This paper studies both the dissemination structure characteristics of urban disaster online public opinion by utilizing complex networks and social network analysis methods, and the changes of public opinion mood by using machine learning. The research finds that the dissemination of urban disaster event has the characteristics of life curve, with short incubation period and the outbreak period, as well as wide and fast multi-level dissemination paths. Government, media and opinion leaders are the key nodes in the public opinion dissemination network, but differences exist among them in public opinion moods and changes, which are easy to produce secondary public opinion. The research reveals that information is diverse and transparent in the context of new media, accompanied by rational thinking of the public and progress of civil society. The research puts forward some recommendations: adopting corresponding guidance strategy in stages, releasing authoritative messages continuously at a hierarchical level, and eliminating information blind spots. Moreover, it is necessary to strengthen the control of harmful information, respond effectively to the online rumors and severely punish rumor disseminators, and to maintain the online public opinion ecology of urban disasters.