Design and Evaluation of a Remote Sensing-Based Landslide Susceptibility Model

Thesis Abstract

Landslides pose significant threats to infrastructure, ecosystems, and human safety in mountainous and hilly regions, necessitating the development of reliable risk assessment tools to inform mitigation strategies. This study aims to design and evaluate a remote sensing-based landslide susceptibility model to enhance early warning capabilities and land-use planning, thereby reducing societal and economic vulnerabilities associated with landslide hazards. The specific objectives include (1) identifying and integrating relevant environmental and anthropogenic variables from satellite remote sensing data, (2) developing a geospatial model employing statistical and machine learning techniques, and (3) validating the model's performance through empirical field data and comparative analysis with existing susceptibility models. The research adopts a mixed-methods approach with a predominantly quantitative design. The study population comprises geographic regions characterized by varying degrees of landslide occurrence within the mountainous terrains of the study area, with a focus on soils, slope, land cover, and rainfall data. A representative sample of 150 landslide inventory points, collected through field surveys and secondary databases, serve as the primary dataset for model training and validation. Remote sensing data are acquired from multispectral satellite imagery, specifically Landsat 8 and Sentinel-2, supplemented by digital elevation models (DEMs) from the Shuttle Radar Topography Mission (SRTM). Data extraction and pre-processing involve image correction, land cover classification via supervised maximum likelihood algorithms, and derivation of slope, aspect, and drainage features. To ensure data validity, the instruments and procedures are subjected to accuracy assessment and calibration using ground control points and existing geospatial datasets. The modeling framework incorporates logistic regression and Random Forest algorithms, with the selection of the best-performing model based on multiple validation metrics including the receiver operating characteristic (ROC) curve, area under the curve (AUC), and confusion matrices. Analytical procedures utilize Geographic Information System (GIS) software integrated with statistical packages such as R and Python scripting for machine learning. The model's performance demonstrates high predictive accuracy, with expected AUC scores exceeding 0.85, indicating robust identification of susceptible zones. Sensitivity analysis examines variable importance and interactions, while cross-validation assesses model stability. Expected findings reveal that slope, land cover, and proximity to geological faults are the most significant predictors influencing landslide susceptibility in the study area. The developed model substantially improves upon existing susceptibility approaches by integrating multi-temporal remote sensing data and machine learning techniques, leading to more precise spatial risk delineation. The study contributes to knowledge by providing a comprehensive framework for remote sensing-based landslide modeling, demonstrating the effectiveness of combining earth observation data with advanced analytical algorithms for hazard assessment. The study concludes that the integrated remote sensing and machine learning approach offers a valuable tool for local authorities and land planners in prioritizing intervention zones and designing mitigation measures. Recommendations include adopting the model within regional planning policies, enhancing remote sensing data resolution for finer-scale assessments, and extending the methodology to incorporate climate change projections for dynamic risk evaluation. Future research should explore the integration of socio-economic factors and real-time monitoring systems to further refine landslide risk prediction models, thereby supporting sustainable development in landslide-prone regions.