Integrating Machine Learning for Flood Impact Prediction in Swedish Operational Forecasting and Warning Services

Flooding is a natural hazard arising from complex and non-linear interactions between hydrometeorological forcing and landscape characteristics, and therefore cannot be reliably represented using simple empirical relationships. The objectives of this study are (1) to identify which hydrological and physiographic variables, or which combinations of them, are most strongly associated with flood-related consequences, and (2) to develop a national-scale flood susceptibility framework for Sweden that can be integrated with forecast information to support operational warning decisions.

The novelty of this work lies in the use of a large, nationwide impact dataset consisting of road closure records from 2000–2023, provided by the Swedish Traffic Agency, as the target for training and validation of a data-driven impact model. Each road closure location is characterized using a comprehensive set of predictors derived from the SHYPE hydrological model — including precipitation, runoff, soil moisture, groundwater storage, and short-term intensity metrics (e.g. 3-hour maxima) — together with topographic and environmental descriptors such as slope, elevation range, upstream contributing area, distance to water bodies and culverts, and land-use classes.

An Extreme Gradient Boosting (XGBoost) classifier was used to learn the relationship between these predictors and observed impacts. The model achieves strong predictive skill (accuracy = 0.977), with a balanced confusion matrix indicating strong ability to distinguish impacted and non-impacted areas. Feature importance analysis reveals that short-term hydrological response dominates model behavior. Surface runoff is the most influential predictor, followed by local runoff and groundwater storage, highlighting the critical role of near-surface hydrological dynamics in translating meteorological forcing into damaging outcomes. Topographic and land-use variables, such as slope and industrial land cover, further modulate susceptibility, emphasizing the influence of local terrain and exposure.

The resulting framework enables the generation of a dynamic flood susceptibility map for Sweden. When driven by real-time or forecast hydrometeorological inputs, the model can function as a “copilot” for forecasters, indicating where events are most likely to produce consequences. This would support more targeted warnings, reduces false alarms, and strengthens proactive risk communication in vulnerable areas.