Method for predicting landslide susceptibility by using machine learning model and SBAS-InSAR deformation

Abstract

The invention relates to a method for predicting landslide susceptibility by using a machine learning model and SBAS-InSAR deformation, and the method comprises the steps: obtaining an environment static factor and landslide catalog in a sample region, obtaining a dynamic factor based on an SBAS-InSAR technology, building a landslide prediction model based on a random forest algorithm, and carrying out the prediction of the landslide susceptibility. Inputting the static factor, the dynamic factor and the landslide record into a landslide prediction model for training; and inputting the environmental static factor and the dynamic factor of the to-be-detected area into the trained landslide prediction model to obtain a landslide susceptibility map. According to the landslide susceptibility mapping method, the earth surface deformation rate serves as a dynamic factor to be fused into landslide susceptibility evaluation, a scientific basis is provided for landslide prevention and disaster reduction, the earth surface rail rising deformation rate and the earth surface rail falling deformation rate are obtained, the earth surface deformation rate is integrated into a static factor data set, and the landslide susceptibility mapping accuracy is improved. After uniform data preprocessing is completed, static factor and dynamic factor data are input into the landslide prediction model for joint training and evaluation, and the method has the advantage of being high in precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of landslide disaster prediction, and particularly relates to a method for predicting landslide susceptibility by using a machine learning model and SBAS-InSAR deformation. Background Art

[0002] A landslide is a natural phenomenon in which rock or soil mass on a slope slides as a whole along a specific sliding surface under the action of gravity and natural or artificial stresses. Landslides are one of the main geological disasters commonly existing in mountainous areas around the world, posing a major threat to human life and property safety and the natural environment.

[0003] Landslide susceptibility mapping (LSM) is an effective method to reveal the spatial distribution and occurrence probability of potential landslides, and its accuracy is crucial for minimizing casualties and property losses. The occurrence and deformation failure of landslides usually start from the creep stage and are closely related to surface deformation. However, traditional landslide susceptibility evaluation mainly relies on static factors such as topography, geology, and hydrology, ignoring the temporal and dynamic characteristics of landslides, resulting in difficulty in accurately identifying active landslides that are deforming, thus affecting the reliability of the evaluation. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for predicting landslide susceptibility by using a machine learning model and SBAS-InSAR deformation in order to solve the above problems.

[0005] The present invention achieves the above purpose through the following technical solutions:

[0006] A method for predicting landslide susceptibility by using a machine learning model and SBAS-InSAR deformation includes the following steps:

[0007] Obtain the environmental static factors, dynamic factors, and landslide catalog in the sample area, where the dynamic factors include the surface deformation rates under ascending and descending orbits in the sample area;

[0008] Establish a landslide prediction model based on the random forest algorithm, and input the static factors, dynamic factors, and landslide catalog into the landslide prediction model for training;

[0009] Obtain the environmental static factors and dynamic factors of the area to be measured, input the environmental static factors and dynamic factors of the area to be measured into the trained landslide prediction model, obtain the landslide susceptibility evaluation index, and determine the prediction result according to the evaluation index.

[0010] As a further optimization scheme of the present invention, the environmental static factors include: elevation, slope, aspect, curvature, distance from road, land use, distance from fault, lithology, normalized difference vegetation index, distance from river, annual rainfall, terrain wetness index.

[0011] As a further optimization scheme of the present invention, the dynamic factor is obtained based on the SBAS-InSAR technology, and the obtaining process includes:

[0012] Obtain a plurality of ascending-track SAR images and descending-track SAR images in the sample area;

[0013] Process the ascending-track SAR images and descending-track SAR images to obtain a plurality of groups of ascending-track interferograms and descending-track interferograms;

[0014] Process the ascending-track interferograms and descending-track interferograms based on a preset screening rule to obtain ascending-track PS points and descending-track PS points;

[0015] Unwrap the interferometric phase of the ascending-track PS points and descending-track PS points based on the 3D unwrapping technology to obtain the ascending-track surface deformation rate and the descending-track deformation rate.

[0016] As a further optimization scheme of the present invention, after obtaining the dynamic factor, the method further includes verifying the results of the ascending-track deformation rate and the descending-track deformation rate:

[0017] Randomly select several points with the same longitude and latitude as homologous points, respectively extract the deformation rates of the homologous points in the ascending-track radar line-of-sight and descending-track radar line-of-sight directions, establish a linear regression model with the ascending-track deformation rate as the independent variable x and the descending-track deformation rate as the dependent variable y for linear fitting, and use the coefficient of determination R 2 Evaluate the correlation between the ascending-track deformation rate and the descending-track deformation rate, and verify the reliability of the results based on a preset correlation threshold.

[0018] As a further optimization scheme of the present invention, the process of processing the ascending-track SAR images and descending-track SAR images is specifically: set a time baseline threshold and a spatial baseline threshold, and eliminate the interferograms with poor coherence and obvious errors with respect to the baseline threshold;

[0019] The preset screening rule is: by setting a coherence threshold, screen out the PS points in the ascending / descending-track interferograms that are higher than the coherence threshold.

[0020] As a further optimization scheme of the present invention, the landslide catalog includes historical landslide samples and non-landslide samples.

[0021] As a further optimization scheme of the present invention, the process of inputting the static factor, the dynamic factor, and the landslide catalog into the landslide prediction model for training is:

[0022] The sample area is divided into several grid cells according to a preset resolution, and a dataset is established based on the grid cells. The dataset includes a feature matrix and a landslide catalog. Among them, the process of obtaining the feature matrix is as follows: overlay the landslide samples and non-landslide samples with the layers of static factors and dynamic factors, and extract the factor values of each grid cell as the feature matrix;

[0023] The grid cells are divided into a training set and a validation set. With the feature matrix as the input and the landslide samples / non-landslide samples as the output, the landslide prediction model is trained based on the training set and verified based on the validation set.

[0024] As a further optimized solution of the present invention, obtaining the landslide susceptibility evaluation index and determining the prediction result according to the evaluation index includes:

[0025] Taking the predicted landslide susceptibility index as the evaluation index, its value range is from 0 to 1. Among them, the larger the value, the higher the possibility of landslide occurrence;

[0026] The natural breakpoint method is used to divide the landslide susceptibility index into different grades, and the corresponding color blocks are selected based on different grades and marked on the map for visual mapping.

[0027] The beneficial effects of the present invention are as follows:

[0028] The present invention integrates the surface deformation rate as a dynamic factor into the landslide susceptibility evaluation, providing a scientific basis for landslide prevention and disaster reduction. By obtaining the surface ascending orbit deformation rate and descending orbit deformation rate, and integrating the surface deformation rate into the static factor dataset, after completing unified data preprocessing, the static factor and dynamic factor data are input into the landslide prediction model for joint training and evaluation, which has the advantage of high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is the flowchart of the landslide susceptibility mapping of the present invention;

[0030] Figure 2 is the simulation schematic diagram of the ROC curve and AUC value of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] The following further describes the present application in detail with reference to the drawings. It is necessary to point out here that the following specific embodiments are only used to further illustrate the present application and cannot be understood as limiting the protection scope of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application according to the above application content.

[0032] Example 1

[0033] As Figure 1As shown in the figure, a method for landslide susceptibility prediction using a machine learning model and SBAS-InSAR deformation includes the following steps:

[0034] S1: Obtain the environmental static factors and landslide catalog within the sample area. The environmental static factors include: elevation, slope, aspect, curvature, distance to road, land use, distance to fault, lithology, normalized difference vegetation index, distance to river, annual rainfall, and terrain wetness index. The landslide catalog includes historical landslide samples and non-landslide samples;

[0035] S2: Obtain dynamic factors based on the SBAS-InSAR technology. The dynamic factors include the surface deformation rates of ascending and descending orbits within the sample area;

[0036] The process of obtaining dynamic factors based on the SBAS-InSAR technology specifically includes:

[0037] Obtain several ascending-orbit SAR images and descending-orbit SAR images of the sample area, and process the ascending-orbit SAR images and descending-orbit SAR images based on the S1IP software. Set the time baseline threshold to 37d and the spatial baseline threshold to 150m. By removing the interferograms with poor coherence with the baseline threshold and obvious errors, obtain several groups of ascending-orbit interferograms and descending-orbit interferograms;

[0038] Obtain ascending-orbit PS points and descending-orbit PS points based on the preset screening rules. The preset screening rules are: by setting the coherence threshold to 0.3, screen out the PS points in the ascending / descending-orbit interferograms higher than the coherence threshold. Use MATLAB based on the Ubuntu 14.04 system to process the ascending-orbit PS points and descending-orbit PS points. Use 3D phase unwrapping technology to unwrap the interferometric phase to obtain the surface ascending-orbit deformation rate and descending-orbit deformation rate; Check the results of the ascending-orbit deformation rate and descending-orbit deformation rate. First, resample the deformation data of the ascending and descending orbits to the same spatial resolution and unify the projection coordinate system to ensure that the geographic coordinate grids of the ascending and descending orbit data are strictly matched and avoid mis-matching of the same-named points due to pixel offset. Second, the selected same-named points should have good spatial distribution representativeness, covering different terrain and land cover types. Select by stratifying according to terrain (flat area, slope, ridge) and surface cover type (city, farmland), and give priority to selecting stable ground points (such as bedrock outcrop areas, permanent buildings), avoiding areas with dense water bodies and vegetation cover. The spatial distribution satisfies an interval of ≥1km in the horizontal direction, and the elevation span covers 80% of the study area. Then, extract the deformation amounts of the same-named points in the ascending-orbit radar line-of-sight and descending-orbit radar line-of-sight directions respectively. Take the ascending-orbit deformation rate as the independent variable x and the descending-orbit deformation rate as the dependent variable y to establish a linear regression model, and use the coefficient of determination R 2 To measure the linear correlation between the two data sets, and based on the preset correlation threshold to test the reliability of the results. The preset correlation threshold is 0.8, and a value greater than 0.8 is a strong correlation;

[0039] Among them, the coefficient of determination R 2 is obtained through formula (1), where y i is the actual value, is the predicted value, and is the mean value.

[0040]

[0041] S3: Establish a landslide prediction model based on the random forest algorithm. In a joint training manner, input static factors, dynamic factors, and landslide catalogs into the landslide prediction model for training, specifically including:

[0042] Divide the sample area into several grid cells according to a resolution of 30m, and establish a dataset based on the grid cells. The dataset includes a feature matrix and a landslide catalog. Among them, the process of obtaining the feature matrix is as follows: overlay the landslide samples and non-landslide samples with the static factor and dynamic factor layers, and extract the factor values of each grid cell as the feature matrix;

[0043] Divide the grid cells into a 70% training set and a 30% validation set. With the feature matrix as the input and the landslide samples / non-landslide samples as the output, train the landslide prediction model based on the training set and verify it based on the validation set;

[0044] S4: Obtain the environmental static factors and dynamic factors of the area to be measured, input the environmental static factors and dynamic factors of the area to be measured into the trained landslide prediction model, obtain the landslide susceptibility evaluation index, and determine the prediction result according to the evaluation index. In order to visualize the result, use the predicted landslide susceptibility index as the evaluation index, and its value range is from 0 to 1. Among them, the larger the value, the higher the possibility of landslide occurrence; use the natural break point method to divide the landslide susceptibility index into different levels, and select corresponding color blocks based on different levels to mark on the map for visual mapping.

[0045] In this solution, the surface deformation rate is incorporated as a dynamic factor into the landslide susceptibility evaluation, providing a scientific basis for landslide prevention and disaster reduction. By obtaining the ascending orbit deformation rate and descending orbit deformation rate of the surface and integrating the surface deformation rate into the static factor dataset, after completing unified data preprocessing, the static factor and dynamic factor data are input into the landslide prediction model for joint training and evaluation, which has the advantage of high accuracy.

[0046] The random forest model integrates the voting results of multiple decision trees to obtain the optimal classification result. Each decision tree in the model depends on randomly selected feature vectors, and the distribution of these trees in the RF is consistent. During the classification prediction process, the RF constructs several decision trees, and each decision tree participates in the voting to select the best classification. Finally, the final classification is determined through the voting mechanism. The RF model enhances the difference between individual classification trees by randomly extracting a specific number of feature vectors during node classification, improving the model performance and avoiding overfitting. The basic principle formula is as follows:

[0047]

[0048] In Equation (2), f(x) is the prediction result of the RF for the input sample x; k is the number of decision trees; t i (x) is the prediction result of the i-th decision tree for the input sample x.

[0049] Comparative Example 1

[0050] Different from Example 1, in this comparative example, a landslide prediction model is established based on the extreme gradient boosting model.

[0051] The extreme gradient boosting model iteratively integrates decision trees, and each new tree fits the residuals of the previous model, thereby gradually improving the prediction accuracy of the model. Its basic principle formula is as follows:

[0052]

[0053] In Equation (3) represents the predicted value of the final tree ensemble; represents the predicted value of the previous tree ensemble; x i represents the eigenvalue corresponding to sample i; f t (x i ) is the predicted value of the newly generated tree; t is the total number of basic tree models.

[0054] The objective function of the extreme gradient boosting model consists of two parts: the loss function and the regularization term. The formula is as follows:

[0055]

[0056] In Equation (4), L (t) represents the objective function; represents the loss function; Ω(f t ) represents the regularization term; n represents the number of samples; y i represents the true value of i; T represents the number of leaf nodes; γ and λ are parameters; w j represents the predicted value of the j-th leaf node of the decision tree in the t-th round.

[0057] Comparative Example 2

[0058] Different from Example 1, in this comparative example, in steps S3 and S4, a weighted superposition method is used to input the static factors and dynamic factors into the landslide prediction model. Specifically, the analytic hierarchy process is used to assign corresponding weights to the static factors and dynamic factors, and the effective coupling of the static factors and dynamic factors is completed through the weighted superposition method.

[0059] Comparative Example 3

[0060] Different from Comparative Example 2, this comparative example establishes a landslide prediction model based on the extreme gradient boosting model.

[0061] Based on the above Example 1 and Comparative Examples 1-3, the present invention evaluates the trained landslide prediction model through evaluation indicators. The sample area is Jinping Miao, Yao and Dai Autonomous County, which is located in the south of Honghe Autonomous Prefecture, Yunnan Province, China, and its geographical coordinates are 22°26′40″ - 23°04′32″ north latitude and 102°31′33″ - 103°38′44″ east longitude.

[0062] The evaluation indicators include: ROC curve, AUC value, recall rate, precision rate, F1 value, mean value (MV) and standard deviation (SD).

[0063] Among them, the ROC curve is widely used in the binary classification problem in the landslide field because it can comprehensively evaluate the performance of the classifier. The ROC curve is drawn based on two coordinate axes of specificity FRP and sensitivity TPR. Among them, specificity represents the proportion of misclassified landslides, while sensitivity represents the proportion of correctly classified landslides. The calculation methods of these two indicators are given by equations (5) and (6) respectively. In the two formulas, TP represents the number of samples correctly identified as landslides; TN represents the number of samples correctly identified as non-landslides; FP represents the number of non-landslide samples misidentified as landslides; FN represents the number of landslide samples misidentified as non-landslides.

[0064]

[0065] The AUC value, that is, the area under the ROC curve, reflects the probability that a randomly selected positive sample ranks higher than a randomly selected negative sample. The closer the AUC value is to 1, the better the prediction performance of the model. Its calculation formula is (7). In formula (7), n0 represents the number of negative samples, n1 represents the number of positive samples, and r i represents the ranking position of the i-th negative sample in the entire test sample set.

[0066]

[0067] Recall, which represents the proportion of positive samples correctly identified by the model among all actual positive samples, is calculated as shown in Equation (8).

[0068]

[0069] Precision represents the proportion of positive samples actually being positive among the positive samples predicted by the model.

[0070] It is calculated as shown in Equation (9).

[0071]

[0072] Since there is usually a trade-off between recall and precision and it is difficult to achieve the optimal values simultaneously, in practical applications, the F1 value is usually used to comprehensively consider these two metrics. The F1 value is the harmonic mean of precision and recall and is calculated as shown in Equation (10).

[0073]

[0074] The mean value (MV) is a metric for measuring the central tendency of data and is calculated as shown in Equation (11), where n is the total number of grid cells, and X n is the landslide susceptibility index value of the nth grid cell, which reflects the overall trend and average level of the landslide susceptibility index. The standard deviation (SD) is a metric for measuring the dispersion degree of data distribution and is calculated as shown in Equation (12), where n is the total number of grid cells, and X i is the landslide susceptibility index value of the ith grid cell, which reflects the variability and discriminability of the landslide susceptibility index.

[0075]

[0076] The evaluation results are as follows: The ROC curve and the AUC value are as Figure 2 shown, and the other precision metrics are shown in Table 1, where JTRF is the random forest model based on joint training in Example 1, JTXGBoost is the extreme gradient boosting model based on joint training in Comparative Example 1, WSRF is the random forest model based on weighted overlay in Comparative Example 2, and WSXGBoost is the extreme gradient boosting model based on weighted overlay in Comparative Example 3.

[0077] Table 1. Experimental data of each precision metric

[0078]

[0079] According to Figure 2It can be seen that the AUC value of the random forest model based on joint training is the highest. According to Table 1, the statistical results of precision, recall, F1 value, mean, and standard deviation accuracy indicators show that the F1 value of JTRF is relatively high, and the JTRF model has the lowest mean, the highest standard deviation, and the highest F1 value, indicating that the model performs outstandingly in terms of prediction accuracy and integrity.

[0080] The above-described embodiments merely represent several implementation manners of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A method for landslide susceptibility prediction using a machine learning model and SBAS-InSAR deformation, characterized in that: Including the following steps: Obtain the environmental static factors, dynamic factors, and landslide catalog within the sample area, where the dynamic factors include the surface deformation rates under ascending and descending orbits within the sample area; Establish a landslide prediction model based on the random forest algorithm, and input the static factors, dynamic factors, and landslide catalog into the landslide prediction model for training; Obtain the environmental static factors and dynamic factors of the area to be measured, input the environmental static factors and dynamic factors of the area to be measured into the trained landslide prediction model, obtain the landslide susceptibility evaluation index, and determine the prediction result according to the evaluation index.

2. A method for landslide susceptibility prediction using a machine learning model and SBAS-InSAR deformation according to claim 1, characterized in that: The environmental static factors include: elevation, slope, aspect, curvature, distance from road, land use, distance from fault, lithology, normalized difference vegetation index, distance from river, annual rainfall, and terrain wetness index.

3. A method for predicting landslide susceptibility using a machine learning model and SBAS-InSAR deformation, as described in claim 1, characterized in that: The dynamic factors are obtained based on the SBAS-InSAR technology, and the obtaining process includes: Obtain several ascending-orbit SAR images and descending-orbit SAR images of the sample area; Process the ascending-orbit SAR images and descending-orbit SAR images to obtain several groups of ascending-orbit interferograms and descending-orbit interferograms; Process the ascending-orbit interferograms and descending-orbit interferograms based on a preset screening rule to obtain ascending-orbit PS points and descending-orbit PS points; Unwrap the interferometric phase of the ascending-orbit PS points and descending-orbit PS points based on 3D unwrapping technology to obtain the ascending-orbit surface deformation rate and descending-orbit surface deformation rate.

4. A method for predicting landslide susceptibility using a machine learning model and SBAS-InSAR deformation according to claim 3, characterized in that: After obtaining the dynamic factors, the method further includes verifying the result reliability of the ascending-orbit deformation rate and descending-orbit deformation rate: Randomly select several points with the same longitude and latitude as homologous points, and respectively extract the deformation rates of the homologous points in the ascending orbit radar line-of-sight and descending orbit radar line-of-sight directions. Take the ascending orbit deformation rate as the independent variable x and the descending orbit deformation rate as the dependent variable y to establish a linear regression model for linear fitting, and use the coefficient of determination R 2 to evaluate the correlation between the ascending orbit deformation rate and the descending orbit deformation rate, and test the reliability of the results based on a preset correlation threshold.

5. A method for predicting landslide susceptibility using a machine learning model and SBAS-InSAR deformation according to claim 3, characterized in that: The process of processing the ascending-orbit SAR images and descending-orbit SAR images is specifically: set a time baseline threshold and a spatial baseline threshold, and remove the interferograms with poor coherence and obvious errors with respect to the baseline threshold; The preset screening rule is: by setting a coherence threshold, screen out the PS points of the ascending / descending orbit interferograms higher than the coherence threshold.

6. A method for predicting landslide susceptibility using a machine learning model and SBAS-InSAR deformation, as claimed in claim 1, wherein: The landslide catalog includes historical landslide samples and non-landslide samples.

7. A method for predicting landslide susceptibility using a machine learning model and SBAS-InSAR deformation, as claimed in claim 6, characterized in that: The process of inputting the static factors, dynamic factors, and landslide catalog into the landslide prediction model for training is: Divide the sample area into several grid cells according to a preset resolution, establish a data set based on the grid cells, the data set includes a feature matrix and a landslide catalog, and the process of obtaining the feature matrix is: overlay the landslide samples and non-landslide samples with the static factor and dynamic factor layers, and extract the factor values of each grid cell as the feature matrix; Divide the grid cells into a training set and a validation set, use the feature matrix as the input and the landslide samples / non-landslide samples as the output, train the landslide prediction model based on the training set, and verify based on the validation set.

8. A method for predicting landslide susceptibility using a machine learning model and SBAS-InSAR deformation, as claimed in claim 1, wherein: The obtaining of the landslide susceptibility evaluation index and determining the prediction result according to the evaluation index include: Use the predicted landslide susceptibility index as the evaluation index, and its value range is from 0 to 1. Among them, the larger the value, the higher the possibility of landslide occurrence; The landslide susceptibility index is divided into different levels by using the natural breakpoint method, and corresponding color blocks are selected based on different levels for annotation on the map for visual mapping.

Images