Quantitative landslide susceptibility assessment of the Waikato region using GIS

Abstract

Areas that have experienced landslide events in the past and the conditioning factors present at these sites can be used to identify areas of the same or similar susceptibility. This can be achieved through a landslide susceptibility assessment using a landslide inventory, a set of predictor variables and specialised computer software. A quantitative landslide susceptibility assessment was conducted for the Waikato Region using two statistical approaches and eleven predictor variables. A landslide inventory map was constructed from the GNS QMap landslide spatial data and GeoNet landslide catalogue. Parameter maps for slope, elevation, aspect, lithology, land cover, soil order, mean monthly rainfall, maximum monthly rainfall, distance from roads, distance from faults and distance from rivers, were constructed and compiled into a database with the landslide inventory. The compiled data underwent both bivariate (weights of evidence) and multivariate (logistic regression) statistical analysis, and a landslide susceptibility map was derived for each. In the weights of evidence approach, the presence and absence of each class in relation to landslide occurrence and non-occurrence was individually assessed for each predictive factor. Logistic regression involves fitting a generalised non-linear model to the data based on a binary predictor (presence or absence of a past landslide event). For each method, a landslide susceptibility map was derived, and the model fit assessed using the landslide inventory. Both susceptibility maps underwent an evaluation to determine the better predictive model. An independent landslide data set was compiled from observations made in Google Earth, and used to establish a set of prediction rate curves and cumulative area curves. Both susceptibility maps resulted in very similar prediction rate curves. Weights of evidence gave a better prediction rate in the 10, 20 and 30% most susceptible pixels, but not in the 40% most susceptible pixels. Neither susceptibility map could be justified as being better than the other based on the prediction rate curves alone. The cumulative area curves for each susceptibility map resulted in very different outcomes. Logistic regression gave the best result with a large proportion of the landslide area within a small proportion of the total area in the high susceptibility classes. Weights of evidence had a larger proportion of the landslide area in high susceptibility classes than logistic regression, but this was associated with a large proportion of total area. Based on the evaluation, the susceptibility map derived using logistic regression was determined to be superior.