AI reads the past to protect the future
Google researchers have developed a new technique that uses large language models (LLMs) to transform qualitative descriptions in old news reports into quantitative flood data that can feed flash-flood prediction systems. By extracting details such as water depth, affected locations, timing, and infrastructure damage from archived articles, the approach creates structured historical records where instrumented data are sparse or missing.
The core innovation is a pipeline that prompts and fine-tunes an LLM to identify and standardize flood-related measurements buried in narrative text. These machine-extracted datapoints can augment existing hydrological datasets, enabling researchers and emergency services to train and validate predictive models with richer historical context—without the time and expense of new sensor deployments.
Practically, this means improved coverage for regions that lack dense monitoring networks, and faster enhancements to early-warning systems. Scalable text-to-data conversion allows responders and planners to uncover decades of local flood history from newspapers and reports, helping communities better anticipate flash-flood risks and prioritize mitigation or evacuation strategies.
Looking ahead, the technique complements—not replaces—physical observations and local knowledge. When integrated into forecasting pipelines and decision-support tools, the approach offers a low-cost, high-leverage way to strengthen resilience to extreme rainfall and sudden floods, putting historical insight to work for safer communities.