Abstract The idea behind any validation scheme in landslide susceptibility studies is to test whether a model calibrated on a certain data can predict an unknown dataset of the same nature (landslide presences/absences and covariates). Almost the entirety of landslide susceptibility studies are validated by subsetting a single dataset into a training and test sets. This dataset usually corresponds either to event-specific or to historical inventories. Very rarely, a multi-temporal inventory is available and, in the few cases where this condition is met, the validation practices involve training a model on a specific landslide inventory, deriving a single predictive equation and validating it on a subsequent landslide inventory. This commonly leads landslide predictive studies, even those with a strong statistical rigor, to neglect the uncertainty estimation in their modeling scheme. In statistics, validation can also be performed via statistical simulations. This means that after fitting a given model, one can generate any number of predictive functions and test their predictive skills on any type and number of unknown datasets. In this work, we take a similar direction and we apply it to model and validate three separate co-seismic inventories, including an uncertainty estimation phase. We mapped these inventories within the same area in Indonesia, for three earthquakes occurred in 2012, 2017 and 2018. Specifically, we build three event-specific Bayesian Generalize Additive Models of the binomial family. From each model we then simulate 1000 predictive realizations over the remaining two inventories, by using a plug-in scheme where all the morphometric covariates are kept fixed and only the ground motion is replaced according to the prediction target. By doing so, we introduce a new analytical tool for near-real-time landslide predictive purposes, which is able to produce a probabilistic model which stands in between the definitions of susceptibility and hazard. In fact, our model is able to accurately estimate “where” and “when” – although not “how frequently” – landslide have occurred by featuring the multitemporal information of the trigger. In our findings, the simulations are quite similar to the fitted models; and the nine combinations we analyse produce excellent performance. This result confirms the assumption that “the past is the key to the future”, as we show that the relative contribution of each variable and their interactions in each probabilistic model remains practically the same across temporal replicates. This information is not trivial because it supports the routines implemented in global near-real-time applications.

中文翻译：

---

通过插件统计模拟对近实时滑坡敏感性模型进行时间验证

摘要 滑坡敏感性研究中任何验证方案背后的想法是测试基于特定数据校准的模型是否可以预测相同性质的未知数据集（滑坡存在/不存在和协变量）。几乎所有的滑坡敏感性研究都是通过将单个数据集子集到训练和测试集来验证的。该数据集通常对应于特定事件或历史库存。很少有可用的多时间清单，并且在满足此条件的少数情况下，验证实践涉及在特定滑坡清单上训练模型，导出单个预测方程并在随后的滑坡清单上对其进行验证。这通常会导致滑坡预测研究，即使是那些具有很强统计严谨性的研究，在他们的建模方案中忽略不确定性估计。在统计学中，也可以通过统计模拟进行验证。这意味着在拟合给定模型后，可以生成任意数量的预测函数，并在任意类型和数量的未知数据集上测试其预测技能。在这项工作中，我们采用了类似的方向，并将其应用于建模和验证三个独立的同震清单，包括不确定性估计阶段。我们绘制了印度尼西亚同一地区的这些清单，分别针对 2012 年、2017 年和 2018 年发生的三场地震。具体而言，我们构建了二项式族的三个特定事件的贝叶斯广义可加模型。然后，我们从每个模型中模拟剩余两个库存的 1000 个预测实现，通过使用插件方案，其中所有形态测量协变量保持固定，并且根据预测目标仅替换地面运动。通过这样做，我们引入了一种用于近实时滑坡预测目的的新分析工具，该工具能够生成介于敏感性和危险性定义之间的概率模型。事实上，我们的模型能够通过触发的多时信息来准确估计滑坡发生的“地点”和“时间”——尽管不是“发生的频率”。在我们的发现中，模拟与拟合模型非常相似；我们分析的九种组合产生了出色的性能。这个结果证实了“过去是未来的钥匙”的假设，正如我们所展示的，每个变量的相对贡献及其在每个概率模型中的相互作用在时间重复中几乎保持相同。此信息并非微不足道，因为它支持在全局近实时应用程序中实现的例程。