Flash floods, a disaster known as the "number one killer of meteorological disasters," claim over 5,000 lives worldwide every year. The most challenging aspect is that it acts like an unpredictable "assassin," with short duration and limited impact area, often "failing" in predictions even with the most advanced satellite monitoring. However, Google has come up with a creative solution: having Gemini enter "quantum reading" mode to predict disasters by reading news.

According to reports, Google researchers used Gemini to sort through global 5 million news reports, accurately extracting records of 2.6 million different historical flood events. These textual descriptions were converted into "groundsource" data with geographic tags. This is the first time that a large language model has been used to build an extremely valuable set of geophysical quantitative data from unstructured text information.

image.png

With these "text-based" historical experiences, researchers trained a neural network model specifically for predicting flash floods. Currently, this model has been launched on Google's Flood Hub platform, marking risk levels for cities in 150 countries around the world. Emergency officials in Southern Africa have reported that this model has significantly improved their response speed to floods.

The core logic of this project is deeply humanitarian: it is designed specifically for poor areas that cannot afford expensive radar systems and monitoring stations. Even without local radar, as long as there are past news records, AI can provide probability alerts based on weather forecasts.

The head of Google's disaster prevention project said that this "dimension reduction" approach of converting qualitative text into quantitative predictions will also be applied to the prediction of other transient disasters such as heatwaves and landslides in the future. When AI no longer only writes poetry and paints pictures but becomes a "prophet" protecting life, this may be the warmest side of technology.