Machine learning driven by environmental covariates to estimate high-resolution PM2.5 in data-poor regions

MODIS AOD data

We gathered AOD information throughout Xinjiang from January 2015 to February 2021 utilizing the Moderate Resolution Imaging Spectroradiometer (MODIS) instruments carried by the Terra and Aqua satellites. Terra and Aqua provide daily observations of the global equator at local crossing times of 10:30 am and 1:30 pm, respectively. Based on a new advanced algorithm, the atmospheric correction is realized from multiple angles. By analyzing time series images of bright and dark vegetation surfaces, AOD estimates over both types of vegetation surfaces can be accurately retrieved with a resolution of 1 km (Zhang et al., 2019b). The Terra and Aqua combined aerosol product MCD19A2 (550 nm, downloaded from https://aeronet.gsfc.nasa.gov) consists of various data layers. Our research is based on the terrestrial 550 nm (green band) AOD.

Ground-level PM_2.5 monitoring

The PM_2.5 ground measurement data we collected from January 2015 to February 2021 were downloaded from China’s National Air Quality Real-time Release Platform (http://106.37.208.233:20035/). The platform has published hourly PM_2.5 concentrations from air quality monitoring stations in more than 1,600 major Chinese cities since 2013. In the main analysis, we limited the study area to the Xinjiang region of China. Xinjiang contains 41 ground air quality monitoring stations in 16 cities spread across an area exceeding 1.6 million km². The locations of the research area and the ground monitoring stations are shown in Fig. 1. The hourly PM_2.5 concentrations on this platform are measured and reported in accordance with China’s National Ambient Air Quality Standard, and the accuracy and quality control of these PM_2.5 measurements have been previously reported (Chen et al., 2018). The ground PM_2.5 monitoring information is shown in Table 1.

Meteorological data

The meteorological data were obtaind from the second edition of the National Centers for Environmental Prediction (NCEP) climate forecasting system, which provides meteorological data every 6 h. Fifteen meteorological variables were considered herein: the dew point temperature at 2 m, maximum/minimum temperature at 2 m, maximum/minimum humidity, land surface temperature (LST), relative humidity (RH), precipitation, potential evapotranspiration rate, upward/downward longwave radiation (ULR/DLR), upward/downward shortwave radiation (USR/DSR), surface pressure (SP), and wind speed (WS). The above data can be downloaded from Google Earth Engine (https://earthengine.google.com/).

Auxiliary data

The land use parameters employed herein include the normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), MODIS NDVI (product name: MYD09GA) and MODIS EVI (product name: MOD09GA) provided by NASA. The predictive factors of the model include the population density and the drought index (DI), namely, the Keetch-Byram drought index (KBDI), which is estimated from the ground temperature and precipitation at a weather station and is manually interpolated and refined by experts. The above data can be downloaded from Google Earth Engine (https://earthengine.google.com/).

Bagging algorithm

The bagging algorithm is a popular, simple and effective ensemble learning algorithm. Proposed by Breiman in 1996, the bagging algorithm can be used for both classification and regression. The algorithm selects from the original data set with replacement; each sample is independent of each other, each subset in the training set is used to train the classifier, and finally, the results of each classifier are combined by voting. The model is guaranteed to achieve high performance, as is the statistically reliable estimation arising from the generalization ability of the model, and the risk of overfitting is avoided.

Random forest

Random forest is a popular machine learning method that was proposed by Breiman (Breiman, 2001). This method has been used for prediction and categorization tasks in a variety of applications. The RF technique consists of a succession of computer-generated decision trees that can extract information from complex input data and learn the highly nonlinear relationship between input and goal variables. Specifically, RF randomly selects multiple attributes from each node’s attribute set to form a subset and then establishes rules to determine the best attribute from this subset and finally predicts based on the average value of all leaf nodes among all trees.

This study aimed to construct an RF model for estimating the PM_2.5 concentrations across Xinjiang and for studying its spatial distribution and regional differences. Considering that the air quality situation is likely to vary slightly over a six-year period, based on the valid data we obtained, the average value of the data was taken every 8 days, and null values and outliers were excluded; the final number of valid samples obtained by all 41 stations in Xinjiang for each of the six years during 2015–2020 was (in order) 1,258, 1,276, 1,365, 1,360, 1,398, and 1,338. In addition, we used two indicators of the RF model to measure variable importance, namely, the percent increase in the mean squared error (%IncMSE) and the increase in node purity (IncNodePurity), to rank the importance of the factors. Finally, based on the six-year overall situation, we selected 9 factors, which were entered as predictors: AOD, DEM, NDVI, SI, SP, DLR, USR, WS, and RH. A RF model was established for the AOD–PM_2.5 relationship, as shown in Eq. (1). This model was implemented in the R3.6.3 language, and the modeling data involved in this paper are shown in Table 2.

The RF model constructed in this study can be abbreviated as: (1)${\displaystyle \mathrm{PM}2.5=\mathrm{RF}\left(\text{AOD, DEM, NDVI , SI, SP, DLR, USR, WS, RH}\right)}$ where AOD is the aerosol optical depth; DEM is the elevation; NDVI is the normalized difference vegetation index; SI is the drought index; SP is the surface pressure; DLR denotes downward longwave radiation; USR denotes upward shortwave radiation; WS is the wind speed; and RH denotes relative humidity.

To validate the model, 10-fold cross-validation (CV) was used in this work. The complete training data set was randomly divided into ten subgroups for each CV, nine of which were utilized as training subsets and the remaining one was employed as a validation subset. The average value was taken as the final accuracy of the RF model, and the Pearson correlation coefficient (R), R² and RMSE were calculated over the entire data set consisting of 10 test subsets to evaluate the correlations between the estimated and observed PM_2.5 concentrations.

Spatial distribution characteristics of AOD

The applicability of MCD19A2 data has been verified in numerous studies (Chen et al., 2021; Li et al., 2021b; Tao et al., 2019), so verification is not performed here. The spatial distribution of AOD in Xinjiang from 2015 to 2020 is shown in Fig. 2. There is an important difference between the southern and northern parts of Xinjiang. The AOD in northern Xinjiang has remained at a low level (<0.2) for six years. This may be due to the extensive vegetation coverage and weaker sand and dust interference in northern Xinjiang than in southern Xinjiang, where the AOD has remained high for six years due to the contribution of dust particles from the Taklimakan Desert, with the AOD values in some areas during 2018 and 2020 exceeding 0.4. There was an AOD inflection point in 2018 (0.222). During the preceding three years (2015–2017), AOD gradually decreased, which may have been linked to the “coal to gas” policy that Xinjiang began to implement in 2012; in contrast, since 2018, AOD has been volatile, which may be related to the advancement of Xinjiang’s economic development strategy (Li et al., 2021b).

Characteristics of the seasonal distribution of AOD

In this study, March, April, and May are grouped into spring, June, July, and August are grouped into summer, September, October, and November are grouped into autumn, and December and January and February are grouped into winter. Figure 3 shows the seasonally averaged spatial distributions of AOD from 2015 to 2020. The AOD magnitude decreased in the order of spring (0.342) > summer (0.219) > autumn (0.166) (AOD was not analyzed in the winter because of the difficulties in the inversion, the absence of AOD data, and the overall low AOD). Because spring brings more sand and dust, the area’s overall AOD value was higher than in the other seasons. In contrast, AOD was lower in summer; however, some regions in southern Xinjiang still exhibited high AOD values due to the influences of summer dust and storms. The climate was dry in autumn, with regular Arctic air intrusions and a stable air structure. Overall, the AOD has declined, indicating a state of diffusion (Zhang et al., 2016).

Annual spatiotemporal variations

According to Fig. 5, the concentration of PM_2.5in the study area was higher in southern Xinjiang than in northern Xinjiang from 2015 to 2020, and high PM_2.5 values were concentrated in the Tarim Basin. At the same time, high concentrations extended outwards, affecting neighboring areas and establishing secondary high-value zones. The Taklimakan Desert, the largest desert in China, is located in the Tarim Basin. Correspondingly, studies have revealed that during high dust periods, wind can blow up dust on the ground and suspend it in the atmosphere. If no precipitation falls, this dust can persist for several days after the dust has passed, causing an increase in PM_2.5. As a result, the Tarim Basin has become a hotspot for high PM_2.5 concentrations (Hui et al., 2010). The estimated average PM_2.5 concentrations in Xinjiang during these six years were (in order) 71.69 µg m⁻³, 73.19 µg m⁻³, 56.68 µg m⁻³, 66.35 µg m⁻³, 60.30 µg m⁻³, and 55.70 µg m⁻³. The lowest and highest values were recorded in 2020 and 2016, respectively.

In addition, the PM_2.5 concentrations in northern Xinjiang were lower than those in southern Xinjiang. Northern Xinjiang occupies a crucial strategic location: it is not only an important aspect of China’s western development strategy but also Xinjiang’s most prosperous region, as the vegetation coverage rate in northern Xinjiang is relatively high, which alleviates certain air pollution to a large extent. However, the population of northern Xinjiang is much denser than that of southern Xinjiang, which corresponds to more human activities in northern Xinjiang, that is, the increased burning of fossil fuels (coal, gasoline, diesel) and biomass (straw, firewood) and increased levels of building dust (Wang et al., 2020; Wang et al., 2019). Therefore, human activity is another source of PM_2.5 in northern Xinjiang.

Seasonal spatiotemporal variations

Figure 6 shows the spatial distributions of the PM_2.5 concentration in Xinjiang from 2015 to 2020 estimated by the RF model. PM_2.5 decreased in the following order: winter (71.95 µg m⁻³) > spring (64.76 µg m⁻³) > autumn (46.01 µg m⁻³) > summer (43.40 µg m⁻³). The highest values in winter are attributed to the heating of homes throughout Xinjiang, as heating in winter increases energy consumption and leads to higher concentrations of PM_2.5 (Zhang et al., 2019a). However, because the underlying surface is covered by snow, it is difficult for dust to be entrained into the atmosphere, and snow increases the surface albedo, which increases the difficulty of AOD inversion and even leads to data gaps in northern Xinjiang, resulting in the overall low PM_2.5 in winter mentioned above.

In spring, frequent sandstorms lead to higher PM_2.5 levels overall. When strong winds blow over the bare desert or bare soil, it causes wind erosion and entrains dust, aggravating air pollution in the surrounding area (Chai et al., 2017; Liu et al., 2020). In spring, the PM_2.5 concentrations in most areas of Xinjiang were greater than 56.64 µg m⁻³. The Tarim Basin in southern Xinjiang has always been the main source of dust aerosols, but due to topographical reasons, the internal wind speed is relatively low, and thus, horizontal diffusion is relatively weak, which concentrates PM_2.5 therein.

Summer was the season with the lowest PM_2.5 concentration. Because the surface air temperature in summer is higher, the atmospheric transparency is better, and the surface vegetation coverage is considerably enhanced. In addition, precipitation in summer is greater than that in other seasons, and rainfall has an obvious ability to remove dust from the atmosphere (Zhang, Ding & Wang, 2017b). Therefore, the PM_2.5 concentrations in summer were lower than those in the other seasons.

Among the four seasons, the average PM_2.5 value in autumn was the second-lowest. The low concentrations in autumn differ from those in summer because the atmosphere in the research area is characterized by locally high pressures, and the atmospheric system is rather stable in October (Mogo et al., 2017). As shown in Fig. 6, the PM_2.5 levels in most areas of Xinjiang remained low in autumn.