Same area, different timeframes: unveiling the evolution of landslide patterns and their impact on susceptibility maps

Landslide susceptibility maps (LSMs) depict the spatial probability of landslide occurrence, typically using data-driven models based on past event distributions. However, these models often overlook temporal changes, assuming that past patterns reliably predict future events. In a rapidly changing world, influenced by factors such as climate change and urbanization, this assumption may lead to inconsistencies in LSMs over time, reducing their predictive accuracy. To investigate this issue, we analyzed a multi-temporal inventory of 1518 landslides over the past 70 years in Wanzhou District, China, a region impacted by significant environmental changes. We first examined the temporal and spatial variations in landslide patterns. Then, we employed random forest (RF) and extreme gradient boosting (XGBoost) algorithms to create five LSMs, each based on landslide data from different periods. Finally, we conducted a multi-temporal validation to evaluate the effectiveness of these maps in predicting landslides occurring in subsequent decades. Our findings reveal temporal shifts in landslide spatial patterns and changes in the model-derived importance of causal factors. These variations may be associated with changing environmental conditions in the study area. Both RF and XGBoost models produced LSMs with varying patterns depending on the temporal baseline of the landslide data used. Importantly, the accuracy of these maps decreases as the time span of the inventory data increases. Maps based on short-term data showed higher accuracy, but their predictive power declined rapidly over time, suggesting that in the context of significant environmental changes, LSMs need continuous updates and further spatiotemporal analysis to remain effective.