Predictability of fog using surface observation-based machine learning models: evaluation at Argentinean airport

Low visibility conditions due to the presence of fog can lead to delays or cancellations of arriving and departing flights to avoid possible incidents or accidents at airports. Therefore, accurate visibility forecasts are required to keep airport capacity as high as possible. Recently, some studies explored the use of machine learning algorithms and routine surface meteorological observations to produce fog forecasts. This study explores the potential of machine learning (ML) models trained solely on surface observations for short-term fog nowcasting at Ezeiza International Airport in Argentina, within 6 hours from initialization time. Three simple ML methods—a logistic regression, a decision tree, and a random forest—were evaluated using 20 years of data. Results show that nonlinear methods (decision tree and random forest) required fewer, simpler predictors to match or exceed the forecasting skill of the logistic regression, which relied more heavily on complex variable transformations. All methods significantly outperformed climatological and persistence-conditional reference models. In all three methods, the most important predictor was the fog condition at initialization time. A seasonal and daily data-splitting strategy improved performance at almost all forecasting time horizons. The models trained in this study represent a promising operational tool for aviation forecasting and could also complement operational numerical models, whose ability to resolve fog-scale processes remains limited.