Risk management of nonstationary floods depends on an understanding of trends over a range of flood frequencies representing small (frequent) to large (infrequent) floods. Quantile regression is applied to the annual peak streamflow distributions at 2683 sites in the contiguous United States to test for trends in the 10th quantile (floods with a 0.9 annual exceedance probability), the 50th quantile (median annual flood), and 90th quantile (floods with a 0.1 annual exceedance probability). Trends are most common (36% of sites) for the median annual flood (50th quantile) and often coherent with trends in both frequent small floods (10th quantile) and infrequent large floods (90th quantile). Changes in the at-site variance of annual peak streamflow, indicated by convergence (decreasing variance) or divergence (increasing variance) of the 10th and 90th quantiles over time, are primarily in response to reservoir operation or urban development rather than climate. An analysis of synthetic series generated from nonstationary distributions demonstrates that quantile regression and standard trend tests used in flood frequency analysis have limited power and high rates of false negatives (>70%) when a test has a significance of p = 0.05. Quantile regression and tests with lower significance complement standard trend testing to inform flood risk management.

中文翻译：

---

年峰值流量趋势的频域评估及其意义

非平稳洪水的风险管理取决于对代表小（频繁）到大（罕见）洪水的一系列洪水频率趋势的理解。分位数回归应用于美国本土 2683 个站点的年度峰值流量分布，以测试第 10 分位数（每年超过概率为 0.9 的洪水）、第 50 分位数（年度洪水中位数）和第 90 分位数（洪水）的趋势年超标概率为 0.1）。年度洪水中位数（第 50 分位数）的趋势最常见（占站点的 36%），并且通常与频繁的小洪水（第 10 分位数）和罕见的大洪水（第 90 分位数）的趋势一致。年峰值流量的现场方差变化，由第 10 分位数和第 90 分位数随时间的收敛（减少方差）或发散（增加方差）表明，主要是对水库运营或城市发展的反应，而不是气候。对非平稳分布生成的合成序列的分析表明，当测试具有显着性时，洪水频率分析中使用的分位数回归和标准趋势检验具有有限的功效和较高的假阴性率 (>70%)p  = 0.05。具有较低显着性的分位数回归和测试补充标准趋势测试，以告知洪水风险管理。