Natural disasters such as floods, cyclones, and earthquakes cause damage to life and property. India’s climate is prone to various natural disasters. Disaster situation monitoring is essential for efficient disaster management. Social media is one of the quickest ways to share real-time information, making it crucial for analysing and monitoring natural disasters. Relying solely on information found online is inadequate due to its lack of credibility and the risk of misinformation. The main challenge lies in extracting disaster-related information from text data. Text feeds can provide disaster type, location, and urgency sentiment. With the location and disaster type extracted, relevant authorities can make informed decisions for relief and rescue missions. Instead of conventional three-class sentiment analysis, the proposed system performs binary urgency classification, in which the “negative” class is treated as a proxy for urgent disaster information. The “positive” and “neutral” classes are combined into a single “not negative” class, representing non-urgent information related to preparedness, recovery, and general updates. The proposed solution integrates transformer-based Natural Language Processing (NLP) models, news article and weather station data retrieval and performs automated information extraction. The proposed operational natural disaster monitoring system also validates its performance through case studies against official meteorological and disaster-reported data. Models such as Bidirectional Encoder Representations from Transformers (BERT) and Robustly Optimized BERT Pretraining Approach (RoBERTa) are fine-tuned on custom datasets created from historical news articles and tweets for disaster news extraction in this study. The BERT model is used to apply a relevance filter over text (binary classification) and achieves an accuracy of 91.52%. The RoBERTa model is fine-tuned for urgency sentiment analysis (binary classification) into “negative” and “not negative,” resulting in an accuracy of 90.90%. The question-answering (QA) model is used to extract the location and disaster type from online text feeds. The system has been running continuously since 16 July 2024. From July to October 2024, the news was validated for the tasks it performs. This yielded an accuracy of 92.04% for the text relevance filter, 90.42% for binary sentiment analysis, 92.71% for location extraction, and 94.52% for disaster type extraction. Further analysis covers one-year datasets from July 2024 to July 2025. The data processed by the system is validated against official meteorological data for rainfall in Delhi and Gujarat over two weeks, flooding in Assam, and two cyclone tracks in India. The following is the repository link for the code and the datasets used: <a href="https://drive.google.com/drive/folders/1tFHTHYKLntKaWyjz-2yIeRVSzuhsD2Z4">https://drive.google.com/drive/folders/1tFHTHYKLntKaWyjz-2yIeRVSzuhsD2Z4</a>. Graphical Abstract The AI-based natural disaster information monitoring system is illustrated in a graphical abstract. It consists of four modules. The first module focuses on model development. The disaster data are collected from social media tweets and historical news articles. The data are manually labelled to create two datasets, one for urgency sentiment analysis and another for relevance filtering. For urgency analysis, data are classified into two classes: “negative” and “not negative.” The neutral and positive are combined into a single class (“not negative”) to improve task efficiency and accuracy. For relevance filtering, data with natural disaster keywords were labelled as relevant or irrelevant. The BERT model is fine-tuned for the relevance filtering task, while RoBERTa is fine-tuned for binary urgency sentiment analysis. The second module fetches news from Really Simple Syndication (RSS) feeds of Indian news channels, pre-processes the data to extract relevant features, and then stores the data in a relational database. The news feed information extracted by the second module is further validated by the third module. The third module is designed to retrieve data published by the India Meteorological Department’s weather stations. The dataset includes meteorological warnings and data on rainfall, wind speed, and temperature over the study area, all stored in a relational database (PostgreSQL). The disaster information processed by the deep learning model is superimposed onto the real-time IMD data for validation. The final module is the development of a web-based spatial information system for data visualisation and analysis. It features a spatial dashboard that enables users to view and analyse information collected from the second and third modules for disaster situation assessment and monitoring. Urgency sentiment analysis plays a significant role in assessing disaster situations. It provides a detailed and synoptic view of the disaster situation across India in near real-time by analysing data from over 220 RSS feed links. This timely disaster information supports the planning of satellite data acquisition for flood mapping and damage assessment. It enables efficient utilisation of Indian Remote Sensing (IRS) satellites through adaptive sensor tasking over affected areas identified by the AI-based natural disaster information monitoring system.

Automated Natural Disaster Monitoring from Online News & Social Media Posts Using Sentiment Analysis & GIS

期刊论文
发表日期：

2026年7月22日

查看全文

Vinod Kumar Sharma,
Kaustubh Hari Sethi,
Sameer Saran,
Abhinav Kumar Shukla,
Shilpa Mahajan

摘要

Natural disasters such as floods, cyclones, and earthquakes cause damage to life and property. India’s climate is prone to various natural disasters. Disaster situation monitoring is essential for efficient disaster management. Social media is one of the quickest ways to share real-time information, making it crucial for analysing and monitoring natural disasters. Relying solely on information found online is inadequate due to its lack of credibility and the risk of misinformation. The main challenge lies in extracting disaster-related information from text data. Text feeds can provide disaster type, location, and urgency sentiment. With the location and disaster type extracted, relevant authorities can make informed decisions for relief and rescue missions. Instead of conventional three-class sentiment analysis, the proposed system performs binary urgency classification, in which the “negative” class is treated as a proxy for urgent disaster information. The “positive” and “neutral” classes are combined into a single “not negative” class, representing non-urgent information related to preparedness, recovery, and general updates. The proposed solution integrates transformer-based Natural Language Processing (NLP) models, news article and weather station data retrieval and performs automated information extraction. The proposed operational natural disaster monitoring system also validates its performance through case studies against official meteorological and disaster-reported data. Models such as Bidirectional Encoder Representations from Transformers (BERT) and Robustly Optimized BERT Pretraining Approach (RoBERTa) are fine-tuned on custom datasets created from historical news articles and tweets for disaster news extraction in this study. The BERT model is used to apply a relevance filter over text (binary classification) and achieves an accuracy of 91.52%. The RoBERTa model is fine-tuned for urgency sentiment analysis (binary classification) into “negative” and “not negative,” resulting in an accuracy of 90.90%. The question-answering (QA) model is used to extract the location and disaster type from online text feeds. The system has been running continuously since 16 July 2024. From July to October 2024, the news was validated for the tasks it performs. This yielded an accuracy of 92.04% for the text relevance filter, 90.42% for binary sentiment analysis, 92.71% for location extraction, and 94.52% for disaster type extraction. Further analysis covers one-year datasets from July 2024 to July 2025. The data processed by the system is validated against official meteorological data for rainfall in Delhi and Gujarat over two weeks, flooding in Assam, and two cyclone tracks in India. The following is the repository link for the code and the datasets used: https://drive.google.com/drive/folders/1tFHTHYKLntKaWyjz-2yIeRVSzuhsD2Z4.

Graphical Abstract

The AI-based natural disaster information monitoring system is illustrated in a graphical abstract. It consists of four modules. The first module focuses on model development. The disaster data are collected from social media tweets and historical news articles. The data are manually labelled to create two datasets, one for urgency sentiment analysis and another for relevance filtering. For urgency analysis, data are classified into two classes: “negative” and “not negative.” The neutral and positive are combined into a single class (“not negative”) to improve task efficiency and accuracy. For relevance filtering, data with natural disaster keywords were labelled as relevant or irrelevant. The BERT model is fine-tuned for the relevance filtering task, while RoBERTa is fine-tuned for binary urgency sentiment analysis. The second module fetches news from Really Simple Syndication (RSS) feeds of Indian news channels, pre-processes the data to extract relevant features, and then stores the data in a relational database. The news feed information extracted by the second module is further validated by the third module. The third module is designed to retrieve data published by the India Meteorological Department’s weather stations. The dataset includes meteorological warnings and data on rainfall, wind speed, and temperature over the study area, all stored in a relational database (PostgreSQL). The disaster information processed by the deep learning model is superimposed onto the real-time IMD data for validation. The final module is the development of a web-based spatial information system for data visualisation and analysis. It features a spatial dashboard that enables users to view and analyse information collected from the second and third modules for disaster situation assessment and monitoring. Urgency sentiment analysis plays a significant role in assessing disaster situations. It provides a detailed and synoptic view of the disaster situation across India in near real-time by analysing data from over 220 RSS feed links. This timely disaster information supports the planning of satellite data acquisition for flood mapping and damage assessment. It enables efficient utilisation of Indian Remote Sensing (IRS) satellites through adaptive sensor tasking over affected areas identified by the AI-based natural disaster information monitoring system.