Massey, C.I.; Lukovic, B.; Wolter, A.; Rosser, B.J. 2025 A rainfall-induced landslide-susceptibility model for northern New Zealand, Version 1.0. Lower Hutt, NZ: Earth Sciences New Zealand. GNS Science report 2023/27. 28 p.; doi: 10.21420/SKHQ-3R36
Abstract
This report presents Version 1.0 of a quantitative statistically based rainfall-induced landslide(RIL) susceptibility algorithm (model) for the northern New Zealand climate zone. It is our first model developed for this region and is based on landslide inventories containing landslides that were triggered by two storm events. The landslides triggered by Cyclone Gabrielle were still being mapped when this model was developed and so were not available to be used. Subsequently, we have created models for other climate zones around New Zealand, and it is our intent to create RIL susceptibility models for all remaining climate zones over the next few years. The overall objective of this larger project is to test whether quantitative, statistically based RIL susceptibility models can be trained for the different New Zealand climate zones at the regional scale. This would allow landslide probability – here defined as the location, likelihood and scale/number of landslides reflecting ‘real-world’ conditions – to be linked to the amount of rainfall and other susceptibility factors that influence landslide occurrence. The outcome this work hopes to achieve is that end users will have knowledge of where rainfall-induced landslides have and/or could occur in the future, their likelihood and scale of occurrence and the associated uncertainties. The landslide types represented by the susceptibility models described in this report are predominantly earth, soil, debris and rock slides and flows, which tend to be the more mobile types of landslides that occur during rain events. The results presented in this report indicate that geology – and the inferred shear-strength variation between the different geological groupings represented by GeolCode– is an important controlling factor on landslide occurrence. Models were developed for each GeolCode (categorical factor), which were used to group the continuous factors and land-cover categorical factor. Of all the susceptibility factors included in the best-fitting models, the results show that slope, elevation and rain have themost explanatory power on landslide probability but that their relative importance varies between the different GeolCode groupings. Overall, the relatively low pseudo R2 values of all models ranging from 0.07 to 0.17 – which describes the strength of the relationship between the factors used in the model to predict landslide occurrence – indicate that the models cannot give accurate hindcasts for individual grid cells. However, their mean ROC:AUC (areas under the receiver operating characteristic curve) values range from 0.73 to 0.88, suggesting that the models are ‘acceptable’ to ‘excellent’.While the ROC:AUC values suggest that the models presented in this report produce overall acceptable results, their relative performance could be improved, especially given the relatively low pseudo-R2 values. This might be achieved by including additional landslide datasets in model training and by varying the type of machine-learning classifier used, for example, to explore non-linearity between landslide probability and the input factors. It would be useful to further explore these statistical relationships through a physics-based lens to understand them better (auths)
