Deep learning algorithms for prediction of PM10 dynamics in urban and rural areas of Korea

• Published in: Earth science informatics Vol. 15; no. 2; pp. 845 - 853
• Main Authors: Choi, Han-Soo, Song, Kyungmin, Kang, Myungjoo, Kim, Yongcheol, Lee, Kang-Kun, Choi, Hanna
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.06.2022; Springer Nature B.V
• Subjects: Air pollution; Algorithms; Artificial neural networks; Carbon monoxide; Deep learning; Earth and Environmental Science; Earth Sciences; Earth System Sciences; Health problems; Information Systems Applications (incl.Internet); Machine learning; Neural networks; Nitrogen dioxide; Ontology; Particulate emissions; Particulate matter; Pollutants; Research Article; Rural areas; Secondary aerosols; Simulation and Modeling; Space Exploration and Astronautics; Space Sciences (including Extraterrestrial Physics; Suburban areas; Sulfur dioxide; Deep learning; prediction; Contribution ratio; PM; Particulate matter
• ISSN: 1865-0473; 1865-0481
• DOI: 10.1007/s12145-022-00771-1

High concentrations of particulate matter (PM) are frequently associated with serious health problems, underlining the importance of accurate PM prediction. This study aimed to predict PM 10 concentrations by analyzing air pollutant data in Korea (Seoul, Incheon, Daejeon, and Busan) using convolutional neural networks (CNNs) and long short-term memory (LSTM) deep learning methods. Real-time data from January 2014 to December 2020 were organized as hourly averages. The SO 2 , NO 2 , CO, O 3 , and PM 10 data from 2014 to 2018 were used for training, and data from 2019 to 2020 were used as test data. The highest prediction accuracy was accomplished using all observations. The contribution ratio of each model component to the predictions was verified using SHapley Additive exPlanations (SHAP), and PM 10 showed the greatest contribution. The other components, as secondary aerosol precursors, were divided by area. CO and O 3 were found to be high in Seoul (Gwanak), which has been highly urbanized. On the other hand, CO and NO 2 were found to be high in Incheon (Namdong), Daejeon (Yuseong), and Busan (Sasang), which are relatively suburban areas. The deep learning results indicated that the predicted PM 10 concentration was most affected by past and present concentrations of PM 10 . It is considered that the atmospheric PM 10 at the study sites mainly originated from direct emissions. We compared the proposed method with recent prediction methods using algorithms, machine learning, and deep learning. The R 2 , root mean square error, and mean absolute error evaluation indices supported the suitability of the proposed method for analyses at the study site.