Identifying the controls on coastal cliff landslides using machine-learning approaches

Transformations are underway in our ability to collect and interrogate remotely sensed data. Here we explore the utility of three machine-learning methods for identifying the controls on coastal cliff landsliding using a dataset from Auckland, New Zealand. Models were built using all available data with a resampling approach used to evaluate uncertainties. All methods identify two dominant landslide predictors (unfailed cliff slope angle and fault proximity). This information could support a range of management approaches, from the development of 'rules-of-thumb' to detailed models that incorporate all predictor information. In our study all statistical approaches correctly predict a high proportion (>85%) of cases. Similar 'success' has been shown in other studies, but important questions should be asked about possible error sources, particularly in regard to absence data. In coastal landslide studies sign decay is a vexing issue, because sites prone to landsliding may also be sites of rapid evidence removal. Machine-learning based approaches successfully identify coastal cliff landsliding controls.Possible management approaches vary from hazard mapping to heuristic tools.Difficult issues surround the nature of absence data.Sign decay may be a prevalent issue in earth-science applications.