西北太平洋台风是最具破坏力的灾害性天气系统之一，给我国东南沿海受 台风影响区带来极端强降水，并引发洪涝等次生灾害，造成巨大的经济损失。 对台风强降水的危险性分析以及风险评估，一直是减灾防灾工作的重点。在全 球暖化的背景下，极端降水事件的强度和频率都在增加。台风产生的强降水是 极端降水的重要组成部分，其危险性也极有可能会随着全球暖化而升高。然而， 传统的危险性分析是基于平稳极值分布模型进行的，未能考虑全球暖化造成的 极值分布的时间非平稳性，也未能有效描述区域极值分布的空间非均一性。本 研究尝试应用广义极值分布的广义加性模型，同时描述区域台风降水极值的时 间非平稳性和空间非均一性，并应用于危险性分析。 本研究以福建为研究区域，基于 1961—2019 年福建台风降水极值观测，结 合台风最佳路径数据集，以各站点的年最大台风过程降水作为年台风降水极值， 并分析其时空特征。利用基于薄板样条函数的广义加性模型来描述地形的影响， 以得到精细化的危险性分析。应用极值分布的广义加性模型模型，分别建立了 台风降水极值的非平稳时空广义加性模型、变系数线性趋势的广义加性模型和 空间广义加性模型，以重现期的形式计算了福建台风降水极值的非平稳危险性 区划，并分析其过去趋势和趋势延续之下的未来情景，为福建地区的台风强降 水危险性评估提供科学依据。 理论分析和应用结果都表明，在线性时间趋势的假设条件之下，台风降水 极值危险性具有非线性、非对称的变化趋势。通常所关注的“极值的平均状态”， 其危险性变化可能不足以造成重大风险，而“极值的极值”的危险性变化则可 能带来灾难性的后果，但这一变化并不能反映在“极值的平均状态”中，有时 甚至与“极值的平均状态”变化趋势相反，即“极值的极值”的危险性呈现上 升趋势时，“极值的平均状态”的危险性呈现微弱下降趋势，从而忽略了对“极 值的极值”的风险防范。特别是在气候变化背景下，分布极值远比分布均值对 气候变化更为敏感，对于极值分布也不例外，而这种变化造成的后果往往是灾 难性的。因此，在与气候变化相关的风险防范研究中，需要对极值危险性的非 平稳趋势给予特别的关注。

Typhoons in the Northwest Pacific are one of the most destructive severe weather systems, bringing extremely heavy precipitation to typhoon-affected areas along the southeast coast of China, and causing secondary disasters such as floods, causing huge economic losses. Hazard analysis and risk assessment of typhoon heavy precipitation have always been the focus of disaster reduction and prevention work. In the context of global warming, the intensity and frequency of extreme precipitation events are increasing. Heavy precipitations from typhoons is an important component of extreme precipitation, and their hazards are likely to increase with global warming. However, the traditional hazard analysis is based on the stationary model of extreme value distribution, which fails to consider the temporal non-stationarity of the extreme value distribution caused by global warming, and fails to effectively describe the spatial inhomogeneity of the regional variation of extreme value distribution. This study tries to apply the generalized additive models of generalized extreme value distribution to describe simultaneously the temporal non-stationarity and spatial inhomogeneity of regional typhoon precipitation extremes, and carries out hazard analysis based on the developed models. Based on the observation of extreme typhoon precipitation in Fujian Province from 1961 to 2019 and combined with the best track dataset of typhoons, this study took the annual maximum typhoon precipitation at each station as the annual typhoon precipitation extreme, and analyzed its spatiotemporal characteristics. A generalized additive model based on thin plate spline function is used to describe the influence of terrain to obtain refined hazard analysis. The non-stationary spatiotemporal generalized additive model, the generalized additive model with linear trend of varying coefficient, and the spatial generalized additive model of typhoon precipitation extremes were established, respectively. By using the generalized additive models of typhoon precipitation extremes, the past trends and the extended future scenarios of typhoon precipitation extremes were calculated in terms of return periods, and were analyzed, so as to provide a scientific basis for the risk assessment of typhoon heavy precipitation in Fujian. Both theoretical analysis and application results show that under the assumption of linear temporal trend, the hazards of typhoon precipitation extremes have nonlinear and asymmetric trends. The change in the hazards of the "average state of the extreme" may not be enough to cause significant risks, while the change in the hazards of the "extreme value of the extreme" may bring catastrophic consequences, but this change cannot be reflected in the "average state of the extreme". Sometimes the two states even show opposite trends, that is, when the hazards of the "extreme value of the extreme value" shows an upward trend, the hazards of the "average state of the extreme value" shows a slight downward trend, thus the risk prevention of the "extreme value of the extreme value" might be ignored. Especially in the context of climate change, the extreme values of the distribution are far more sensitive to climate change than the mean of the distribution, and the generalized extreme value distributions are no exception, and the consequences of such changes are often catastrophic. Therefore, special attention needs to be paid to the non-stationary trend of extreme value hazards in the study of climate change-related risk prevention.