A novel sample-enhancement framework for machine learning-based urban flood susceptibility assessment

Abstract

The commonly used random sampling method in machine learning-based flood susceptibility studies has two major issues: a default invalid assumption of spatial homogeneity and an inadequate number of non-flood samples. To address these issues, this study proposed a novel sample-enhancement framework to improve the quality of training samples on both flood and non-flood sides. Three one-way enhancements (two flood and one non-flood) and two joint enhancements were designed. The enhancements were evaluated against random sampling using four mainstream machine learning algorithms (ANN, RF, SVM, and XGBoost) across two heterogeneous urban regions in Guangzhou, China. The highest performances are achieved by the joint enhancements, which are followed by one-way enhancements and random sampling (no enhancement). Another important conclusion is that one-way enhancements exhibit divergent yet complementary effects. Flood enhancements primarily affect susceptibility distribution (mean value and standard deviation), while non-flood enhancements mainly influence binary classification performance (AUC).