Contrast or Diversity: Non-Flood sampling in urban flood susceptibility modelling

Flood susceptibility modeling is a typical imbalanced problem in which the amount of flood data is much smaller than that of non-flood data. To ensure balanced learning, only a small fraction of non-flood data is selected for machine learning. Traditional sampling methods, such as Random Sampling (RS) and Stratified Sampling (SS), neglect abundant information within non-flood data and its relationship with flood data. This neglect leads to insufficient binary classification performance and biased susceptibility estimation, both of which are influenced by sample contrast and diversity, respectively. Unfortunately, these two objectives cannot be achieved simultaneously due to the trade-off between sample contrast and diversity. This dual-objective optimization requires a trade-off between contrast and diversity in sample quality. This study proposed a Distance-Based Sampling (DBS) framework that connects non-flood samples to flood data using Euclidean distance. Ten DBS scenarios with varying contrast and diversity levels (from 0.0 to 1.0 to 0.9–1.0) were designed for systematic evaluation. The best DBS scenario was further compared with RS, SS, and Inverse-Occurrence Sampling (IOS). To derive robust results, four machine learning techniques—Artificial Neural Network (ANN), Random Forest (RF), Support Vector Machine (SVM), and eXtreme Gradient Boosting (XGBoost)—were employed in two heterogeneous urban districts in Guangzhou, China, along with multiple indices, i.e. Area Under the Curve (AUC), mean susceptibility, standard deviation of susceptibility, and flood density order. The main findings of this study were as follows: (1) High sample contrast led to excellent binary classification performance but resulted in an overestimation of flood susceptibility. (2) High sample diversity resulted in insufficient binary classification performance and an underestimation of flood susceptibility. (3) Under the DBS framework, the performance curve of the dual-objective problem is unimodal. The best performance was achieved at a trade-off between contrast and diversity, specifically in the DBS scenario 0.3–1.0. (4) The DBS scenario 0.3–1.0 outperformed RS, SS, and IOS. Finally, this study underscores the critical role of non-flood sample quality and the balance between sample contrast and diversity in flood susceptibility modeling. The proposed DBS framework is objective and flexible, and can be applied to negative sampling in susceptibility modeling for other hazards, such as landslides and wildfires.