Estimation of the Visibility in Seoul, South Korea, Based on Particulate Matter and Weather Data, Using Machine-learning Algorithm

Abstract

Visibility is an important indicator of air quality and of any consequent meteorological and climate change. Therefore, visibility in Seoul, which is the most polluted city in South Korea, was estimated using machine learning (ML) algorithms based on meteorological (temperature, relative humidity, and precipitation) and particulate matter (PM10 and PM2.5) data acquired from an automatic weather station, and the estimated visibility was compared with the observed visibility. Meteorological data, observed at 1-h intervals between 2018 and 2020, were used. Through learning and validation of each ML algorithm, the extreme gradient boosting (XGB) algorithm was found to be most suitable for visibility estimations (bias=0 km, root mean square error (RMSE)=0.08 km, and r=1 for training data set). Among the meteorological and particulate matter data used for learning the XGB algorithm, the relative importance of PM2.5 and relative humidity variables were high (51% and 19%, respectively), whereas precipitation and wind speed had the low relative importance (approximately 1%). The estimation accuracy for the test dataset was good (bias=-0.11 km, RMSE=2.08 km, and r=0.94); the estimation accuracy was higher in the dry season (bias=-0.06 km, RMSE=1.79 km, and r=0.96) than in the rainy season (bias=-0.17 km, RMSE=2.34 km, and r=0.91). The results of this study indicated a higher correlation than the results of previous visibility estimation studies. The proposed method promotes accurate estimation of visibility in areas with poor visibility, and thus, it can be used to assess public health in areas with poor air quality.

FGS-Net: A Visibility Estimation Method Based on Statistical Feature Stream in Fog Area

“…Among the deep learning-based visibility estimation methods, most researchers apply classical CNN models to visibility estimation directly. For example, Ortega, Mu, and Wang et al [17,18,19,20] used CNN to estimate the visibility of images, and their method directly leverage the power of feature learning of CNN models and then put them into fully connected layers to map the learning features to different visibility levels. Giyenko [5] constructed a CNN model for visibility estimation and achieved an 84% accuracy on the visibility dataset where captured by Korean CCTV.…”

STCN-Net: A Novel Multi-Feature Stream Fusion Visibility Estimation Approach

“…Ortega, Mu, and Wang [17,18,19,20] used CNN to estimate the visibility range of a single image, but their methods directly applied deep learning to obtain "learned features" which lead to low accuracy of visibility estimation. Chincholkar [21] adds some "engineering features" to his method, such as the brightness and variance of the image, but the accuracy of visibility estimation is still low because the brightness and variance of the image are "common features" of the image rather than "specific features" of visibility.…”

Estimating visibility and understanding factors influencing its variations at Bangkok airport using machine learning and a game theory-based approach

“…Aman et al (2022) reported two visibility events in the winter of 2014 and 2015 affected by the synergetic effect of particulate pollution and meteorology. Recently, visibility-related studies across the world have started using machine learning (ML) models for visibility prediction using particulate matter and meteorological variables as input data (Kim et al 2022a; Kim et al 2022b; Penov and Guerova 2023) and the importance of different features is identi ed by feature importance plots (Kim et al 2022a; Penov and Guerova 2023). However, feature importance has limited insights, as it does not give information about their directional relationship with visibility and interactive effects of different variables.…”