Google is using old news reports and AI to predict flash floods

Flash floods, notorious for their rapid onset and devastating impact, claim over 5,000 lives annually, making them one of the most perilous weather phenomena. Despite advances in meteorology, accurately predicting these events remains a formidable challenge. Google believes it has found a groundbreaking solution by leveraging historical news reports. Traditional weather data collection often falls short when it comes to flash floods, which are typically brief and localized. This limitation hampers the ability of advanced deep learning models to forecast such events effectively. To bridge this gap, Google researchers tapped into their Gemini language model to analyze an extensive database of 5 million news articles worldwide. They identified reports on 2.6 million distinct flood events, creating a geo-tagged time series known as "Groundsource." This innovative approach marks the first instance of Google employing language models for flood prediction, according to Gila Loike, a product manager at Google Research. The research findings, along with the Groundsource data set, were made publicly available on Thursday. Using Groundsource as a foundational reference, the researchers developed a model based on a Long Short-Term Memory (LSTM) neural network. This model integrates global weather forecasts to assess the likelihood of flash floods in specific areas. Currently, Google's forecasting system is operational in urban regions across 150 countries through its Flood Hub platform, providing crucial data to emergency response teams worldwide. António José Beleza, an emergency response official with the Southern African Development Community, noted that Google’s model significantly enhanced his organization’s ability to respond swiftly to flooding incidents. However, the model does have its limitations. It operates at a relatively low resolution, assessing risks over 20-square-kilometer regions, and lacks the precision of systems like the US National Weather Service's flood alerts. This is partly because Google’s model does not utilize localized radar data necessary for real-time precipitation tracking. The initiative aims to address the needs of regions where local governments cannot afford advanced weather-sensing technologies or lack comprehensive meteorological records. "By aggregating millions of reports, the Groundsource data set helps balance the information landscape," explained Juliet Rothenberg, a program manager on Google’s Resilience team. "It allows us to extrapolate data to less-informed regions." Looking ahead, Rothenberg expressed hopes that the application of large language models to generate quantitative data from qualitative sources could extend beyond flash floods to other critical weather phenomena like heat waves and mudslides. Marshall Moutenot, CEO of Upstream Tech, emphasized the importance of data in geophysics, noting that while there is an abundance of Earth data, accurate validation remains a challenge. He praised Google’s initiative as a creative solution to gathering vital data for deep learning weather forecasting models.