Comparing methods of landslide data acquisition and susceptibility modelling: Examples from New Zealand

Abstract

The acquisition of landslide inventory data remains an important challenge for landslide susceptibility modelling. For rainfall-induced landslides, comprehensive mapping may be hindered by the size of storm-affected areas and large number of landslides generated as well as the time and costs involved in preparing multi-temporal inventories. In New Zealand, storm events trigger hundreds to thousands of shallow landslides, causing significant damage to land and infrastructure as well as impacts on freshwater and marine environments. Despite this, there are few quantitative assessments of shallow landslide susceptibility to inform targeting of control measures. Here, we compare the effect of using landslide inventories assembled from a) manual versus semi-automated mapping and b) event versus multi-temporal records on the performance of two widely applied methods for landslide susceptibility modelling, namely logistic regression and random forest classification. Evaluation of object-based image analysis (OBIA) for semi-automated landslide mapping showed mixed results, where producer's and user's accuracies ranged 62–81 and 45–55%, respectively, without manual refinement. However, the relative reduction of 6–11% in susceptibility model predictive performance based on area under receiver operating characteristic curves (AUC) using OBIA (AUC = 0.63–0.75) versus manual (AUC = 0.67–0.81) inventories with different variable combinations was low in comparison, and the spatial patterns in modelled susceptibility were generally similar. The random forest model produced slightly better prediction performance compared with logistic regression based on cross-validation within the same study area. However, this was reversed and logistic regression mostly outperformed random forest when the models were fitted and tested with data from different study areas. Model predictive performance for event versus multi-temporal records was comparable. Our results highlight both the challenges associated with semi-automated landslide detection over large areas as well as the opportunity to use OBIA for efficient data collection without necessarily compromising the resulting susceptibility maps. This potentially overcomes significant time and cost impediments to the preparation of landslide inventories that continue to hinder quantitative landslide susceptibility assessment.