Drought impacts and drought framing in news and social media: an LLM-based approach

Drought is a slow-onset hazard with impacts that develop over time and space, making it particularly suitable for an impact-based monitoring approach compared to more sudden hazards. This study explores how large language model (LLM)–classified newspaper articles and Twitter messages can be used for drought impact monitoring in the Netherlands, and what needs to be considered when applying such approaches.

Results show that both data sources are valuable for extracting drought impact information and broadly align with temporal drought patterns. Different impact categories exhibit distinct temporal peaks, suggesting that reported impacts may function as early signals of drought development. At the same time, clear differences emerge between data sources. Spatial impact patterns derived from newspapers show greater variation in reported impact counts, while Twitter-based patterns are more strongly shaped by population density and platform-specific usage. The most frequently reported impact categories per region reflect underlying land-use characteristics.

Importantly, impact reporting is not neutral. The type of news outlet and social media actor influences which drought impacts are emphasised, and drought attention in both newspapers and social media is subject to memory effects and competition with other societal events. Building on these insights, the study additionally explores how drought is framed in social media discourse, distinguishing between diagnostic, prognostic, and motivational framing, and examining how these framing types evolve over time and across drought phases.

Overall, the results highlight that developing a drought impact monitoring system requires explicit choices regarding data sources, classification methods, impact definitions, and interpretative lenses, as these choices directly shape how drought impacts, vulnerabilities, and societal responses are represented.