Regional flood frequency analysis (RFFA) is widely used to quantify flood risk at ungauged and sparsely gauged locations. There are minimal attempts to use partial duration series (PDS) for RFFA, though the use of PDS instead of widely used annual maximum series (AMS) can offer some advantages. This article contributes two novel random/regression forests (RFs)-based methodologies, namely generalized pareto distribution (GPD)-based distributional RFs (DRFs) and multivariate RFs (MVRFs), for RFFA with PDS. The RFs facilitate modeling interactions between predictors and their complex relationships with the predictands without explicitly specifying them. The DRFs and MVRFs comprise an ensemble of corresponding regression trees, each constructed by recursive binary partitioning of the feature space into meaningful segments. The proposed DRFs account for the sampling uncertainty of PDS in the partitioning and parameter estimation. In DRFs (MVRFs), quantile estimates for an ungauged site are obtained using maximum likelihood estimates (expected values) of GPD parameters corresponding to the segments to which the site belongs. The potential of DRFs and MVRFs relative to two recently proposed techniques (univariate RFs-based quantile regression, generalized additive model based on GPD) is demonstrated through Monte-Carlo simulation experiments and a study on 1,031 watersheds in the United States. The key features influencing scale and shape parameters of GPD fitted to PDS of the watersheds are identified as drainage area and 24-hr rainfall intensity corresponding to 2-year return period, respectively. Those identified for shape parameter differ from key features known based on analysis with AMS and generalized extreme value distribution.

中文翻译：

---

具有部分持续时间序列的区域频率分析的分布回归森林方法

区域洪水频率分析 (RFFA) 被广泛用于量化未测量和测量稀疏位置的洪水风险。对 RFFA 使用部分持续时间序列 (PDS) 的尝试很少，尽管使用 PDS 代替广泛使用的年度最大序列 (AMS) 可以提供一些优势。本文为带有 PDS 的 RFFA 贡献了两种基于随机/回归森林 (RF) 的新方法，即基于广义帕累托分布 (GPD) 的分布 RF (DRF) 和多变量 RF (MVRF)。RF 有助于对预测变量之间的交互以及它们与预测变量之间的复杂关系进行建模，而无需明确指定它们。DRFs 和 MVRFs 包含一组相应的回归树，每个回归树都通过将特征空间递归二进制划分为有意义的段来构建。所提出的 DRF 考虑了 PDS 在分区和参数估计中的采样不确定性。在 DRF (MVRF) 中，未测量站点的分位数估计是使用对应于站点所属段的 GPD 参数的最大似然估计（预期值）获得的。DRFs 和 MVRFs 相对于最近提出的两种技术（基于单变量 RFs 的分位数回归、基于 GPD 的广义加性模型）的潜力通过蒙特卡罗模拟实验和对美国 1,031 个流域的研究得到证明。影响流域 PDS 拟合的 GPD 尺度和形状参数的关键特征分别为流域面积和对应于 2 年重现期的 24 小时降雨强度。