Date of Graduation
5-2025
Document Type
Thesis
Degree Name
Master of Science in Geology (MS)
Degree Level
Graduate
Department
Geosciences
Advisor/Mentor
Aly, Mohamed H.
Committee Member
Befus, Kevin M.
Second Committee Member
Lamb, Andrew P.
Keywords
Geohazards; Geology; Landslides; Machine Learning; Science & Technology
Abstract
Landslides pose significant hazards to human safety, infrastructure, and the environment, particularly in regions of high elevation that experience extended periods of heavy rainfall. This research focuses on preparing and evaluating landslide susceptibility maps (LSMs) for the Blue Ridge Mountains, a portion of the Appalachian Mountains in western North Carolina, utilizing three machine learning algorithms: Logistic Regression, Random Forest, and Gradient Boosting Regression. Sixteen landslide conditioning factors, reflecting topographic, geological, environmental, and anthropogenic influences, were identified for model input. The landslide inventory database, comprising 7,350 locations, was randomly divided into training (80%) and testing (20%) sets. The performance of each model was evaluated and compared using confusion matrices. The Random Forest algorithm had the highest performance in predicting landslide locations, with a strong emphasis on elevation, slope, and proximity to roads as the most influential factors. The Logistic Regression model provided useful insights into the linear relationships between the conditioning factors and landslide susceptibility, performing well in areas where the relationship between predictors and landslide occurrence is more straightforward. In contrast, the Gradient Boosting model, known for its ability to capture complex nonlinear relationships, identified similar critical factors as the Random Forest model but with a higher sensitivity to variations in slope and distance to drainages. The differences in model outcomes can be attributed to the inherent characteristics of the algorithms: Logistic Regression’s simplicity and linearity, Random Forest’s ability to handle complex interactions through ensemble learning, and Gradient Boosting’s strength in optimizing weak learners for more nuanced pattern recognition. Overall, the LSMs produced by these models suggest that 20% of the study area is highly susceptible to landslides. These LSMs can serve as a valuable tool for land use planning, disaster preparedness, and risk mitigation efforts at a large scale.
Citation
Edmonds, A. (2025). Leveraging Machine Learning Models for Enhanced Landslide Prediction in Western North Carolina. Graduate Theses and Dissertations Retrieved from https://scholarworks.uark.edu/etd/5781
