RAPID FLOOD INUNDATION MAPPING AT LARGE SCALE USING A MULTI-DISCIPLINARY MACHINE LEARNING APPROACH

J. Albee (Stantec) , T. Lewis (Stantec), R. Els (Stantec)

ABSTRACT

At the forefront of stormwater management technology, our team introduces a state-of-the-science flood inundation mapping system, called Floor Predictor that harnesses the power of a multidisciplinary machine learning approach. This innovative solution delivers swift, accurate analysis and can be uniquely accessible via a web browser. Proven highly effective in the United States, this technology is now recommended for adoption in New Zealand to transform stormwater modelling and significantly elevate emergency response efficiency. This approach can empower local councils to make data-driven decisions, fostering sustainable, flood-resilient communities well-prepared to face the challenges posed by a changing climate.

Climate change is increasing the frequency and severity of extreme flooding events, causing challenges for land managers who lack rapid, reliable, high-resolution riverine and pluvial flood risk assessments and forecasts at scales relevant to their management region(s). Managers need a common way to understand the likelihood of flooding across disparate areas with divergent physical and climatic attributes. A common approach to sharing this knowledge, however, is often missing, making it hard to extrapolate flood predictions to new areas and timeframes.

The flood predictor machine-learning approach allows us to show how physics, hydrologic and hydraulic domain knowledge, and machine learning be used to extrapolate flood knowledge (modelling and observations) to new areas on a rapid basis. We leveraged Buckingham Pi Theory to integrate hydrologic, climate, soil, and land use data from local and national sources into dimensionless features that capture the similarity of the flooding process across different regions and climates. These dimensionless features were then used as inputs for a machine learning algorithm trained to detect flooding extents. Not only did our predictions have high agreement with physics-based modelling results (up to 95%) but when compared to satellite-derived observations of flooding extents, our model accurately predicted flooding extents over 90% of the time. This method is anticipated to significantly contribute to geohazard risk assessment and serve as a foundational element for the development of rapid flood warning systems, ultimately aiding in the reduction of flood-related damages. By focusing on the practical applications of this approach, our aim is to support communities and local councils across New Zealand in enhancing their flood resilience strategies.

Our paper will describe the Buckingham Pi theorem formulation of flood mapping and discuss how machine learning can expand observations