Comparing physical and statistical landslide susceptibility models at the scale of individual trees

Abstract

Physically or statistically based approaches are widely used to quantify shallow landslide susceptibility. Despite the underlying data, methods, and assumptions being significantly different, there has been little quantitative work to evaluate the differences in model outcomes. Therefore, we compare a previously developed physical and a statistical shallow landslide susceptibility model for a study area in New Zealand's silvopastoral hill country. Both models include individual tree effects on slope stability. To identify individual trees for both models, we use convolutional neural networks to delineate tree crowns and classify species at the scale of individual trees with regional aerial photography. SlideforMAP, a physically based probabilistic model that includes both lateral and basal root reinforcement, is adapted for application in silvopastoral landscapes. Root reinforcement models for physical modelling are subsequently developed for the most abundant tree classes based on allometric relationships to remotely sensed, above-ground metrics. In contrast, the statistical landslide susceptibility model uses binary logistic regression, including empirically derived individual tree influence models alongside topographic and lithological explanatory variables. Binary logistic regression displays a better performance (AUC of 0.911 and 0.936 vs. 0.871 and 0.826) for our study area, resulting from the model being trained and tested on an inventory of past shallow landslide observations, whereas SlideforMAP makes predictions largely independent of past slope failures. SlideforMAP has the advantage of computing shallow landslide susceptibility for different rainfall scenarios coupled with specific return periods. The two predictions of landslide susceptibility within the study area were > 70 % in agreement. Differences in the model predictions are due to the different assumptions and datasets used by the models. Important considerations for land management include recognizing the advantages and limitations of both approaches in relation to erosion control objectives. We argue that the contrast afforded by the two different approaches to slope stability modelling in terms of assumptions, methods, and spatial predictions can increase process understanding, since – while being different – both approaches are geomorphologically plausible. However, to increase performance of erosion and sediment mitigation, we suggest land managers prioritize tree planting in areas where the statistical and physical approach agree within the “high” susceptibility classification (8 % of the study area). Comparing shallow landslide susceptibility scenarios based on potential trees and a treeless baseline can further improve targeted planting.