Machine Learning Reveals Increasing Flood Risk from Atmospheric Rivers

Atmospheric rivers (ARs) are essential components of the Earth's hydrological cycle, transporting vast amounts of moisture from the tropics to midlatitudes. While vital for sustaining water resources, ARs are also major drivers of extreme precipitation and flooding, with profound impacts on human populations and ecosystems. In this study, we develop advanced Machine Learning models to predict flood risk associated with ARs by integrating AR data with additional meteorological variables, topographic information, and other relevant factors. Achieving an Area Under the Receiver Operating Characteristic Curve (ROC AUC) score of 0.96, our models demonstrate exceptional capability in forecasting flood events and uncovering key predictors of AR-induced flood risk. For comparison, a logistic regression model tested on this prediction task using the exact same data achieved a significantly lower score of 0.79. We further validate these results on an additional flood dataset and maintain the high predictive performance, underscoring the robustness and generalization of the Machine Learning model.

Next, we evaluate our model on a case study of the 2018 California flood, an event driven by a persistent Atmospheric River that triggered severe rainfall, flooding, and mudslides. The model predictions show a substantial increase in flood risk starting from the morning of the day before the event. The predicted danger level remained high for three consecutive days, accurately capturing the prolonged impact of the Atmospheric River and highlighting the model’s capability to forecast extended flood risk windows.

We find that Atmospheric Rivers drive a third of midlatitude floods, emphasizing their critical role in flood prediction systems. Building on these models, we analyze the primary meteorological conditions driving AR flood risk. Our findings reveal an increase of over 10% in AR-related flood risk over the last four decades, driven by the rising intensity and frequency of Atmospheric Rivers. We find that the majority of the global population is exposed to ARs to some extent, including those that pose a potential flood risk. These results underscore the potential of integrating Machine Learning and AR data to enhance early warning systems and support effective flood preparedness and response in a changing climate.