The article "Global prediction of extreme floods in ungauged watersheds," published in Nature, showcases the significant advancements made possible by machine learning (ML) technologies in global-scale flood forecasting, particularly in regions where flood-related data is scarce. By leveraging AI-based technologies, the authors were able to extend the reliability of existing global nowcasts from zero to five days and achieve forecast improvements comparable to those available in Europe across regions in Africa and Asia. The evaluation of these models was conducted in collaboration with the European Center for Medium-Range Weather Forecasting (ECMWF).
One of the practical applications of these technologies is demonstrated through the Flood Hub, which provides real-time river forecasts up to seven days in advance, covering river reaches across more than 80 countries. This information serves as a valuable resource for individuals, communities, governments, and international organizations, enabling anticipatory action to protect vulnerable populations.
The ML models powering the FloodHub tool are the result of extensive research conducted over many years in collaboration with various partners, including academics, governments, international organizations, and NGOs.
The initiative began with the launch of a pilot early warning system in the Ganges-Brahmaputra river basin in India in 2018. The pilot expanded in the following years, incorporating an inundation model, real-time water level measurements, elevation mapping, and hydrologic modeling. Collaborating with academic institutions such as the JKU Institute for Machine Learning, the authors explored ML-based hydrologic models, demonstrating that LSTM-based models could produce more accurate simulations compared to traditional methods. This research facilitated flood forecasting improvements, allowing for expanded coverage to include all of India and Bangladesh. Collaboration with researchers at Yale University also explored technological interventions to enhance the reach and impact of flood warnings.
The hydrological models used in this research predict river floods by analyzing publicly available weather data such as precipitation and physical watershed information. However, the challenge arises in regions lacking streamflow gauging stations, which provide crucial data for model calibration. ML addresses this challenge by enabling a single model to be trained on available river data globally, allowing predictions to be made for ungauged basins where no data are available. This approach overcomes the limitations posed by the lack of infrastructure and data in vulnerable regions, ultimately contributing to more effective flood risk management on a global scale.
Source of the research.
