Please note: The algorithm descriptions in English have been automatically translated. Errors may have been introduced in this process. For the original descriptions, go to the Dutch version of the Algorithm Register.

Predicting floods in the Selzerbeek catchment area using DeepWave

DeepWaive is an AI model that helps to predict potential flooding in the Selzerbeek area. The model analyses data such as rainfall, elevation, land use and the water system. The model is currently being researched and tested. It has not yet been used for warnings or crisis management.
Last change on 6th of August 2026, at 6:47 (CET) | Publication Standard 1.0
Publication category
Impactful algorithms
Impact assessment
Field not filled in.
Status
In development

General information

Theme

  • Public Order and Safety
  • Nature and Environment
  • Space and Infrastructure

Begin date

2026-03

End date

2027-02

Contact information

algoritme@prvlimburg.nl

Link to publication website

https://Limburg.nl/bestuur/open-overheid/algoritmeregister

Responsible use

Goal and impact

The aim of the research is to investigate whether DeepWaive can produce flood imagery more quickly and accurately than traditional models. This imagery can help experts to assess flood risks and prepare appropriate measures. The model is currently only being tested. The results are not yet being used to make decisions or to warn people and businesses. Consequently, people and businesses are not currently directly affected by the algorithm.

This research forms part of the WRL programme, for which the Province of Limburg, Spatial Planning Cluster, is the formal commissioning body.

Should Deepwave be put into production in the future, the contract will be awarded by the Water Management Centre of the Limburg Water Board.

Considerations

Traditional hydraulic models can make accurate flood predictions, but they take a long time to run. DeepWaive can produce these predictions more quickly. In future, this could help us identify the risk of flooding earlier when there is heavy rainfall. One drawback is that it is more difficult to explain how an AI model works than a traditional model based on physics. We also need to establish when the predictions are reliable. That is why DeepWaive is currently only being researched and tested. The results are being compared with historical measurement data and traditional simulations. We will only consider putting it into use once we are certain that it is reliable.

Human intervention

Experts review the model’s results. During the research phase, the results are not immediately translated into warnings or decisions. There is no automated decision-making. If the model is to be used in practice, the responsibilities and controls relating to human assessment must be clearly defined.

Risk management

The main risk is that the model may fail to predict a flood, predict it too late, or predict it incorrectly. Accuracy may also vary depending on the location, precipitation conditions or the quality of the input data. Furthermore, the functioning of the AI model is less straightforward to explain than that of a traditional hydraulic model.

To manage these risks, the model is currently only being tested. The results are compared with historical measurement data, real-time observations and traditional hydraulic simulations. Experts assess the results and any discrepancies are investigated. The model is not yet linked to automatic alerts or measures.

Before any operational use takes place, aspects such as reliability, margins of error, availability, information security, responsibilities and procedures for human oversight will be assessed. The necessary periodic monitoring and reassessment will also be determined at that stage.

Elaboration on impact assessments

The model does not process any personal data. A DPIA is therefore not required. The model is currently only being investigated, and the findings do not yet have any implications for members of the public or businesses. Consequently, no formal impact assessments have yet been carried out.

Prior to any operational use, a fresh assessment will be carried out to determine whether an impact assessment, including an IAMA, is required. This assessment will, in any case, examine the consequences of incorrect predictions, the model’s explainability, human oversight and the allocation of responsibilities.

Operations

Data

Operational:

Precipitation radar: DWD RADOLAN (5-minute intervals)

Precipitation radar: KNMI Radar (real-time, 5-minute intervals)

Precipitation forecast: DWD ICON-D2 deterministic


Algorithm used:

Digital elevation model: derived from AHN

Profile data for watercourses and hydraulic structures: WL

Land use: LGN2023 (NL) and ALKIS (DE)

Water level and discharge measurements: WL

Technical design

DeepWaive uses elevation data, land-use information, soil data, observed precipitation and precipitation forecasts to simulate the development of floods in the Selzerbeek catchment area. Parameters for surface roughness and flow resistance are derived from the available terrain and land-use data. Runoff after infiltration is estimated using the Curve Number method and then processed by DeepWaive, a physics-based AI model trained on simulations from hydraulic models. The model predicts water depths and flow development much faster than traditional numerical flood models. The operational forecast is triggered when predefined thresholds for precipitation forecasts are exceeded, after which 2D flood maps and water depths can be generated for the 48-hour forecast period.

External provider

FloodWaive Predictive Intelligence GmbH