How to predict floods in small basins where models fail

Ravines, seasonal watercourses, and streams that remain dry most of the time are often the ones that respond fastest during intense rainfall.

Flood prediction in small basins is, in fact, one of the most difficult problems to solve in water risk management, and one of the least visible to the administrations that must make decisions.

Not because the threat is lower, but because the data required to evaluate it, in many cases, do not exist.

Why small basins concentrate the risk

In a large basin, water takes hours or days to travel through the system. There is time to observe, process, and issue warnings. During one of the floods recorded in the Ebro River in 2024, the peak flow recorded in Castejón (Navarra) took a day and a half to reach Zaragoza.

Flooding of the Ebro River as it passes through Zaragoza (04/04/2007). Escarlati CC BY-SA 2.5

In a small basin (⪅ 500 km²), that margin disappears. The lead time is measured in hours, sometimes in minutes, and the conditions that determine whether or not a flood will occur change abruptly and locally.

Several comparative studies agree that regional-scale hydrological models tend to underestimate or shift peak flows in these systems, and that their performance degrades the smaller the basin and the more arid the climate.

The reasons are structural:

• Short concentration times: flash floods develop at a local scale in a matter of minutes or hours, triggered by high-intensity convective rainfall, a type of event whose frequency is increasing due to climate change. There is no time for conventional warning systems to detect, process, and communicate the threat before the water arrives.
• Dependence on local rainfall: convective precipitation is highly heterogeneous in space and time. Radar and rain gauge networks often have insufficient spatial resolution to capture the impact of a storm on a specific ravine.
• Sensitivity to antecedent soil moisture: the moisture level of the ground before the episode decisively determines how much rain becomes surface runoff. Without prior data, models operate with high uncertainty in one of the parameters that most influences the flood peak.

Absence of historical series: in ungauged basins, knowledge of hydrological behavior is based on physical inferences and statistical regionalizations, rather than direct observation. This limits the calibration of any model and increases the risk of false negatives in warnings.

Flood prediction in small basins: the role of gauging

A basin is gauged when it has at least one station that records the river level or flow, not necessarily in continuous real-time. These measurements are what allow hydrological models to be calibrated with greater reliability, establishing warning thresholds based on observed data and detecting if water is rising faster than expected.

Without that information, hydrological monitoring of minor basins is practically impossible to perform with any rigor.

An academic paper from 2025 on the Zapardiel River basin in Ávila illustrates this with precision. Despite the fact that this river and its tributaries have caused recurring floods in several population centers, official flood zone mapping does not include them due to the lack of adequate records.

The fact that the problem of basins without hydrological data is structural is confirmed, from another angle, by Google itself. Groundsource, the methodology presented in March 2026 to improve flash flood prediction, was designed to fill this gap by analyzing public reports to create a history of events where none existed before.

What to do when gauging data is missing

The absence of historical flow series does not mean there are no options. Today, early warning systems for floods in small basins combine several strategies that allow this gap to be partially bridged.

The first is direct instrumentation with low-cost level sensors installed at critical points along the riverbed. In very small basins, one or two strategic sensors can provide the information the model needs to function.

The second is the use of high-resolution digital terrain models, obtained with LiDAR, which allow for precise characterization of the riverbed topography and anticipation of where and at what depth the water would spread under different scenarios.

The third is the integration of hybrid models that combine terrain physics with machine learning algorithms, capable of learning from the system’s behavior even with limited data, as documented by recent research on requirements for prediction systems in small and medium-sized basins.

What distinguishes the most advanced systems is not just the predictive capacity, but the complete chain that goes from data to decision.

The process integrates meteorological conditions and basin characteristics into a rainfall-runoff model that generates impact scenarios for specific infrastructures, and these scenarios are translated into operational alerts for decision-makers.

This is the approach applied by the FAST flood prediction platform.

Based on the information collected by hydrometeorological stations and river level sensors, the system generates 2-D hydraulic models that show how the flow could evolve in different scenarios, updating that prediction every ten minutes in less than ten seconds.

But prediction alone doesn’t close the loop. FAST incorporates emergency management modules that allow users to define protocols linked to specific alert levels, centralize key contacts, and log every incident to improve future response.

The risk is local and the warning must be local as well.

Does your municipality or infrastructure have this type of exposure? The IntelliAlert team can help you identify it.