Author(s): Heather McGrath; Victor Alhassan
Linked Author(s):
Keywords: Deep learning; Flood susceptibility; Machine learning; Spatial grids; Spatial indices
Abstract: Flood susceptibility mapping increasingly uses deep learning (DL) for spatial dependencies, yet machine learning (ML) remains popular for interpretability and efficiency. This study evaluates whether ML models enhanced with spatial indices (SI) can match convolutional neural networks (CNN) for national-scale flood mapping across Canada. Four models were tested on 268,049 balanced samples: XGBoost, XGBoost+SI, CNN, and CNN+SI. XGBoost+SI achieved the highest accuracy (0.94) and AUC (0.99), outperforming both CNN models. Misclassification analysis showed ML models had lower false positive and negative rates, yielding more balanced predictions. These results suggest well-engineered ML models, especially with SI, can rival or exceed CNN performance for large-scale flood risk assessment.
Year: 2026