Physically based dimensionless features for pluvial flood mapping with machine learning

Each year, flash floods are responsible for 80%–90% of all flood‐related
deaths, with approximately 40% of these fatalities linked to pedestrian stream crossings or vehicles. Quickly predicting where flash floods will occur is crucial for saving lives and property. Traditional methods for predicting floods are slow and require significant computational resources, but machine learning (ML) can help speed up this process. However, existing ML models often have difficulty making accurate predictions for new areas with different climate patterns or landscape conditions. In this study, we improve ML flood prediction by
representing the watershed with dimensionless numbers, which capture the relationships between key factors like rainfall and terrain shape, without relying on specific units (e.g., meters or seconds). By removing these units, the dimensionless numbers help ML models focus on the underlying patterns of flooding, making the ML more adaptable to different regions. We tested this approach with an ML model to predict flood risk, and the results showed that it performed well, even in regions where the ML had not been trained. This method could help create faster and more accurate flood maps for a wide range of areas, allowing communities to better
prepare for floods. More here.