Abstract:

Based on social media text data, the flood loss estimation model (ISFRD) was constructed that combines the core factors causing flooding, disaster-bearing vectors and real-time sentiment data. First, text information related to flooding on the Sina Weibo platform was extracted based on natural language processing technology to achieve data preprocessing. Geolocation enrichment was then performed and the validity of the Weibo data was verified using the example of the exceptionally heavy rainfall in Henan province. Afterwards, loss estimation was made for several flood events in China based on a multi-source set of factors such as flood causation and sentiment, and the accuracy of the loss estimation was verified against the actual losses. The results are as follows. (1) In social media, the peak sentiment mutation points of heavy rainfall and flooding are mainly concentrated in June to August each year. Also, the peak sentiment change and the discussion of hot flood events have a strong synchronous relationship. (2) Flood losses have an inverse relationship with average sentiment, i.e., the lower the average sentiment value is, the more serious the disaster damage is in general. (3) The ISFRD flood damage model can effectively assess heavy rainfall and flooding events at the provincial (municipal) scale with different degrees of damage, and the estimation results have high accuracy (average accuracy>90%, MAE=27.04, RMSE=45.26). Under the increasingly complex climate environment, the model can provide a certain reference for rapid determination of flood damage, disaster prevention and mitigation, and public opinion guidance.

Key words: Social media, Sentiment analysis, Multi-source data, ISFRD model, Loss estimation