Predictive Modeling of Envelope Flood Extents Using Geomorphic and Climatic‐Hydrologic Catchment Characteristics

Abstract

A topographic index (flood descriptor) that combines the scaling of bankfull depth with morphology was shown to describe the tendency of an area to be flooded. However, this approach depends on the quality and availability of flood maps and assumes that outcomes can be directly extrapolated and downscaled. This work attempts to relax these problems and answer two questions: (1) Can functional relationships be established between a flood descriptor and geomorphic and climatic-hydrologic catchment characteristics? (2) If so, can they be used for low-complexity predictive modeling of envelope flood extents? Linear stepwise and random forest regressions are developed based on classification outcomes of a flood descriptor, using high-resolution flood modeling results as training benchmarks, and on catchment characteristics. Elementary catchments of four river basins in Europe (Thames, Weser, Rhine, and Danube) serve as training data set, while those of the Rhône river basin in Europe serve as testing data set. Two return periods are considered, the 10-and 10,000-year. Prediction of envelope flood extents and flood-prone areas show that both models achieve high hit rates with respect to testing benchmarks. Average values were found to be above 60% and 80% for the 10-and the 10,000-year return periods, respectively. In spite of a moderate to high false discovery rate, the critical success index value was also found to be moderate to high. It is shown that by relating classification outcomes to catchment characteristics, the prediction of envelope flood extents may be achieved for a given region, including ungauged basins. Plain Language Summary Topographic features can be extracted from a digital terrain to identify floodplains. In turn, the classification of these features can be used to represent flood extents. The classification, however, depends on existing flood maps to be used as the reference for its calibration in a small portion of the study area, before flood extents can be extrapolated. This work seeks to improve this dependence on the existence of a reference and the assumption that extrapolation can be performed without considering the physical differences between areas. To do so, a machine learning model was developed using the classification outcomes and characteristics of four river basins in Europe (Thames, Weser, Rhine, and Danube) and was tested in the Rhône river basin in Europe, showing promising results. This development should help stakeholders to allocate time and money better, by giving them nearly instantaneous views of areas that can potentially be affected by floods, with an associated probability of occurrence, a physical basis for extrapolation, and without needing a reference flood map for calibration for each case.

Direct Integration of Numerous Dams and Reservoirs Outflow in Continental Scale Hydrologic Modeling

“…More resolute DEM data may overcome the limitations of the NED and more sophisticated hydraulic modeling may overcome AutoRoute's limitations. However, more resolute DEM data only exist for portions of the MRB and more sophisticated hydraulic software options require large amounts of reliable geometry and hydrologic observations (Tavares da Costa et al., 2020) and they are not as computationally efficient as AutoRoute. All of these considerations are important for continental scale hydraulic modeling.…”

Unraveling Long-Term Flood Risk Dynamics Across the Murray-Darling Basin Using a Large-Scale Hydraulic Model and Satellite Data

“…Several methods have been developed both at local and global scales to allow for a detailed assessment of flood hazard, either based on traditional hydrological and hydraulic models (Bates et al, 2010;Yamazaki et al, 2011;Pappenberger et al, 2012;Winsemius et al, 2013;Rudari et al, 2015;Sampson et al, 2015;Dottori et al, 2016;Schumann et al, 2016) or innovative DEM-based (digital elevation model) techniques (Lee et al, 2017;Samela et al, 2017;Tavares da Costa et al, 2020). Typically, flood models simulate inundated areas based on the probability of exceedance of a particular discharge value (i.e., by considering a particular return period) or based on longterm time series of discharge, without accounting for detailed topographic features along floodplains.…”

Machine Learning Analysis of Hydrologic Exchange Flows and Transit Time Distributions in a Large Regulated River

“…Previous studies have proved that ML models can achieve high prediction on accuracy in hydrologic applications (Hsu et al, 1995;Tesoriero et al, 2017;Barzegar et al, 2018;Mo et al, 2019;Nearing et al, 2020). In this study, we adopted two ML models, Random Forest (RF) and Extreme Gradient Boosting (XGB), given their competitive capabilities to deal with system high-dimensionality, nonlinearity, mixed numerical and categorical variables, highly correlated predictor variables, as well as overfitting, and they have been proven to be superior to traditional ML methods in various case studies (Prieto et al, 2019;Yan et al, 2019;Li et al, 2020;Tavares da Costa et al, 2020;Xenochristou et al, 2020). Bagging/boosting techniques have been used to get ensemble learners in which each individual member of the ensemble is trained using a different training data set subsampled randomly in both rows and columns from the full training data set.…”