Landfalling Atmospheric Rivers in Ireland: Statistical and Machine Learning Insights

Atmospheric rivers (ARs) transport high concentrations of water vapor in narrow bands from the tropical and subtropical Atlantic to western European coasts. Ireland frequently falls in their paths and receives ~50 ARs annually. To better understand AR-specific synoptic states and behaviour in Ireland, we examine how statistical and machine learning analysis fares in identifying and characterising ARs at both 6-hourly and daily resolutions. We use a dataset based on 1949 landfalling Irish ARs detected using the “tARget” AR detection tool for a 42-year period of 1980-2021, and which are linked to ~80% of Ireland’s daily extreme precipitation events.

Notably, traditional statistical analyses (e.g., correlations, PCA) of daily weather parameters (e.g., 10-m wind, 10-m highest wind gust, air temperature) loosely identify AR days for different landfall regions, but 6-hourly reanalysis variables such as Integrated Vapor Transport (IVT), 850 hpa vertical velocity (ω), and 500 hpa geopotential height strongly distinguish ARs. K-means clustering shows that persistent ARs with high IVT and long overland durations are most common with southern and western Ireland landfalls, whereas northern and eastern landfall sites receive weaker ARs. When trained with daily observational data, machine learning models (Random Forest, XGBoost, and LSTM) identify AR vs. non-AR days with 75-85% F1 scores (precision/recall efficiency). With reanalysis data, the models score ~75% at multi-class classification for AR ranks detection but are less successful for high-intensity ARs (ranked 4 and 5). The Random Forest model performs the best at predicting daily maximum precipitation (R²: 0.63), with key predictors being the 850 hpa upward motion of air (-ω, in %) and maximum IVT. The important reanalysis and observation variables identified above can be selected to reduce model complexity and to train specialized hybrid models for future AR studies.