Deep-seated landslide and earthflow detection (DSLED): A suite of automated algorithms for mapping landslide-prone terrain with digital topographic data

In addition to constituting a significant natural hazard, large landslides often serve as the primary sediment production mechanism in mountainous landscapes. The first step in deciphering their role in landscape modification is mapping their spatial extent, a task often involving air photo and/or topographic map interpretation. This approach has several disadvantages, including: 1) it is highly time-consuming, 2) vegetation often obscures relevant features, and 3) resulting maps are often highly subjective (as demonstrated by recent studies in which the same area was mapped by 6 or more `experts'). Here, we outline a suite of tools for the objective delineation of terrain prone to deep-seated landslides and earthflows using digital topographic data. Importantly, these algorithms are intended to identify terrain that has already experienced slope instability, and thus may be the locus of renewed movement. The topographic signature of prior landslide activity depends on the spatial resolution of available data; benchy terrain associated with large landslides typically appears rough and jumbled at the meter-scale, but relatively smooth and uncrenulated at the 50-100m scale. The algorithm DSLED-Bench uses a combined criterion of slope and curvature values estimated from greater than 10m DEMs to identify bench-like landforms associated with large landslides (Roering et al., GSA Bulletin, 2006). DSLED-Drain uses spatially-averaged values of drainage area per unit contour width (a/b) calculated with topographic data from airborne LiDAR to identify large poorly-drained landforms commonly associated with slope instability (Mackey, AGU Fall Mtg. 2005). Lastly, DLSED-Rough uses a directional eigenvalue matrix calculation to estimate local terrain roughness from airborne LiDAR topographic data (McKean and Roering, Geomorphology, 2004). We tested these algorithms against independently mapped landslides and field observations at several sites in Northern California and the Cascade Range of Oregon. Each of the methods provides predictive power in identifying slide-prone terrain, although site-specific calibration is necessary for the DSLED-Bench algorithm. The rate of false positive and false negative error depends on the data source, as sites characterized by airborne LiDAR data tended to overpredict the extent of slide-prone terrain. These methods are best utilized as a reconnaissance tool in combination with field, geologic, and air photo mapping.