Tom tat tieng anh nghiên cứu nâng cao Độ chính xác trong xây dựng bản Đồ phân vùng nguy cơ sạt lở Đất cho khu vực miền núi tỉnh quảng ngãi

Today, the development modern statistical models such as machine learning helps improve the accuracy of landslide spatial prediction models.. Therefore, landslide susceptibility assessme

THE UNIVERSITY OF DANANG

UNIVERSITY OF SCIENCE AND TECHNOLOGY

DOAN VIET LONG

IMPROVING THE ACCURACY OF LANDSLIDE SUSCEPTIBILITY MAPPING IN THE MOUNTAINOUS

REGION OF QUANG NGAI PROVINCE

Major: Water Resources Engineering

Code: 9580202

SUMMARY OF DOCTORAL THESIS

Da Nang - 2024

INTRODUCTION

1 Critical Review of Literatures

Landslides have been known as one of the most dangerous natural disasters in the world, they are distributed on 5 continents, causing a lot of damage In Vietnam, landslides are mainly concentrated in the Northern mountainous provinces and the Central and Highlands provinces Landslides occur in these areas due to many reasons, of which rainfall is considered a triggering factor Under the impact of climate change, the number of extreme rainfall events are expected to increase and become the main factor causing landslides in these areas

In landslide research field, landslide susceptibility maps are a useful tool to prevent this type of natural disaster Today, the development modern statistical models such as machine learning helps improve the accuracy of landslide spatial prediction models Therefore, landslide susceptibility assessment using machine learning methods is becoming a major topic in current studies However, for new study areas, determining a suitable machine learning model for landslide spatial prediction is very important

In landslide susceptibility modeling-based machine learning method, data is a very important factor Today, the development of remote sensing technology along with modern analytical techniques allows us to restore and improve statistical data In addition, selecting appropriate data for machine learning models is very important, especially the time-variant data such as rainfall and normalized difference vegetation index (NDVI) Regarding rainfall data, many previous studies used annual average rainfall on landslide susceptibility assessment This approach, however, is less reasonable because this rainfall type cannot represent the triggering factor that

causes landslide occurrences, especially in regions where slope failures just occur after long heavy rainfall In addition to the rain factor, NDVI is also a time-variant factor However, many previous studies only used a specific NDVI map to build landslide prediction models

Thus, there are still many problems related to data and prediction models that need to be improved in landslide susceptibility assessment The mountainous region of Quang Ngai province is a locality where landslides occur every year However, available data related to landslide assessment are very limited Therefore, this thesis selected this study area to improve the accuracy of landslide susceptibility mapping, applying it for natural disaster prevention, project planning and design

2 Problem Definition

- For a new study area, it is necessary to apply many different machine learning models to select the best fit landslide susceptibility model

- For areas with limited measurement data, study on applying remote sensing techniques can be carried out to restore data related to landslides and improve input data for machine learning model

- It is necessary to find a more reasonable approach in using time-variant data such as rainfall and NDVI in landslide susceptibility assessment

- It is necessary to use a more reasonable rainfall type for landslide susceptibility mappingfor areas with limited detailed data

3 Objective

3.1 Overall objectives

Improving the accuracy of landslide susceptibility mapping for areas with limited data

3.2 Detailed objectives

- Data enrichment for the landslide spatial prediction model

- Selecting a suitable and highly accurate landslide spatial prediction model for a specific study area

- Generating landslide susceptibility maps corresponding to different rainfall frequency

4 Subject and scope of the research

4.1 Research subject

- Landslide events have occurred

- Rainfall triggered landslides

- Landslide conditioning factors

- Landslide susceptibility models

- Model and map elevating methods

- Hydraulic constructions in study area

4.2 Scope of the research

5 mountainous districts region of Quang Ngai province

5 Methodology

- Analytical and statistical methods combined with theoretical research

- Investigation, survey, and inheritance methods

- Remote sensing techniques

- Methods of data synthesis, analysis and processing

- Machine learning methods

- Methods of testing, evaluating, and comparing models

- Method of building maps using GIS techniques

6 Scientific and practical significance

6.1 Scientific significances

- The thesis has supplemented and enriched data using for

landslide susceptibility mapping in the mountainous region of Quang

Ngai province

- The thesis has successfully applied machine learning models

to build landslide susceptibility maps corresponding to different

rainfall frequences in the mountainous region of Quang Ngai province

6.2 Practical significances

Landslide susceptibility maps are useful tools for natural disaster

prevention, land use planning, constructions design and management in

the mountainous region of Quang Ngai province

7 General layout of the dissertation

The thesis has a structure including an Introduction, 4 chapters,

Conclusion, and recommendations The complete thesis is presented

in 133 pages of A4 size, excluding the appendix

Chapter 1: CRITICAL REVIEW OF LITERATURES 1.1 Overview of landslides

1.1.1 Landslide Definitions

Landside is the movement of materials from upslope to downslope under the effect of gravity This is a complex phenomenon because it is affected by the interaction of numerous natural factors (geology, geomorphology, hydrology, meteorology) and anthropogenic factors

1.1.2 Landslide types

The study of L Highland has classified landslides into different types: slides, falls, topples, flows, spreads, and complexes Among them, slides are the most common type of landslides around the world (accounting for 55.2%), followed by flows (19.3%), falls (9.4%), the remaining types account for about 13.2% In Vietnam, landslides occur mostly due to rainfall, the types of landslides mainly include slides, and debris flow

1.1.3 Causes and landslide affecting factors

The study of Crozier et al divides the landslide affecting factors into two main groups: (i) Conditioning factors, including: slope degree, aspect, altitude, fault, lithology, drainage density, land use, and soil type, and (ii) group of triggering factors include: rainfall, earthquake, and human activity

1.1.4 Damage of landslides

According to the World Bank statistics, about 3.7 million km2 of land covers and about 300 million people (about 5% of the world's population) are affected by landslides In Vietnam, landslides occur frequently in the Northern, Central and Highland mountainous provinces, causing many damages to people and property From 2000

to 2020, natural disasters of flash floods and landslides killed 1,117 people and disappeared, 671 injured, 12,038 houses collapsed

1.2 Landslide Assessment and Mapping

1.2.1 Landslide Assessment

The study of D J Varnes and F Guzzetti has divided the assessment of landslides into the following types: (i) Landslide zonation, (ii) Landslide spatial prediction, (iii) Landslide hazard assessment, (iv) Landslide vulnerability assessment, (v) Landslide risk assessment

1.2.2 Landslide Mapping

The study of D J et al has divided the landslide maps into the following types: (i) Landslide inventory map, (ii) Landslide susceptibility map, (iii) Landslide hazard map, and (vi) Landslide risk map Among them, the landslide hazard map and landslide risk map are difficult to generate mainly due to limitation of availability of site-specific data For areas with limitation data, this study approaches the landslide spatial prediction and landslide susceptibility mapping

1.3 Landslide susceptibility modeling methods

The study of Shano et al has shown 3 basic methods for landslide susceptibility modeling, including: (i) Qualitative method, (ii) Semi -quantitative method and (iii) Quantitative method In particular, qualitative methods and semi -quantitative methods depend on the experts’ opinions, so it is often less accurate Meanwhile, the quantitative method is considered to be objective in nature and better accuracy The quantitative methods include deterministic approach and statistical approach Of which, the deterministic approach is restricted to small-scale areas Meanwhile, statistical methods are often used for large-scale areas

Statistical methods include traditional statistics and machine learning Currently, the learning machine method with many advantages has been widely applied and replaced with traditional statistical methods The study of Liu et al has listed 3 methods of learning machine learning used for landside susceptibility modeling, including: (i) Conventional machine learning, (ii) Ensemble techniques and hybrid machine learning, and (iii) Deep learning This study has indicated that ensemble techniques and hybrid machine learning are more effective than conventional machine learning, and deep learning methods are more effective than other machine learning methods However, deep learning-based methods are only effective in most data-rich situations, so conventional machine learning methods are still applied in areas where there is not that much data in this field

to train a deep learning network perfectly

The literature review shows that statistical methods are suitable for landslide susceptibility In particular, machine learning methods with superior advantages compared to traditional statistical methods are proposed for use in this study

Chapter 2: BACKGROUND THEORY OF METHODS USED FOR LANDSLIDE SUSCEPTIBILITY MAPPING

2.1 Process of landside susceptibility mapping using machine learning

The flowchart of the methodology is shown in Figure 2.1

Figure 2.1 Methodology of landslide susceptibility mapping

2.2 Landslide detection using remote sensing techniques

2.2.1 Satellite image database

For mountainous areas of Vietnam, optical image data sources that can be used in this study include Google Earth images, Landsat satellite images and Sentinel-2 satellite images In particular, Sentinel 2 images have just been released since 2015 with better resolution than Landsat images, helping to identify landslides better Therefore, this study uses

a combination of Sentinel 2 satellite images and Google Earth images for landslides detection

2.2.2 Analytical techniques

This study proposes an analysis method of combination Google Earth images and Sentinel 2 images to overcome the limitations of

each specific method This method helps to increase the effectiveness

of landslide detection and time series classification of landslide events

Figure 2.3 Flowchart of combining two image sources: Sentinel 2

and Google Earth to detect landslide sites

2.3 Method of generating time series of rainfall data and time series of NDVI data on landslide susceptibility assessment

Rainfall and NDVI factors are the time-variant However, previous study only used a specific map of each factor to evaluate a set

of landslides that were collected over many years To improve this problem, the study proposes to use time series of data of rainfall and NDVI on landslide susceptibility assessment The methodology of constructing these data is shown in Figure 2.4 and Figure 2.5

Figure 2.4 Flowchart of the method for generating time series of rainfall data

Figure 2.5 Flowchart of the method for generating time series of NDVI

2.4 Regional rainfall frequency method

In rainfall frequency analysis, the regional frequency analysis (RFA) approach is considered more effective than local approaches Therefore, this study proposes to use the RFA method to create the annual maximum rainfall maps These maps are used to build landslide susceptibility maps according to rainfall frequency scenarios

Figure 2.6 Flowchart of generating rainfall database using RFA method

2.5 Methods of analyzing and feature selection data

This study uses two types of data evaluation, including: multicollinearity analysis using the variance inflation factor (VIF) method and feature selection using the Boruta method The multicollinearity analysis process will eliminate factors that have dependent relationships when applying linear models Meanwhile, the feature selection process will eliminate factors that are not necessary in the landslide spatial prediction model

2.6 Theory of landslide spatial prediction models

2.6.1 LR model

In landslide susceptibility assessment, LR is used as a regression model with the independent variables being the values of the conditioning factors and the dependent variable describing the landslides (value = 1) non-landslides (value = 0) This study uses the SGD method to optimize the loss function to find the best parameters of the landslide susceptibility model

2.6.2 SVM model

Assume we have a training dataset (Xi, yi), where Xi are the input values of the landslide conditioning factors and yi are the output values (landslide or non-landslide) SVM can be used for non-linear data with kernel functions The objective of SVM is to find an optimal hyper-plane that discriminate this training dataset into two classes by solving the classification function:

SVM with kernel functions can solve classification problems such

as non-linear input data Some common kernel functions include PL,

sigmoid, RBF The parameters of the SVM model depend more on the kernel function applied

2.6.3 DT model

This is a non-parametric classification method, which includes clustering and classification based on decision rules The DT model can be developed using many algorithms such as: ID3, C4.5, CART, CHAID, MARS Three commonly indexes widely used in these algorithms are the Gini index, twoing-rule, and cross-entropy For classification case, the Gini index is often chosen because it provides the best overall classification accuracy The main parameters of the DT model include complexity parameter (cp), max_depth, minsplit, minbucket

2.6.4 RF model

RF is a data mining algorithm that uses the ensemble technique of the bagging method This model is capable of accurately classifying data using a set of decision trees (DT) The parameters of the RF model include (i) Number of trees to grow (ntree) and (ii) Number of variables randomly sampled as candidates at each split (mtry) The task of the study is to find the best fit of model’s parameters

2.6.5 XGBoost model

XGBoost is a scalable machine learning system for tree boosting, developed by Chen (2016) XGBoost creates many classification and regression trees (CART) and integrates them using the gradient boosting algorithm The goal of XGBoost algorithm is to minimize the following regularized objective function:

where 𝑦̂ and 𝑦𝑖 𝑖 are the predicted value and observed value respectively; Ω(f) is the penalty term that helps to smooth the final

learnt weights to avoid overfitting; γ, λ are the regularization degrees respectively; w is the score of each leaf

The main parameters of XGBoost algorithm applied in landslide susceptibility assessment: nrounds, max_depth, eta, gamma, colsample_bytree, min_child_weight, subsample

2.7 Evaluation and comparison methods

This study uses some statistical indexed method and receiver operating characteristic curve (ROC) method for landslide susceptibility evaluation

2.7.1 Statistical indexed method

Six statistical indexes are chosen to evaluate the performance of the landslide susceptibility model: accuracy, sensitivity, specificity, kappa, positive predictive value, negative predictive value

2.7.2 ROC method

ROC method is usually used to assess the quality of landslide susceptibility models ROC curve represents the value of sensitivity on the y-axis, and (1-specificity) value on the x-axis The Area Under the ROC Curve (AUC) can be used as an index to evaluate the overall performance of a model The larger the area, the better the performance

of the model The AUC value can be divided into many intervals with the model quality respectively, including: 0.6 – 0.7 (poor), 0.7 – 0.8 (fair), 0.8 – 0.9 (good), and 0.9 – 1.0 (very good)

2.8 Landslide susceptibility corresponding to rainfall frequency mapping method

The landslide susceptibility mapping method is presented as diagram in Figure 2.11

Figure 2.11 Flowchart of building landslide susceptibility maps

according to rainfall frequency