Role of baseline landslide inventory and grid rainfall precision on the sensitivity of susceptibility or hindcast models

Multivariate logistic regression models are the most popular in estimating landslide susceptibility by assessing various landslide causes—covariates—in mapped landslides or hindcasting landslides by including landslide triggering information such as rainfall. Although the sensitivity of these models to the variety of input data is frequently tested, the influence of data quality on the model accuracy is rarely discussed. For example, accurately representing spatial rainfall variability that triggered landslides may be essential in hindcasting models. Additionally, the properties of the mapped landslides, such as sample size, location, or time, are crucial to set a robust susceptibility model. Using an inventory that predominantly covers larger landslides would hinder a model by broadly covering the diversity of the factors leading to slope instability. Whereas smaller landslides could fail to capture sufficiently the range of values in the covariate space, likely decreasing the model performance. Another aspect is whether the number of mapped landslides is enough to estimate the susceptibility accurately or does more data means a better model. We developed several simple logistic regression models to answer all the above-listed questions relevant to assessing the model sensitivity. The model first demonstrated that global grid rainfall products could not accurately represent spatial rainfall distribution, which has a major influence on a landslide hindcast model. We have further found out that using only part of the individual landslides surprisingly may suffice to make accurate susceptibility estimates. Using smaller landslides in a susceptibility model outperforms a model that relies on larger landslides. Lastly, the model performance marginally varied after progressively adding more landslide data in a pilot study.