Fine dust, which directly threatens people's health and life, is now emerging as a matter of national disaster as well as a common concern of the whole people.

NIRS signed a business cooperation agreement (MoU) with the UN Global Pulse* Jakarta Research Institute (April 19th, 2018), and has forecasted fine dust in Northeast Asia and analyzed data on major factors.

* A program under the direct control of the UN Secretary-General, designed to protect the vulnerables from crises and disasters using Big Data. Currently operating Pulse Labs in New York, Jakarta, Indonesia, and Kampala, Uganda.

In order to accurately identify domestic and foreign factors, Incheon, a densely populated area of the west coast, was selected as the subject of analysis.Unlike existing numerical prediction models, machine learning was used to 1) develop a fine dust prediction model to predict fine dust tomorrow and, 2) identify main factors affecting fine dust.

This analysis includes,
1) fine dust and air pollution data in Incheon from January 2015 to March 2019 (Ministry of Environment, 28,464 cases), 2) satellite sensor data in Northeast Asia provided by NASA* and, 3) Aeronet (AERONET)**'s ground observation sensor data was used, and the UN Global Pulse Jakarta Research Center provided technical advice based on its experience in Indonesian air pollution related data analysis. ***

* Observation of aerosols, small particles suspended in the air, such as fine dust, with sensor data from the MODIS (Medium Resolution Imaging Spectrometer) sensor of the NASA Aqua satellite

** Observations from the ground with the International Collaborative Aerosol Observation Network operated by NASA

*** Nowcasting Air Quality by Fusing Meteorological Data, Insights from Satellite Imagery and Photos shared on social media using Deep Learning (2018)

First, we implemented a gradient boosting*-based prediction model that showed optimal performance for fine dust forecasting. As a result of predicting the first quarter of 2018, it was confirmed that the accuracy of the forecast was about 15% higher than that of the existing domestic fine dust forecast.
- fine dust (PM10) 84.4% and ultrafine dust (PM2.5) 77.8

* Machine learning models that improve predictiveness by combining weak predictive models

The main predictors were wind direction, rainfall, and aerosol concentration in the west coast and Shandong Province of China for fine dust, and wind speed, wind direction, and aerosol concentration in inner Mongolia, Beijing and Hebei Province in China for ultrafine dust.

As a result of the detailed analysis, it was verified that when the fine dust was 'bad', the wind direction was westerly, and the aerosol concentration was very high in Shandong, Shaanxi, Beijing and Hebei provinces in China.

As a result of comparing the fine dust prediction correlations of 20 observatories in Incheon in particular, fine dust and nitrogen dioxide (NO2) in the Baengnyong Island area, not in the downtown area of Incheon, showed the highest correlation, which indicates that foreign factors are relatively higher than domestic ones.

In addition, as a result of predicting the first quarter of 2018 after removing foreign factors from the data, the 'good' grade increased by 50% from 20 days to 30 days.

In the future, NIRS plans to secure additional data from domestic geostationary satellites (Cheonrian 2A, 2B) which are with excellent aerosol analysis performance, and to improve prediction accuracy by combining with other analysis models as well.