Groundsource: Google uses AI for better flood forecasts

Floods are a deadly danger and often difficult to predict. Artificial intelligence could improve predictions. However, the problem is that suitable training material has not been available in sufficient quantities until now. This is where Google now wants to contribute with a new project called Groundsource by analyzing news articles from around the world in over 80 languages and extracting suitable data material from them. 2.6 million historical flood events have already been published as an open access dataset.

Google has already demonstrated with its WeatherNext forecasting model that AI models can make useful predictions for the future based on past knowledge – its successor WeatherNext 2 already achieves significantly higher speeds. That Google's AI models can even outperform human experts in hurricane forecasting was also confirmed by independent researchers recently.

Chaotic data situation

Unlike weather data, the data situation for flood events was significantly more chaotic. According to Google, a standardized observation infrastructure is lacking. Existing databases such as the satellite-based Global Flood Database (GFD) and the Dartmouth Flood Observatory (DFO) primarily record large, long-lasting disasters and have physical limitations. Other data collections are too small to train AI models on a global scale.

The relevant news articles first had to be collected by a bot and translated into English via the Cloud Translation API. In the next step, a classification was carried out using the Gemini-LLM: the AI had to filter out reports of actual events from those about warnings or political debates.

The remaining articles were precisely located in time and space and, after comparison with the Google Maps Platform, entered into a database. Manual checks showed that 60 percent of the extracted events were exact in location and time; 82 percent were sufficiently accurate for practical analysis, as Google writes in a blog post.

Forecasts up to 24 hours in advance

The data comprises 2.6 million flood events in more than 150 countries and covers the period from 2000 to the present day. There is a particularly large amount of data for the period from 2020 to 2025, as the increase in digital news has led to a higher data density here. Google states that events can be predicted up to 24 hours in advance. The forecasts are provided via Google's Flood Hub, which provides risk alerts for urban areas in more than 150 countries and also shares its data with disaster management authorities in the affected regions.

However, there are some limitations: The model currently still has a coarse spatial resolution. In addition, there is no interface to local radar data on precipitation. For regions that do not have access to such infrastructure, the AI model is still better than nothing. Prospectively, AI is also to be used for the prediction of landslides and heatwaves.

(mki)