Integrated Landslide Susceptibility Modeling Using Remote Sensing and In-Situ Data

Thesis Abstract

This study addresses the critical challenge of predicting landslide susceptibility in mountainous environments by integrating remote sensing data with detailed in-situ observations to improve spatial accuracy and temporal relevance for hazard mitigation. The aim is to develop a hybrid, data-driven susceptibility model that combines multi-sensor remote sensing outputs with geotechnical and hydrological field measurements to produce a transferable risk map for emergency planning and land-use decision-making. Specific objectives include (i) to compile a geodatabase of terrain, vegetation, rainfall, soil properties, and landslide inventories for a 1,200 km2 study area, (ii) to extract and harmonize high-resolution remote sensing variables (e.g., landsat/Sentinel-derived slope, aspect, NDVI, soil moisture indices, and cumulative rainfall) with in-situ soil shear strength, piezometric pressure, and groundwater dynamics from 60 boreholes and 120 shallow piezometers, (iii) to develop and compare multiple modeling approaches—random forest, gradient boosting, and logistic regression—augmented by a physically informed framework linking hydrological thresholds and slope stability equations, and (iv) to validate the integrated model against an independent 15% sample of landslide events using spatial k-fold cross-validation and receiver operating characteristic (ROC) analysis. The study adopts a pragmatic positivist paradigm and draws on the theory of rainfall-triggered slope instability anchored in infinite slope models, while integrating conceptual constructs from the theory of environmental risk assessment to justify variable selection and model integration. Data for modeling will emanate from a population comprising landslide-affected hillside units, resident crisscrossed by 14 catchments; the sample comprises 800 grid cells (30 m resolution) with recorded landslide events and 1,400 grid cells without events, selected to ensure balanced class representation. Instrumentation includes high-resolution satellite imagery (Landsat 8 and Sentinel-2), airborne LiDAR-derived terrain metrics, ground-based soil shear strength tests (direct shear and triaxial tests) on 200 samples, groundwater levels from 120 piezometers, and continuous rainfall data from 18 meteorological stations. Validity and reliability will be ensured through calibration with 2015–2019 landslide inventories and reliability checks of in-situ measurements using duplicate sampling and inter-instrument cross-validation. Analytical procedures encompass data preprocessing (radiometric calibration, atmospheric correction, topographic correction), feature extraction and selection via recursive feature elimination, and model training with nested cross-validation to prevent overfitting. The primary analytical techniques include random forest, gradient boosting (XGBoost), and logistic regression, with performance evaluated via ROC-AUC, precision-recall curves, and calibration plots. Additionally, partial dependence plots and SHAP (SHapley Additive exPlanations) values will be employed to interpret variable contributions and interactions, ensuring model transparency for decision-makers. Hypothesis testing will examine the incremental predictive value of in-situ hydrological variables over remote-sensing-derived predictors, using likelihood ratio tests and information criteria (AIC/BIC). Anticipated findings suggest that the integrated model will outperform single-source models, achieving ROC-AUC improvements of 0.08–0.12 and higher true-positive rates at a given false-positive level, particularly in regions with rapid hydrological fluctuation. The study contributes to knowledge by operationalizing an interoperable landslide susceptibility framework that fuses remotely sensed indicators with geotechnical measurements, enabling more accurate hazard mapping and scalable transferability to other mountainous regions. It also advances methodological integration of machine learning with physically based thresholds in a geotechnical-hydrological context and provides a replicable workflow for open data sharing and decision-support tools for local authorities. The main conclusion anticipates that embracing a hybrid modeling strategy with rigorous in-situ validation substantially enhances predictive skill and reliability of landslide hazard maps under changing rainfall regimes. Recommendations include adopting the integrated model for regional disaster risk reduction planning, institutionalizing continuous in-situ data collection to update model parameters, expanding the sensor network to capture predisposing soil moisture anomalies, and refining the model with near-real-time rainfall and soil moisture monitoring to enable proactive warning and land-use management.