Application of stacking hybrid machine learning algorithms in delineating multi-type flooding in Bangladesh

Floods are among the most devastating natural hazards in Bangladesh. The country experiences multi-type floods (i.e., fluvial, flash, pluvial, and surge floods) every year. However, areas prone to multi-type floods have not yet been assessed on a national scale. Here, we used locally weighted linear regression (LWLR), random subspace (RSS), reduced error pruning tree (REPTree), random forest (RF), and M5P model tree algorithms in a hybrid ensemble to assess multi-type flood probabilities at a national scale in Bangladesh. We used historical flood data (1988-2020), remote sensing images (e.g., MODIS, Landsat 5-8, and Sentinel-1), and topography, hydrogeology, and environmental datasets to train and validate the proposed algorithms. According to the results, the stacking ensemble machine learning LWLR-RF algorithm performed better than the other algorithms in predicting flood probabilities, with R² = 0.967-0.999, MAE = 0.022-0.117, RMSE = 0.029-0.148, RAE = 4.48-23.38%, and RRSE = 5.8829.69% for the training and testing datasets. Furthermore, true skill statistics (TSS: 0.929-0.967), corrected classified instances (CCI: 96.45-98.35), area under the curve (AUC: 0.983-0.997), and Gini coefficients (0.966-0.994) were computed to validate the constructed (LWLR-RF) multi-type flood probability maps. The maps constructed via the LWLR-RF algorithm revealed that the proportions of different categories of flooding areas in Bangladesh are fluvial flooding 1.50%, 5.71%, 12.66%, and 13.77% of the total land area; flash floods of 4.16%, 8.90%, 11.11%, and 5.07%; pluvial flooding: 5.72%, 3.25%, 5.07%, and 0.90%; and surge flooding, 1.69%, 1.04%, 0.52%, and 8.64% of the total land area, respectively. These percentages represent low, medium, high, and very high probabilities of flooding. The findings can guide future flood risk management and sustainable land-use planning in the study area.