Zhice Fang, Yi Wang, Cees J. van Westen , Luigi Lombardo 

Portraying spatiotemporal variations in landslide susceptibility patterns is crucial for landslide prevention and management. In this study, we implement a space-time modeling approach to predict the landslide susceptibility on a yearly basis across the main island of Taiwan, from 2004 to 2018. We use a Bayesian version of a binomial generalized additive model, which assumes that landslide occurrences follow a Bernoulli distribution. We generate 46,074 slope units to partition the island of Taiwan and divided the time domain into 14 annual units. The binary landslide label assigned to each slope unit and their temporal replicates come from an available landslide database, that contains an inventory for every year. We only consider new landslides or reactivations of previous mass movements in the yearly inventories. This information and its absence counterpart are regressed against a set of static and dynamic covariates.
Our modeling strategy features an initial explanatory model to test the goodness-of-fit and interpret the effect of covariates. Then, five cross-validation (CV) schemes are tested to provide a full spectrum of the predictive capacity of our model. Specifically, we implement a fully randomized 10-fold CV, a spatially constrained CV, two temporal CV (a leave one year out and a sequential temporal aggregation), together with a spatio-temporal CV. We summarize the performance in each of these tests, through their pure numerical expression as well as their residual representation in space and time.
Overall, our space-time model produces excellent and interpretable results. We consider this type of dynamic prediction the new direction to take to finally move away from the static view provided by traditional susceptibility models. And, we consider such analyses just a stepping stone for further improvements, the most natural of which would lead to statistical simulations for future scenarios.