Short-term prediction of runoff events in a blanket bog catchment using various cost-sensitive machine learning algorithms with feature selection

Accurate prediction of runoff events in blanket peat catchments provides an important means of understanding their hydrological functioning, managing downstream flood risk and targeting sampling routines for water quality assessment. This study evaluates the performance of four machine learning classifiers — Support Vector Machine, Decision Tree, Random Forest and RUSBoost — for instantaneous (Ev1) and 12-hour forecast (Ev12) prediction of rainfall-runoff events in the near-intact Letterunshin Blanket Bog Catchment (Co. Sligo, Ireland). From two years of hourly hydrometeorological data, 56 rainfall-runoff events were extracted and characterised, revealing a responsive system with variable lags and seasonal differences in rainfall thresholds. Twelve instantaneous and aggregated hydrometeorological predictors were evaluated using wrapper and embedded feature-selection methods, alongside misclassification cost adjustments to address class imbalance. Ensemble learners consistently outperformed single classifiers. RUSBoost achieved the strongest overall performance (Ev1 MCC = 0.85; Ev12 MCC = 0.83), providing the best balance of precision and recall, while Random Forests yielded the highest precision (>89%) but lower sensitivity to event detection. Feature selection reduced predictor dimensionality without compromising accuracy, with 12-hour aggregated hydrometeorological variables emerging as dominant controls on event occurrence. Cost-sensitive learning increased recall across learners, though with reductions in precision, and produced the most consistent gains for Random Forest models. Although predictive skill declined modestly for 12-hour forecasting relative to instantaneous classification, all Ev12 models achieved high accuracy (>91%) and MCC (>0.74). These findings demonstrate the potential of cost-sensitive ensemble learning to support high flow early warning, adaptive sampling and hydrological monitoring in blanket peat systems.