Extracting Geospatial Information from Social Media Data for Hazard Mitigation, Typhoon Hato as Case Study (Short Paper)

Li Guoqing, Jibo Xie, Tengfei Yang, Guoqing LiGiscience 2018 · DOI

TL;DR

This paper introduces a novel method, the Social Media based Hazard information Recognition and Classification (SHRC) model, to extract fine-grained geospatial information about hazard damages from social media data. By analyzing text and sentiment, the model aims to provide near real-time, valuable data for disaster mitigation efforts, as demonstrated through a case study of Typhoon Hato.

What Problem Does This Paper Address?

In the aftermath of natural disasters, timely and accurate information about the extent of damage is crucial for effective mitigation and relief efforts. While social media platforms generate a massive amount of real-time, user-generated content, this data is often unstructured, fragmented, and challenging to process. Traditional geospatial data acquisition methods can be slow and may not capture the granular details of damage experienced by affected populations. This paper addresses the challenge of extracting actionable geospatial information and classifying hazard losses from the complex and noisy landscape of social media data to supplement existing disaster management tools.

Methods

The core of the paper is the Social Media based Hazard information Recognition and Classification (SHRC) model. The workflow involves several key steps:

  1. Event-driven hazard information acquisition: Utilizing social media platform APIs to retrieve data based on time, location, and event-specific keywords.
  2. Data cleaning and storage: Filtering out redundant or irrelevant messages and storing the cleaned data for analysis.
  3. Definition of hazard loss categories: Proposing eleven categories to classify hazard damages, including loss of life, building damage, and infrastructure damage.
  4. Creation of a classification knowledge base: Developing a lexicon of feature words and their collocations, augmented by word vector models and external corpora, to aid in identifying hazard-related information.
  5. Hazard information interpretation and extraction: Employing Chinese language processing tools (NLPIR) for word segmentation, part-of-speech tagging, and matching feature words against the knowledge base to extract disaster loss information.
  6. Sentiment analysis: Analyzing the emotional tone (positive, neutral, negative) of social media posts using a sentiment lexicon, extended with feature words from social media.
  7. Spatio-temporal visualization: Geo-locating extracted information using GPS, address matching, and place name identification to create visualizations and maps of affected areas.
  8. Evaluation and validation: Using precision, recall, and F-measure to assess the accuracy of the extracted information.

Datasets

  • Training Data: The paper mentions using a dataset of 2017 typhoon events (approximately 20,000 records) for training the model.
  • Case Study Data: For the case study of Typhoon Hato, approximately 1600 records of hazard-related information were collected from "Sina-Weibo" (https://weibo.com/) after cleaning and filtering.

Key Insights and Ideas

  • Social media as a valuable, near real-time data source: The paper strongly emphasizes that social media, despite its complexities, offers a unique and timely perspective on disaster impacts that traditional methods often miss.
  • Granular damage classification is possible: By developing a detailed eleven-category classification system, the authors demonstrate that specific types of hazard losses can be identified and quantified from social media.
  • Sentiment analysis adds crucial context: Understanding public sentiment alongside damage reports provides a richer picture of the disaster's impact on individuals and communities, revealing how people react to relief efforts even amidst damage.
  • Spatio-temporal dynamics of hazards: Visualizing hazard information over time and space reveals how impacts evolve and spread, offering critical insights for dynamic response planning.

Key Results

  • The SHRC model was tested on the 2017 Typhoon Hato event.
  • The paper reports that the comprehensive evaluation index for different disaster loss categories was greater than 0.74, and for sentiment categories, it was greater than 0.83.
  • Analysis of "Sina-Weibo" data for Typhoon Hato revealed that power outage and interruption of water supply were the most frequently reported damages.
  • Sentiment analysis showed varying emotional responses to different hazard types, with positive sentiment sometimes expressed towards relief efforts even when damage was present.
  • Spatio-temporal analysis illustrated the progression of hazard impacts, with power and water outages becoming prominent as the typhoon made landfall.

Limitations and Future Work

  • Data Complexity and Fragmentation: The authors acknowledge that social media data is inherently complex, fragmented in format and semantics, posing ongoing challenges for extraction.
  • Language Processing: The current model is specifically designed for Chinese language processing.
  • Future Work: The authors are developing a near real-time social media-based hazard information acquisition and analysis system. They plan to use more hazard events to further train and validate the model, aiming for its practical application in hazard mitigation.

Summary updated

BibTeX
@inproceedings{guoqing2018extracting,
  author = {Li Guoqing and Jibo Xie and Tengfei Yang and Guoqing Li},
  title = {{Extracting Geospatial Information from Social Media Data for Hazard Mitigation, Typhoon Hato as Case Study (Short Paper)}},
  booktitle = {Giscience},
  year = {2018},
  doi = {10.4230/lipics.giscience.2018.65},
}
Figure