Evaluation of flood susceptibility prediction based on a resampling method using machine learning

Abstract

AbstractThe largest recorded flood loss occurred in the study area in 2013. This study aims to examine resampling methods (i.e. cross-validation (CV), bootstrap, and random subsampling) to improve the performance of seven basic machine learning algorithms: Generalized Linear Model, Support Vector Machine, Random Forest (RF), Boosted Regression Tree, Multivariate Adaptive Regression Splines, Mixture Discriminate Analysis, and Flexible Discriminant Analysis, and found the factors causing flooding and the strongest correlation between variables. The model is evaluated using Area Under the Curve, Correlation, True Skill Statistics, and Deviance. This methodology was applied in Kendari City, an urban area that faced destructive floods. The evaluation results show that CV-RF has a good performance in predicting flood susceptibility in this area with values, AUC = 0.99, COR = 0.97, TSS = 0.90, and deviance = 0.05. A total of 89.44 km2 or equivalent to 32.54% of the total area is a flood-prone area with a dominant area of lowland morphology. Among the 17 parameters that cause flooding, this area is strongly influenced by the vegetation density index and the Terrain Roughness Index (TRI) in the 28 models. The strongest correlation occurs between the TRI and the Sediment Transport Index (STI) = 0.77, which means that flooding in this area is strongly influenced by elements of violence.