How robust are landslide susceptibility estimates?

Abstract

Much of contemporary landslide research is concerned with predicting and mapping susceptibility to slope failure. Many studies rely on generalised linear models with environmental predictors that are trained with data collected from within and outside of the margins of mapped landslides. Whether and how the performance of these models depends on sample size, location, or time remains largely untested. We address this question by exploring the sensitivity of a multivariate logistic regression—one of the most widely used susceptibility models—to data sampled from different portions of landslides in two independent inventories (i.e. a historic and a multi-temporal) covering parts of the eastern rim of the Fergana Basin, Kyrgyzstan. We find that considering only areas on lower parts of landslides, and hence most likely their deposits, can improve the model performance by >10% over the reference case that uses the entire landslide areas, especially for landslides of intermediate size. Hence, using landslide toe areas may suffice for this particular model and come in useful where landslide scars are vague or hidden in this part of Central Asia. The model performance marginally varied after progressively updating and adding more landslides data through time. We conclude that landslide susceptibility estimates for the study area remain largely insensitive to changes in data over about a decade. Spatial or temporal stratified sampling contributes only minor variations to model performance. Our findings call for more extensive testing of the concept of dynamic susceptibility and its interpretation in data-driven models, especially within the broader framework of landslide risk assessment under environmental and land-use change.

Temporal Variations in Landslide Distributions Following Extreme Events: Implications for Landslide Susceptibility Modeling

“…This result is similar to that of Ozturk et al. (2021) who found that the accuracy of a logistic regression based susceptibility model saturated after 0.01% of the study region had failed and was used to train the model. The reason for this saturation likely relates to the averaging of landslide distributions.…”

“…This suggests that BLR-based landslide susceptibility modeling needs to move away from time-independence, which assumes landslide distributions are static, toward more dynamic or time-dependent modeling that can account for expected or unexpected temporal variations in landslide distributions, particularly following extreme events. This suggestion follows a growing number of similar recommendations in the literature for landslide susceptibility to be considered temporally dynamic (e.g., Gorsevski et al, 2006;Lombardo et al, 2020;Meusburger & Alewell, 2009;Ozturk et al, 2021).…”

Landslide Geometry Reveals its Trigger

“…Although event‐based inventories include clear traces of the triggering mechanisms of landslides (von Specht et al., 2019), many other inventories, such as satellite‐based, lack crucial information linking a given landslide to a specific triggering mechanism (Behling et al., 2014). The missing information about triggering mechanisms decreases the efficacy of these inventories in landslide hazard analyses, as this could introduce biases, for instance, inadvertently using earthquake‐triggered landslides to assess landslide hazard for extreme rainfall (Ozturk et al., 2020). Hence, there is a need to identify triggers of landslides in exiting databases to make them usable in hazard models.…”

Can global rainfall estimates (satellite and reanalysis) aid landslide hindcasting?

“…The base model includes only elevation, hillslope inclination, and total curvature along with the geology as landslide predictors (Table 1). These topographic features and geology are commonly ranked very high in landslide susceptibility studies (e.g., Schicker and Moon 2012;Althuwaynee et al 2014;Meyer et al 2014;Ozturk et al 2020) Additional topographic, land cover, or landuse related covariates potentially increase the performance of the base model, yet may also lead to overfitting. As we are primarily interested in how far incorporation of rainfall products increases predictive performance, we thus avoided including any other metrics.…”