Characterization of the shallow structure of fog using categorical machine learning models.

This study characterizes the shallow structure of fog using categorical machine learning models based on horizontal visibility and the Runway Visual Range (RVR). Given the critical role of RVR in determining approach minima during low visibility landings, accurate estimation and interpretation of visibility parameters are essential for aviation safety.

Fog events over a 20-year period were studied using METAR data from the International Airport of Lisbon. The fog events were characterized by duration and intensity, with intensity defined by the combination of horizontal visibility and RVR. Each fog event was classified into one of five fog types (precipitation, advection, radiation, cloud base lowering, and evaporation) according to well-defined criteria based on preconditioning features, using machine learning decision- ree-based models: random forest (RF), extreme gradient boosting (XGB), light gradient boosting machine (LGBM), adaptive boosting (ADB), and categorical boosting (CB).

The study was divided into two stages. In the first stage, the model’s assessment focused on fog occurrence prediction. The results have shown that the XGB model significantly outperforms the other models, with nearly 85% reproducibility of fog events, compared with 48% to 60%. The second stage assessed fog-type prediction. The best model’s ability to learn from fog-classification criteria was shared among the RF, LGBM, and XGB models. All models showed the highest performance on the most common fog type, the advection type. Despite lower all-types reproducibility, LGBM predicted 58% on average, followed by RF, with 57%, and XGB, with 52%. Overall, XGB is excellent at fog-occurrence prediction, but LGBM, RF, and XGB are only decent at fog-type prediction. Because the models predict radiation fog instead of advection fog, this misclassification will negatively impact low-visibility operations, since the radiation fog type lasts longer than the advection type and can persist for up to 24 hours.