Floods stand as the most prevalent natural calamity, inflicting approximately $50 billion in annual global financial damages and affecting nearly 1.5 billion people or 19% of the world’s population with significant flood event risks.
The escalation of flood-related disasters, doubling since the year 2000 mainly due to climate change, underscores the critical need for advanced global flood forecasting. In response, a concerted effort since 2017 has led to the development of a real-time operational global flood forecasting system, leveraging machine learning (ML) technologies to enhance forecast accuracy and timeliness, particularly in regions lacking local data.
The journey in improving global flood forecasting has been marked by significant research advancements, culminating in the publication “Global prediction of extreme floods in ungauged watersheds” in Nature. This study showcases the efficacy of ML in enhancing flood forecasts on a global scale, notably in areas where flood data is scarce. By employing AI-based technologies, the reliability of nowcasts has been extended from zero to five days on average, with significant improvements in forecasts across Africa and Asia, bringing their accuracy on par with those available in Europe. The Flood Hub, powered by these technologies, now offers real-time river forecasts up to seven days in advance for over 80 countries, a tool that communities, governments, and organizations can utilize for anticipatory actions to safeguard vulnerable populations.
The foundation of this global flood forecasting capability is a suite of ML models developed through extensive research and collaboration with partners across academia, governments, and NGOs. A notable pilot early warning system was launched in the Ganges-Brahmaputra river basin in India, exploring ML’s potential to address the complexities of reliable flood forecasting at scale. This initiative has expanded, incorporating an inundation model, real-time water level measurements, hydrologic modeling, and ML-based hydrologic models, such as LSTM-based models, which have shown superior accuracy compared to traditional approaches.
These hydrological models, essential to the forecasting process, analyze publicly available weather data and watershed information, addressing the challenge posed by the low percentage of river watersheds with streamflow gauges. ML’s capability to train a model on available data and apply it to ungauged basins significantly enhances the predictive accuracy for any river location, thereby democratizing access to critical flood forecasts, especially in lower GDP countries more vulnerable to flood risks.
The advanced river forecast model employs two sequential LSTMs to process historical and forecasted weather data, alongside static watershed attributes, to generate probabilistic forecasts. This novel approach, validated against the current state-of-the-art global flood forecasting system, GloFAS version 4, demonstrates ML’s capacity to offer earlier and more accurate warnings for major and impactful events.
Looking forward, the flood forecasting initiative aligns with Google’s broader Adaptation and Resilience efforts, reflecting a commitment to leveraging AI and ML in advancing science and research towards effective climate action. Collaborations with international aid organizations and the World Meteorological Organization (WMO) are ongoing, aiming to further expand flood forecasting coverage globally and to encompass additional flood-related events, including flash floods and urban floods. This effort not only showcases a significant leap in flood forecasting technology but also underscores a collective aspiration to harness the power of AI for the greater good, making communities around the world more resilient in the face of climate change-induced challenges.
https://blog.research.google/2024/03/using-ai-to-expand-global-access-to.html