Gaussian process regression on multiple drivers and attributes for rapid prediction of maximum flood inundation extent and depth

Abstract

Traditional high-resolution flood models are too slow for real-time predictions. The most common industry practice is to use a lookup table or interpolation algorithm to derive flood extents from a pre-generated library of flood maps. For the library interpolation approach to be effective, the input flood data need to closely match those in the map library. To effectively emulate complex and dynamic flood behaviour, the interpolation approach should be able to account for multiple flood drivers (such as rivers, tributaries and tides) and attributes (such as shape and timing of the hydrographs). However, a simple extension of existing interpolation algorithms would make them overly complex and need a very large map library to deal with these drivers and attributes. To address this challenge, this study investigates the capability of a Gaussian Process (GP) modelling approach to accommodate the complex influence of multiple flood drivers and attributes, and thus to provide robust, accurate and fast predictions. By training the GP model, it learns the underlying relationships between flood depths and multiple flood drivers and attributes. Model accuracy and speed in predicting maximum flood extents and depths are examined. In a case study of a floodplain in Port Fairy, Australia, the GP model is found to generally outperform the library interpolation approach in accuracy, particularly for complex floods, while being similar efficient. The GP model is a promising approach for real-time flood predictions.