During crises, people rely on Google for accurate information to protect themselves and their loved ones. Floods are a common natural disaster affecting approximately 1.5 billion people, or 19% of the global population, with economic damages amounting to around $50 billion annually.
Historically, accurate flood forecasting on a large scale was challenging due to the complexity of the issue and limited resources and data availability. The lack of streamflow gauges in most rivers posed an additional safety obstacle for communities in developing and vulnerable regions.
A recent study published in Nature reveals how AI technology has enabled enhanced flood forecasting globally, benefiting areas most affected by climate change. By utilizing AI, more accurate riverine flood predictions have been made up to 7 days in advance in 80 countries, covering 460 million residents. These forecasts are made accessible through Google Search, Google Maps, and Android notifications.
The study, detailed in our AI blog, showcases the Google Research-built global hydrological technology, which significantly improves flood forecasting compared to existing methods. This advancement is particularly beneficial in regions with limited flood data, broadening the scope of global flood forecasting. Early warnings provided by these systems play a crucial role in reducing casualties, and having more lead time proves invaluable for communities. Through AI-based forecasting, the reliability of current global predictions has been extended from zero to five days on average, with enhanced accuracy in Africa and Asia akin to that in Europe.
This information empowers individuals, communities, governments, and aid organizations to take proactive measures and safeguard vulnerable populations. Overcoming the challenges in regions with scarce data and disproportionately high flood impact has been a significant accomplishment. As we present our latest findings, we reflect on pivotal moments that have shaped our journey in utilizing AI for precise riverine flood forecasting:
Insights From Our Initial Trial in India
Our research journey began with a pilot project in India’s Patna region, known for its susceptibility to devastating floods. Collaborating with local authorities and leveraging real-time data, we developed flood forecasts integrated into Google Public Alerts in 2018. By incorporating various factors such as historical events, river levels, and terrain characteristics, we created accurate river flood models through extensive simulations for specific locations.
While our initial focus was on specific locations, our aim was to resolve global flood forecasting challenges using machine learning techniques. The hypothesis was that machine learning could address the scale requirements of global flood forecasting.
Collaborations Propel Advancements
In 2019, we enhanced our flood forecasting coverage by 12 times and issued 800,000 alerts to affected individuals, advancing our forecasting methodologies. As we explored machine learning’s potential for better flood modeling, we initiated partnerships with academic researchers to combine hydrological physics-based simulations with AI approaches.
Based on our progress and the promise shown by Long Short-Term Memory networks (LSTMs) in accurate predictions, we envisioned a comprehensive global flood forecasting platform delivering trusted information, especially in regions lacking streamflow gauges.
Scaling Up Flood Forecasts With Constraints
Following the successful pilot in India, we expanded our forecasts to cover 360 million people in India and Bangladesh, providing predictions up to 48 hours in advance. Technological advancements significantly contributed to this progress, although local streamflow data availability remained crucial, limiting forecast scalability.
Transition to a Global AI-Based Forecasting Model
Recognizing the data limitations for flood forecasts and leveraging AI advancements, we shifted our focus to a comprehensive global model. This transition necessitated training the model on global data using LSTM networks to predict floods even in regions lacking local streamflow data.
In 2022, the launch of the Flood Hub platform provided forecast access in 20 countries, including 15 in Africa, where forecasting had been previously constrained by limited global data.
By 2023, we extended coverage to 60 new countries across Africa, Asia-Pacific, Europe, South, and Central America, benefiting 460 million people globally. Free real-time forecasts are now available on the Flood Hub, aiding vulnerable communities in developing nations. With advancements in our global AI-based model, Africa now enjoys flood forecasting levels comparable to Europe.
Embracing Collaborative Partnerships
To continue advancing science and positively impacting communities in need, collaboration with academia, local governments, and international organizations remains critical.
Our cooperation with international aid organizations facilitates actionable flood forecasts. Collaborating with the World Meteorological Organization (WMO) supports early warning systems, including the Early Warnings for All initiative, aiming to provide global early warnings for climate hazards by 2027. Ongoing studies explore leveraging AI to address challenges faced by national flood forecasting agencies.
Our close collaboration with academics and hydrological organizations through workshops and initiatives like the Caravan project facilitates data standardization and aggregation, promoting scientific progress.
Continuing the Journey
As climate change intensifies, unexpected floods pose increasing threats. Our commitment is to leverage our research and technology to expand coverage, predict various flood events, such as flash floods and urban inundations, and explore AI’s role in addressing climate adaptation challenges and broader climate and sustainability initiatives.
