科学者に聞く：AIは洪水をどのように予測できるか？Ask a Scientist: How can researchers use AI to predict a flood?

Flooding ranks among the most frequent and destructive natural hazards, affecting hundreds of millions of people each year and causing significant economic damage. In a recent interview, a Google researcher explained how artificial intelligence tools such as Flood Hub and Groundsource are being used to forecast floods earlier and more accurately, with the goal of helping communities around the world prepare and respond.

Flood Hub is Google's public-facing flood forecasting platform, which grew out of a research effort the company began several years ago. The system applies machine learning to a combination of data sources, including historical flood records, real-time river level measurements, rainfall data, and satellite imagery, to predict when and where rivers are likely to overflow. According to Google, the platform can issue riverine flood forecasts up to seven days in advance in many locations, giving residents, emergency services, and aid organizations more lead time to act.

One of the central challenges the researcher describes is data scarcity. Reliable flood prediction has traditionally depended on dense networks of river gauges and long records of local hydrological data, infrastructure that many vulnerable regions simply do not have. A key advantage of the AI approach is its ability to make forecasts for so-called ungauged or poorly monitored basins by learning patterns from areas where more data is available and transferring that knowledge. This is particularly relevant in parts of Africa, Asia, and Latin America, where flooding is common but monitoring resources are limited.

Groundsource appears to complement this work by focusing on ground-level information. As described, it is aimed at gathering local, on-the-ground observations that can validate model outputs and improve their reliability. Combining large-scale model predictions with localized data is a recurring theme in flood science, because a forecast is only useful if it reflects real conditions on the ground and reaches the people who need it in a form they can act on.

The broader context is a rapid expansion of AI in weather and climate applications. Google DeepMind, for example, has published research on machine learning weather models such as GraphCast and GenCast, which aim to produce forecasts faster and, in some cases, more accurately than traditional physics-based methods. Established institutions like the European Centre for Medium-Range Weather Forecasts and national meteorological agencies have also been experimenting with AI-driven models. Flood forecasting sits downstream of weather prediction, since accurate rainfall estimates are a prerequisite for anticipating how rivers will respond.

It is important to note the limits of these systems. AI models are only as good as the data they are trained on, and flood dynamics can be influenced by local factors such as dams, urban drainage, land use changes, and rapidly shifting weather that are difficult to capture. Forecasts are probabilistic rather than certain, and effective early warning still depends on communication channels, local authorities, and community preparedness. The researcher frames the technology as a tool to extend and improve existing efforts rather than a replacement for on-the-ground expertise.

Access and reach are also part of the story. Google has worked to make Flood Hub forecasts freely available and to distribute warnings through channels such as Search, Maps, and Android notifications in supported regions, as well as through partnerships with humanitarian organizations. The stated aim is to close the gap between a technically accurate forecast and an alert that actually reaches an at-risk household in time.

As climate change is expected to intensify extreme rainfall and flooding in many parts of the world, interest in scalable, low-cost forecasting is likely to grow. Tools like Flood Hub and Groundsource illustrate how AI is being applied not only to consumer products but to disaster resilience, though their long-term impact will depend on continued data collection, validation, and integration with local warning systems. For now, the technology is best understood as one layer in a larger response system, where accurate prediction is a necessary but not sufficient step toward reducing the human cost of floods.