Google and AI Revolutionize Flood Forecasting

Google and AI: A Groundbreaking Advancement in Predicting Flash Floods

Flash floods, although brief, are among the deadliest natural disasters, claiming the lives of over 5,000 people each year. Their unpredictability makes them particularly dangerous. However, Google believes it has found an innovative solution to this problem by turning to the news.

Despite the abundance of weather data collected by humans, flash floods often evade precise measurement due to their localized and transient nature. Unlike phenomena such as temperature or river flow, these events leave few measurable traces. This lack of data complicates the use of deep learning models, which have proven effective in other areas of weather forecasting.

To circumvent this challenge, Google researchers leveraged Gemini, their advanced language model, to analyze a vast amount of textual data. They sifted through 5 million news articles from around the world, identifying reports related to 2.6 million distinct floods. This information was transformed into a chronological and geolocated database, named "Groundsource." According to Gila Loike, product lead at Google Research, this is the first time the company has used language models for such a project. The results of this research were recently made public.

With Groundsource, researchers were able to train a model based on a Long Short-Term Memory (LSTM) neural network. This model incorporates global weather forecasts to estimate the likelihood of flash floods in specific areas.

Google's flash flood forecasting model is now operational in 150 countries via the company's Flood Hub platform. It also shares its data with emergency management agencies around the globe. António José Beleza, an emergency response official at the Southern African Development Community, testified to the model's effectiveness, stating that it has allowed his organization to respond more quickly to floods.

However, the model is not without flaws. Its resolution remains relatively low, only allowing for risk identification over areas of 20 square kilometers. Furthermore, it does not achieve the accuracy of the flood alert system from the National Weather Service in the United States, particularly because it does not incorporate local radar data, which is essential for real-time monitoring of precipitation.

One of the project's goals is to provide a solution where local governments lack the resources to invest in costly weather infrastructure or do not have sufficient historical weather data.

Juliet Rothenberg, program lead within Google's Resilience team, explained that aggregating millions of reports in the Groundsource dataset helps rebalance the map of available information. This paves the way for extrapolations to less documented regions.

Rothenberg also expressed hope that the use of LLMs to convert qualitative sources into quantitative datasets could be extended to other difficult-to-predict natural phenomena, such as heatwaves and landslides.

Marshall Moutenot, CEO of Upstream Tech, a company specializing in river flow forecasting using deep learning models, praised Google's initiative. According to him, this effort is part of a broader movement to gather data to improve deep learning-based weather forecasting models. Moutenot, also a co-founder of dynamical.org, an organization dedicated to compiling weather data for research and startups, emphasized that the lack of data poses a major challenge in geophysics. He described Google's approach as "truly creative" in addressing this deficit.