Google has unveiled a groundbreaking machine learning model that significantly improves global flood forecasting capabilities, offering earlier and more accurate warnings for communities around the world. The new system, developed as part of Google's commitment to climate change solutions, outperforms current state-of-the-art systems like GloFAS version 4, especially for larger and more impactful flood events.
The model leverages data from 5,680 streamflow gauges across the globe, trained on records from 1980 to 2023 provided by the Global Runoff Data Center. By using a dual-LSTM architecture that processes both hydrological and meteorological data, the system achieves reliability scores comparable to or better than GloFAS nowcasts (0-day lead time) up to 5 days in advance.
Key innovations include the ability to predict flood events with high precision and recall over various return periods, from 1-year to 10-year events, and notably, to provide more accurate forecasts for rare but severe events such as 5-year return period floods. This advancement is crucial for early warning systems, particularly for vulnerable populations in regions prone to flooding.
The model's impact extends beyond research, with Google actively collaborating with international aid organizations like the Centre for Humanitarian Data and the Red Cross to deliver actionable flood forecasts. In partnership with the World Meteorological Organization (WMO), Google is also conducting studies to understand how AI can enhance national flood forecasting systems and support climate resilience efforts globally.
Despite these advances, Google emphasizes that there is still work to be done to expand flood forecasting coverage to more locations and to include other types of flood-related disasters such as flash floods and urban floods. The company looks forward to continued collaboration with academic institutions, local governments, and industry partners to achieve broader global impact.
This development marks a significant step forward in applying artificial intelligence to address climate change challenges and strengthen global resilience against natural disasters. The full findings are published in a paper available through Nature.
Source: Google Research Blog
