Urban flooding is becoming more difficult to predict because rainfall is increasingly intense, cities are expanding over permeable land, and drainage systems are often aging or poorly documented. Conventional hydrological and hydraulic models remain essential, but they can be computationally expensive and may struggle to represent the complex interaction between rainfall, land cover, terrain, soil moisture, roads, buildings and underground drainage.

Artificial intelligence (AI) and machine learning (ML) offer a complementary approach. By learning from historical rainfall, flood observations, drainage characteristics, topography and land-cover data, models can estimate where flooding is likely to occur, how severe it may become and which people or infrastructure could be affected. This application provides a practical connection between environmental risk analysis and the wider field of AI and predictive analytics.

Why Urban Flooding Is Difficult

Urban flooding usually occurs when rainfall exceeds the capacity of soil, roads, drains, culverts, channels or storage facilities. A storm that produces manageable runoff in a rural catchment can create dangerous surface water in a city because buildings, pavements and roads prevent infiltration and rapidly route water into low-lying areas.

Several processes may occur simultaneously:

• Pluvial flooding: rainwater accumulates on streets and properties when drainage capacity is exceeded.
• Fluvial flooding: rivers, streams or canals overflow after catchment rainfall.
• Tidal or coastal flooding: high tides, storm surges or sea-level rise push water into urban drainage networks.
• Drainage failure: blocked, undersized or damaged pipes and culverts prevent water from leaving an area.
• Compound flooding: rainfall, river overflow and tidal conditions interact.

The forecasting challenge is therefore not simply to predict rainfall. A useful system must estimate how rainfall will be transformed into runoff, how that runoff will move through the urban landscape and where it will encounter constraints.

Data for Flood Prediction

A reliable AI flood-risk model depends on the quality, coverage and compatibility of its input data. Four data groups are particularly important.

1. Rainfall records

Rainfall is the primary forcing variable for stormwater prediction. Useful rainfall data include:

• Rain-gauge observations at short time intervals.
• Weather-radar estimates of rainfall intensity and movement.
• Satellite precipitation products.
• Numerical weather forecasts.
• Historical extreme-rainfall records.
• Antecedent rainfall, which indicates whether the ground is already saturated.

The model should normally use more than total daily rainfall. Peak five-minute or one-hour intensity, storm duration, cumulative rainfall and the time between storms can be more informative for urban flooding.

Rainfall records can be converted into features such as:

where represents accumulated rainfall, is maximum rainfall intensity, is storm duration and represents previous rainfall or soil wetness.

2. Drainage and hydraulic data

Drainage information helps a model distinguish between places that receive similar rainfall but experience different flood outcomes. Relevant variables include:

• Pipe diameter, length, slope and material.
• Inlet and manhole locations.
• Pumping-station capacity.
• Culverts, channels and detention ponds.
• Drainage catchment boundaries.
• Known blockages, maintenance records and failure reports.
• River, tide and water-level observations.
• Locations of previous sewer or surface-water flooding.

In practice, drainage data are often incomplete. Some cities have detailed asset databases, while others possess only paper plans or partial records. Machine learning can help identify likely drainage bottlenecks, but it should not be treated as a substitute for field surveys or hydraulic engineering.

3. Land cover, terrain and urban form

Land-cover information describes how much rainfall can infiltrate and how quickly water can travel across a surface. Typical features include:

• Impervious-surface percentage.
• Buildings, roads, parking areas and pavements.
• Parks, wetlands, agricultural land and tree canopy.
• Soil type and infiltration rate.
• Elevation, slope and local depressions.
• Distance to rivers, coastlines and drainage channels.
• Building density and population distribution.

Digital elevation models are especially valuable because small changes in terrain can determine whether water flows toward a drain, remains trapped in a depression or enters a building. Land-cover maps can be derived from satellite imagery, aerial photography, cadastral data or classified street-level imagery.

A 2024 study from North Carolina State University used Sentinel-2 satellite imagery, elevation and land-cover data with a Random Forest model to map flooding in southeastern Pennsylvania after Hurricane Ida. The model achieved 99% overall accuracy, while the mapped flood class achieved user’s and producer’s accuracies above 97%. The study also found that substantial flooding occurred outside designated 500-year flood zones, demonstrating why urban flood assessment should not rely only on traditional regulatory maps. Read the NC State case study.news.ncsu

4. Flood observations and exposure data

Machine-learning models require examples of what happened during previous storms. Training labels may come from:

• Water-level sensors.
• Emergency-service and road-closure reports.
• Insurance claims.
• Crowd-sourced photographs and mobile-phone reports.
• Social-media posts.
• Satellite-derived flood footprints.
• Residents’ observations.
• Historical flood maps.

Exposure data then indicate the consequences of flooding. These data can include population, hospitals, schools, transport routes, electricity substations, businesses and vulnerable households.

The distinction between hazard and risk is important. Hazard describes the probability, depth or extent of flooding. Risk combines hazard with exposure and vulnerability:

An AI system may therefore produce both a flood-likelihood map and an impact map showing which assets are most likely to be affected.

How the Machine-Learning Workflow Works

An urban flood-prediction system typically follows six stages.

1. Data integration

Rainfall, drainage, terrain, land cover and flood observations are converted into a common geographic grid or set of drainage catchments. The data must also be synchronized in time. For example, rainfall recorded at 10:00 should be linked to water levels, road closures and flood reports occurring during the following minutes or hours.

2. Feature engineering

Raw data are transformed into variables that a model can interpret. Examples include:

• Rainfall accumulated over multiple time windows.
• Distance to the nearest drain.
• Percentage of impervious surface within a catchment.
• Local elevation relative to surrounding streets.
• Drainage density.
• Soil saturation.
• River or tidal level.
• Number of previous floods in the same location.

3. Model selection

Different models suit different prediction tasks:

Model	Suitable application	Main strength
Random Forest	Flood susceptibility and satellite-image classification	Handles mixed variables and nonlinear relationships
Gradient Boosting	Flood probability and risk ranking	Often performs well with   structured tabular data
Support Vector Machine	Classification with limited training data	Effective in high-dimensional  feature spaces
Long Short-Term Memory network	Rainfall, river-level and water-level forecasting	Learns temporal patterns
Convolutional Neural Network	Flood maps and spatial prediction	Recognizes patterns in images and grids
Hybrid physical–AI model	Operational forecasting and scenario simulation	Combines physical consistency with computational speed

For time-series forecasting, an LSTM model may use rainfall and water-level sequences to predict the level at a future time:

Here, is the predicted water level, is rainfall, is previous water level, represents soil or wetness conditions, represents drainage characteristics and represents land-cover or spatial variables.

4. Training and validation

A model should be trained using some storms and tested on separate storms. Randomly splitting individual observations can produce misleadingly high accuracy because observations from the same storm may appear in both the training and test datasets. A stronger approach is to hold out entire storm events, seasons or geographic areas.

Common evaluation measures include:

• Precision: the proportion of predicted flood events that occurred.
• Recall: the proportion of observed flood events that the model detected.
• F1 score: a balance between precision and recall.
• AUC: the model’s ability to distinguish flooded from non-flooded locations.
• Root mean square error: useful for predicted water depth or water level.
• Critical success index: useful when flood events are relatively rare.

Research reviewed by the International Association for Hydro-Environment Engineering and Research reported commonly used flood-model performance ranges of approximately 0.70–0.98 for precision, 0.64–0.98 for accuracy and 0.69–0.99 for AUC, although these figures vary considerably with location, data quality, flood definition and validation method. See the review summary.iahr

5. Probabilistic prediction

A responsible system should not present a single deterministic answer when uncertainty is high. It might report:

• A 75% probability of street flooding.
• An expected flood depth of 15–30 centimetres.
• A high-risk period between 18:00 and 20:00.
• A confidence level reduced because rainfall-radar data are incomplete.

This information is more useful to emergency managers than an unexplained “flood” or “no flood” label.

6. Operational alerts

The final output may be delivered through a dashboard, SMS, mobile application, traffic-management platform or emergency-control centre. Alerts should be linked to predefined actions, such as inspecting drains, closing roads, moving vehicles, protecting electrical equipment or opening temporary shelters.

Case Studies

South Korea: AI water-level forecasting

South Korea’s Ministry of Environment has installed an AI-based flood-forecasting system at approximately 223 locations, particularly near large rivers and tributaries. The system combines rainfall intensity, water levels and soil moisture with an LSTM model trained on historical and real-time observations.

It predicts river levels at 10-minute intervals and combines statistical learning with hydrological and hydraulic modelling. When predicted levels exceed thresholds, alerts can be distributed to authorities and the public through SMS and cell-broadcast systems. The World Meteorological Organization describes this programme as part of a broader effort to combine AI with digital-twin technologies and high-resolution spatial data. Read the WMO report.wmo

Although the example is primarily river-focused, the same architecture can be adapted to urban stormwater systems by replacing or supplementing river gauges with rain gauges, drain-level sensors, pump telemetry, traffic cameras and citizen reports.

New Zealand: rapid flood-inundation maps

New Zealand’s National Institute of Water and Atmospheric Research has used machine learning to produce flood-inundation forecasts much faster than conventional physical simulations. According to PreventionWeb, the AI system can generate flood maps in approximately one to two minutes, compared with as long as 24 hours for some physical-model workflows.

The project used Westport as a case study and validated its model against the town’s real 2021 flooding. Instead of forecasting only river height, the system aims to show which streets, properties and infrastructure are likely to be affected. This distinction is important because residents and emergency managers generally need an impact forecast rather than a water-level value alone. Read the PreventionWeb case study.preventionweb

Urban flash floods: global-scale forecasting

In March 2026, Google Research announced an urban flash-flood forecasting system designed to provide up to 24 hours of advance notice in covered areas. The model uses time-series meteorological inputs alongside urbanization density, topography and soil-absorption characteristics.

A major challenge is the lack of “ground truth” for flash floods, since many events occur away from stream gauges. Google’s approach used an AI-assisted dataset derived from public news reports to identify the location and timing of historical flood events. The system currently works at approximately 20-by-20-kilometre resolution, so it provides broad warning rather than the street-level precision that a city-specific sensor and drainage model could offer. Read the Google Research announcement.research

The same source reports that flash floods account for approximately 85% of flood-related fatalities worldwide and that even a 12-hour warning can substantially reduce damage. These figures should be interpreted as global estimates, but they illustrate why improving warning time is a public-safety priority. research

United Kingdom: hybrid AI and hydraulic modelling

Machine-learning systems are also being developed for UK flood-management applications. JBA Consulting reports an LSTM model for the River Doe Lea in Derbyshire that was trained using data from 66 catchments across England and then fine-tuned with local observations. During a one-year test period, the system reportedly avoided more than 15 false alerts compared with traditional calibrated rainfall-runoff methods while still identifying all significant flood events.

The same organisation describes an LSTM model for Thames tidal forecasting that produced errors within ±0.1 metres for 92% of high tides at Silvertown, compared with 68% for the existing model. It also reports a convolutional-neural-network emulator for Eastbourne capable of producing flood forecasts in seconds, whereas the detailed integrated catchment model required approximately 16 days to simulate 24 hours. These are project-specific results rather than universal benchmarks, but they show how AI can accelerate existing engineering models. Read the JBA examples.jbaconsulting

Benefits and Limitations

AI-based stormwater prediction can provide several benefits:

• Faster forecasts for emergency response.
• More detailed street- or neighbourhood-level risk maps.
• Better identification of drainage bottlenecks.
• Improved prioritization of maintenance and infrastructure investment.
• Continuous updating as new sensor and rainfall data arrive.
• Rapid testing of land-use or drainage-improvement scenarios.
• Lower computational cost when AI emulates a detailed hydraulic model.

However, AI does not eliminate uncertainty. A model may fail when future storms differ substantially from historical events, when a new development changes runoff patterns or when drainage assets are missing from the database. Satellite imagery can also be affected by clouds, building shadows and the short duration of urban flooding. The NC State study, for example, noted that buildings, concrete surfaces and complex drainage systems make urban flood mapping particularly difficult.news.ncsu

There are also governance concerns. A warning system must explain why an area has been classified as high risk, protect personal and location data, avoid systematically under-warning poorly monitored communities and communicate uncertainty clearly. Model accuracy should be assessed separately for different neighbourhoods rather than reported only as a city-wide average.

A Practical Implementation Framework

A city beginning an AI flood-prediction programme could proceed in stages:

1. Create a baseline map: combine elevation, land cover, drainage boundaries and historical flood reports.
2. Improve observation networks: install rain gauges and water-level sensors at known hotspots and drainage outfalls.
3. Standardize data: establish common geographic coordinates, timestamps, metadata and quality-control procedures.
4. Build a pilot model: begin with interpretable methods such as Random Forest or Gradient Boosting.
5. Add time-series forecasting: use LSTM or related models where continuous rainfall and water-level data are available.
6. Link predictions to actions: define alert thresholds and response procedures with emergency managers.
7. Validate after every major storm: compare forecasts with observed depths, road closures and resident reports.
8. Combine AI with physics: use hydraulic models to test unusual scenarios and constrain implausible predictions.
9. Publish uncertainty: show confidence intervals, missing-data warnings and the model’s geographic limits.
10. Update the system: retrain models as land cover, drainage infrastructure and rainfall patterns change.

Conclusion

AI-based stormwater and flood-risk prediction is most effective when it combines four elements: high-quality rainfall observations, detailed drainage information, land-cover and terrain data, and carefully validated machine-learning models. Random Forest models are useful for mapping flood extent and susceptibility, while LSTM and convolutional models can forecast water levels or rapidly emulate complex hydraulic simulations.

The strongest systems are not designed to replace hydrology, civil engineering or emergency management. Instead, they use AI to process large and changing datasets, identify hidden relationships and deliver forecasts quickly enough to support action. In that sense, urban flood prediction is a practical example of predictive analytics: historical and real-time data are transformed into probabilities, maps and decisions that can reduce damage before the water arrives.