Application of tree-based machine learning models for winter-spring visibility forecasting at Noi Bai International Airport: A combined regression and classification approach

This study aims to develop and evaluate a machine learning (ML)-based visibility forecasting method for the winter-spring months (January, February, and March) at Noi Bai International Airport, Vietnam. We developed tree-based machine learning models, including Decision Tree (DT), Random Forest (RF), and XGBoost (XGB), using surface meteorological variables and Global Forecast System (GFS) data to forecast visibility at 1-, 2-, and 3-hour lead times. The results indicate that many predictors did not improve forecasting performance. In contrast, DT models with just 3–5 key predictors (selected via RFECV and SHAP) achieved optimal results, particularly in detecting low- and rare-visibility events. The most influential predictors included lagged visibility, time of day, and the vertical gradient of relative humidity within the near-surface and atmospheric boundary layers. The optimal DT models consistently achieved skill scores for the Probability of Detection (POD) of low- and rare-visibility classes ranging from +10% to +35% compared to the persistence forecast, while operating stably without evidence of overfitting.