| Abstract: | The growing availability of digital topographic data and the increased reliability of precipitation forecasts invite modelling efforts to predict the timing and location of shallow landslides in hilly and mountainous areas in order to reduce risk to an ever‐expanding human population. Here, we exploit a rare data set to develop and test such a model. In a 1·7 km2 catchment a near‐annual aerial photographic coverage records just three single storm events over a 45 year period that produced multiple landslides. Such data enable us to test model performance by running the entire rainfall time series and determine whether just those three storms are correctly detected. To do this, we link a dynamic and spatially distributed shallow subsurface runoff model (similar to TOPMODEL) to an in?nite slope model to predict the spatial distribution of shallow landsliding. The spatial distribution of soil depth, a strong control on local landsliding, is predicted from a process‐based model. Because of its common availability, daily rainfall data were used to drive the model. Topographic data were derived from digitized 1 : 24 000 US Geological Survey contour maps. Analysis of the landslides shows that 97 occurred in 1955, 37 in 1982 and ?ve in 1998, although the heaviest rainfall was in 1982. Furthermore, intensity–duration analysis of available daily and hourly rainfall from the closest raingauges does not discriminate those three storms from others that did not generate failures. We explore the question of whether a mechanistic modelling approach is better able to identify landslide‐producing storms. Landslide and soil production parameters were ?xed from studies elsewhere. Four hydrologic parameters characterizing the saturated hydraulic conductivity of the soil and underlying bedrock and its decline with depth were ?rst calibrated on the 1955 landslide record. Success was characterized as the most number of actual landslides predicted with the least amount of total area predicted to be unstable. Because landslide area was consistently overpredicted, a threshold catchment area of predicted slope instability was used to de?ne whether a rainstorm was a signi?cant landslide producer. Many combinations of the four hydrological parameters performed equally well for the 1955 event, but only one combination successfully identi?ed the 1982 storm as the only landslide‐producing storm during the period 1980–86. Application of this parameter combination to the entire 45 year record successfully identi?ed the three events, but also predicted that two other landslide‐producing events should have occurred. This performance is signi?cantly better than the empirical intensity–duration threshold approach, but requires considerable calibration effort. Overprediction of instability, both for storms that produced landslides and for non‐producing storms, appears to arise from at least four causes: (1) coarse rainfall data time scale and inability to document short rainfall bursts and predict pressure wave response; (2) absence of local rainfall data; (3) legacy effect of previous landslides; and (4) inaccurate topographic and soil property data. Greater resolution of spatial and rainfall data, as well as topographic data, coupled with systematic documentation of landslides to create time series to test models, should lead to signi?cant improvements in shallow landslides forecasting. Copyright © 2003 John Wiley & Sons, Ltd. |
