Comparative explainable ensemble learning for flood susceptibility mapping using Sentinel-1-derived flood inventory and multi-source geospatial data: A case study of the Semarang-Demak coastal area

Abstrak

Flood susceptibility mapping is increasingly supported by remote sensing and machine learning, yet pixel-level flood inventories remain difficult to obtain in many Indonesian coastal regions where official disaster records are commonly reported at administrative scale. This study develops a comparative explainable ensemble-learning framework for flood susceptibility mapping in the Semarang-Demak coastal area by integrating a Sentinel-1-derived flood inventory with multi-source geospatial predictors from Google Earth Engine. Sentinel-1 SAR was used to derive the binary flood label from pre- and post-event change detection of the March 2024 flood, while conditioning variables represented topographic, hydrological, meteorological, spectral, land-cover, and built-environment factors. Seven ensemble models were evaluated on a balanced dataset of 1,870 samples, consisting of 935 flood and 935 non-flood observations: Random Forest, Extra Trees, Gradient Boosting, XGBoost, LightGBM, CatBoost, and a stacking ensemble. Random Forest achieved the best overall discrimination with ROC-AUC = 0.852, accuracy = 0.791, recall = 0.932, and F1-score = 0.816 at the optimized threshold. DeLong tests and paired bootstrap resampling showed that the differences among the leading models were not statistically significant, whereas spatial block cross-validation revealed that the random-split performance estimates are optimistic for spatial generalization. SHAP interpretation, computed on training data, indicated that distance to permanent water, antecedent rainfall, NDVI, upstream drainage area, MNDWI, NDBI, river width, and distance to river were the dominant predictors, implying that flood susceptibility in the study area is governed by interacting hydrological proximity, rainfall accumulation, surface moisture, vegetation condition, and built-up intensity. The susceptibility map highlights high-risk patterns in low-lying, water-adjacent, and urbanized coastal zones. The proposed framework contributes a reproducible workflow for transforming Sentinel-1 flood detection into an explainable susceptibility model for data-limited coastal urban environments.

Kata kunci

• Coastal Flooding
• Ensemble Learning
• Flood Susceptibility
• Google Earth Engine
• Sentinel-1
• SHAP