orbit refinement and geocoding to invert the ground target height, remove the surface phase related to the terrain in the differential interferometry phase map, and obtain the deformation extraction result; delineate the deformation area based on the deformation extraction result to form an InSAR deformation rate map; S3. Perform radiation correction, orthorectification, image enhancement, and image fusion on high-resolution remote sensing image data to produce image base maps and perform optical remote sensing identification. According to the spectral characteristics and spatial distribution characteristics of the target area's own objects, combined with geological data and the internal structural characteristics of the disaster body, establish identification marks based on high-resolution remote sensing image data; S4. Clean, integrate, standardize and regulate the data on the identification of the obvious landslide hazards in the target area; S5, inputting the processed explicit landslide hazard identification data into a convolutional neural network model for convolution operation; the convolutional neural network model includes an input layer Input, a convolution layer Conv, a pooling layer Pool, a fully connected layer FC and an output layer Output; the convolution kernel in the convolution layer is used as a feature extractor to perform a convolution operation on the feature map of each layer input, and then the convolution layer data is output through the convolution kernel to generate a convolution feature map; S6, inputting the InSAR deformation extraction results and high-resolution remote sensing image data into the convolutional neural network model, producing explicit landslide hazard identification samples, and training the produced samples; S7, placing the pooling layer in the convolutional neural network model between the convolutional layers, transferring the convolutional feature map to the fully connected layer, and outputting the recognition result; Then the recognition results are fused into the original input size to obtain the recognition results of the explicit landslide hazard; S8. Correct and verify the identification results of dominant landslides and build a database of dominant landslide hazards; S9. Obtain characteristic factor data of hidden landslide hazards in the target area; the characteristic factors of hidden landslide hazards include elevation factor, slope factor, aspect factor, plane curvature factor, profile curvature factor, vegetation coverage factor, lithology factor, structural factor and water system factor; S10. Use the Pearson correlation coefficient to perform correlation analysis and absolute value processing on the data of explicit landslide hazards and the characteristic factors of implicit landslides, and obtain the absolute value display results of the Pearson correlation in which each characteristic factor is sorted according to importance; according to the absolute value display results of the Pearson correlation, select the characteristic factors of implicit landslides with high correlation as the main controlling factors of implicit landslide hazards for identification, and use the random point method to exclude cloud areas and landslide areas in the target area to generate negative samples with the same number of landslides.

2. The method for identifying explicit and implicit landslides based on deep learning according to claim 1, characterized in that: The specific method of step S4 includes the following sub-steps: S4-1 The interpolation method is used to eliminate missing values ​​and the elimination method is used to remove duplicate values ​​for the explicit landslide hazard identification data in the target area, and the jump values ​​are treated as missing values ​​or duplicate values; S4-2 Use ArcGIS software to process topographic data, geological data and remote sensing data of the target area; S4-3 uses the minimum-maximum normalization method to standardize the explicit landslide hazard identification data in the target area and normalize different types of data to the same value range; S4-4 uses ArcGIS software to divide the topographic data, geological data, and remote sensing data into grid cells of the same size, and exports the topographic data, geological data, and remote sensing data as CSV files for model training.

3. The method for identifying explicit and implicit landslides based on deep learning according to claim 1, characterized in that: In step S5, the convolution kernel of the convolution layer in the convolutional neural network model transforms and maps the input explicit landslide hazard identification data to extract local feature information and output it to the pooling layer; the pooling layer uses a filter to downsample each data volume to change its spatial size; All neurons in the fully connected layer are connected through weights to complete the classification task.

4. The method for identifying explicit and implicit landslides based on deep learning according to claim 3 is characterized in that In step S5, the calculation formula used for the convolutional layer output is: Among them, k is the number of convolution kernels; C j represents the output of the jth convolution kernel; f represents the nonlinear activation function; i is the spatial position of the convolution operation; x i Represents the input data corresponding to the convolution window; w j and b j denote weight and bias respectively.

5. The method for identifying explicit and implicit landslides based on deep learning according to claim 3 is characterized in that In step S5, the calculation formula for downsampling in the pooling layer is: Among them, O i is the output of the pooling layer; i is the spatial position of the pooling; It is the input data corresponding to the spatial position i of this pooling process.

6. The method for identifying explicit and implicit landslides based on deep learning according to claim 1, characterized in that: Step S10 includes the following sub-steps: S10-1 used the Pearson correlation coefficient to conduct correlation analysis and absolute value processing on the data of explicit landslide hazards and the characteristic factors of implicit landslide hazards, and analyzed the control effect of each characteristic factor on implicit landslides; the characteristic factors with high correlation in implicit landslide identification were screened and trained using the random forest model, and the decision tree was constructed using the Bootstrap replacement resampling method, and the final result was obtained by voting based on the training results of each decision tree; S10-2 After the characteristic factors with high correlation in the identification of hidden landslides were trained using the random forest model, SHAP was used to analyze the characteristic factors used for training to quantify the control effect of each characteristic factor on the formation of hidden landslides; the predicted value of the SHAP interpretation model is the sum of the contribution values ​​of each characteristic factor, which is defined as: In the formula, g(z') is the explanatory model; Φ0 is the model baseline value; k is the total number of characteristic factors input into the model; Φ i Represents the SHAP value of the i-th feature factor, which is used to indicate the contribution to the model prediction results; S10-3 draws the relationship between the changes of each characteristic factor value and its SHAP value, and obtains the mechanism of the interaction of characteristic factors on the formation of landslides; S10-4 uses the data of the remaining samples that did not participate in the model to test the accuracy of the model; according to the field verification situation, a confusion matrix of the comprehensive identification results of explicit landslides and implicit landslides is prepared, and the recognition accuracy is quantitatively evaluated using the accuracy rate ACC, recall rate Recall, precision rate Pre and comprehensive evaluation F1 indicators.

Images