Comparative Performance of Tree-Based Machine Learning Algorithms for Horizontal Visibility Classification at Soekarno–Hatta International Airport Using Meteorological and PM2.5 Data

Abstract

Low horizontal visibility phenomena are a major threat to aviation safety and operational efficiency, particularly at Soekarno–Hatta International Airport (WIII), which is vulnerable to fog and haze arising from complex interactions between meteorological conditions and air quality. This study develops and comparatively evaluates four tree-based machine learning algorithms as independent base learners, Random Forest, Extra Trees, XGBoost, and LightGBM, for horizontal visibility classification, without employing any ensemble learning scheme. The dataset comprises AWOS meteorological observations integrated with PM2.5 concentrations from 2021 to 2025. Results show that XGBoost achieved the best performance, with 97.62% accuracy, 97.65% F1-score, and an AUC-ROC of 0.9926, followed by LightGBM, Random Forest, and Extra Trees. Feature importance analysis identified PM2.5 as the most dominant predictor, followed by wind-related variables. These findings indicate that a well-optimized standalone base learner can achieve competitive predictive performance while offering greater interpretability and computational efficiency for operational implementation. Keywords: horizontal visibility, machine learning, PM2.5, XGBoost, aviation meteorology