Development of Two-Dimensional Visibility Estimation Model Using Machine Learning: Preliminary Results for South Korea

Abstract

A two-dimensional visibility estimation model was developed, based on random forest (RF), a machine learning-based technique. A geostatistical method was introduced into the visibility estimation model for the first time to interpolate point measurement data to gridded data spatially with a pixel size of 10 km. The RF-based model was trained using gridded visibility data, as well as meteorological and air pollution input variable data, for each location in South Korea, which were characterized by complex geographical features and high air pollution levels. Generally, relative humidity was the most important input variable for the visibility estimation (average mean decrease accuracy: 35%). However, PM2.5 tended to be the most crucial variable in polluted regions. The spatial interpolation was found to result in an additional visibility estimation error of 500 m in locations where no adjacent visibility observations within 0.2° were available. The performance of the proposed model was preliminarily assessed. Generally, the best detection performance was achieved in good visibility conditions (visibility range: 10 to 20 km). This study is the first to demonstrate a visibility estimation model based on a geostatistical method and machine learning, which can provide visibility information in locations for which no observations exist.