Analyzing Spatial-Temporal Distribution of Natural Hazard Events in Iran (1390-1400 SH) Automatically Extracted from News Stories

Analyzing natural hazards due to economic, environmental, and social effects is necessary for crisis management. However, good analysis comes from good data. In this case, natural hazard databases or inventories that contain all spatial, temporal, and other relevant information for each event are required. To realize that, news and social media, which provide detailed information about the hazards, are two precious resources. Nevertheless, most works manually extracted events from these text resources in the inventory development process. This paper presents a framework for extracting natural hazard events automatically from news stories by leveraging text mining techniques. By implementing the framework for the study region of Iran, we analyzed the spatial-temporal distribution of natural hazards.

According to spatial and temporal coverage of Mehr news agency, we selected this website as the main resource, and for training machine-learning-based models, we used ISNA's news articles. The process starts by mining web pages. All irrelevant records such as the news stories about maneuvers, conferences, and contradicts are removed automatically. Then, standardizing the texts of news articles should be accomplished. In the next stage, the text classification technique determines whether a news story is about a newly occurred event or a current event, in which context about the event the news story is published, and which natural hazards are pointed out in the text. For the first and the second text classification task, we used machine-learning-based models which achieved 0.875 and 0.716 for F-score, respectively. For the third task, we developed leveraged a rule-based model. The results of text classifications are used in the proceeding steps, including toponym recognition and resolution, information extraction, and topic detection and tracking.

The results show that although most natural hazards have a specific temporal distribution, the highest total frequency is for a Solar Hijri year's initial and final months. Spatially, the storm occurs more in eastern than western provinces. The diversity of other meteorological hazards, except dust, in northern provinces, is more than in southern ones. Regardless of the fatalities, the highest frequency of reported earthquake events is for southern provinces.

This paper presents a framework for automatically extracting natural hazard events from news stories. The framework leverages several text mining techniques such as text classification and information extraction to develop an inventory of the hazards. We implemented the framework to analyze the natural hazards of Iran from 1390 to 1400 Solar Hijri. Comparing the number of extracted events with the thematic maps published by the National Disaster Management Organization shows that the differences vary with the province. Based on that, the frequency of extracted events for some provinces equals the official statistics; hence the analysis for mined events is generalizable to real-world situations.