Google has developed a novel approach to predicting flash floods by leveraging old news reports and artificial intelligence (AI). Flash floods, known for their unpredictability, have posed significant challenges for traditional forecasting methods. Google’s solution involves transforming qualitative news reports into quantitative data using Gemini, Google’s large language model (LLM).
By analyzing 5 million news articles worldwide and extracting data on 2.6 million floods, Google created a geo-tagged time series called ‘Groundsource.’ This innovative methodology marks Google’s first application of language models for such weather-related tasks.
Using the Groundsource dataset as a foundation, Google’s researchers developed a predictive model based on a Long Short-Term Memory (LSTM) neural network. This model integrates global weather forecasts to estimate the likelihood of flash floods in specific regions.
The impact of Google’s flash flood prediction model is already evident, with urban areas in 150 countries benefiting from risk assessments on Google’s Flood Hub platform. Emergency response agencies worldwide are leveraging this data to enhance their flood response strategies.
While the model has limitations, such as lower resolution and precision compared to existing systems, it represents a significant step forward in leveraging AI and data to improve flood prediction capabilities.
Source: TechCrunch