随着全球范围内的城镇化发展，城市水循环受到较大干扰，城市内涝频繁发生，科学解析并解决城市内涝迫在眉睫。武汉市是受到内涝侵袭的典型城市之一，内涝发生机理复杂，影响因子众多，科学地分析内涝影响因子，量化其对内涝形成的影响程度对城市内涝管理具有重要意义。本文以武汉市及中心城区为研究区域，对影响城市内涝的水循环及下垫面要素演变规律进行多角度时空分析，探求内涝发生的空间分布特征，运用机器学习模型对城市内涝影响因子进行了定量化研究。主要结论如下：（1）武汉市水循环及下垫面要素的时空演变规律近40年武汉市降雨量整体呈现下降趋势，但是短历时强降雨强度增大，频次上升，空间上呈现出明显的“雨岛效应”；气温整体呈现显著上升趋势，存在明显的“热岛现象”，且有显著加强趋势；潜在蒸散发整体呈现上升趋势，中心城区及武汉市东部显著增加；土地利用变化剧烈，建设用地扩张明显，中心城区湖泊等水体面积逐渐下降，调蓄能力减弱；不透水面面积逐年扩大，暴雨下渗受阻严重。（2）武汉市中心城区内涝空间分布特征基于一般降雨和极端降雨两种情景下的内涝空间分布图，武汉市中心城区的内涝点在空间上显示出显著的强相关性，整体空间上呈现出稳定的分布格局，在局部地区存在不同程度上的空间差异性，可以对强相关性的地区进行全局统一规划管理，再局部分块，削弱内涝灾害的空间相关性，做好内涝的预险管理工作。（3）基于机器学习的内涝影响因子研究基于逻辑回归、随机森林和支持向量机三种机器学习算法，综合考虑选择了11个内涝因子，通过五个评价指标对三种机器学习模型进行评价，结果表明随机森林表现最优。选择随机森林模型进一步进行参数优化，计算得出中心城区内涝易感性分布图，量化各个影响因子对内涝形成的影响程度，并通过公开发布的内涝风险点对模型的预测能力进行了验证。研究结果可为武汉市内涝风险预测提供方法支撑。关键词：城市内涝；城市水循环；时空分布；机器学习

With the development of global urbanization, the urban water cycle is greatly disturbed, which causes frequent urban waterlogging. It is urgent to analyze and solve the urban waterlogging scientifically. Wuhan is one of the typical cities suffered from waterlogging. Because of the complicated mechanism and numerous influencing factors of waterlogging, it is of great significance to analyze and quantify the influencing factors of waterlogging scientifically. Based on the traditional GIS analysis tools, the research focuses on the whole scope of Wuhan and its central urban area, searching for the rules of urban water cycle and the underlying surface elements’ changes. Then, the spatial distribution characteristics of waterlogging are explored. The machine learning models are used to quantify the weight of the influencing factors of urban waterlogging. The main conclusions are summarized as follows:(1) Temporal and spatial evolution law of water cycle and underlying surface factors in Wuhan cityFirstly, the precipitation in Wuhan shows a downward trend as a whole over the past 40 years. However, the intensity and frequency of occurrence of short-term heavy precipitation increase, with a significant “rain island effect” in space. Secondly，the overall temperature shows a significant upward trend, with an obvious “heat island effect” and an obvious strengthening trend. Then，the potential evapotranspiration shows an overall upward trend, the center and eastern part of Wuhan changes significantly. Finally，the land use has changed drastically. The construction land has expanded significantly, the area of water bodies such as lakes in the central urban area has gradually decreased, causing the storage capacity weakened. The impervious surface expands year by year, the rainstorm infiltration is blocked seriously.(2) Spatial distribution characteristics of waterlogging in the central urban area of Wuhan Based on the data of waterlogging under different precipitation scenarios, the waterlogging points in the central urban area of Wuhan show a significant and strong spatial correlation. The overall spatial distribution presents a stable distribution pattern, and spatial differences exist in some local areas. Overall planning and management of areas with strong correlations can be more effectively helpful for the pre-risk management of waterlogging.(3) Quantitative research on the influence factors of waterlogging based on machine learning modelBased on the comprehensive analysis, eleven factors of waterlogging are selected to build the three machine learning models of logistic regression, random forest and support vector machine. Five evaluation indicators are used to evaluate the simulation effect of three models. The results show that the random forest model has the best performance. After further parameter calibration, the random forest model is used to draw the waterlogging susceptibility distribution map of the central urban area. The influence of each factor on the formation of waterlogging is quantified. The model is validated by the publicly released waterlogging risk points. These research results can provide a reference for the management and control of urban waterlogging. Key words: urban waterlogging; urban water cycle; temporal and spatial distribution；machine learning