KR102145250B1 - Method and apparatus for landslide susceptibility mapping using machine-learning architecture - Google Patents

Abstract

In the method of creating a landslide vulnerability map of the present invention, receiving data from a preliminary survey result for a target area for creating a landslide vulnerability map, selecting an environmental factor necessary to calculate the landslide vulnerability, and the preliminary survey result data and the above Considering environmental factors, establishing a landslide vulnerability model using a machine learning algorithm, and creating a landslide vulnerability map of the target area based on this.
According to the present invention, it is possible to improve accuracy and prediction performance by applying various machine learning techniques in creating a landslide vulnerability map.

Description

A method and apparatus for creating a landslide vulnerability mapping using machine learning techniques {Method and apparatus for landslide susceptibility mapping using machine-learning architecture}

The present invention relates to a landslide-related mapping technology, and more particularly, to a landslide vulnerability mapping technology using machine learning technology.

Landslides are very complex natural phenomena that cause significant damage to people, property and transportation networks. Landslides generally occur in mountainous areas due to steep slopes. Korea has a large mountainous area that accounts for about 70% of the total area. In particular, mountainous areas composed of granite or gneiss have a steep slope and are most vulnerable to landslides.

In Korea, average monthly rainfall during the rainy season is recorded as 300 mm in July and 280 mm in August. These conditions not only have low soil strength due to weathering and unstable slopes, but also high rainfall during the rainy season has a significant impact on slope stability and susceptibility to landslides. Crazy. In response, many government agencies have sought solutions to mitigate the disastrous consequences of landslides by educating people about the serious impacts of landslides and by developing appropriate planning and decision-making tools. This process generally involves identifying and mapping areas where a landslide may occur based on a landslide vulnerability assessment that evaluates the spatial distribution of the probability of a landslide occurring in a given area based on regional geographic and environmental factors. Include.

Landslide vulnerability analysis is a method of predicting the possibility of landslides by considering rainfall and earthquakes that directly cause landslides, and analyzing the spatial distribution of factors related to the occurrence of landslides.

Landslide vulnerability analysis methods include statistical analysis methods and geotechnical analysis methods. The statistical analysis method is a method of analyzing vulnerability by considering the statistical correlation between the data of various landslide triggers in a wide area and the location of the landslide. In the case of statistical analysis, it is limitedly applicable to the area where landslides occurred in the past, and has the disadvantage of insufficient consideration of the engineering properties of slope materials.

On the other hand, the geotechnical analysis method assumes a landslide model regardless of the occurrence of a landslide, and analyzes the stability of the slope by considering the geometric characteristics of the slope and the geotechnical characteristics of the slope material, which can take into account the mechanism and process of the landslide. It is one of the effective vulnerability analysis.

In the case of the geotechnical analysis method, the Infinite slope model is most commonly used in order to consider the physical slope model, the shape of the landslide fracture surface is similar and easier to analyze than other models. In particular, recently, researches have been actively conducted to perform vulnerability analysis on a wide area by combining an infinite slope model and a geographic information system (GIS).

However, the existing landslide vulnerability analysis method using the infinite slope model has limitations in considering the hydraulic characteristics of a wide area research area. In other words, in the existing technology, the groundwater level in the slope was assumed to be an arbitrary value and used for analysis without considering the hydraulic characteristics of the ground or the effect of rainfall. In this case, information related to rainfall characteristics such as rainfall intensity and rainfall is not considered at all, and it cannot be considered that the groundwater level is fluidly saturated according to the hydrogeological characteristics of the study area during rainfall.

Republic of Korea Patent Publication 10-2015-0131801

The present invention has been conceived to solve the above problems, and an object of the present invention is to provide a method for creating a map of a landslide vulnerability using a machine learning technique in order to predict and prevent a landslide risk in advance.

The object of the present invention is not limited to the above-mentioned object, and other objects not mentioned will be clearly understood by those skilled in the art from the following description.

In the method of creating a landslide vulnerability map according to the present invention for achieving the above object, the step of receiving data from a preliminary survey result for a target area to be prepared for the landslide vulnerability map, selecting an environmental factor necessary to calculate the landslide vulnerability, and And establishing a landslide vulnerability model using a machine learning algorithm in consideration of the preliminary survey result data and the environmental factors, and creating a landslide vulnerability map of the target area based on this.

In the step of selecting the environmental factor, the slope, side, maximum curvature, profile curvature, convexity, texture, mid-slope position (MSP), topographic position index (TPI) of the target area, Terrain ruggedness index (TRI), flow accumulation, stream power index (SPI), topographic wetness index (TWI), surface marker, soil forest type, forest age, forest density, forest diameter, geology, Environmental factors including the distance from the fault can be selected, and a spatial database for each environmental factor can be constructed.

In the step of creating the landslide vulnerability map in an embodiment of the present invention, a landslide vulnerability model may be established using the AdaBoost algorithm, and a map of the landslide vulnerability may be created based on this.

In the step of creating the landslide vulnerability map in an embodiment of the present invention, a landslide vulnerability model may be established using the LogitBoost algorithm, and a map of the landslide vulnerability in a target area may be created based on this.

In the step of creating the landslide vulnerability map in an embodiment of the present invention, a landslide vulnerability model may be established using a multiclass classifier algorithm, and a map of the landslide vulnerability in a target area may be created based on this.

In an embodiment of the present invention, in the step of creating the landslide vulnerability map, a landslide vulnerability model may be established using a Bagging algorithm, and a map of the landslide vulnerability in a target area may be created based on this.

In the step of creating the landslide vulnerability map in an embodiment of the present invention, a landslide vulnerability model may be established using an artificial neural network (ANN) algorithm, and a map of the landslide vulnerability in a target area may be created based on this.

In the step of creating the landslide vulnerability map in an embodiment of the present invention, a landslide vulnerability model may be established using a Support Vector Machine (SVM) algorithm, and a map of the landslide vulnerability in a target area may be created based on this.

The apparatus for creating a landslide vulnerability map according to the present invention includes a data input unit for providing an interface for receiving data from a preliminary survey result on a target area for creating a landslide vulnerability map, and an environmental factor for selecting environmental factors necessary to calculate the landslide vulnerability. And a map preparation unit that establishes a landslide vulnerability model using a machine learning algorithm, taking into account the determination unit and the preliminary survey result data and the environmental factors, and creates a landslide vulnerability map of a target area based on this.

The environmental factor determination unit includes the slope, side, maximum curvature, profile curvature, convexity, texture, mid-slope position (MSP), topographic position index (TPI), and terrain ruggedness (TRI) of the target area. index), flow accumulation, stream power index (SPI), topographic wetness index (TWI), surface marker, soil forest type, forest age, forest density, forest diameter, geology, distance from fault It is possible to select environmental factors including, and build a spatial database for each environmental factor.

The map maker may establish a landslide vulnerability model using the AdaBoost algorithm, and create a landslide vulnerability map of a target area based on this.

The map preparation unit may establish a landslide vulnerability model using the LogitBoost algorithm, and create a landslide vulnerability map of a target area based on this.

The map preparation unit may establish a landslide vulnerability model using a multiclass classifier algorithm, and create a landslide vulnerability map of a target area based on this.

The map preparation unit may establish a landslide vulnerability model using the Bagging algorithm, and create a landslide vulnerability map of a target area based on this.

The map preparation unit may establish a landslide vulnerability model using an artificial neural network (ANN) algorithm, and create a map of the landslide vulnerability in a target region based on this.

The map preparation unit may establish a landslide vulnerability model using a Support Vector Machine (SVM) algorithm, and create a landslide vulnerability map of a target area based on this.

According to the present invention, it is possible to improve accuracy and prediction performance by applying various machine learning techniques in creating a landslide vulnerability map. Therefore, it is possible to predict and prevent landslides in advance through the landslide vulnerability map of the present invention, and there is an effect of preventing national disasters in advance.

In addition, it is expected that the technique of creating a map of a landslide vulnerability of the present invention will be very useful in an area prone to a landslide.

1 is a block diagram conceptually showing the configuration of an apparatus for creating a landslide vulnerability map using a machine learning technique according to an embodiment of the present invention.
2 is a flowchart illustrating a method of creating a landslide vulnerability map using a machine learning technique according to an embodiment of the present invention.
3 is a diagram showing the location of a study area according to an embodiment of the present invention on a Google map.
4 is a satellite image before and after a landslide occurs in a study according to an embodiment of the present invention.
5 shows a digital altitude model and a landslide occurrence in a study area according to an embodiment of the present invention.
6 to 8 show spatial databases according to the cause of landslides in a study according to an embodiment of the present invention.
9 is a map of a landslide vulnerability in a study area according to each machine learning technique according to an embodiment of the present invention.
10 is a graph showing the distribution of vulnerability ratings in a landslide vulnerability map composed of four machine learning models in a study according to an embodiment of the present invention.
11 is a diagram showing the area under the curve (AUC) of four machine learning models in a study according to an embodiment of the present invention.
12 is a graph showing a receiver operating characteristic (ROC) curve of a landslide vulnerability map using four algorithms of a meta classifier in a study according to an embodiment of the present invention.
13 is a graph showing a receiver operating characteristic (ROC) curve of a landslide vulnerability map generated using a conventional algorithm and an AUC test result variable.

In the present invention, various modifications may be made and various embodiments may be provided, and specific embodiments will be illustrated in the drawings and described in detail. However, this is not intended to limit the present invention to a specific embodiment, it is to be understood as including all changes, equivalents, and substitutes included in the spirit and scope of the present invention.

The terms used in the present application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In the present application, terms such as "comprise" or "have" are intended to designate the presence of features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and one or more other features It is to be understood that the presence or addition of elements or numbers, steps, actions, components, parts, or combinations thereof does not preclude in advance.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person of ordinary skill in the art to which the present invention belongs. Terms as defined in a commonly used dictionary should be interpreted as having a meaning consistent with the meaning in the context of the related technology, and should not be interpreted as an ideal or excessively formal meaning unless explicitly defined in this application. Does not.

In addition, in the description with reference to the accompanying drawings, the same reference numerals are assigned to the same components regardless of the reference numerals, and redundant descriptions thereof will be omitted. In describing the present invention, when it is determined that a detailed description of a related known technology may unnecessarily obscure the subject matter of the present invention, a detailed description thereof will be omitted.

The ensemble technique is a machine learning technique in which a predictive model is formed by a combination of various basic classifiers. Up to now, various ensemble frameworks such as stacking, random subspace, random forests, rotation forests, Bagging, AdaBoost, and MultiBoost have been proposed.

The present invention proposes a method and apparatus for predicting landslide vulnerability using machine learning, including an ensemble-based machine learning technique, in order to improve the performance of a predictive model.

1 is a block diagram conceptually showing the configuration of an apparatus for creating a landslide vulnerability map using machine learning according to an embodiment of the present invention.

Referring to FIG. 1, an apparatus for mapping a landslide vulnerability using machine learning according to an embodiment of the present invention includes a data input unit 110, an environmental factor determination unit 120, and a map preparation unit 130.

The data input unit 110 provides an interface necessary for receiving data from a landslide survey result.

The environmental factor determination unit 120 determines an environmental factor required to determine the vulnerability of a landslide.

The map preparation unit 130 establishes a landslide vulnerability model using a machine learning algorithm in consideration of the landslide survey result data and environmental factors, and creates a landslide vulnerability map based on this.

In an embodiment of the present invention, the environmental factor determination unit 120 includes a slope, a side surface, a maximum curvature, a profile curvature, a convexity, a texture, a mid-slope position (MSP), and a terrain position index ( topographic position index, TPI), terrain ruggedness index (TRI), flow accumulation, stream power index (SPI), topographic wetness index (TWI), surface marker, soil forest type, forest age, forest density, forest Environmental factors including diameter, geology, and distance from faults can be selected, and a spatial database for each environmental factor can be constructed.

In one embodiment of the present invention, the map creation unit 130 includes various machine learning such as AdaBoost algorithm, LogitBoost algorithm, Multiclass Classifier algorithm, Bagging algorithm, Artificial Neural Network (ANN) algorithm, and Support Vector Machine (SVM) algorithm. -learning) algorithm is used to establish a landslide vulnerability model, and based on this, a landslide vulnerability map can be created.

2 is a flowchart illustrating a method of creating a landslide vulnerability map using a machine learning technique according to an embodiment of the present invention.

Referring to FIG. 2, first, an investigator selects an area to create a landslide vulnerability map, and conducts an investigation on a landslide in the selected area.

In addition, the landslide vulnerability map preparation device receives the pre-investigation result data of the landslide (S210).

Then, the landslide vulnerability mapping apparatus determines an environmental factor necessary to determine the landslide vulnerability (S220).

In addition, the landslide vulnerability mapping apparatus establishes a landslide vulnerability model using a machine learning algorithm, taking into account the landslide survey result data and environmental factors, and creates a landslide vulnerability map based on this (S230).

In an embodiment of the present invention, in step S220, the slope, side, maximum curvature, profile curvature, convexity, texture, mid-slope position (MSP), topographic position index (TPI) of the target area ), TRI (Terrain ruggedness index), flow accumulation, SPI (stream power index), topographic wetness index (TWI), surface marker, soil forest type, forest age, forest density, forest diameter, geology ), environmental factors including distance from faults can be selected, and a spatial database for each environmental factor can be constructed.

In an embodiment of the present invention, in step S230, various machine-learning algorithms such as AdaBoost algorithm, LogitBoost algorithm, Multiclass Classifier algorithm, Bagging algorithm, Artificial Neural Network (ANN) algorithm, and Support Vector Machine (SVM) algorithm Using this, a landslide vulnerability model can be established, and a map of the landslide vulnerability in the target area can be created based on this.

The ANN algorithm is one of the widely used classifiers, and has a characteristic consisting of one input layer, at least one input layer hidden layer, and one output layer. In addition, each layer is a nonlinear processing unit, neurons and connections In a continuous layer, weights are assigned between neurons, and only transfer connections are allowed.

The ANN algorithm is not uniquely specified according to its own processing unit, or the characteristics of a selected training rule or running rule. And, the network topology (i.e., the number of hidden layers and units and their interconnections) also affects the classifier *?* performance.

And, SVM is a state-of-the-art machine learning algorithm, based on statistical learning theory, and was developed by Vapnik.

The SVM algorithm was first used to classify hyperspectral images, but later it was used to classify multispectral images.

SVM separates the classes using a crystal surface that maximizes the difference between the classes. For a linear case, these surfaces are called optimal hyperplanes that allow the maximum difference between the two classes, and the data points closest to the surface are called Support vectors (SVs). These vectors are a key element of the educational set. When implementing SVM, it is important to choose the appropriate kernel. The SVM classifier provides four types of common kernel: linear, polynomial, radial basis function (RBF) and sigmoid kernel.

Since the RBF kernel works well in most cases and performs excellent nonlinear classification, it is preferable to use this kernel in the embodiment of the present invention.

In the present invention, a certain area is selected, a landslide-related survey is performed on the area, and a landslide vulnerability map is created based on this. Now, in the preparation of the landslide vulnerability map of the present invention, a process of conducting research on an actual area will be described as follows.

3 is a diagram showing the location of a study area according to an embodiment of the present invention on a Google map.

3, a study area according to an embodiment of the present invention is located in Sacheon-myeon, Gangneung-si, Gangwon-do (37° 47'50 "N, 128° 48'08"E). Gangneung has frequent landslides during typhoons and rainfall, and typhoon Rusa hit Gangneung due to storms and heavy rains on August 30 and 31, 2002. The daily rainfall in the Gangneung area was recorded as 609 mm, and the hourly rainfall was 80 mm. In addition, typhoons and natural disasters caused by typhoons and heavy rains resulted in 266 deaths and property damage of about 8 billion, of which 81 were killed in landslides and the collapse of sheer slopes. Among them, in the Gangneung area, Sacheon-myeon was one of the largest landslides.

In this study, landslides in Sacheon-myeon were analyzed using web-based aerial photography. The first step in any landslide vulnerability assessment consists of collecting existing information and data within the area of interest. Since the reliability and accuracy of the collected data affects the success of the applied method, this step is accepted as the most important part of landslide risk mitigation studies. Therefore, the relationship between landslide occurrence and the control parameters used is important for landslide sensitivity mapping.

Preparation of a landslide survey plan is considered a major step in the assessment of landslide susceptibility.

In this study, a landslide survey map was created based on the locations of 762 landslides identified in aerial photographs using images without ground control points (GCPs), and images taken after the landslide event were selected from each region. And it is applied to the characteristics of digital features using ArcMap ver.

4 is a satellite image before and after a landslide occurs in a study according to an embodiment of the present invention.

In FIG. 4, a, c, e, and g are satellite images before a landslide occurs, and b, d, f, and g are satellite images after a landslide occurs.

5 shows a digital altitude model and a landslide occurrence in a study area according to an embodiment of the present invention.

Referring to FIG. 5, the spatial distribution of landslides was determined using a remote sensing and spatial analysis method of a geographic information system (GIS). The detected landslide sites were displayed on aerial photographs, and the landslide sites were entered into digital elevation model (DEM) data for spatial processing.

[Table 1]

Environmental factors are important when determining the vulnerability of a landslide.

Referring to Table 1, in the study of the present invention, 20 conditional landslides were considered. These factors are important controls, but in practice it is difficult to determine the spatial distribution. The locations of landslides and environmental factors are expressed in 30-m × 30-m pixels, and terrain and hydrological elements are constructed from DEM data using the topographic analysis of the SAGA GIS module. Meanwhile, soil, forest and geological factors were extracted from soil, forest and geological maps, respectively.

6 to 8 show spatial databases according to the cause of landslides in a study according to an embodiment of the present invention.

In Figures 6 to 8, slope (a), side (b), maximum curvature (c), profile curvature (d), convexity (e), texture (f), mid-slope position (MSP) (g), topographic position index (TPI)(h), TRI(Terrain ruggedness index)(i), flow accumulation(j), SPI(stream power index)(k), topographic wetness index, TWI)(l), surface marker(m), soil(n) type of forest(o), forest age(p), forest density(q), forest diameter(r), geology(s) , The distance t from the fault is shown.

Topography and hydrology influence the initiation of debris flow through gradient effects on slope stability with rainfall. These factors also determine the concentration of the material and the mass balance of the gradient related to the dispersion and gradient stability.

In FIGS. 6 to 8 (a) to (i), the extracted terrain coefficients are slope, surface, maximum curvature, profile curvature, convex surface, texture, intermediate slope position (MSP), terrain position index (TPI), and terrain. It is a durability index (TRI). Here, the slope means the slope of the hill, and the side means the steepest downhill road. Since the MSP is correlated with the slope, a value of 1 is designated as the maximum vertical distance from the middle slope to the valley direction. And, the curvature function including the maximum curvature and the profile curvature indicates the shape of the terrain. Here, some of the surface may be concave or convex, which affects the divergence and convergence of the flow across the surface. Convexity is described as a positive surface curvature and represents the proportion of convex-upward cells. Texture or terrain surface texture correlates with TRI and TPI. TRI is used to express the amount of elevation difference between the surface and surrounding areas, including concave and convex areas, and TPI is an algorithm used to measure the location of the terrain slope.

The hydrological elements extracted from (j) to (l) of FIGS. 6 to 8 are flow accumulation, river power index (SPI), and terrain wetness index (TWI). The flow (flow) accumulation defines the flow accumulation value through the valley, and the valley shape of the watershed is obtained from the flow accumulation. SPI refers to the erosion power of water flows, and TWI is a steady state wetting index commonly used to quantify topographic control for hydrological processes.

In Table 1, the attribute column of digital land signs can be represented by the land use shown in (m). Land use was classified as lowland and settlement, grass, farm, forest, rice field and road.

In the soil map of FIGS. 6 to 8 (n), the grade of soil thickness is divided into four grades. That is, it is divided into very shallow grade (0-19 cm), shallow grade (20-49 cm), medium grade (50-99 cm) and deep grade (> 100 cm). Slope soil affects the spatial distribution of debris. These factors are important control factors and can be represented as spatial distributions in digital elevation models (DEM) and soil maps and surface cover maps.

In addition, the strength of plant roots substantially affects the occurrence of slope failure. Accordingly, forest factors such as forest type, forest diameter, forest density, and forest age in FIGS. 6 to 8 (o) to (r) have a close relationship with the strength of soil root bonds. This means that high-density wood maintains water pressure and has a high capacity to sink the soil in heavy rains. Older forests, generally composed of trees with large diameters, tend to have strong roots that contribute to slope stability. Root strength and wood diameter are affected by forest type or existing tree type.

And, since various limestone units have different sensitivity to active topographic processes such as landslides, lime plays an important role in the occurrence of landslides.

Therefore, as shown in (s) in FIGS. 6 to 8, limestone in this study is regarded as a geological parameter.

And, in (t), strata is also considered to be an important factor causing landslides in tectonic regions. Therefore, the strength of the fracture and shear stress decisively affects the slope instability, so the relationship between the linearity and the occurrence of a landslide is investigated by the distance to the defect.

In the study of the present invention, in order to construct a landslide model and evaluate its performance, the landslide inventory and 20 condition factor maps were converted into ASCII format with a cell size of 30 m. In addition, all determined landslide locations in the ArcGIS version were divided into a training set (70%) and a verification set (30%). In addition, since the number of landslide pixels is much smaller than the total number of pixels in the irradiated area, the undersampling method was used. Therefore, the number of landslide pixels other than landslide was randomly extracted from the free landslide area. Here, a value of 1 is assigned to a landslide pixel, and a value of 0 is assigned to a non-landslide pixel. Finally, landslide control factor values were extracted to create a landslide condition coefficient training and validation data set.

Weka was used as an open source tool in this study. Weka is a data mining tool that supports easy application of learning algorithms to data sets. Weka consists of a variety of machine learning algorithms for various data mining applications. So, without writing program code, you can pre-process the data set, place it in a learning system, and analyze the generated classes and performance. And, all algorithms used a single relational table in ARFF format as input. The ARFF format was derived from the values of landslide control factors collected from statistical files and converted to ARFF format using Weka. The main focus of this study is the performance and evaluation of the Meta (i.e., ensemble learning) algorithm for landslide vulnerability analysis. In the study of the present invention, a training set was used as a test option for classifying ARFF data and executing an ensemble algorithm.

In this study, the four ensemble learning algorithms used to establish an ensemble model of landslide susceptibility such as AdaBoost, LogitBoost, Multiclass Classifier and Bagging are briefly described as follows.

AdaBoost ("adaptive boosting") is a machine learning ensemble algorithm proposed by Freund and Schapire (1997). AdaBoost is one of the most widely used amplification algorithms in which a series of individual classifiers are iteratively generated, and each classifier in an ensemble tries to accurately classify training data. The classifier selects training samples using an adaptive resampling technique. For example, misclassified data sets created by previous classification criteria are selected more often than correctly classified classifications, so new classifications can perform better on new data sets. Each iteration assigns a weight to the data set, so the next integration focuses on the previously misclassified reconstructed data set. The final model is obtained from the sum of the weights of all classification criteria models. AdaBoost can also be used to evaluate the relevance of variables by examining the frequency of choices made by weak learners. In this study, Weka was used to process the AdaBoost model. The class used for boosting is a nominal class classifier using the AdaBoost method, and only nominal class problems can be considered. This method greatly improves performance, but sometimes overfitting.

In the field of machine learning and computational learning theory, LogitBoost is a boosting algorithm formulated by Friedman, Hastie and Tibshirani. The AdaBoost model was regarded as a generalized additional model to which the cost function of logistic regression was applied, and therefore the LogitBoost algorithm is an extension of the AdaBoost algorithm and changes the exponential loss of the AdaBoost algorithm to a conditional Bernoulli-likelihood loss. LogitBoost performs classification using a regression method as a basic learner, and uses additional logistic regression for classification that can handle multi-class problems.

In machine learning, multiclass or multiclass classifier solves the problem of classifying instances into one of three or more classes, unlike binary classification, which classifies instances into one of two classes. Some classification algorithms can of course use more than one class, while others are essentially binary algorithms. However, they can be transformed into polynomial classifiers according to various strategies. Multi-class classification should not be confused with multi-label classification, in which multiple labels are predicted for each instance, and is a meta classifier capable of handling multi-class data sets with two class classifiers.

Bagging is a machine learning ensemble meta-algorithm designed to improve the stability and accuracy of machine learning algorithms used for statistical classification and regression. Bagging was one of the earliest ensemble methods using bootstrap sampling techniques. The bootstrap technique involves performing random sampling with substitution to generate several samples that form a training set. Each generated subset is used to construct the decision tree and later aggregated into the final model. This improves classification accuracy by reducing the variance of classification errors. In this study, the use of Bagging proved to be useful for a landslide vulnerability model for the sensitivity to small changes in training data, resulting in a more robust and accurate landslide model, and improved predictive ability. Bagging can use classifiers to reduce variance and perform classification and regression according to basic learners.

In this study, a landslide model was constructed using a training data set. The performance of the landslide model is highly dependent on the choice of computational parameters. Therefore, optimization was carried out to obtain the optimal parameters for obtaining the optimal landslide model performance, and Table 2 below lists the optimal parameters for the landslide model in this study.

[Table 2]

9 is a map of a landslide vulnerability in a study area according to each machine learning technique according to an embodiment of the present invention.

In FIG. 9, (a) is AdaBoost, (b) is LogitBoost, (c) is a multiclass classifier, and (d) is a map of landslide vulnerability based on the Bagging model.

Landslide vulnerability maps were created after successful completion of the model training course. First, a landslide sensitivity index was generated for every pixel in the study area, and each pixel was assigned a unique vulnerability index. Then, the vulnerability index was reclassified using the quantile method based on finding adjacent feature pairs with relatively large differences. Based on this reclassification of the Landslide Sensitivity Index, the Landslide Vulnerability Map uses five vulnerability ratings: Very low, Low, Moderate, High, and Very high. Was constructed.

10 is a graph showing the distribution of vulnerability ratings in a landslide vulnerability map composed of four machine learning models in a study according to an embodiment of the present invention. 10 shows the distribution of vulnerability levels in a landslide vulnerability map.

Referring to FIG. 10, in the landslide vulnerability map produced using the AdaBoost model, 59.65% of the surveyed area had very low vulnerability to landslides, and 26.23%, 0.33%, and 13.79% of the area were low light, medium and high, respectively. Showed vulnerability. The AdaBoost model can only classify 4 classes. Therefore, very high vulnerability ratings were excluded.

In a landslide vulnerability map based on the LogitBoost model, 73.43% of the target area is very vulnerable to landslides, while 15.05% indicates low vulnerability, 2.01% indicates moderate vulnerability, 6.43% indicates high vulnerability, and 3.08% indicates very high vulnerability. Shows high vulnerability.

In the landslide vulnerability map based on the Multiclass Classifier model, 70.38% of the target area is very vulnerable to landslides, while 14.76% of the area indicates low vulnerability, 7.42% indicates moderate vulnerability, and 3.46% indicates high vulnerability. 3.98% indicates a very high vulnerability.

Finally, in the landslide vulnerability map based on the Bagging model, 77.45% of the target area is very vulnerable to landslides, while 11.43% of the area indicates low vulnerability, 5.91% indicates moderate vulnerability, and 2.39% indicates high vulnerability. And 2.82% indicates a very high vulnerability.

Landslide vulnerability maps should be able to effectively predict future landslide areas and validate when existing landslide location data and landslide locations are combined. Therefore, the results of the analysis of landslide vulnerability can be confirmed using the validation data. Landslide vulnerability maps generated by various meta-classifier models were verified by comparing the validation data consisting of 229 landslides (30% of total landslides).

11 is a diagram showing the area under the curve (AUC) of four machine learning models in a study according to an embodiment of the present invention.

Referring to FIG. 11, an area under the curve (AUC) is used to quantitatively compare model performance for four models, and a model having the highest AUC as a result of the comparison is considered to be the best model. And, the success rate curve is obtained by comparing the training data and the landslide vulnerability map. As a result, the Multiclass Classifier algorithm showed the highest AUC (0.859), followed by Bagging algorithm (0.854), LogitBoost algorithm (0.848), and AdaBoost algorithm (0.840). These results are shown in the graph of FIG. 12 below.

12 is a graph showing a receiver operating characteristic (ROC) curve of a landslide vulnerability map using four algorithms of a meta classifier in a study according to an embodiment of the present invention.

All models showed sufficient performance to be applied to spatial prediction of landslide risk in the study area, but it can be seen that the landslide vulnerability map generated by the Multiclass Classifier algorithm showed the best performance.

As described above, the new machine learning technique proposed in the present invention has proven to be effective in terms of prediction performance. Therefore, in the present invention, a method of applying four machine learning ensembles such as AdaBoost, LogitBoost, Multiclass Classifier, and Bagging to the evaluation of landslide vulnerability is proposed.

The research results of the present invention can confirm that the performance of the landslide model is improved using the machine learning ensemble. For comparison, a conventional model frequency ratio with an AUC of 0.817 was also considered, and a graph thereof is shown in FIG. 13 below.

13 is a graph showing a receiver operating characteristic (ROC) curve of a landslide vulnerability map generated using a conventional algorithm and an AUC test result variable. Here, the AUC of the frequency ratio method is 0.817.

In the study of the present invention, the prediction of the ensemble model improved by 2.3% in AdaBoost, 3.1% in LogitBoost, 3.7% in Bagging, and 4.2% in Multiclass Classifier. In particular, the Multiclass Classifier model showed the greatest performance improvement among the four machine learning ensembles. . This can be explained by the fact that feature extraction is used to optimize the training set used to learn the basic classifier, which improves the predictive function compared to other ensembles (AdaBoost, LogitBoost and Bagging).

The biggest advantage of the four machine learning algorithms is that they automate the process of examining multiple databases to gather valuable information. It can provide specific assumptions that can be used to support planning decisions along with automating the analysis of large data sets.

The present invention has been described above using several preferred embodiments, but these embodiments are illustrative and not limiting. Those of ordinary skill in the art to which the present invention pertains will understand that various changes and modifications can be made without departing from the spirit of the present invention and the scope of the rights set forth in the appended claims.

110 Data input section
120 Environmental factor determination section
130 Map Maker

Claims (16)

In the landslide vulnerability mapping method in the landslide vulnerability mapping device,
Receiving data as a result of a preliminary survey on a target area where a landslide vulnerability map should be created;
Selecting an environmental factor necessary to calculate a landslide vulnerability; And
In consideration of the pre-examination result data and the environmental factors, establishing a landslide vulnerability model using a machine learning algorithm, and creating a landslide vulnerability map of a target region based on this,
In the step of selecting the environmental factor,
Select environmental factors including topographic coefficients, hydrological factors, properties of land signs, soil thickness grades, forest factors, geological information, and strata information of the target area, and establish a spatial database for each environmental factor,
The terrain coefficients include slope, side, maximum curvature, profile curvature, convexity, texture, mid-slope position (MSP), topographic position index (TPI), and terrain durability index of the target area. (Terrain ruggedness index, TRI), the slope means the slope of the hill, the side surface means the steepest downhill road, the intermediate slope position represents the highest point value at the intermediate slope, the maximum curvature and the profile curvature Is a curvature function representing the shape of the topography, the convexity is described as a positive surface curvature and represents the ratio of convex-upward cells, and the texture is correlated with TRI and TPI, and TRI is Represents the amount of elevation difference between the surface and the surrounding area, including concave and convex, TPI represents the terrain slope location,
The hydrological element is a flow accumulation, a stream power index (SPI), and a topographic wetness index (TWI), and the flow accumulation represents a flow accumulation value through a valley, and the river power index ( SPI) represents the erosion power of the water flow, and Terrain Wetness Index (TWI) represents the steady state wetting index used to quantify terrain control for hydrological processes,
The property of the land sign indicates land use classified into settlement, grass, farm, forest, rice field, and road,
The soil thickness grade is divided into four grades: very shallow grade (0-19 cm), shallow grade (20-49 cm), medium grade (50-99 cm) and deep grade (> 100 cm),
The forest factors are forest type, forest diameter, forest density, and forest age,
The geological information is unit information of limestone,
The stratum information is the distance from the fault,
In the step of creating the landslide vulnerability map,
A landslide vulnerability map creation method characterized by establishing a landslide vulnerability model using a multiclass classifier algorithm, and creating a landslide vulnerability map of a target area based on this.
delete The method according to claim 1,
In the step of creating the landslide vulnerability map,
A method of creating a landslide vulnerability map, characterized by establishing a landslide vulnerability model using the AdaBoost algorithm instead of the Multiclass Classifier algorithm, and creating a landslide vulnerability map of a target area based on this.
The method according to claim 1,
In the step of creating the landslide vulnerability map,
A landslide vulnerability map creation method characterized by establishing a landslide vulnerability model using the LogitBoost algorithm instead of the Multiclass Classifier algorithm, and creating a landslide vulnerability map of a target area based on this.
delete The method according to claim 1,
In the step of creating the landslide vulnerability map,
A landslide vulnerability map creation method characterized by establishing a landslide vulnerability model using a Bagging algorithm instead of the Multiclass Classifier algorithm, and creating a landslide vulnerability map of a target area based on this.
The method according to claim 1,
In the step of creating the landslide vulnerability map,
A landslide vulnerability map creation method characterized by establishing a landslide vulnerability model using ANN (Artificial Neural Network) algorithm instead of Multiclass Classifier algorithm, and creating a landslide vulnerability map of a target area based on this.
The method according to claim 1,
In the step of creating the landslide vulnerability map,
A method of creating a landslide vulnerability map, characterized by establishing a landslide vulnerability model using a SVM (Support Vector Machine) algorithm instead of a multiclass classifier algorithm, and creating a landslide vulnerability map of a target area based on this.
A data input unit for providing an interface for receiving pre-investigation result data for a target area to be prepared for a landslide vulnerability map;
Environmental factor determination unit to select environmental factors necessary to calculate landslide vulnerability; And
In consideration of the preliminary survey result data and the environmental factors, including a map creation unit that establishes a landslide vulnerability model using a machine learning algorithm, and creates a landslide vulnerability map of a target area based on this,
The environmental factor determining unit,
Select environmental factors including topographic coefficients, hydrological factors, properties of land signs, soil thickness grades, forest factors, geological information, and strata information of the target area, and establish a spatial database for each environmental factor,
The terrain coefficients include slope, side, maximum curvature, profile curvature, convexity, texture, mid-slope position (MSP), topographic position index (TPI), and terrain durability index of the target area. (Terrain ruggedness index, TRI), the slope means the slope of the hill, the side surface means the steepest downhill road, the intermediate slope position represents the highest point value at the intermediate slope, the maximum curvature and the profile curvature Is a curvature function representing the shape of the topography, the convexity is described as a positive surface curvature and represents the ratio of convex-upward cells, and the texture is correlated with TRI and TPI, and TRI is Represents the amount of elevation difference between the surface and the surrounding area, including concave and convex, TPI represents the terrain slope location,
The hydrological element is a flow accumulation, a stream power index (SPI), and a topographic wetness index (TWI), and the flow accumulation represents a flow accumulation value through a valley, and the river power index ( SPI) represents the erosion power of the water stream, and Terrain Wetness Index (TWI) represents the steady state wetting index used to quantify terrain control for hydrological processes,
The property of the land sign represents land use classified as settlement, grass, farm, forest, rice field, and road,
The soil thickness grade is divided into four grades: very shallow grade (0-19 cm), shallow grade (20-49 cm), medium grade (50-99 cm) and deep grade (> 100 cm),
The forest factors are forest type, forest diameter, forest density, and forest age,
The geological information is unit information of limestone,
The stratum information is the distance from the fault,
The map creation unit,
A landslide vulnerability mapping device, characterized by establishing a landslide vulnerability model using a multiclass classifier algorithm, and creating a landslide vulnerability map of a target area based on this.
delete The method of claim 9,
The map preparation unit establishes a landslide vulnerability model using an AdaBoost algorithm instead of a Multiclass Classifier algorithm, and creates a landslide vulnerability map of a target area based on this.
The method of claim 9,
The map preparation unit establishes a landslide vulnerability model using a LogitBoost algorithm instead of a Multiclass Classifier algorithm, and creates a landslide vulnerability map of a target area based on this.
delete The method of claim 9,
The map preparation unit establishes a landslide vulnerability model using a bagging algorithm instead of a multiclass classifier algorithm, and creates a landslide vulnerability map of a target area based on this.
The method of claim 9,
The map creation unit establishes a landslide vulnerability model using an artificial neural network (ANN) algorithm instead of a multiclass classifier algorithm, and creates a landslide vulnerability map of a target area based on this.
The method of claim 9,
The map preparation unit establishes a landslide vulnerability model using a Support Vector Machine (SVM) algorithm instead of a multiclass classifier algorithm, and creates a landslide vulnerability map of a target area based on this.

Images