A Multidimensional Study of the 2023 Beijing Extreme Rainfall: Theme, Location, and Sentiment Based on Social Media Data

Xun Zhang, Xin Zhang, Yingchun Zhang, Ying Liu, Rui Zhou, Abdureyim Raxidin, Min LiISPRS Int. J. Geo Inf 2025 · DOI

TL;DR

This paper introduces a novel framework for analyzing extreme rainfall events using social media data, focusing on theme evolution, spatial distribution, and public sentiment. By leveraging large language models and a custom Bert-BiLSTM model, the study provides a nuanced understanding of how public discourse shifts before, during, and after a disaster, offering valuable insights for disaster management.

What Problem Does This Paper Address?

Extreme rainfall events, exacerbated by climate change, pose significant threats to urban infrastructure and public safety. Effectively managing these disasters requires timely and comprehensive information. While social media offers a rich source of real-time data, extracting meaningful insights from the sheer volume and complexity of posts remains a challenge. Existing research often focuses on single aspects (e.g., topic analysis or sentiment analysis) or uses broad sentiment categories. This paper aims to bridge this gap by developing a multidimensional approach to analyze disaster-related social media content, capturing the dynamic evolution of themes, the spatial clustering of impacts, and the fine-grained emotional landscape of public and official responses.

Methods

The study employs a multi-pronged approach:

  • Disaster Theme Extraction: Large language models (LLMs) and prompt engineering are used to extract disaster-related keywords from Weibo hashtags. These keywords are then clustered using k-means to identify distinct themes across pre-disaster, mid-disaster, and post-disaster phases.
  • Disaster Spatial Information Extraction: The framework identifies primary and secondary disaster events, categorizing nine types of secondary disasters triggered by extreme rainfall. It then analyzes the geographical locations of these events to understand spatial clustering.
  • Disaster Sentiment Classification: A novel Bert-BiLSTM multi-sentiment classification model is developed. This model uniquely incorporates three key features: comment content, emoticons, and post content. It refines sentiment categories beyond the typical positive/neutral/negative to eight specific disaster-related emotions (Gratitude, Optimism, Neutral, Sympathy, Fear, Anger, Anxiety, Helplessness). The model also considers contextual information from posts and comments for more accurate sentiment analysis.

Datasets

The primary dataset used in this study is social media data from Sina Weibo.

  • Official Media Posts: 497 posts from official media outlets were collected.
  • Public Posts and Comments: 5,791 posts from ordinary users and 60,546 public comments were gathered.

The data was collected for the period of the 2023 Beijing extreme rainfall event, specifically from July 28 to August 10, 2023, and divided into three stages: pre-disaster, during-disaster, and post-disaster.

Key Insights and Ideas

  • Temporal Shift in Public Discourse: The study reveals a clear evolution of social media themes, moving from warnings and prevention (pre-disaster) to specific events and rescue actions (during disaster), and finally to expressions of gratitude and social cohesion (post-disaster). This highlights how public attention and emotional focus adapt to the disaster lifecycle.
  • Spatial Aggregation of Secondary Disasters: Key secondary disaster events, such as trapped personnel and building damage, are found to be spatially clustered, particularly in mountainous and low-lying areas like Mentougou and Fangshan. This underscores the critical role of geographical factors and infrastructure in disaster impact and necessitates targeted resource allocation.
  • Nuanced Sentiment Analysis: The development of an eight-category sentiment classification model, incorporating emoticons and post context, provides a richer understanding of public emotions than traditional models. This allows for the differentiation of subtle emotional expressions during a crisis.
  • Divergent Communication Styles: A notable difference exists between official media (neutral, fact-based reporting) and public sentiment (more diverse, emotional). This highlights the need for tailored communication strategies that acknowledge and address public emotions while maintaining factual accuracy.

Key Results

  • The Bert-BiLSTM model achieved an accuracy of 96.09% in multi-emotion disaster classification, outperforming both the standalone Bert model (95.34%) and the Qwen2.5-7B-Instruct LLM (93.11%).
  • The ablation study demonstrated that incorporating emoticons improved accuracy by 1.43%, and adding post content as context improved accuracy by 3.85%. The combined model achieved a 5.65% overall increase in accuracy compared to using only comment content.
  • The analysis identified Mentougou and Fangshan as areas with significant spatial clustering of disaster events and public attention.
  • Sentiment analysis revealed that neutrality and sympathy were the most prevalent sentiments in affected districts, with anxiety and fear also being significant during the mid-disaster phase.

Limitations and Future Work

  • Data Confinement: The study is limited to Sina Weibo data, which may not capture the full spectrum of public opinion or disaster impact, especially in areas with lower social media penetration.
  • Generalizability: The model's generalizability across different languages and types of disasters is not fully explored.
  • Future Work: The authors suggest expanding the analysis to multilingual platforms, integrating multimodal data (images, videos), and developing dynamic models for long-term sentiment tracking and real-time disaster response. They also propose exploring broader impacts on socioeconomics and ecological environments.

Summary updated

BibTeX
@article{zhang2025multidimensional,
  author = {Xun Zhang and Xin Zhang and Yingchun Zhang and Ying Liu and Rui Zhou and Abdureyim Raxidin and Min Li},
  title = {{A Multidimensional Study of the 2023 Beijing Extreme Rainfall: Theme, Location, and Sentiment Based on Social Media Data}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2025},
  doi = {10.3390/ijgi14040136},
}
Figure