A Framework for Integrating Remote Sensing Data into Landslide Susceptibility Models

Thesis Abstract

Landslides pose a significant hazard to infrastructural development, environmental stability, and human safety in many mountainous and hilly regions worldwide. Despite advancements in remote sensing technology, integrating these geospatial data effectively into reliable landslide susceptibility models remains a challenge, often limiting the accuracy and applicability of hazard predictions. This study aims to develop a comprehensive framework for the integration of remote sensing data into landslide susceptibility modeling to improve predictive accuracy and operational utility. The specific objectives include identifying the most relevant remote sensing datasets for landslide analysis, developing an integrated data processing workflow, evaluating various machine learning algorithms for susceptibility prediction, and formulating a standardized framework adaptable to different geographic contexts. The research adopts a mixed-methods approach, combining quantitative spatial analysis with qualitative assessment of data integration processes. The study area encompasses a 1500 km² mountainous region in Central Asia, selected due to its high frequency of landslide events and availability of diverse remote sensing datasets. The population of the study comprises geographic and geotechnical data layers, including Normalized Difference Vegetation Index (NDVI), Digital Elevation Models (DEM), and Landsat 8 satellite imagery, collected from national geospatial agencies and satellite repositories. A stratified random sampling technique was employed to select 200 landslide occurrence sites validated through field surveys and historical records, alongside control sites for susceptibility modeling. Data collection instruments include high-resolution satellite imagery, UAV-derived orthophotos, and GIS-based geological and hydrological maps. Preprocessing of remote sensing data involves atmospheric correction, image classification, and derivation of terrain attributes such as slope, aspect, and curvature using ArcGIS and ERDAS Imagine software. The study employs supervised classification techniques and NDVI thresholding to extract land cover features, integrating these datasets into a geospatial database. Machine learning classifiers, including Random Forest, Support Vector Machines, and Gradient Boosting, are trained and validated through k-fold cross-validation, with model performance assessed via Area Under the Receiver Operating Characteristic Curve (AUC), Precision-Recall curves, and confusion matrices. Analytical procedures utilize GIS-based susceptibility modeling frameworks, integrating multi-criteria decision analysis (MCDA) and statistical techniques such as multivariate logistic regression and spatial autocorrelation analysis. A novel integrated model framework is developed, combining remote sensing-derived variables with geotechnical and environmental factors to enhance land susceptibility predictions. The study hypothesizes that models incorporating remote sensing data will outperform traditional approaches relying solely on field measurements, as validated through performance metrics. Expected findings include identification of key remote sensing indicators contributing to landslide susceptibility, improved prediction accuracies (anticipated AUC exceeding 0.80), and a replicable integration framework adaptable to diverse geomorphological settings. The research advances knowledge by providing a systematic approach to combine diverse remote sensing datasets within machine learning models, supported by theoretical underpinnings from the models of susceptibility and the theory of spatial data integration. The proposed framework emphasizes scalability, reproducibility, and operational Utility. The study concludes that integrating remote sensing data significantly enhances landslide susceptibility assessments, thereby offering stakeholders a reliable tool for hazard mitigation planning. Recommendations include adopting the proposed framework in regional hazard management agencies, expanding the dataset to include real-time monitoring, and exploring the use of emerging remote sensing technologies such as Synthetic Aperture Radar (SAR) and LiDAR for further refinement of models. Future research directions may involve dynamic susceptibility modeling incorporating temporal remote sensing data and expanding the framework's applicability to other natural hazard assessments. This research contributes to the literature by bridging the gap between remote sensing technology and landslide risk modeling, supporting evidence-based decision-making in geohazard management.