Prediction of Daily PM10 Concentration for Air Korea Stations Using Artificial Intelligence with LDAPS Weather Data, MODIS AOD, and Chinese Air Quality Data

Abstract

PM (particulate matter) is of interest to everyone because it can have adverse effects on human health by the infiltration from respiratory to internal organs. To date, many studies have made efforts for the prediction of PM10 and PM2.5 concentrations. Unlike previous studies, we conducted the prediction of tomorrow's PM10 concentration for the Air Korea stations using Chinese PM10 data in addition to the satellite AOD and weather variables. We constructed 230,639 matchups from the raw data over 3 million and built an RF (random forest) model from the matchups to cope with the complexity and nonlinearity. The validation statistics from the blind test showed excellent accuracy with the RMSE (root mean square error) of 9.905 ㎍/㎥ and the CC (correlation coefficient) of 0.918. Moreover, our prediction model showed a stable performance without the dependency on seasons or the degree of PM10 concentration. However, part of coastal areas had a relatively low accuracy, which implies that a dedicated model for coastal areas will be necessary. Additional input variables such as wind direction, precipitation, and air stability should also be incorporated into the prediction model as future work.

1. Introduction

Aerosol refers to the fine solids and liquids floating in the air. PM (particulate matter) is a kind of aerosol which is very fine and invisible. According to the size of the diameter, the PM is divided into two categories: PM10 with a diameter below 10 μm and PM2.5 with a diameter below 2.5 μm (Seinfeld and Pandis, 2006). PM2.5 is usually formed by the secondary reactions of the precursor such as NH3, SO2, NOx, and VOC (volatile organic compound) under a specific atmospheric condition (Kim, 2006). PM10 is formed by the chemical mechanism similar to that of PM2.5 and by the soil constituents such as Al, Si, Ca, Ti, and Fe (Kim et al., 2016).

PM10 and PM2.5 become a matter of interest for everyone because it can have adverse effects on human health by the infiltration from respiratory to internal organs.WHO (World Health Organization) mentioned that approx. 50% of acute lower respiratory infections, 43% of chronic obstructive pulmonary disease, 29% of lung cancer, 25% of ischemic heart disease, and 24% of stroke can be caused by air pollution (WHO, 2018). Because Korea is located on the leeward side of China that suffers from severe air pollutions (Kim et al., 2006), accurate information about PM10 and PM2.5 becomes more critical. However, the prediction of PM10 and PM2.5 in Korea is not yet a satisfactory level in terms of accuracy (Cho et al., 2020).

The previous studies in the prediction of PM10 and PM2.5 are primarily divided into process-based modeling and statistical modeling. Process-based approaches employed numerical models such as Adam(AsianDust Aerosol Model) (Cho et al., 2007; Kim et al., 2011), WRF-Chem (Weather Research and Forecasting Model with chemistry)(Saide et al., 2011; Moon et al., 2014), and CMAQ (Community Multi-scale Air Quality) (Kim et al., 2016; Jo et al., 2017) for the simulation of PM10 and PM2.5 concentrations, but it requires post-processing to reduce its unique systematic errors(Sohn et al., 2016). Statistical approaches aimed to mitigate the systematic errors of process-based models and used the meteorological variables such as temperature, humidity, precipitation, solar radiation, and wind that influence the PM10 and PM2.5 concentrations (Cha and Kim, 2018).

The AI (artificial intelligence) approaches using meteorological data with in-situ observations became an alternative to statistical models. Prediction models were constructed by employing RF (random forest) and GBM (gradient boosting algorithm), a decision tree-based ensemble machine learning technique (Kwon et al., 2015; Choubin et al., 2020; Lee and Lee, 2020), and MLP (multi-layer perceptron) and ANN (artificial neural network), a structured weighting network consisting of multiple neurons (Grivas and Chaloulakou, 2006; Gupta and Christopher, 2009; Asadollahfardi et al., 2016; Cho et al., 2019a). Recently, more intensive optimization of the learning network can be performed in a thicker and deeper network system. The deep learning models such as DNN (deep neural network) and RNN (recurrent neural network) were also employed for the prediction of PM10 and PM2.5 concentrations (Athira et al., 2018; Cho et al., 2019b; Cho et al., 2020). However, these AI models have been applied to only a few cities.

Also, a long-term prediction of PM10 is a challenging task because many uncertainties are involved in the determination of long-term atmospheric physics and chemistry. Various efforts have been made to mitigate the uncertainties, such as seasonal models (Miri et al., 2019), low- and high-concentration models(Cho et al., 2020), and more advanced time-series models based on the LSTM (long short-term memory) and GRU (gated recurrent unit) (Athira et al., 2018; Wu et al., 2020). 

Indeed, satellite-derived AOD (aerosol optical depth) is closely associated with the PM10 and PM2.5 concentrations (Lee et al., 2006), but few studies have combined the satellite and meteorological data for the prediction of PM10 and PM2.5 concentrations. Korea can be affected by the air quality of China, but the information about the PM10 and PM2.5 concentrations of Chinese cities is rarely utilized for ancillary data. Although Air Korea has more than 300 stations, only a few stations were used in the previous studies.

Thus, we aim to conduct an AI modeling for prediction of PM10 concentration using the entire data sets from the Air Korea stations for the recent five years (January 2015 to December 2019) and the air quality data of Chinese cities, in addition to satellite AOD and meteorological data. We gathered the 3 million records for the hourly PM10 concentration from the 331 Air Korea stations. Also, the daily PM10 concentration data for three Chinese cities, the MODIS (Moderate Resolution Imaging Spectroradiometer) AOD (aerosol optical depth) images, and the LDAPS (Local Data Assimilation and Prediction System) meteorological variables were obtained to construct a matchup database. We built an RF model for the prediction of the PM10 concentration of tomorrow for the 331 Air Korea stations. We conducted 10-fold cross-validation for the blind test to evaluate the accuracy of our prediction model.

2. Data and Methods

1) Data

The in-situ measurements, satellite images, and meteorological variables for this study are summarized in Table 1.

Table 1. Data used in this study

The hourly PM10 measurements for the 331 Air Korea stations were obtained from the website (https://www.airkorea.or.kr/) managed by the Ministry of Environment, and the data was aggregated into the daily mean PM10 concentration for each station. Daily mean PM10 concentration of the three cities (Beijing, Tianjin, and Weihai) on the east coast of China has been gathered from the AQICN (Air Quality Index– China) website (https://aqicn.org/). Korea is under the influence of the westerlies, so the atmosphere of the east coast of China often moves to the Korean peninsula(Fig. 1). Fig. 2 shows the changes in the hourly PM10 concentration for Beijing, Tianjin, and Weihai in China, and Chuncheon in Korea. On May 7, the PM10 concentration of Chuncheon surpassed 500 μg/m³, and a governmental alert wasissued for the PM10. Twentyfour hours before that day, the PM10 concentration of Weihai was almost 500 μg/m³, and 12 hours before that time, the PM10 concentration of Beijing and Tianjin came close to 1000 μg/m³. It was presumably because of the influence of westerlies from China. The atmosphere has moved from Beijing and Tianjin through Weihai to the Korean peninsula, which could transport the PM10 to Korea. So, we calculated the daily average PM10 concentration for the three Chinese cities, which was used as one of the input variables for the prediction model.

Fig. 1. Wind maps for the east coast of China and the Korean peninsula, May 6 and 7, 2016

Fig. 2. Hourly PM10 concentration in Beijing, Tianjin, and Weihai in China, and Chuncheon in Korea, May 6 and 7, 2016.

MODIS is a sensor onboard Terra and Aqua satellites that were launched by NASA (National Aeronautics and Space Administration)in December 1999 and May 2002. The two satellites are scanning the entire Earth twice a day at an altitude of approx. 700 km and are gathering data for 36 spectral bands. We used the 550- nm AOD data from the MYD04_L2 product by Aqua MODIS. Part of solar radiation is absorbed and scattered by aerosols before reaching the ground. The aerosol extinction coefficient is calculated as the sum of the absorption and scattering coefficients. The 550- nm AOD is the vertical integral of the aerosol extinction coefficient at the wavelength of 550 nm. Because the MYD04_L2 product is a granule data along the orbit, 2 to 4 scenes per day are created for the region around Korea. We merged the multiple scenes into one image in the geographic coordinate system of latitude and longitude.

LDAPS is a numerical weather prediction model by KMA (Korea Meteorological Administration). It provides 3-hourly forecasts for the Korean peninsula with a spatial resolution of 1.5 km in the LCC(Lambert Conforma lConic) projection.We extracted the data for air temperature, relative humidity, wind speed, and boundary layer height at 03 UTC (12 KST). The boundary layer is the lowest part of the atmosphere, i.e., 1 to 2 km above the ground directly influenced by the Earth’s surface. In this layer, the temperature, humidity, and wind can showa relatively rapid fluctuation with a strong vertical mixing because of the turbulent transfer of air mass (Pielke and Hayden, 2020). The high boundary layer means a broader space for the mixing and vertical diffusion of air pollution, so it is closely related to the PM10 concentration on the ground (Du et al., 2013;Geiß et al., 2017;Li et al., 2017; Miao and Liu, 2019). Today’s weather variables were obtained from the LDAPS reanalysis, and tomorrow’s weather data were acquired from the LDAPS forecast. The LDAPS reanalysis and forecast in the LCC projection were transformed into the geographic latitude and longitude for use in the collocation with the Air Korea stations.

2) Methods

RF is an ensemble method that utilizes a number of decision trees derived from random samples. If necessary, a bootstrap process is performed for resampling by accounting for the sample distribution. A bagging (bootstrap aggregating) process creates a final solution by averaging the results from the bootstrapped trees (Ali et al., 2012). We used the h2o framework (https://www.h2o.ai/) to optimize the hyper parameters for the tree numbers and the variable numbers for splitting nodes in our experiment.

We built an RF model for the PM10 concentration of tomorrow for the 331 Air Korea stations (Fig. 3). The RF model included 11 input variables: today’s PM10 concentration for each station, today’s PM10 concentration of Beijing, Tianjin, and Weihai, today’s MODIS AOD, and air temperature, relative humidity, wind speed, and boundary layer height for today and tomorrow. LDAPS reanalysis was used for today’s meteorological variables, and the forecast was used for tomorrow’s meteorological data. MODIS AOD can have missing pixels because of the cloud. Except for the missing values, we construct a matchup consisting of 230,639 valid records. This is a much larger database when compared with previous studies, which can help AI models achieve a good performance.

Fig. 3. Random forest model for the prediction of PM10 concentration of tomorrow for Air Korea stations.

We carried out 10-fold cross-validation for the blind test to evaluate the accuracy of our RF model (Fig. 4). First, the 230,639 matchups were divided into ten groups by random sampling. For round 1, group 1 was set to the validation target, and the other nine groups were used for model calibration. For round 2, group 2 becomes a validation target, with the other nine groups for model calibration.In this way,ten-round experiments were iterated, and the result was summarized for validation statistics.The 10-fold cross-validation enables a more stable model by using the training data with a less-biased sampling. The indices such as MBE (mean bias error), MAE (mean absolute error), RMSE (root mean square error), NRMSE (normalized mean square error), and CC (correlation coefficient) were used for the validation statistics.

Fig. 4. Concept of 10-fold cross-validation with random sampling for the prediction of PM10 concentration of tomorrow (Kim et al., 2020).

3. Results and Discussions

Table 2 shows a summary of the result of the 10-fold cross-validation for our RF model using the 230,639 cases. The MAE of 6.846 μg/m³ indicates that the differences between the in-situ observations and our predictions were small. The CC of 0.918 shows very high predictability with the scatter plot concentrated around 1:1 line (Fig. 5). Because the PM10 of Korea is somewhat severe in winter and spring, we examined how the performance of our model changed according to seasons (Table 3). The number of matchups in summer was relatively small, which is because of the missing pixels of MODIS AOD due to the cloud. The accuracy did not show a significant difference by season. However, the summer PM10 has small values, so the RMSE was also small despite the lower CC than other seasons. For more objective validation statistics, we added NRMSE, a normalized index of the RMSE divided by the mean. The NRMSE of the summer PM10 was similar to that of other seasons, so we made sure that the performance of our prediction model did not depend on seasons.

Table 2. Validation statistics of the random forest model for the prediction of PM10 concentration of tomorrow during 2015-2019

Fig. 5. Observed vs. predicted daily PM10 concentration for 331 Air Korea stations, 2015-2019.

Table 3. Validation statistics of the random forest model for the prediction of PM10 concentration of tomorrow by season

Korean government presents a classification of PM10 concentration: good (less than 30 μg/m³), normal (30 to 80 μg/m³), bad (80 to 150 μg/m³), and very bad (greater than 150 μg/m³).We examined if the accuracy of our prediction model can vary according to the degree of PM10 concentration. We divided the result of the 10-fold cross-validation into four groups according to the classification of today’s PM10 concentration, because it is an already known value. The number of cases for the group “Good” was 48,965; the group “Normal” 161,635 cases; the group “Bad” 18,685 cases; and the group “Very bad” 1,353 cases. Approx. 10% of the data belonged to the group “Bad” or “Very bad.” Table 4 shows the validation statistics for each group. Small values usually result in minor errors, and large values often lead to significant errors. So, the RMSE of the group “Bad” and “Very bad” was higher, but the NRMSE was similar for all groups. The CC for all groups was also similar, so we supposed that our model produced a stable prediction for the PM10 concentration of tomorrow irrespective of the degree of PM10.

Fig. 6. Observed vs. predicted daily PM10 concentration for 331 Air Korea stations by season, 2015-2019.

Fig. 7. Observed vs. predicted tomorrow’s PM10 concentration for 331 Air Korea stations grouped by today’s PM10 classification (good, normal, bad, and very bad).

Table 4. Validation statistics of the random forest model for the prediction of PM10 concentration of tomorrow according to today’s PM10 classification (good, normal, bad, and very bad)

Each Air Korea station can have as many as 1,826 matchups for the five years. We examined how are the characteristics of accuracy according to the 331 stations. The stations were mainly located in the big cities, and the agricultural land and forest area have fewer stations. It is because the issue of PM10 is of interest to big cities. Fig. 8 shows that the MBE did not have a significant pattern, according to the region. However, the MAE, RMSE, and CC in Fig. 9, Fig. 10, and Fig. 11 indicates that part of coastal areas had relatively low accuracy, particularly for Jeju and Kangwon. It implies that an additional model for coastal areas will be required for future work.

Fig. 8. Map of mean bias error for the prediction of tomorrow’s PM10 concentration by stations, 2015-2019.

Fig. 9. Map of mean absolute error for the prediction of tomorrow’s PM10 concentration by stations, 2015-2019

Fig. 10. Map of root means square error for the prediction of tomorrow’s PM10 concentration by stations, 2015-2019.

Fig. 11. Map of the correlation coefficient for the prediction of tomorrow’s PM10 concentration by stations, 2015-2019.

Unlike previous studies, we used Chinese PM10 data in addition to the satellite AOD and weather variables. The validation statistics from the blind test using 230,639 cases showed excellent accuracy with the RMSE of 9.905 μg/m³ and the CC of 0.918. It is because the complexity and nonlinearity were coped with by the AI models and because the quantity and quality of the training data were sufficient. Table 5 shows the variable importance; that is, the contribution of each variable to the model result. Today’s PM10 concentration of Air Korea station was 34.3%, and today’s PM10 concentration of the three Chinese cities was 26.0%. The other variables contributed to the RF model by approx. 40%; MODIS AOD and LDAPS meteorological variables have similar importance.

In particular, the two essential variables(i.e., today’s PM10 concentration of Korea and that of China) showed unique seasonal characteristics. The PM10 concentration of Korea is usually higher in winter and spring, presumably because of the westerlies from China. Also, the variable importance of today’s PM10 concentration of the three Chinese cities was higher in winter and spring but lower in summer and fall(25.5% for winter and 22.7% for spring, but 12.2% for summer and 18.3% for fall). The importance of AOD and meteorological variables did not significantly vary according to seasons. However, the AOD showed higher importance in summer than in other seasons; the boundary layer height was more important in fall than in other seasons; the wind speed was relatively important in winter than in otherseasons.It implies that individual seasonal models will be necessary for amore delicate prediction of the PM10 concentration in Korea.

4. Conclusions

We conducted the prediction of tomorrow’s PM10 concentration for the 331 Air Korea stations using Chinese PM10 data in addition to the satellite AOD and weather variables. We constructed 230,639 matchups from the raw data over three million and utilized them for the RF prediction model to cope with the complexity and nonlinearity. As a result, our model produced excellent accuracy from the blind test with the RMSE of 9.905 μg/m³ and the CC of 0.918. Moreover, our prediction model showed a stable performance without the dependency on seasons or the degree of PM10 concentration. However, part of coastal areas had a relatively low accuracy, which implies that a dedicated model for coastal areas will be necessary. Moreover, the variable importance by season indicates the possible needs for seasonal models. Quality control for the LDAPS weather variables using the ground measurements of ASOS (Automated Surface Observing Systems) may be necessary. Additional input variables such as wind direction, precipitation, and air stability should also be incorporated into the prediction model as future work.