Signals from the Floods: AI-Driven Disaster Analysis through Multi-Source Data Fusion

Xian Gong, Paul X. McCarthy, Lin Tian, Marian-Andrei Rizoiu · 2025

Read Signals from the Floods: AI-Driven Disaster Analysis through Multi-Source Data Fusion by Xian Gong, Paul X. McCarthy, Lin Tian, Marian-Andrei Rizoiu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-17038.txt.

Introduction

Xian Gong and co-authors mine two streams of public text to understand behavior during the 2022 floods in New South Wales, Australia. They analyzed more than 55,000 flood tweets and 1,450 submissions to a public inquiry. Tweets are short and scattered, while submissions are long and structured. The team combined topic modeling with large language models, using the submissions as a reference to judge which tweets were truly flood-relevant.

This relevance filter cut noise and surfaced content that responders could act on. Topic patterns also varied by place and viewpoint. The authors present the fusion of social media with formal submissions as a way to sharpen real-time awareness during disasters. They argue the same approach can inform long-term planning for resilience, giving agencies a cleaner signal from the flood of crisis posts.

Catalog record

CREATORXian Gong, Paul X. McCarthy, Lin Tian, Marian-Andrei Rizoiu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-17038

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Signals from the Floods: AI-Driven Disaster Analysis through Multi-Source Data Fusion Xian Gong, Paul X. McCarthy, Lin Tian, Marian-Andrei Rizoiu Abstract Massive and diverse web data are increasingly vital for government disaster response, as demonstrated by the 2022 floods in New South Wales (NSW), Australia. This study examines how X (formerly Twitter) and public inquiry submissions provide insights into public behaviour during crises. We analyse more than 55,000 flood-related tweets and 1,450 submissions to identify behavioural patterns during extreme weather events. While social media posts are short and fragmented, inquiry submissions are detailed, multi-page documents offering structured insights. Our methodology integrates Latent Dirichlet Allocation (LDA) for topic modelling with Large Language Models (LLMs) to enhance semantic understanding. LDA reveals distinct opinions and geographic patterns, while LLMs improve filtering by identifying flood-relevant tweets using public submissions as a reference. This Relevance Index method reduces noise and prioritizes actionable content, improving situational awareness for emergency responders. By combining these complementary data streams, our approach introduces a novel AI-driven method to refine crisis-related social media content, improve real-time disaster response, and inform long-term resilience planning. 1 Introduction This study presents new research that extends analytical work originally commissioned by the NSW Government to examine the 2022 floods. It builds on those insights by integrating machine learning techniques to improve real-time disaster response and long-term resilience planning. While the independent NSW Flood Inquiry provided key recommendations, our work aligns with both its findings and the government’s subsequent response, contributing to efforts to enhance disaster communication, emergency coordination, and flood risk management. By leveraging machine learning for tweet filtering, this study ensures that relevant disaster-related content is prioritized, further demonstrating the potential of AI-driven methods in crisis management and policy planning. Floods account for 43% of weather disasters (1995–2015), affecting 2.3 billion people UNISDR and CRED 2015 . Between February and April 2022, catastrophic floods devastated Northern New South Wales (NSW) and parts of South-East Queensland, impacting over 70 local government areas. The disaster tragically claimed 22 lives and caused over $5 billion in insured losses, making it the costliest flood in Australian history Insurance Council of Australia 2022 . Sydney, Australia’s most populous city, was also significantly affected by related storms and flooding during this period. In the most affected areas, such as Lismore and Richmond Valley, the devastating flood disaster compelled thousands of people to seek emergency accommodation, damaged thousands of homes, and disrupted business activities, underscoring its profound social, financial, and economic impacts Lismore City Council 2022 ; Richmond Valley Council 2022 ; Queensland Reconstruction Authority 2022 ; Mary O’Kane, Michael Fuller 2022 . The NSW Premier commissioned the independently-led NSW Floods Inquiry to investigate the causes, assess emergency responses, and recommend future flood management strategies. The inquiry conducted extensive consultations—holding 144 meetings and receiving 1,494 submissions—to inform recommendations on future flood risk management and coordination among government agencies Mary O’Kane, Michael Fuller 2022 . With climate change intensifying natural disasters, understanding public responses, crisis communication, and societal impact is increasingly critical. Previous studies on disaster response and resilience face two major limitations. First, they often rely solely on social media posts—mostly Twitter—which are inherently short and fragmented, limiting the breadth of analysis. Second, their text analysis typically uses lexicon-based methods that focus mainly on sentiment while neglecting topic extraction. As a result, the full range of insights available from these posts is not fully exploited. This paper introduces a novel analytics framework to analyse the flood disaster by integrating social media and public submissions. Social media complement data from public submissions to flood inquiries by providing real-time data that captures immediate reactions, concerns, and trends during the event. Multiple data sources analysis enables us to explore public behaviour patterns through different lenses, making the results more comprehensive and robust. We separately analyse the differences in public topics reflected in tweets and public submissions and leverage semantic and syntactic information from large language models (LLMs) and public submissions to quantify the relevance of tweets to flood disasters. We aim to answer the following research questions: 1. How do topics discussed in social media compare to those in public submissions regarding floods? 2. How can machine learning (ML) and large language models (LLMs) improve filtering and relevance assessment of flood-related tweets? Our findings highlight that each medium provides unique perspectives due to its inherent nature and interacts with each other to provide a more comprehensive analysis. RQ1. There are distinct differences between social media discussions and public submissions regarding floods. Public submissions, primarily from flood-affected residents, emergency personnel, and organizations, focus on personal experiences, infrastructure failures, emergency response effectiveness, and long-term policy recommendations. In contrast, social media discussions are more immediate, diverse, and politically charged, covering real-time weather updates, rescue efforts, donation appeals, emotional reactions, and government criticism. While submissions provide structured, in-depth reflections, tweets offer rapid, high-volume insights but contain significant noise and off-topic content. RQ2. To enhance relevance filtering of flood-related tweets, we developed a machine learning-based Relevance Index using LongT5 embeddings. By leveraging public submissions as a reference corpus, we trained an LLM to score tweets based on their semantic similarity to verified flood-related content. This approach effectively filters out unrelated tweets, distinguishing meaningful crisis information from general online chatter. Our results demonstrate that ML and LLMs significantly improve disaster-related information by reducing misinformation, prioritizing critical content, and enhancing real-time situational awareness. These findings underscore the importance of integrating structured public submissions with unstructured social media data to create a more reliable and actionable framework for flood disaster response and policy planning. This paper makes several key contributions to flood disaster response and crisis management. It introduces a multi-source framework to analyse public behavior during floods. Using Latent Dirichlet Allocation (LDA), it compares real-time, diverse social media discussions with structured, policy-focused public submissions. A machine learning-based Relevance Index is developed to filter flood-related tweets, leveraging public submissions as a reference corpus to improve the signal-to-noise ratio in social media data. By fusing traditional and digital data sources, this study enhances disaster preparedness, real-time crisis response, and long-term resilience planning, providing actionable insights for policymakers, emergency responders, and urban planners. 2 Related Work Literature on flood disaster response and management spans emergency response management, resource allocation and decision-making strategies during floods, and community resilience and recovery, which address rebuilding efforts and psychosocial support Burstein et al. 2008 ; Cutter et al. 2008 . Especially related to our work, the role of social media and crowdsourced information in providing real-time situational awareness and enhancing public communication has also been a significant area of investigation Vieweg et al. 2010 ; Fohringer et al. 2015 ; Kryvasheyeu et al. 2015 . The tools and technologies used in flood disaster response and management depend largely on the nature of the data sources. Therefore, we focus on discussing the data sources, social media, public submissions, and corresponding analytical methods involved in our research. 2.1 Social Media Analysis Social networks, particularly X (Twitter), has been recognised as a valuable tool for real-time disaster response in spatiotemporal and sentiment analysis. During Hurricane Sandy, Kryvasheyeu et al. Kryvasheyeu et al. 2016 found that social media data could help rapidly assess spatial distribution damage. Subsequent research by Kankanamge et al.

Read the complete text