92-3 Mapping the Unseen: Machine Learning Approaches for Landslide Susceptibility and Runout Modeling in East Tennessee

Session: Landslide Inventory Mapping and Next Steps: Assessing Susceptibility, Hazard Models, Risk, and Policy

Presenting Author:

Authors:

Abstract:

Landslides are a persistent natural hazard in mountainous regions, particularly in the southern Appalachian Mountains, where steep terrain, complex geology, and increasingly intense rainfall intersect. In late September 2024, multi-day rainfall from Hurricane Helene triggered numerous landslides in eastern Tennessee. These events, often initiating as translational and rotational slides, evolved into debris flows as they moved downslope.

This study applied ensemble learning techniques to model landslide susceptibility and runout areas. A dataset of approximately 270 landslide locations in the Nolichucky River Basin, East Tennessee, was analyzed using a Random Forest classification model trained on environmental predictors, including climatic variables, bedrock geology, soil type, a post-storm digital elevation model (DEM), and terrain derivatives such as slope angle, aspect, curvature, topographic wetness index (TWI), and roughness. All data were harmonized to a consistent resolution and coordinate system.

Model tuning was conducted via randomized parameter search to optimize tree count, depth, and predictor sampling, enhancing generalizability and reducing bias. The final model, comprising 90 decision trees, achieved a median validation accuracy of 93.2% and F1-scores above 0.90 for landslide classes. Variable importance rankings consistently identified slope angle, aspect, rainfall, and bedrock geology as the most influential factors. Diagnostic analyses confirmed the model's stability across training and validation sets.

Landslide locations and the DEM were input into USGS Grfin Tools to delineate landslide-induced inundation zones based on source areas. Guided by field evaluations and predictor thresholds, source areas were defined by slopes between 25° and 60° and a minimum area of 100 m². Growth zones were assigned to drainage channels with slopes ≥5°, and debris volume was capped at 3,000 m³ from upslope zones.

Results demonstrate that both the Random Forest susceptibility model and Grfin Tools runout modeling offer reliable, scalable approaches for assessing landslide risk in data-limited, geologically complex terrains. These tools support future efforts in predictive risk mapping, emergency response planning, and climate-resilient infrastructure development in the southern Appalachians.

Geological Society of America Abstracts with Program. Vol. 57, No. 6, 2025

doi: 10.1130/abs/2025AM-7191

© Copyright 2025 The Geological Society of America (GSA), all rights reserved.