Data-driven flood classification to assess climate change effects on riverine compound floods in pre-Alpine catchments

Abstract. Flood events are some of the most impactful natural hazards globally, and in most cases, they result from combinations of different hydrometeorological drivers, defined as riverine compound floods. Climate change is projected to alter the frequency and interplay of these drivers, posing a challenge for flood risk assessment. In this study, we develop an unsupervised machine learning approach to identify distinct riverine compound flood types and to classify individual flood events based on hydrometeorological features. These classifications are then used to assess climate change effects on flood risk. We use hydrological large-ensemble simulations in four neighbouring Bavarian pre-Alpine catchments, encompassing historical and future periods. Results project systematic shifts in the frequency and magnitude of flood regimes under warming. The frequency of rainfall-driven and rain-on-snow floods increases with warming, rising by 22% and 8%, respectively, between +1 K and +4 K. Conversely, snowmelt-driven flood frequency declines substantially (approximately 42% between +1 K and +4 K), consistent with a diminishing role of snow in pre-Alpine flood generation. Frequencies of saturation excess floods, characterised by antecedent wet conditions, show declines of about 10%. For 100-year return levels, rainfall-driven events show the largest increases in flood magnitude, though with strong variability between catchments (1% to 48% between +1 K and +4 K). Conversely, the magnitude of 100-year return-level snowmelt-driven events decreases 10–13%. Seasonally, pre-Alpine flood risk shifts from a spring, snowmelt-dominated regime towards rainfall-driven flooding throughout the year, with implications for seasonal preparedness, infrastructure design, and forecasting systems. We compare our unsupervised machine learning approach with an established threshold-based classification method, with the two classifications converging for 64% of events. By identifying which flood types increase or decrease under warming scenarios, our compound flood classification framework supports the development of process-specific adaptation strategies.