Explainable Machine Learning for Rare Fog Prediction at Ivato Airport
Fog is a serious hazard for aviation, as it reduces visibility and causes delays and diversions at airports. At Ivato International Airport (FMMI), Madagascar, fog forecasting relies on forecaster experience and general Numerical Weather Prediction (NWP), and fog remains difficult to predict. This study aims to predict fog three hours ahead using only routine Meteorological Aerodrome Report (METAR) observations. The dataset contains 96,067 hourly reports from 2015 to 2025. Fog was observed in only 0.4% of them. Four methods were compared: a persistence baseline, Logistic Regression, Random Forest, and XGBoost. The years 2015 to 2021 were used for training, 2022 for selecting the decision thresholds, and 2023 to 2025 for testing. XGBoost performed best, with a Critical Success Index (CSI) of 0.17 compared with 0.09 for persistence, a Probability of Detection (POD) of 0.53 and a False Alarm Ratio (FAR) of 0.80. It also detected 46% of fog onsets, when no fog was present at forecast time but fog formed three hours later. SHapley Additive exPlanations (SHAP) analysis showed that the model relies mostly on pre-dawn hours, high humidity, low temperature and winter months, which are the conditions in which radiation fog forms. Because the method uses only free public METAR data, it could be applied at other airports with limited resources.
Authors
- Valisoa Nantenaina Iangotiana Andriafalimanana
Institutions
- Asia Pacific University of Technology & Innovation (MY)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24362147
- Primary Topic
- Meteorological Phenomena and Simulations
- Type
- preprint