Google’s new tool predicts flash flood risk

Google announced that it has developed a new Gemini-based data analysis system that aims to help detect flash flood events earlier. This tool, called “Groundsource,” uses the company’s artificial intelligence model to extract data from past news and combine this information with current weather forecasts to assess the risk of flash floods in certain regions. It is known that flash floods are among the natural events that are difficult to predict meteorologically. However, Google aims to take a new approach by turning past recorded events into large-scale data sets.

According to information provided by Google, Gemini was tasked with extracting records of flood events by examining approximately 5 million news articles published around the world. In this process, events in the news were classified chronologically and each was marked with geographical location information. Thus, a comprehensive data set containing flood events in different regions was created. In addition, the researchers trained a separate forecast model that can evaluate this data set together with current meteorological forecasts. The system tries to calculate how high the chance of flash flooding is by analyzing the weather conditions expected in a particular area.

How does the Google Gemini based Groundsource system work?

The Groundsource platform is not limited to using past flood events only as historical records. In addition, it creates a model that produces risk assessment by associating meteorological forecasts with these data. According to the information provided by Google, the system is now available through the Flood Hub platform, which offers risk indicators for cities in 150 countries. On the other hand, the company plans to share the obtained data and analyzes with emergency management institutions. Thus, it is aimed to enable local governments and rescue teams to react more quickly to possible flood risks.

However, the current system also has some technical limitations. The area in which the model can perform risk analysis is limited to regions of approximately 20 square kilometers. In addition, it may offer lower sensitivity than the flood warning systems used by the US National Weather Service. The main reason for this is that local radar data is not used in Google’s model. Local radar data often makes it possible to monitor rainfall amount and movement in real time. However, the Groundsource platform is specifically designed to work in regions that do not have developed meteorological infrastructure.

Juliet Rothenberg, who works as a program manager in the Google Resilience team, states that the system can also be used to predict different natural events in the future. According to Rothenberg, by bringing together millions of reports, inferences can be made for regions where there is a lack of data. In addition, it is stated that this approach can also be used to predict heat waves, landslides and similar risky events.

This is not the first project where Google has used artificial intelligence in weather forecasting. The company’s WeatherNext 2 model, developed by DeepMind, has achieved a high accuracy rate in tests conducted in recent years. However, Groundsource offers a different method, especially in terms of producing meteorological risk assessment using language model-based data analysis. The results of this approach in the long term will be understood more clearly with real-world usage data of the system.