Projected seasonal flooding in Canada under climate change with statistical and machine learning

Abstract

Study region: Canada Study focus: Floods are among the costliest and deadliest natural hazards in the world. To date, little is known about future seasonal flooding across all Canada. In this paper, data-driven models for flood occurrence (i.e., happening of a flood) and impact (i.e., displaced population) were calibrated for spring and summer seasons in 14,000 watersheds across Canada. Generalized Additive Models (GAM), Random Forests (RF) and Gradient Boosting Machines (GBM) were considered to model seasonal floods. The best-performing flood models were then used with regional climate models to assess the effect of climate change on flooding for three time horizons: historical, medium-term (∼2050) and long-term (∼2080). New hydrological insights for the region: GAM offered the best out-of-sample performance trade-off in both seasons for predicting flooding in Canada. Projections with GAM showed a general increase in summer flooding occurrence and impact in 2050 and 2080, mainly in the Yukon, western British Columbia, southern Prairies, Ontario and Quebec and some Atlantic provinces. Results for spring flooding were more mixed, but there seemed to be a slight decrease in the impact of spring flooding in the southern Prairies, particularly in 2080. The combination of statistical/machine learning and climate models have provided a more detailed and contrasted picture of the projected seasonal flooding situation over Canada that will help authorities better mitigate future flood risks.

Keywords

• Spring flooding
• Summer flooding
• Generalized additive model
• Random forest
• Gradient boosting
• Regional climate models