Influence of landslide inventory timespan and data selection on slope unit-based susceptibility models

Key advantages of modelling landslide susceptibility at the level of slope units—homogeneous landscape elements bound by drainage and divide lines—instead of grid cells have recently been highlighted. However, there has been limited investigation into the sensitivity of a slope unit landslide susceptibility approach to the characteristics of the landslide inventory used for calibration and the modelling approach. Here, a slope unit landslide susceptibility assessment is conducted for the Da Bac district, Vietnam, based on a multi-temporal landslide inventory, using logistic regression and support vector machine classification algorithms and a set of environmental and anthropogenic controlling factors. A landslide inventory for the period 2013–2020 was created using Google Earth© imagery, including large landslide events in 2018 and 2019. Results highlight that models calibrated from a sample of a single-year inventory and validated with a later year have the same accuracy as those calibrated with a random sample of the entire inventory. Regardless of the calibration data used, the support vector machine algorithm consistently outperforms logistic regression. This is evident from the lower standard deviation of susceptibility values observed when compared to those obtained using logistic regression. The landslide susceptibility models for slope units remain reliable, even when calibrated using a temporally short and event-specific landslide inventory.