Methods

We developed an approach to landslide-susceptibility mapping combining machine-learning and geomorphic statistics (Crawford and others, 2021). The susceptibility model is based upon high-resolution airborne lidar-derived data sets from a 1.5-m regional digital elevation model and a detailed landslide inventory for Magoffin County, Kentucky. We used a logistic regression model to establish a connection between specific slope variables and landslide occurrence. We evaluated slope angle, aspect, elevation, terrain roughness, curvature, and plan curvature to determine what influences landslide occurrence. Next, we modeled the probability of occurrence and used those values to create a landslide-susceptibility map. Our model and resulting map estimate the probability of an event (a landslide) determined by the slope variables.

The probability and map classification are not a landslide prediction result from a scenario-based event (a rainfall event, for example), or a probability with a time component. This means the map classification will not predict how or when a landslide might occur, only the likelihood of a past or future occurrence. Data that occurs on less than a 3-degree slope were excluded because these areas are mostly flat. Although determining landslide susceptibility has inherent uncertainty, our map results, and distribution of higher probabilities (moderate, moderate-high, high) effectively reflect the geomorphic variables that are indicative of unstable ground conditions and potential landslide activity. The low-moderate and low susceptibility classes do not indicate that landslides have not or cannot occur in these areas.

Methods

We developed an approach to landslide-susceptibility mapping combining machine-learning and geomorphic statistics (Crawford and others, 2021). The susceptibility model is based upon high-resolution airborne lidar-derived data sets from a 1.5-m regional digital elevation model and a detailed landslide inventory for Magoffin County, Kentucky. We used a logistic regression model to establish a connection between specific slope variables and landslide occurrence. We evaluated slope angle, aspect, elevation, terrain roughness, curvature, and plan curvature to determine what influences landslide occurrence. Next, we modeled the probability of occurrence and used those values to create a landslide-susceptibility map. Our model and resulting map estimate the probability of an event (a landslide) determined by the slope variables.

The probability and map classification are not a landslide prediction result from a scenario-based event (a rainfall event, for example), or a probability with a time component. This means the map classification will not predict how or when a landslide might occur, only the likelihood of a past or future occurrence. Data that occurs on less than a 3-degree slope were excluded because these areas are mostly flat. Although determining landslide susceptibility has inherent uncertainty, our map results, and distribution of higher probabilities (moderate, moderate-high, high) effectively reflect the geomorphic variables that are indicative of unstable ground conditions and potential landslide activity. The low-moderate and low susceptibility classes do not indicate that landslides have not or cannot occur in these areas.