Flash floods are one of the deadliest meteorological disasters in the world, killing more than 5,000 people every year. However, they have long been difficult to accurately predict due to their rapid onset, small scope, and short duration. Faced with this problem, Google’s new answer is to “let AI read news reports.”
Traditional meteorological monitoring has accumulated a large amount of data such as temperature, rainfall, and river flow. However, for sudden and extremely strong flash flood events, humans do not have as complete and continuous observation records as other meteorological elements. This has led to the fact that even though deep learning is becoming more and more powerful in the field of weather forecasting, it has been unable to perform at the same level in flash flood prediction because there is a lack of enough "true value" data to train the model.
To fill this data gap, the Google research team used its large language model Gemini to screen about 5 million news reports from around the world, automatically identify and extract about 2.6 million different flood events, and then convert these text reports into a sequence data set "Groundsource" with time and geographical tags. Gila Loike, Google's research product manager, said that this is the first time the company has used a large language model to complete this type of quantitative data construction work. The relevant research results and data sets were publicly released early Thursday morning.
After obtaining this "real-world baseline", the researchers trained a new flash flood prediction model based on a long short-term memory (LSTM) neural network, allowing it to input global weather forecast data and output the probability of flash floods in a specific area. Currently, Google's flash flood prediction model has provided risk tips for urban areas in 150 countries on its Flood Hub platform, and has opened data to many emergency management agencies around the world. António José Beleza, an emergency response officer at the Southern African Development Community (SADC), said in a trial with Google that the model helped his team respond to floods faster.
However, this system still has obvious limitations. On the one hand, its spatial resolution is relatively low and it can currently only provide risk assessment on a scale of about 20 square kilometers; on the other hand, because it does not incorporate real-time precipitation monitoring data such as local radar, its accuracy is not as accurate as the existing flood warning system of the National Weather Service in the United States.
Google emphasized that one of the original intentions of this project was to play a role in developing areas that lack expensive weather observation infrastructure and have no long-term weather records. By aggregating millions of news reports from around the world, the Groundsource dataset "rebalances the map" to a degree that allows models to extrapolate predictions to areas where data is otherwise scarce. Juliet Rothenberg, a program manager on Google's resilience team, said this approach allowed the team to cover areas where information was previously severely lacking.
Rothenberg also said that the idea of using large language models to convert text narratives into structured quantitative data is not limited to flash floods. In the future, similar technologies are expected to be used to build data sets on equally short-lived but extremely important natural phenomena such as heat waves and mudslides, providing a basis for the prediction of more extreme weather and geological disasters.
According to industry insiders, Google’s attempt is an important step in promoting the development of deep learning weather forecasting through creative data collection. Marshall Moutenot, CEO of Upstream Tech, a company that also uses deep learning to predict river flow for clients such as hydropower companies, pointed out that the current field of earth science is facing the persistent problem of "data scarcity": on the one hand, earth observation data is extremely complex, and on the other hand, there are very limited high-quality "truth values" that can be used to calibrate and validate models. Moutenot is also the co-founder of dynamical.org, an organization dedicated to organizing weather data sets for researchers and startups that can be directly used in machine learning. He believes that Google’s work is a typical example of obtaining valuable data through “very creative methods.”