EVALUATION OF QUANTITATIVE PREDICTION MODELS FOR LANDSLIDE HAZARD

C. F. Chung (Geological Survey of Canada, 601 Booth Street. Ottawa, Ontario, Canada K1A 0E8)

S. Obayashi and H. Kojima (Science University of Tokyo, 2641 Yamazaki, Noda, Chiba, 278 Japan)

Quantitative prediction models for the occurrences of future landslides are based on the quantitative relationships between the past landslide occurrences and several layers of spatial map data which are known to be "latent" factors of landslides. To evaluate a performance of a prediction model for future landslides, the corresponding quantitative relationship from which the model has been built, should be robust with respect to time. We have investigated the time-robustness of prediction models in Rio Chincina Study area in Colombia, by dividing the distribution of the occurrences of the landslides in the study area into two groups, (i) Pre-1960 consisting of the landslides occurred prior to the year 1960; and (ii) Post-1960 consisting of the landslides occurred from 1961 to 1988, and then studying quantitative relationships of these two groups and the spatial map data. We have considered three prediction models, "Direct method", "Fuzzy set theory using algebraic sum operator" and "Bayesian probability method". For each model and each group, "success" rates indicating how well the model performed with respect to the occurrences in the group and "prediction" rates with respect to the occurrences in the other group which has not been used to construct the model. The prediction rates of the models are highly correlated with the time-robustness of the models.