Flash floods, recognized as some of the deadliest and most challenging weather events globally, cause over 5,000 fatalities annually and are notoriously difficult to predict. Google researchers believe they have found an innovative solution to this problem by leveraging news articles from around the world.

Unlike other weather phenomena such as temperature or river flows, flash floods are too transient and localized for comprehensive measurement through traditional means. This data scarcity has hindered deep learning models, despite their growing capabilities in weather forecasting, from effectively predicting flash flood occurrences.

To bridge this critical data gap, Google researchers utilized Gemini, the company's large language model (LLM), to sift through an extensive collection of 5 million news articles worldwide. This process allowed them to identify and isolate reports of 2.6 million distinct flood events, subsequently transforming these reports into a geo-tagged time series dataset named “Groundsource.” According to Gila Loike, a Google Research product manager, this marks the first instance of Google employing language models for this specific type of work. The research and the Groundsource dataset were publicly shared.

With Groundsource serving as a real-world baseline, the researchers trained a model built upon a Long Short-Term Memory (LSTM) neural network. This model is designed to ingest global weather forecasts and subsequently generate the probability of flash floods in specific areas. Google's flash flood forecasting model is now actively highlighting risks for urban regions across 150 countries via the company’s Flood Hub platform and is sharing its data with emergency response agencies globally.

António José Beleza, an emergency response official at the Southern African Development Community, confirmed that trialing Google's forecasting model significantly improved his organization's speed in responding to flood events.

Despite its advancements, the model does possess certain limitations. Its resolution is relatively low, identifying risk across 20-square-kilometer areas. Furthermore, it does not achieve the same level of precision as the US National Weather Service’s flood alert system, partly because Google’s model currently lacks local radar data for real-time precipitation tracking.

However, a key design principle of this project was its applicability in regions where local governments often lack the financial resources to invest in expensive weather-sensing infrastructure or possess limited historical meteorological data. Juliet Rothenberg, a program manager on Google’s Resilience team, emphasized that by aggregating millions of reports, the Groundsource dataset helps