In statistical applications, choosing a suitable data distribution or likelihood that matches the nature of the response variable is required. To spatially predict the planimetric area of a landslide population, the most tested likelihood corresponds to the Log-Gaussian case. This causes a limitation that hinders the ability to accurately model both very small and very large landslides, with the latter potentially leading to a dangerous underestimation of the hazard. Here, we test a distribution-agnostic solution via a Graph Transformer Neural Network (GTNN) and implement a loss function capable of forcing the model to capture both the bulk and the right tail of the landslide area distribution. An additional problem with this type of data-driven hazard assessment is that one often excludes slopes with landslide areas equal to zero from the regression procedure, as this may bias the prediction towards small values. Due to the nature of GTNNs, we present a solution where all the landslide area information is passed to the model, as one would expect for architectures built for image analysis. The results are promising, with the landslide area distribution generated by the Wenchuan earthquake being suitably estimated, including both zeros, the bulk and the extremely large cases. We consider this a step forward in the landslide hazard modelling literature, with implications for what the scientific community could achieve in light of a future space–time and/or risk assessment extension of the current protocol.

中文翻译：

---

通过 Graph Transformers 进行与分布无关的滑坡灾害建模


在统计应用程序中，需要选择与响应变量的性质相匹配的合适数据分布或似然。为了在空间上预测滑坡种群的平面面积，经过最多测试的可能性对应于 Log-Gaussian 情况。这会导致一个限制，阻碍对非常小和非常大的滑坡进行准确建模的能力，后者可能会导致对危险性的低估。在这里，我们通过图变换神经网络 （GTNN） 测试了一个与分布无关的解决方案，并实现了一个损失函数，该函数能够强制模型捕获滑坡区域分布的主体和右尾。这种类型的数据驱动的灾害评估的另一个问题是，人们通常会从回归过程中排除滑坡面积等于零的斜坡，因为这可能会使预测偏向于较小的值。由于 GTNN 的性质，我们提出了一种解决方案，其中所有滑坡区域信息都传递给模型，正如人们所期望的为图像分析而构建的架构一样。结果是有希望的，汶川地震产生的滑坡面积分布得到了适当的估计，包括零、大块和极大情况。我们认为这是山体滑坡灾害建模文献向前迈出的一步，对科学界根据当前协议的未来时空和/或风险评估扩展可以实现的目标产生影响。