EGU General Assembly 2022
© Author(s) 2022. This work is distributed under
the Creative Commons Attribution 4.0 License.

A regional flood impact prediction tool using machine learning to manage flood risk in real-time. A case study in New Zealand.

Phil Mourot, Nick Lim, Bernhard Pfahringer, and Albert Bifet
Phil Mourot et al.
  • Computer Science, University of Waikato, Hamilton, New Zealand (

Regional resilience has been identified as a key strategic priority for the Waikato Regional Council in New Zealand. Weather extremes are going to impact more our communities and what is important is how the regions can anticipate and respond to the impact of climate change. Flooding is the Waikato region’s most frequent and widespread natural hazard. The council’s priority is to prevent risks to people and property by providing flood protection and flood warnings. The local government works close to emergency services and civil defence to help people at risk. In addition to flood defences, flood impact prediction can help our communities be more resilient. This research is part of the TAIAO project ( that aims to develop new machine learning (ML) methods to provide a robust and fit-for-purpose tool to help New Zealand solve critical environmental problems. Over the past decade, increased research has aimed to develop new hydrological models for flood forecasting using machine learning. A data-driven approach provides the ability to deliver reliable results, especially for short-term forecasts, without the complete and complex knowledge of the physical processes usually required by a physically-based approach. Our research focuses on developing a regional real-time flood forecasting tool for emergency management that can run with low computational effort and a small number of parameters. Our target is to provide a better flood prediction with available information from the observation network. For our pilot study, we focus on the Coromandel Peninsula, a popular destination for the holidays, and where the weather is often challenging to forecast, like in New Zealand in general. We have used and compared the capability of various ML models to provide accurate results with low timing errors. To solve the problem of lagged prediction, we have developed a more holistic approach that combines hydrological state parameters and Long Short-Term Memory networks (LSTM). From these preliminary results, we demonstrate the real challenge to embed our LSTM-based model into operational procedures to predict with a lead time from 1 hour to 6 hours the severity of the impacts of heavy rainfall. The predictions are presented in a helpful way that facilitates decision-making and improves the regional flood response management.

How to cite: Mourot, P., Lim, N., Pfahringer, B., and Bifet, A.: A regional flood impact prediction tool using machine learning to manage flood risk in real-time. A case study in New Zealand., EGU General Assembly 2022, Vienna, Austria, 23–27 May 2022, EGU22-12455,, 2022.


Display link

Comments on the display

to access the discussion