Predicting shallow landslide size and location across a natural landscape: Application of a spectral clustering search algorithm

Predicting shallow landslide size and location across landscapes is important for understanding landscape form and evolution and for hazard identification. We test a recently developed model that couples a search algorithm with 3-D slope stability analysis that predicts these two key attributes in an intensively studied landscape with a 10 year landslide inventory. We use process-based submodels to estimate soil depth, root strength, and pore pressure for a sequence of landslide-triggering rainstorms. We parameterize submodels with field measurements independently of the slope stability model, without calibrating predictions to observations. The model generally reproduces observed landslide size and location distributions, overlaps 65% of observed landslides, and of these predicts size to within factors of 2 and 1.5 in 55% and 28% of cases, respectively. Five percent of the landscape is predicted unstable, compared to 2% recorded landslide area. Missed landslides are not due to the search algorithm but to the formulation and parameterization of the slope stability model and inaccuracy of observed landslide maps. Our model does not improve location prediction relative to infinite-slope methods but predicts landslide size, improves process representation, and reduces reliance on effective parameters. Increasing rainfall intensity or root cohesion generally increases landslide size and shifts locations down hollow axes, while increasing cohesion restricts unstable locations to areas with deepest soils. Our findings suggest that shallow landslide abundance, location, and size are ultimately controlled by covarying topographic, material, and hydrologic properties. Estimating the spatiotemporal patterns of root strength, pore pressure, and soil depth across a landscape may be the greatest remaining challenge.