Google DeepMind Partners with Met Office for AI-Driven Flood Forecasting

Published 2025-03-05 · Environment

Google DeepMind and the UK Met Office have jointly deployed an AI-based flood forecasting system that provides six-hour advance warnings for communities across the Ganges-Brahmaputra-Meghna river basin in South Asia. The system, described in a paper published in Science on February 28, 2025, uses a graph neural network to model river flow dynamics and has been operational since December 2024, covering a region home to approximately 400 million people.

Flooding in the Ganges-Brahmaputra delta is among the most lethal and economically destructive natural hazards on Earth. The 2024 monsoon season alone displaced an estimated 9.6 million people across Bangladesh and eastern India, with damages exceeding $4.2 billion. Existing flood forecasting in the region relies on physics-based hydrological models that require hours of computation on supercomputers — a latency that limits warning lead times and prevents real-time updates as conditions evolve.

The Model Architecture

The forecasting system, named GraphFlood, represents river networks as mathematical graphs — nodes represent gauging stations and river confluences, and edges represent the flow connections between them. The graph structure allows the model to capture the spatial relationships inherent in watershed dynamics: water levels at one point on a river are not independent of water levels upstream and downstream, and the graph architecture explicitly encodes these dependencies.

The model ingests three primary data streams: real-time water level readings from 1,847 gauging stations across the basin (sourced from the Bangladesh Water Development Board and India's Central Water Commission), satellite-based precipitation estimates from the Global Precipitation Measurement mission, and soil moisture data from the ESA's SMOS satellite. These inputs are processed through a 12-layer graph attention network that predicts water levels at every node in the graph six hours into the future.

The system runs inference in approximately 90 seconds on a single Cloud TPU v5, compared with the 45 to 90 minutes required by the Met Office's existing physics-based flood model running on a Cray XC40 supercomputer. This speed advantage allows the system to update forecasts every 15 minutes during active flood events, providing near-real-time situational awareness that was previously impossible.

Validation Results

GraphFlood was validated against observed flood events during the 2024 monsoon season (June through September), using a strict temporal holdout — the model was trained on data from 2005 to 2023 and evaluated on out-of-sample events from 2024 that it had never seen. Across 342 flood events exceeding the "danger level" threshold at one or more gauging stations, the model correctly predicted the event at least six hours in advance in 89% of cases.

For severe flood events — those exceeding the "highest flood level" historically recorded at a given station — the detection rate was 94%. The median lead time between the model's first warning and the observed flood peak was 7.3 hours, with a standard deviation of 2.1 hours. False alarm rates were manageable: approximately 12% of positive predictions were not followed by a flood event within the warning window.

These figures represent a substantial improvement over existing operational forecasts. The Bangladesh Flood Forecasting and Warning Centre, which uses a combination of statistical and physics-based models, achieved a detection rate of approximately 62% for the same events, with a median lead time of 3.8 hours. The Centre's false alarm rate was 24%.

Real-World Deployment

The system's outputs are currently being integrated into the operational workflows of the Bangladesh Meteorological Department and India's National Disaster Management Authority. During the 2024 monsoon season, GraphFlood's predictions were provided to these agencies as an experimental overlay alongside their existing forecasts. By all accounts, the AI predictions were taken seriously.

Dr. Sultana Razia, a hydrologist at the Bangladesh University of Engineering and Technology who served as an independent evaluator of the system, described a specific case from August 2024: "The Brahmaputra at Sirajganj was rising steadily, but the official forecast suggested it would remain below danger level. GraphFlood predicted it would exceed danger level within seven hours. The district commissioner acted on the AI forecast and ordered evacuations of low-lying char islands. The river peaked 8.3 hours later, 0.7 metres above danger level. Without the early warning, several thousand people would have been stranded."

Not every prediction was acted upon. In some districts, local officials were unfamiliar with the AI system and chose to rely on traditional forecasting methods. The Met Office and DeepMind are working with the United Nations Office for Disaster Risk Reduction to develop training programmes for local emergency management officials.

Technical Challenges and Limitations

The system's primary limitation is its dependence on real-time gauging station data. Across the Ganges-Brahmaputra basin, approximately 30% of gauging stations experience intermittent outages during the monsoon season due to flooding, power failures, or vandalism. When a station goes offline, the model must infer its state from neighbouring stations — a process that degrades prediction accuracy in the affected sub-basin.

DeepMind is exploring the use of satellite-based river width measurements as a supplementary data source for stations that go offline. Researchers at the University of Bristol have demonstrated that Sentinel-1 synthetic aperture radar can detect river width changes with sufficient precision to supplement ground-based gauges, though the spatial and temporal resolution of current satellite observations introduces its own latency.

The model also struggles with flash floods in steep, mountainous tributaries where water levels can rise from normal to catastrophic in under two hours — faster than the model's six-hour prediction window. Addressing this limitation would require either shorter-horizon, higher-resolution models or integration with rainfall nowcasting systems that predict precipitation patterns minutes to hours ahead.

Scalability Plans

DeepMind and the Met Office intend to expand the system to other flood-prone regions, starting with the Mekong delta in Southeast Asia, the Niger river basin in West Africa, and the Amazon floodplain in South America. Each expansion requires site-specific model calibration using local hydrological data, but the underlying graph neural network architecture is transferable.

The team has made the model code and pre-trained weights available under an Apache 2.0 licence, and Google.org is funding a $15 million grant programme to support national meteorological agencies in developing countries that wish to deploy the system. The programme includes technical assistance, hardware provisioning, and training for local forecasters.

Dr. Peter Dueben, the Met Office's head of AI and a co-author of the Science paper, framed the partnership as a template for future collaboration between AI labs and national meteorological services: "Weather and climate agencies around the world are sitting on decades of observational data. AI can unlock the value of that data in ways that traditional numerical models cannot. But the deployment has to be led by the agencies themselves — they understand the local context, the communication channels, and the decision-making processes that turn a forecast into a life-saving action."

For more on AI applications in environmental monitoring, explore our AI for Environment research repository. Related coverage includes Microsoft's AI for Earth grants to climate startups and WHO's global guidelines on AI in health.

← Back to Publications