CHARACTERIZING THE LANDSLIDE HAZARD OF NORTHEASTERN KANSAS USING MULTIPLE LOGISTIC REGRESSION

Landslides are a problem even in vertically challenged areas such as Kansas. Earth flows and earth slides in soils developed on Pennsylvanian bedrock and unconsolidated Quaternary deposits have destroyed houses, closed roads, and broken underground utility lines. Characterizing landslide hazards begins with an inventory of existing earth flows and earth slides including fresh landslides and areas of hummocky terrain indicating past landslides. For predictive purposes, the occurrence data were related to a digital geologic map, a digital elevation model, and a digital soil map using multiple logistic regression. Presence or absence of landslides is a binary event, so the landslide map was converted into a continuous variable by a logistic tranformation. Logits are the natural logarithm of odds, where the odds are the ratio of the probability that a landslide occurs to the probability that a landslide does not occur. Regression analysis yields coefficients that can be used to estimate logistic values at all locations on a map; these then can be transformed back into odds, and thence into probabilities of occurrence. The final product is a map showing the probability of occurrence of landslides throughout the mapped area. The power of the technique is its ability to incorporate both continuous data (slope, aspect, etc) and categorical data (geology, soils, etc) into a single model, and to assess the relative importance of each predictor variable. Analysis of the results for northeast Kansas shows that slope and bedrock lithology are the most important predictors of landslides, although soils data can be substituted for geologic data because the two are highly correlated. Slope aspect has no influence on landslide occurrence in this area. This method calculates the long-term (thousands of years) climate-independent probability that a location is either part of a prior landslide or will be involved in a future landslide. The stability of the statistical estimates was examined by random resampling followed by recalculation of coefficients, and comparing the distributions of areas at predicted levels of hazard. The hazard map based on multiple logistic regression compares favorably with a map produced by a binomial model that does not require any of the statistical assumptions of logistic regression.