Automated Landslide Risk Assessment via Drone-Imaging and ML Analytics

Thesis Abstract

This study addresses the increasing frequency and socioeconomic impact of landslides in mountainous regions by integrating drone-based imaging with machine learning analytics to deliver automated, scalable risk assessments for early warning and land-use planning. The aim is to develop a robust, transferable framework that fuses high-resolution drone imagery, geospatial-derived features, and data-driven models to quantify landslide susceptibility and imminent-risk indicators with quantified uncertainty. Specific objectives include (a) to construct a multi-temporal, high-resolution geospatial database from drone-acquired imagery over 12 study sites representing varied lithologies, slopes, and rainfall regimes; (b) to extract and harmonize a comprehensive set of predictors encompassing topographic (elevation, aspect, curvature), hydrological (soil moisture, precipitation intensity, groundwater indicators), spectral (normalized difference indices, vegetation indices), and geomorphological features; (c) to compare the performance of supervised machine learning algorithms (Random Forest, Gradient Boosting, XGBoost) and deep learning approaches (Convolutional Neural Networks) for landslide susceptibility modeling; (d) to develop an automated pipeline for real-time risk scoring and visualization within a GIS platform; (e) to assess model transferability across sites via cross-site validation and transfer learning; and (f) to evaluate decision-relevant metrics including cost–benefit implications for emergency response and land-management policies. A mixed-method research design is employed, combining a quantitative, data-driven modeling paradigm with qualitative expert validation to ensure interpretability and operational relevance. The population comprises 12 landslide-prone catchments across a temperate mountain belt with diverse geology and land cover. A stratified random sampling approach results in a dataset of 720 drone-derived orthoimages (60 per site) captured across pre- and post-storm events, complemented by historical landslide inventories and meteorological records. Data collection instruments include high-resolution UAS sensors (RGB and multispectral cameras, 20–50 cm GSD), differential GNSS ground control points for orthorectification, LiDAR-derived terrain models where available, and field surveys for ground-truthing a representative 10% subset of detections. Model development employs feature engineering to derive topo-hydraulic indices (slope, curvature, flow accumulation, soil moisture proxies) and spectral features (plant stress indices, moisture-sensitive bands). The analysis uses a structured pipeline data preprocessing, feature selection via SHAP-explained importance, model training with k-fold cross-validation (k=10), hyperparameter tuning through Bayesian optimization, and performance evaluation using AUC, F1-score, and calibration curves. Model interpretability is enhanced with SHAP value analyses and partial dependence plots. Transferability is tested through leave-one-site-out validation and fine-tuning with limited site-specific data, incorporating domain adaptation techniques. Ethical considerations address privacy in drone operations, data security, and stakeholder engagement. Expected findings include (i) higher predictive performance for ensemble methods (AUC 0.88–0.92, F1 0.80–0.85) relative to single-model baselines, (ii) identification of the most influential predictors as flow accumulation, planar curvature, soil moisture proxies, and spectral vegetation stress indices, (iii) demonstration that transfer learning reduces site-specific data requirements by 40–60% while maintaining comparable accuracy, and (iv) a reproducible, automated processing pipeline with a user-friendly GIS dashboard capable of nightly refresh for near-real-time risk assessment. The study contributes to knowledge by advancing an integrated, ICT-driven framework for automated landslide risk assessment that bridges remote sensing, geospatial analytics, and machine learning, while providing empirical evidence on model transferability and uncertainty quantification under varying hydro-geomorphological conditions. The implications for theory include empirical validation of a multi-criteria, data-driven susceptibility paradigm that accommodates dynamic rainfall and land-cover changes, aligning with resilience and disaster risk reduction frameworks. Practical contributions encompass a scalable workflow for agencies and communities to automate surveillance, prioritize mitigation investments, and enhance early warning capacity. The main conclusion posits that drone-imaging combined with ML analytics can produce accurate, transferable landslide risk assessments with quantified uncertainty, enabling proactive management and rapid decision-making. Recommendations include (a) adopting the proposed pipeline within regional hazard monitoring programs, (b) expanding data acquisition to include near-real-time meteorological feeds and soil moisture sensors for improved forecasting, and (c) developing standardized protocols for cross-border data sharing to support multi-jurisdictional risk mitigation.