Landslide hazard zoning using genetic programming

Genetic programming (GP) is presented as a technique to induce models that can be used with GIS data to map landslide-hazard zones. GP is a data-mining technique frequently used to solve engineering and scientific problems; in recent years it has been applied to several spatial questions. This paper describes attempts to optimize the GP system, to evolve models that are then tested in different locations. GP map accuracy is evaluated by comparison to landslide-hazard maps of the same California locations created by the USGS (United States Geological Survey; Wentworth et al., 1997). GP maps are compared to maps created using linear discriminant analysis (LIDA). in three out of the four study sites, GP models produce more accurate hazard maps than the LIDA process. The highest accuracy attained by GP models is 84% and by LIDA is 69%. Map accuracies at the training site (Yountville and Capell Valley) are higher than at other locations, highlighting the need for care when choosing a training site. Inaccuracies could be caused by using data of too coarse spatial resolution or by differences in data-processing techniques of GP and the USGS, that is, individual cell versus a regional approach.