LANDSLIDE RISK ASSESSMENT IN THE MOUNTAIN STATE: RANDOM FOREST MODELLING IN MAJOR LAND RESOURCE AREAS HELPS SURMOUNT GAPS IN THE WEST VIRGINIA LANDSLIDE INVENTORY AND BETTER ALIGN LIDAR-BASED MAPPING WITH LOCAL GEOLOGIC KNOWLEDGE

The West Virginia Landslide Risk Assessment was launched in 2018, funded by the FEMA Hazard Mitigation Grant Program and the West Virginia Division of Emergency Management. Step one in the assessment is a statewide digital landslide inventory, initially populated by > 93,000 previously documented landslides: mostly mapped by state and federal agencies from 1970s and 1980s aerial photography. The inventory has been further populated by > 65,000 additional landslides identified since January 2019 using 1 or 2 meter DEMs based on QL2 or QL3 LiDAR. A project goal to cover the whole state with LiDAR-based landslide mapping is ~ 60 % complete. Inconsistencies in the overall inventory stem from disparate types of data, collected with different purposes and methods. Examples of methodological bias include the virtual absence of rock falls and shallow roadside slides in our on-going LiDAR mapping, in sharp contrast with under-reporting of long-runout debris flows and exceptionally large landslides in some older data sets.

Methodological biases and areal coverage gaps have not allowed use of the older data to model landslide susceptibility; however, Maxwell has made slope failure predictions using random forest machine learning derived from the recent LiDAR-based mapping. The landslide prediction models are segregated into individual USDA Major Land Resource Areas, which generally coincide with physiographic divisions. In exception to USDA boundaries, the Allegheny Plateau and Mountains MLRA was divided into northern and southern sections to align with differences in topography and well-documented bedrock facies changes.

Susceptibility modelling has been accomplished for the Northern Appalachian Ridges and Valley (including very small portions of Northern Blue Ridge and Southern Appalachian Ridges and Valley along the Virginia border), the Cumberland Plateau, and the southern section of the Allegheny Plateau and Mountains. Modelling of the Central Allegheny Plateau (the largest and most populous MLRA in West Virginia) and the northern section of the Allegheny Plateau and Mountains are anticipated in the coming year. When susceptibility modelling is coupled to property and infrastructure maps, the resulting Risk Assessment model appears to accurately delineate high risk areas consistent with local landslide knowledge, even in areas where very few landslide incidences occur in the inventory.