Traffic Impact Area Detection and Spatiotemporal Influence Assessment for Disaster Reduction Based on Social Media: A Case Study of the 2018 Beijing Rainstorm

Li Guoqing, Tengfei Yang, Jibo Xie, Guoqing Li, Naixia Mou, Cuiju Chen, Jing Zhao, Zhan Liu, Zhenyu LinIsprs Int. J. geo inf 2020 · DOI

TL;DR

This paper introduces a novel framework that leverages social media data to detect traffic impact areas during urban rainstorms and assess their spatiotemporal influence. By combining fine-grained road condition information with public sentiment, the approach offers a more comprehensive understanding of disaster impacts than previous methods, crucial for effective disaster reduction.

What Problem Does This Paper Address?

Urban rainstorms pose significant threats, disrupting daily life and public travel. While traditional monitoring methods exist, they often have limitations in cost, coverage, and real-time data acquisition. Social media has emerged as a valuable data source due to its timeliness, low cost, and rich information. However, existing research often struggles to integrate specific road condition details with public sentiment to comprehensively assess traffic impacts. This paper aims to bridge this gap by developing a method that extracts both granular road conditions and public emotions from social media to pinpoint traffic impact areas and understand their evolving influence over time and space.

Methods

The proposed framework integrates several key techniques:

  • Data Acquisition and Parsing: Social media data (Sina Weibo) is collected using a custom tool to overcome API limitations. This involves filtering by time, region, and keywords, followed by parsing to extract time, location descriptions, and text content. Address descriptions are geocoded into latitude and longitude coordinates.
  • Fine-Grained Road Condition Extraction: A novel method combining natural language processing (NLP) and a semantic knowledge base of hazard damage information is employed. This approach uses rule templates to extract candidate words related to road damage and matches them against a knowledge base, which is enriched by external resources. This moves beyond general disaster keywords to specific road conditions.
  • Public Emotion Information Extraction: A convolutional neural network (CNN) is utilized to extract emotional information from social media text. This deep learning approach is chosen over traditional methods for its ability to handle colloquial language and capture contextual sentiment, classifying emotions into positive, negative, and neutral categories.
  • Traffic Impact Area Detection: The extracted road condition and public emotional information are combined with location and road network data to identify traffic impact areas.
  • Spatiotemporal Influence Assessment: A complex network-based model is developed to assess the spatiotemporal influence of detected traffic impact areas. This model incorporates attention degree, interaction degree, and weighting degree, considering the co-reference relationship of events and integrating public emotional values to quantify influence.

Datasets

The primary dataset used in this study is social media data from Sina Weibo, collected for the Beijing rainstorm event on July 16, 2018. The data was filtered for keywords like "rainstorm," "heavy rain," and "flood" within the Beijing region from July 15th to July 18th. For training and testing the road condition classification model, approximately 320 labeled texts were used. For emotion classification, nearly 1700 texts were manually labeled, with 1400 used for training and 300 for testing. Road network data for Beijing was obtained from the "National Catalogue Service For Geographic Information."

Key Insights and Ideas

  • Synergy of Road Conditions and Sentiment: The paper highlights that combining specific road condition information (e.g., "flooded," "road closed") with public sentiment (e.g., "worried," "frustrated") provides a much richer picture of disaster impact than either source alone. Negative sentiment often correlates with more severe impacts.
  • Granularity Matters: Moving beyond general disaster mentions to extract fine-grained road conditions is crucial for accurately identifying traffic disruptions.
  • Social Media as a Dynamic Sensor: The study demonstrates how social media can act as a dynamic sensor, revealing the evolution of disaster impacts over time (e.g., morning peak vs. evening peak) and their spatial spread.
  • Network Analysis for Influence: Applying complex network concepts to social media interactions allows for a quantitative assessment of how an event in one location influences other areas, revealing patterns of attention and impact.

Key Results

  • Extraction Accuracy: The road condition extraction achieved a comprehensive evaluation index (F-1) of over 72% across different damage categories. Public emotion classification achieved an F-1 score of over 78%.
  • Traffic Impact Areas: The study successfully identified key traffic impact areas during the 2018 Beijing rainstorm, particularly around transportation hubs like "Xierqi subway station" and "Huilongguan subway station."
  • Spatiotemporal Evolution: Figures 4 and 7 visually demonstrate how disaster information and public sentiment evolved throughout the day, showing concentrated impacts during peak hours and broader, though sometimes less intense, effects at other times.
  • Influence Assessment: The spatial influence assessment model identified "Xierqi subway station" and "Huilongguan subway station" as having significant attention and weighting degrees, indicating their substantial impact on surrounding areas and beyond.

Limitations and Future Work

  • Geographic Entity Recognition: The paper acknowledges that its current method struggles with implicitly mentioned geographic entities (e.g., "滴滴大厦附近" - near Didi Building) and suggests introducing methods for automatic recognition of such entities to improve road condition extraction.
  • Implicit Semantic Information: Extracting implicit semantic information (e.g., "我在西二旗看海" - I watch the sea at Xierqi, implying flooding) remains a challenge, and the authors propose using more relevant corpora and clustering algorithms like LDA.
  • Emotion Classification Nuances: While overall accuracy is good, the precision of neutral emotion and recall of negative emotion could be improved. The authors suggest incorporating artificial feature engineering, such as extracting emojis and their emotional weights, to enhance classification.
  • Data Integration: Future work will focus on integrating more diverse data sources, such as population distribution data and bus route data, to provide a more comprehensive understanding of regional impacts and people's travel patterns.
  • Location Information: The paper notes that not all social media data contains explicit location information, and future research will explore using geographic-related named entities to infer upload locations.

Summary updated

BibTeX
@article{guoqing2020traffic,
  author = {Li Guoqing and Tengfei Yang and Jibo Xie and Guoqing Li and Naixia Mou and Cuiju Chen and Jing Zhao and Zhan Liu and Zhenyu Lin},
  title = {{Traffic Impact Area Detection and Spatiotemporal Influence Assessment for Disaster Reduction Based on Social Media: A Case Study of the 2018 Beijing Rainstorm}},
  journal = {Isprs Int. J. geo inf},
  year = {2020},
  doi = {10.3390/ijgi9020136},
}
Figure