Seeing the Flood Before It Comes: AI for Water Resources and Flood Risk

For many people, climate change is experienced first through water: too much of it during intense rainfall, too little of it during drought, or water in the wrong place when rivers overflow and coasts flood. In Lagos, a coastal megacity laced with lagoons and creeks, heavy rains regularly turn roads into rivers. Elsewhere in Nigeria, riverine flooding along major river systems can displace large numbers of people in a single season.

The World Meteorological Organization and the UN have championed the goal of early warnings for everyone, recognising that timely alerts are among the most cost-effective ways to reduce disaster losses. The challenge is that classical hydrological forecasting depends on dense networks of river gauges and well-calibrated models, which many basins in Africa simply do not have. This is precisely the gap where AI has started to make a real difference, and it is the area of climate tech closest to my own background in geomatics and my work with StrikeClimate.

Learning rainfall-runoff relationships

Traditional hydrological models encode physical processes such as infiltration, evaporation and routing, and they must be calibrated basin by basin. Researchers showed that machine learning can learn these relationships directly from data. The influential study Rainfall-runoff modelling using Long Short-Term Memory networks (Kratzert et al.) demonstrated that LSTMs trained across many catchments could match or outperform established conceptual models. Because a single model learns from many basins at once, it can transfer knowledge to basins with little or no local data.

Forecasting in ungauged basins

Building on that idea, a team at Google Research published Global prediction of extreme floods in ungauged watersheds in Nature, showing that an AI model could extend reliable riverine flood forecasts to regions without gauges, with particular relevance for Africa and other data-scarce regions. Those forecasts are surfaced publicly through Flood Hub.

Inundation mapping from satellites

Knowing that a river will rise is half the story; knowing which streets and farms will flood is the other half. Synthetic aperture radar (SAR) satellites can see through cloud, which is crucial during storms. Segmentation models trained on SAR and optical imagery map flood extent rapidly, and services such as the Copernicus Emergency Management Service produce rapid maps for disaster response.

Urban pluvial flooding

City flooding from intense rainfall is driven by drainage capacity, impervious surfaces and micro-topography. Here, high-resolution elevation models, drainage network data and rainfall nowcasting combine with surrogate ML models that approximate expensive hydrodynamic simulations, allowing many scenarios to be run quickly.

Water resource management

Beyond floods, time-series models forecast reservoir inflows, groundwater levels and irrigation demand, and anomaly detection on smart water meters helps utilities find leaks and non-revenue water.

Flood forecasting services covering many countries now provide river forecasts days in advance, including for parts of Africa that previously had no operational forecasts. Humanitarian organisations have experimented with anticipatory action, releasing cash or supplies to communities before a forecasted flood peaks rather than after the damage is done. Satellite-based flood maps are routinely used by response agencies to target aid and assess damage.

At the city scale, municipalities are combining drainage asset inventories, elevation data and crowdsourced flood reports to prioritise drain clearing and infrastructure upgrades. In my view, crowdsourced reports from residents, including photos geotagged on WhatsApp, remain an underused data source for training and validating urban flood models in places like Lagos.

• The warning is not the outcome. A perfect forecast saves nobody if it does not reach people in time, in a language they understand, with clear guidance on what to do. Dissemination and trust are where many systems fail.
• Extremes are rare in training data. The floods that matter most are the ones least represented in historical records. Models must be evaluated specifically on extreme events.
• Changing landscapes. Urbanisation, deforestation, dam operations and sand mining change how basins respond. Models trained on past behaviour can drift.
• False alarms erode trust. Communities that receive repeated warnings for floods that never come may ignore the next one. Communicating uncertainty honestly is essential.
• Equity of protection. Flood risk maps can influence property values, insurance prices and investment. Used carelessly, they can deepen the disadvantage of already vulnerable neighbourhoods.
• Data sovereignty. National hydrological agencies must remain central. Global AI forecasts should strengthen, not bypass, local institutions with legal mandates for warnings.

1. Partner with hydrological and emergency agencies. In Nigeria, national hydrological and meteorological services and emergency management agencies hold mandates, data and community relationships. Build with them.
2. Combine global models with local data. Use global forecasts and open satellite data as a baseline, then improve them with local gauges, rainfall stations and community reports.
3. Design the warning chain end to end. Map how a forecast becomes an SMS, a radio announcement, a WhatsApp message or a call from a community leader. Test it before the rainy season.
4. Communicate in impact terms. "Water may reach knee height on these streets by Thursday evening" is more actionable than a river level in metres.
5. Invest in elevation and drainage data. For urban flooding, accurate digital elevation models and drainage inventories are often the binding constraint, more than model architecture.
6. Measure lead time and reach. Evaluate not just forecast skill but how many people received a warning, how early, and what they did.

The UN Office for Disaster Risk Reduction provides frameworks and guidance on early warning and risk governance that are worth reading for any team entering this space.

AI has quietly transformed what is possible in flood forecasting, especially for the data-scarce regions that need it most. But the technology is only one link in a chain that runs from satellites and models to a mother in a flood-prone community deciding whether to move her children to higher ground. As engineers, our job is to strengthen every link, not just the one that is most fun to build.

• Floods are among the most damaging climate hazards, and data-scarce regions have historically lacked reliable forecasts.
• LSTM-based hydrological models trained across many basins can forecast floods even in ungauged watersheds.
• SAR satellite imagery and segmentation models enable rapid flood extent mapping through cloud cover.
• The warning chain, trust and communication of uncertainty matter as much as forecast accuracy.
• Effective projects partner with national agencies, blend global models with local data and measure who was actually warned.

1. Rainfall-runoff modelling using Long Short-Term Memory (LSTM) networks — Hydrology and Earth System Sciences (Kratzert et al.)
2. Global prediction of extreme floods in ungauged watersheds — Nature (Nearing et al.)
3. Flood Hub — Google Research
4. Copernicus Emergency Management Service — European Commission
5. World Meteorological Organization — WMO
6. UN Office for Disaster Risk Reduction — UNDRR

Cover image: original illustration created for this article. Cited works remain the property of their respective authors and publishers.