Flood mapping using Res-Q and machine learning on imbalanced data

Abstract

Flood disaster mapping requires accurate methods to support early warning and mitigation planning. To address common issues such as imbalanced data distribution and limited attribute handling, this study proposes an improved approach. The methodology includes: i) modification of the spatial sort filter skyline method with reverse normalization based on attribute preferences, applied when an attribute has minimal preference to ensure balanced consideration during skyline filtering; ii) data labeling and balancing, where initial flood potential labeling is generated using Res-Q, followed by K-Means clustering to group data into four classes (low, moderate, high, and very high) and SMOTE to further balance the dataset with 558 data points per class; iii) model evaluation using the C5.0 algorithm under three schemes, showing high and consistent accuracy with 89.24% on imbalanced data (Schema 2) and 93.3% on balanced data (Schema 3), while Schema 1 shows overfitting due to extreme imbalance; and iv) the main contribution, integrating reverse normalization with skyline filtering combined with clustering and resampling, enhancing both accuracy and robustness in identifying flood-prone areas. This structured approach highlights methodological improvements, reliable results, and practical contributions for effective flood disaster management.