기계학습 분류모형 기반 급경사지 재해위험인자 평가 = Evaluation of Disaster Risk Factors on Steep Slopes using a Machine Learning-Based Classification Models

다국어 초록 (Multilingual Abstract)

Hydrological changes induced by extreme weather events have increased prolonged monsoon periods and localized heavy rainfall, rapidly saturating the ground and reducing the shear resistance of slopes. This, in turn, accelerates the activity along potential failure surfaces and expands weak zones, thereby increasing the likelihood of slope failures.

Such slope hazards cause significant property damage and loss of life. Consequently, the Ministry of the Interior and Safety has been conduction Field surveys to identify steep slopes nationwide and implement management measures to prevent disasters.

Because slope failures occur unpredictably in both time and space, it is essential to evaluate the factors influencing slope stability and to manage them in advance in order to ensure disaster prevention and public safety.

The assessment of unstable steep slopes is typically performed using tools such as the “Disaster Risk Assessnent Table.” Establishing appropriate evaluation criteria and determining key assessment factors are therefore critical for accurately evaluation hazard levels and predicting failures. To improve the reliability of assessment tables, numerous studies have applied diverse methodologies.

Past research on developing assessment tables has utilized probabilistic and statistical approaches such as logistic regression, artificial neural networks, and satellite imagery analysis. Models based on AHP analysis derived from expert surveys have also been widely used. More recently, big data techniques, machine learning, and deep learning have been applied to overcome the limitations inherent in traditional mathematical and engineering approaches.

However, existing models that rely on binary classification(failure vs. non-failure) face limitations in data generation and sampling. In particular, the scarcity of field data on failed slopes has resulted in insufficient consideration of critical risk factors that strongly influence slope failure.

Therefore, this study aims to develop a framework for classifying steep slopes with high failure potential by applying various maching-learing algorithms to field-survey-based hazard-factor data and evaluating their classification performance. Using field investigation data from 7,791 non-failure slopes and 90 failure slopes, along with disaster risk assessment records, we conducted analyses and identified major risk factors influencing slope failure.

Analysis of failed slopes showed that failures predominantly occurred in rock slopes(77.9%), on irregular-shaped transverse profiles(62.2%), and in small-to medium-sized slopes(5-24m) with high gradients(>63°). For soil slopes, rotational failure and surface erosion were strongly associated with failure, whereas rock slopes were primarily affected by rockfall and well-developed joints.

For model training, a portion of the dataset was separated and cross-validation was performed. After hyperparameter tuning, the XGBoost model demonstrated superior performance across most evaluation metrics, with significantly higher sensitivity compared to other models, indication strong predictive capability for slope failures.

To analyze the relative importance of various factors, the XGBoost model was combined with SHAP, an explainable AI technique. The results showed that transverse shape, soil depth, rockfall, surface erosion, herbaceous vegetation density, woody vegetation density, joint development, slope length, and slope angle exhibited the highest variable importance in this order. These findings were further validated through SHAP summary plots, force plots, and waterfall plots for individual slope profiles.

The analysis indicates that the currently used disaster risk assessment table does not adequately capture the hazard characteristics of slopes, as discrepancies between risk factors and evaluation items contribute to underestimation-particularly in terms of social factors and the relative scale of failures. Based on the results of this study, we recommend the development of more refined data exploration techniques and continuous research to enhance the performance and reliability of machine learning models that account for characteristics of slope-failure datasets.

목차 (Table of Contents)

• 제1장 서론 1
• 1.1 연구 배경 및 목적 1
• 1.2 연구동향 5
• 1.3 연구내용 및 범위 7
• 제2장 이론적 배경 9
• 2.1 급경사지 평가 및 관리시스템 9
• 2.2 비탈면 붕괴특성 17
• 2.3 기계학습(ML) 분류모형 20
• 제3장 연구방법 25
• 3.1 급경사지 실태조사 25
• 3.1.1 개요 25
• 3.1.2 조사방법 27
• 3.2 조사결과의 분석 33
• 3.2.1 Logistic regression 34
• 3.2.2 Random Forest 35
• 3.2.3 Extremely randomized tree 36
• 3.2.4 XGBoost 37
• 제4장 실태조사 결과분석 39
• 4.1 미붕괴 비탈면의 실태조사 결과 40
• 4.1.1 미붕괴 비탈면의 특성 40
• 4.1.2 미붕괴 비탈면의 위험요인 44
• 4.2 붕괴 비탈면의 조사 결과 53
• 4.2.1 붕괴 비탈면의 특성 53
• 4.2.2 붕괴 비탈면의 위험요인 57
• 제5장 기계학습 연구결과 69
• 5.1 모델의 적합성 검증 69
• 5.1.1 훈련 및 평가데이터 71
• 5.1.2 평가지표 71
• 5.1.3 하이퍼 파라미터 튜닝 75
• 5.1.4 적합성 검증결과 76
• 5.2 위험요인의 영향력 분석 81
• 5.2.1 SHAP(SHapley Additive exPlanations) 81
• 5.2.2 SHAP 변수 중요도(SHAP feature importance plot) 82
• 5.2.3 SHAP 요약 도표(SHAP summary plot) 83
• 5.2.4 개별 비탈면 붕괴 발생 데이터의 SHAP 분석 86
• 제6장 결론 98
• 참고문헌 100
• 초록(Abstract) 107